Upload
others
View
4
Download
0
Embed Size (px)
Citation preview
Structural reliability analysis of a dike with a sheet pile wall
-Coupling Reliability methods with Finite Elements- A Rippi
Structural reliability analysis of a dike
with a sheet pile wall
Coupling Reliability methods with Finite Elements
by
A RIPPI
in partial fulfilment of the requirements for the degree of
Master of Science
in Civil Engineering
at the Delft University of Technology
to be defended publicly on Wednesday November 25 2015 at 1100
Graduate Aikaterini Rippi Student ID 4325583 E k-rippihotmailcom
Thesis committee Prof dr ir S N Jonkman TU Delft
Dr ir R B J Brinkgreve TU Delft and Plaxis bv
Dr ir T Schweckendiek TU Delft and Deltares
Dr A Teixeira Deltares
An electronic version of this thesis is available at httprepositorytudelftnl
MSc Thesis A Rippi i
MSc Thesis A Rippi ii
Preface
This thesis is the final challenge in the master Hydraulic Engineering at Delft University of
Technology The report ldquoStructural reliability analysis of a dike with a sheet pile wall Coupling
Reliability methods with Finite Elementsrdquo was completed at Deltares as a part of a larger
research project namely TO2 in collaboration with Toegepast Natuurwetenschappelijk
Onderzoek (TNO) I chose that subject first of all because it combines two things that I enjoyed
a lot at TU Delft probabilities and flood defences systems Secondly it was an opportunity for
me to get acquainted with FEM and geotechnics that I was always interested in
Different people have contributed to the successful completion of this thesis First of all I would
like to express my appreciation to my graduation committee for their guidance and particularly
my daily supervisor Timo Schweckendiek Together we had many fruitful and interesting
discussions on the subject that triggered and motivated me for keep searching Especially I
want to express my gratitude and my thanks to Ana Teixeira and Jonathan Nuttall employees of
Deltares as they were also some of the main contributors to this research They stood by me not
only as colleagues and thesis mentors but also as friends I would like also to thank the rest of
Deltares employees whom ensured a friendly and easy going working environment Last but not
least I want to thank my friends and especially Panagiotis Apostolidis and my family for their
love support and advice throughout all my studies
Katerina Rippi
Delft November 2015
MSc Thesis A Rippi iii
MSc Thesis A Rippi iv
Abstract
Some dike sections in the Netherlands failed to comply with the safety standards and one of the
most promising countermeasures is the construction of retaining walls inside the dike The
Dutch design codes for dikes with retaining walls rely on Finite Element Analysis (FEM) in
combination with partial safety factors However this can lead to conservative designs For this
reason in this research a reliability analysis is carried out with FEM calculations aiming to
demonstrate the feasibility of reliability analysis for such a soil-structure interaction problem
The case study concerns a (river) dike with an anchored sheet pile wall modelled in PLAXIS The
sensitivity and reliability analyses were enabled by coupling the uncertainty software package
OpenTURNS and PLAXIS through a Python interface The most relevant (ultimate) limit states
concern the anchor the sheet pile wall and global instability (soil body failure) The case was
used to investigate the applicability of the First Order Reliability Method (FORM) and
Directional Sampling (DS) to analysing these limit states Finally also the system reliability was
evaluated using sampling-based methods (DS)
Due to the considerable number of random variables before starting the reliability analysis a
sensitivity analysis was conducted for each limit state This indicated the most important soil
layers to be accounted for and the variables to be considered as stochastic The sensitivity
analysis and later on the reliability analysis were based on analytical formulations of the limit
state functions The anchor and the sheet pile limit states were formulated in terms of their
yield stress for global instability loss of equilibrium in the FEM analysis was used to define
failure Moreover the systemrsquos reliability was evaluated by taking into account all the three limit
states that were mentioned previously
The goal is to implement the coupling between FEM and reliability methods in order to analyse
the components of such a system (ie anchor sheet pile wall and dikersquos soil body) estimate the
probability of failure and identify the most important soil properties that affect the behaviour of
each component and the system as a whole The results of this research can be used to assess
and optimize the current design procedure for dikes with retaining walls
MSc Thesis A Rippi v
MSc Thesis A Rippi vi
Table of Contents
Preface ii
Abstract iv
List of abbreviations x
1 Introduction 1
11 Project objective and main research questions 2
12 Research approach and outline 3
2 System description and current design concept 7
21 System description and forces configuration 7
22 Current design concept 12
23 Safety standards 17
3 Literature study 21
31 Background 21
311 Finite Element Modeling 21
312 Uncertainties and Sensitivity analysis 24
32 Previous Studies 29
33 Overview 34
4 Structural Reliability Analysis 35
41 Basics of Reliability Analysis 35
42 Overview of Reliability Analysis Methods 36
421 Level III Methods 36
422 Level II Methods 40
423 Level I Methods (semi-probabilistic) 42
424 Response Surface Techniques (RS) 44
43 Coupling Reliability Analysis with FEM 46
431 The functionality and possibilities of OT 46
432 Coupling OpenTURNS-Plaxis 47
44 Overview 49
5 Failure Mechanisms and Limit State Functions 53
51 Introduction to the system analysis and the limit states 53
52 Limit State Functions 55
521 Serviceability Limit State 55
522 ULS for Structural Members 57
MSc Thesis A Rippi vii
523 ULS for Soil Failure 60
53 Overview 66
6 Case Study-Dike with an anchored sheet pile wall 69
61 Case Description 69
62 Soil Parameters 70
63 Finite Element Model 73
64 Deterministic Analysis 74
641 Calculation Scheme and Design Values 75
642 Construction Stages 77
643 Determination of the structural elementsrsquo characteristics 78
65 Overview 81
7 Reliability analysis results with stochastic soil properties 85
71 Method description 85
72 Mean values calculations 88
73 Sensitivity Analysis Results 91
74 Soil Shear Failure 98
75 Anchor Failure 102
76 Sheet pile wall failure 105
8 Conclusions and Recommendations 121
81 Conclusions 121
82 Recommendations 123
References 125
Appendix A 129
OpenTURNS features 129
A1 Fourier Amplitude Sensitivity Test (FAST) 129
A2 Optimization Algorithms in FORM 132
Principles of optimization algorithms 132
Convergence criteria 134
Evaluation of the algorithms performance 135
A3 Distribution Types 143
Uniform Distribution 143
Normal Distribution 144
Lognormal Distribution 144
Truncated Normal Distribution 145
Appendix B 147
MSc Thesis A Rippi viii
Plaxis 2D (2015) features 147
B1 Mohr Coulomb failure criterion 147
B2 φ-c Reduction Technique 150
B3 Initial Stress Generation 151
B4 Interface Strength 151
Appendix C NEN 6740 - Table 1 153
Appendix D 155
Input Files for the Reliability Analysis 155
Appendix E 159
Characteristic and mean values 159
Appendix F 161
Reliability methods 161
F1 Generation of random samples in Monte Carlo 161
F2 Other Sampling Methods 162
F3 First Order Second Moment (FOSM) Method 165
MSc Thesis A Rippi ix
MSc Thesis A Rippi x
List of abbreviations
CDF Cumulative Distribution Function
CoV Coefficient of Variation
CUR Civieltechnisch Centrum Uitvoering Research en Regelgeving
DS Directional Sampling
FAST Fourier Amplitude Sensitivity Analysis
FEA Finite Element Analysis
FEM Finite Element Model
FERM Finite Element Reliability Method
FORM First Order Reliability Method
FOSM First Order Second Moment
LEM Limit Equilibrium Method
LRFD Load and Resistance Factor Design
LSF Limit State Function
MC Monte Carlo
OT OpenTURNS
PDF Probability Distribution Function
RFEM Random Finite Element Method
RS Response Surface
SA Sensitivity Analysis
SLS Serviceability Limit State
SORM Second Order Reliability Method
ULS Ultimate Limit State
VNK Veiligheid Nederland in Kaart
MSc Thesis A Rippi xi
The roots of education are bitter but the fruit is sweet
Aristotle
MSc Thesis A Rippi 1
1 Introduction
In the Netherlands according to the Flood Protection Program (Hoogwater
Beschermingsprogramma1) and the Room for the River program (Ruimte voor de Rivier2)
alternative structural techniques for the reinforcement of existing dikes or for future dike
constructions additional to the conventional ones have been introduced and tend to be
attractive The heightening of the dike crest and the construction of a stability berm are some of
the most common current actions for dike strengthening Additionally filter layers geotextiles
and drainage systems can be applied in advance in order to prevent several failure mechanisms
Recently cantilever or anchored sheet pile walls and diaphragm walls tend to be used as an
alternative procedure for embankment reinforcing (see Figure 11) Such an alternative can be
chosen for an improvement of the macro stability of the dikersquos inner slope and additionally for
saving space in the land area that needs to be habited
Figure 11 Design options for dike reinforcement (source Flood Defences 2015)
For that purpose engineers need a concrete and unequivocal design methodology for such
combined structures In the meantime Deltares in cooperation with the Water Board of
Rivierenland is working on a design guideline which will be applicable to design the so-called
Type II stability fences3 A draft of such a guideline is elaborated in the report of Larsen et al
(2013) In this report suggestions and recommendations are outlined for the design of such
structures with Finite Element Modeling (FEM) and partial factors Remarks are also made for
the advantages and the limitations of such models and how they should be handled so that
reasonable and trustworthy results can be retrieved from the analysis
1 httpwwwhoogwaterbeschermingsprogrammanldefaultaspx 2 httpalbertawatercomhow-is-water-governedwhat-is-room-for-the-river 3 Type II structures are structures which in combination with a soil construction fulfil water retaining functionalities
Introduction
MSc Thesis A Rippi 1
The concept behind the recommended design criteria (Larsen et al 2013) is the definition and
the evaluation of several partial factors with the view to determine overall safety factors These
factors are further discussed in section 22 However validation of these factors shall be carried
out before being used in any case otherwise either the safety of the structure is jeopardized or
the structure will be overdesigned and cost inefficient
For improving the design criteria researchers for a long time focused on enhancing structural
models (beams shells etc) and constitutive laws (elasticity plasticity damage theories etc)
With the development of computer science a great amount of work has been devoted to
numerically evaluate approximated solutions of the boundary value problems describing the
mechanical system FEM is probably nowadays the most applied approach for the solution of
these problems
However the increasing level of detail of the constitutive models and the constant enhancement
of the computational tools do not solve the problem of identification of the model parameters
and the inherent physical and modelling uncertainties Moreover in most civil engineering
applications the intrinsic randomness of materials (soil rock concrete etc) or loads (water
elevation wind earthquake motion etc) is such that deterministic models are using average or
later on characteristic values of the properties at best lead to rough representations of the
reality
As a counteraction a semi-probabilistic methodology has been developed that was based on the
application of characteristic and design values by using partial factors Current design codes
such as Eurocode provide target reliabilities for different types of structures and structural
elements according to the potential consequences of failure However these partial factors are
not always equally suitable and efficient for all types of structural applications since they have
been calibrated under specific conditions Besides the consequences of failure of flood defences
such as dikes can be comparable to the investments in increasing the reliability of such systems
are For that reason it would be advisable for these systems to define target reliabilities based
on a risk assessment (ie tailor-made solution) rather than using the standard partial factors
coming from general geotechnical design codes which may be either too low or too high for a
given flood defence system
One step of such a risk assessment is accounting for randomness and spatial variability of the
mechanical properties of materials is one of the tasks of stochastic or probabilistic mechanics
which has developed fast in the last decade In this master thesis project the uncertainty of soil
properties is going to be treated in terms of its contribution to failure For that purpose
probabilistic methods are going to be implemented for a dike with an anchored sheet pile wall
(see Figure 12) simulated in a FEM software that is specialized in soil mechanics The successful
implementation and in future research the verification of such methods can be considered as
the most preferable and cost efficient way to design structures with high safety requirements
and not only for the validation of the partial factors Of course such a procedure tends to be
time consuming However the gradual improvement of the current probabilistic methods in
combination with the state-of-the-art computer capabilities as well as the scientific knowledge
gained in terms of different systems behaviour and failure modes can introduce a more
optimized way of designing structures with considerable investments
Introduction
MSc Thesis A Rippi 2
Figure 12 Reinforced dike section with an anchored sheet pile wall
11 Project objective and main research questions At the beginning of this research it has been observed that the design procedure being followed
for dikes with retaining walls modeled in FEM can lead to an overestimation of the critical loads
and thus to a potential cost inefficient final structure An example of the magnitude of the design
values in such a case study is given in the next chapter where also a more detailed explanation
of the current design concept is described
The objective of this study is to implement a full probabilistic analysis for evaluating the
reliability of a dike reinforced with an anchored sheet pile wall Such a research analysis can
subsequently come up with valuable recommendations for the improvement of the present
design approach Soilrsquos and structural elementsrsquo behavior should be taken into account both
separately and as a system Some of the reliability methods can deal with system reliability
problems such as Monte Carlo (MC) and Directional Sampling (DS) whereas for others like
FORM and SORM additional methods should be applied that use reliability information for each
individual limit state function to obtain the systemrsquos reliability
In principle the probability of failure of different limit states is to be computed individually
while in the sequence the system reliability is going to be estimated As far as the soil failure
mechanisms are concerned this research will focus mainly on the global stability of the dike
slope while for the sheet pile wall and the anchorrsquos failure the exceedance of the steel yield
stress is going to be considered
The scope of this thesis is not only to test different reliability models robustnessrsquo in conjunction
with FEM simulations but also to get a better insight into the specific system behavior (ie of a
dike with a retaining wall) analyze each component separately and investigate its response
under certain load conditions Furthermore the minimization of the computational effort and
time could also be carried out meaningfully under the constraint of sufficient accuracy The
accuracy should be such that the probability of failure is acceptable for the ultimate limit state
(ULS) in a normal design process and according to the current safety standards Finally the
Sheet pile
wall
Anchor
Dike section
Introduction
MSc Thesis A Rippi 3
robustness of the coupling between the reliability model and the FEM will be tested via their
capability of adapting to new input parameters without encountering convergence errors
during execution
The main research question of this master thesis project is thus formed as follows
How can the probability of failure of a dike with a sheet pile wall due to global instability modeled
by a Finite Element Model be analyzed
Essential questions regarding the soil models and failure criteria as well as the reliability
methods and the systems behavior are generated which are listed below
Subquestion 1 Which reliability methods are computationally tractable in
combination with FEM
Subquestion 2 How robust (convergence) are the tractable methods
Subquestion 3 Which limit states are relevant for the envisaged application of
retaining walls in dikes and how can they be formulated using FEM
analysis outcomes
Subquestion 4 What is the contribution of different uncertainties in the failure
mechanisms of the system
Subquestion 5 Can response surface techniques help to increase the efficiency and
robustness of the reliability model
Subquestion 6 How can the current design approach for dikes with sheet piles be
improved
In the next chapters the above research question and the related subquestions are going to be
answered by following the methodology that is described in the next section
12 Research approach and outline In this section an overview is given of how the aforementioned objective and sub-questions are
approached The thesis is mainly divided into 8 chapters and 6 appendices In Figure 13 the
thesis outline is illustrated and an indication of which subquestions (Sq1-6) are treated to
which chapter is given In Chapter 1 an introduction into this research content is made and a
first illustration of the system under investigation is presented together with the main research
questions
Chapter 2 has been devoted to the description of the system (ie a dike with an anchored sheet
pile wall) and to the case study that has been adopted for being analyzed in terms of its
reliability The forces configuration is also illustrated and the results of the case study according
to the current design concept are presented and evaluated Last but not least the new
recommended safety standards are included and the current required reliability of the specific
dike section is presented
Introduction
MSc Thesis A Rippi 4
In Chapter 3 a literature study related to FEM uncertainties and sensitivity analysis is
presented Moreover previous studies that are associated with the objective of this research are
mentioned whereas some of their results were also taken into account for proceeding with this
research
The main scope of this thesis is the implementation of different reliability methods on a specific
case study with the view to investigate and analyze its behavior Some of these methods are
continuously mentioned through the test and thus in Chapter 4 an overview is given of the
most well-known reliability methods Eventually an evaluation of these methods is made based
on literature and preliminary testing with simple case studies and the procedure of their
coupling with FEM is discussed The evaluation helped to qualitatively answer subquestion 1
while also an introduction of how special reliability methods such Response Surfaces were used
in this thesis that partially covers subquestion 5 In this chapter an introduction into the Limit
State Functions (LSF) concept is made and how they are considered for the reliability analysis of
a system An LSF actually represents a failure mode that can be detected in a structural or soil
element and it is expressed as a function of several variables In Chapter 5 the failure
mechanisms and the corresponding LSF that are related to the specific case study are identified
and formulated which answers subquestion 3
In Chapter 6 the case study whose reliability is to be evaluated is presented as it was modeled
in FEM The boundary conditions are specified and the soil and structural properties are
indicated and illustrated Moreover a deterministic design is also taking place in order to
roughly estimate the structural elementsrsquo properties This will help to make a qualitative
comparison between the current design procedure and the design according to a fully
probabilistic approach by referring thus to subquestion 6
Eventually the results of the aforementioned analysis are presented in Chapter 7 The results
mainly include the estimated probability of failure of each of the system components as well as
of the system as a whole by considering the soil properties as stochastic Moreover an
interpretation of the failure points is made and an assessment of the level of impact of the
random variables on the systemrsquos behavior is carried out At that chapter subquestions 2 4 and
5 are mainly treated
Last but not least in Chapter 8 the general conclusions are presented together with some
valuable recommendations for future consideration and research Furthermore a reflection on
the methodology and how the different research questions were approached is made
Introduction
MSc Thesis A Rippi 5
Figure 13 Thesis outline
Introduction
MSc Thesis A Rippi 6
MSc Thesis A Rippi 7
2 System description and current design concept
In this section the system to be analyzed is described more in detail and the individual
components are identified Moreover the forcesrsquo configuration is explained and illustrated as
they would have been calculated with the conventional way in parallel with a qualitative
evaluation of the potential expected deformations Finally the current design concept and the
safety standards that are referred to such a system are introduced
21 System description and forces configuration The system to be investigated in this thesis consists of a dike section reinforced with a one-layer
anchored sheet pile wall The system has been simulated by FEM and later on it was coupled
with a reliability package for carrying out a reliability analysis In the figure below an
illustration of the system is depicted and the several elements of an anchored sheet pile wall are
showed
Figure 21 System layout and different components of the anchored sheet pile wall
In particular an anchored sheet pile wall consists of the sheet piles which are anchored in the
soil via a tie rod The rod is grouted most of the times in a sandy layer with a cemented grout
body and it both alleviates the sheet piles from the axial forces exerted by the upper structure
and keeps the wall stable in case of excessive developed moments due to the lateral earth
pressure Last but not least a waling system is applied mainly downstream of the sheet piles in
order to transfer the loads from the piles to the anchors in such a manner so as to avoid
excessive local stresses on the intersection between the sheet piles and the tie rod
Cross-Section
Top View
System description and current design concept
MSc Thesis A Rippi 8
The main load configuration acting on a dike section as well as a sheet pile wall is depicted in
Figures 22-24 together with the possible displacements As far as the soil body stability is
concerned most of the slope stability analysis computer programs are based on the limit
equilibrium concept according which a soil mass tends to slide down under the influence of
gravity The transitional or rotational movement is considered on an assumed or known slip
surface below the soil There an equilibrium should be achieved among the driving and the
resisting forces In that case the driving moments consist of the soil weight the water pressures
and the loads around the center of the slip surface such as a possible vertical load on the top of
the dike crest (ie traffic load) The magnitude of the water pressures is controlled by the water
elevation on the river side
In Figure 22(a) the external and internal forces acting on a slice of the slip surface are
illustrated The driving forces are the soil weight W the water pressure U and any additional
load that can contribute to the rotation of the slip surface which in turn activate the lateral
active earth pressure The resisting forces consist of the lateral passive earth pressure
(Rankine 1857) and the shear strength (Terzaghi 1943) Their expressions are written as
follows
[kN] (21)
[kN] (22)
[kPa] (23)
where is the active lateral earth pressure coefficient and the passive lateral earth
pressure [kNm3] is the unit weight of the corresponding soil layer and H [m] is the thickness
of the soil layer that the lateral earth pressure comes from (in case of multiple soil layers there
should be a distinction among the different forces) and [kPa] is the effective cohesion of the
soil layer The safety factor that is written in the Figure 22(a) as FS is defined by the ratio of the
shear strength (excluding the pore water pressure) divided by the weight of the earth body
(including the pore water pressure)
The loss of equilibrium between the driving and the resisting moments lead to the rotation and
instability of a slip surface as it is illustrated in Figure 22(b) A slide surface can take various
shapes according to the Limit Equilibrium Method (LEM) that is followed According to Bishop
method this surface tends to be circular in Uplift-Van method a horizontally compressed zone
can be also considered whereas in Spencer method the shape of the slide body can be arbitrary
The slip surface can be located either in the landside or in the river side of the dike depending
on the load the soil characteristics and the design of the dike (ie inner or outer berm
reinforcing revetment on the outer slope etc) Most of the times the landside area of the dike is
jeopardized due to different failure mechanisms that are described in section 523 The
instability of the outer slope is mainly relevant in case of sea dikes and a potential damage can
be usually reconstructed until the next flooding
System description and current design concept
MSc Thesis A Rippi 9
(a)
(b) Figure 22 (a) External and internal forces acting on a slip surface of a dike section and (b) possible deformation pattern
As a counteraction for the slope instability a sheet pile wall can be installed inside the dike The
special thing about these structures is that they make possible a greater freedom in form and
functionality than a traditional dike design This structure derives its strength from the
materials used such as steel which are able to withstand higher pressures than clay for instance
The general stability is due to friction and wedging in the bottom
After the reinforcement of the inner side of the dike with a retaining wall the strength capacity
of the wall is also important for the global stability of the system The forces to be taken into
account for the sheet pile wall stability are the active and the passive earth pressures (effective
pressures) and the water pressure as they are illustrated in Figure 23(a) In that figure the
System description and current design concept
MSc Thesis A Rippi 10
forces were simplified in a singular triangular shape as in a homogeneous soil body in order to
explain and depict the overall picture of the acting forces However the stresses distribution
over depth can be more complex depending on the variety of the soil layers that are present In
Figure 23(a) it can be noticed that a sheet pile wall which is installed close to the inner berm
might not have significant instability issues as there is the passive side that contributes to
resisting forces However in the passive side the soil can differ and be weaker than this of the
active side Therefore the passive force in that case might not be very supportive and thus the
wall shall be designed cautiously
Additionally the anchor resistance shall be reassured for the sake of the stability of both the
sheet pile wall and the overlying soil body In Figure 23(a) the forces acting on the anchorage
are displayed The distributed load qs comes from the earth pressure of the upper soil layer (it
might not be uniformly distributed as it is depicted in the figure but more in a trapezoidally type
of load) whereas the tension force Ft4 stems both from the earth pressure and the displacement
of the sheet pile wall The bending and the tension capacity of the anchor are determinant for its
stability In Figure 23(b) a possible displacement pattern of the retaining wall and the
anchorage is illustrated
Furthermore corrosion is an additional weakening impact on the wall that depends on the
water level and the pore water pressures near the structure The thickness of the wall and the
material properties are playing a key role to the resistance towards corrosion This will not be a
subject of this thesis however it should be taken into account in case of design purposes
(a)
4 where O is the average circumference of the pile P [kPa] is the shaft friction at depth z and L is the length of the pile
System description and current design concept
MSc Thesis A Rippi 11
(b) Figure 23 (a) Mobilizing forces acting on the sheet pile wall and the anchorage and (b) possible deformation pattern
It should be mentioned that as far as the pore pressures are concerned in Figure 23(a) only the
hydrostatic pore pressures are depicted Generally the pore pressures (or active pore pressures)
are the sum of the steady state pore pressures and the excess pore pressures
[kNm2] (24)
Steady state pore pressures can be hydrostatic (based on a horizontal phreatic line) or non-
hydrostatic (based on steady state groundwater flow) Excess pore pressures are based on
loading of undrained soils In FEM these various pore pressures are taken into account
automatically In Figure 24 an example of a potential distribution of the active pressures
nearby the sheet pile wall is depicted as it was deduced from Plaxis Therefore as it can be
noticed from the figure the distribution can indeed sometimes approximated as triangular
System description and current design concept
MSc Thesis A Rippi 12
Figure 24 Example distribution of the active pore pressures acting nearby the sheet pile wall
The knowledge of the type of forces that are exerted on both the structural components and the
soil body as well as of the most expected form of displacements is valuable in order for
someone to be able to evaluate the results deduced from FEM and detect potential modelling
errors Therefore this validation was necessary before starting with the reliability analysis
In the next sections the current design approach of the system described above is elaborated
Additionally the Dutch recommended safety standards are presented for this type of structures
which show the need of carrying out a reliability analysis For that purpose a case study was
adopted that it is presented first as it was designed according to the current regulations while
finally a comparison is made between the original and the new case study as it was found to be
according to the reliability analysis
22 Current design concept In the Netherlands dike reinforcement is becoming a major issue concerning the design of flood
defenses Special reinforcing structures such as cofferdams anchored sheet piles and diaphragm
walls are used for strengthening the dike Especially in the context of the Flood Protection
Program and the Room for the River program many dike reinforcement projects have been
suggested
In the report of Larsen et al (2013) the design procedure of a dike reinforced with sheet pile
wall is prescribed using FEM The safety philosophy that is followed in this technical report
actually composes the current procedure for designing dikes with sheet pile walls and it is
linked with the usage of partial safety factors Particularly the required overall safety factor
FEM that should be compared with the one from FEM calculations is determined as follows
SFEM
b d m n
(25)
where
System description and current design concept
MSc Thesis A Rippi 13
Partial safety factor indicating the uncertainties of the soil composition and
the water pressures (also called schematization factor)
Partial safety factor which is related to the calculation model and the way the
calculations have been carried out (also called model factor)
Partial safety factor which is related to the material parameters (also called
material factor)
Partial safety factor associated with damage caused during the soil tests (also
called loss factor)
Safety factor of load
Each of these partial factors shall be at first calibrated by conducting a full probabilistic analysis
of the system under consideration In most of the cases these partial factors have been already
calibrated on previous similar projects and then they are reused for any similar case In
Schweckendiek et al (2013) a new approach of using partial factors for flood defences is
proposed whose application is still under consideration At the moment the aforementioned
partial factors for dike reinforced with sheet pile walls modelled in FEM are estimated
according to mostly engineering judgement and political negotiations
In sequence the FEM design procedure that is applied for sheet pile walls inside soil structures
is based on a so-called ldquosafety calculationrdquo (for information for this type of calculation see
Appendix B2) In this calculation an artificial reduction to the soil strength parameters (friction
angle and cohesion) is applied in order to evaluate the overall safety factor and confirm if this is
lower or higher than the required one that was estimated according to Eq 25 Furthermore the
resulting moments and forces developed on the structural elements during the specific
calculation are used for their design
This is a quite helpful and indicative calculation type available in Plaxis that can give a sense of
the structurersquos safety factor and the possible failure mechanisms that can occur under the
predefined load configuration However such a method of calculating the developed stresses
can be misleading in terms of the moments and forces acting on the structural elements and the
total deformations of the system
In this thesis the reliability of a dike with an anchored sheet pile wall is going to be evaluated
and in particular the case study of Breedeveld (2011) is to be used for that purpose In Figure
26 the location of the existing dike section is presented while in Figure 25 the structure as it
was modelled in FEM is illustrated It is essential to mention that the retaining wall does not
exist in reality but it was placed so as to implement and demonstrate the current design
regulations in the report of Breedeveld (2011) The anchorrsquos angle has been taken equal to 30deg
from the vertical which is generally a steep anchor inclination That is expected to reduce the
part of the horizontal force that is undertaken by the anchor and thus the most of it exerts on
the sheet pile wall The angle is usually taken 30deg-45deg from the horizontal axis (CUR 2005) but
this is of course depended on the load conditions and the construction requirements for each
System description and current design concept
MSc Thesis A Rippi 14
case The influence of the anchor orientation and the construction methods that are generally
followed for an anchored sheet pile wall are not considered in this thesis The case study was
taken as it was modelled in the aforementioned project and the reader should be aware that this
serves the proceedings of a reliability analysis and not of a perfectly constructed retaining wall
Figure 25 Original dike section in Varik-Heesselt as it is presented in Breedeveld (2011)
In particular based on this case study and trying to understand the design criteria a
comparison was carried out between the ldquosafety calculationrdquo and a normal ldquoplastic calculationrdquo
(ie elastoplastic drained or undrained analysis) In both cases the boundary conditions the
external loads and the input parameters are the same The difference between them is the
calculation procedure followed in order to come up with final stresses which in the case of the
ldquoplastic calculationrdquo does not include a reduction of the soil strength properties In Figures 27
and 28 the results from both calculations are presented In Figure 27 the total deformations
are exhibited while in Figure 28 the moments and the axial forces on the sheet pile are shown
Figure 26 Location of the Varik-Heeselt dike section (section number TG-078) in the Dike ring 43
BetuweTieler-en Culemborgerwaarden west (source Breedeveld 2011)
System description and current design concept
MSc Thesis A Rippi 15
As it is realised from this figure the discrepancy between the two results is quite high while as
far as the deformations are concerned no realistic conclusions can be drawn from a ldquosafety
calculationrdquo Moreover taking as design moments the moments that are deduced from the
ldquosafety calculationrdquo will lead to a much heavier and more expensive sheet pile cross section than
the ldquoplastic calculationrdquo
It is also essential to mention that the required safety factor was estimated up to FEM =18
according to the partial-factor procedure that is defined in Eq 25 Such a safety factor as
prerequisite for coming up with the design values of the moments and forces of the sheet pile
and the procedure of the ldquosafety calculationrdquo in FEM can lead to a conservative and financially
inaccessible design
(a) ldquoPlastic calculationrdquo Deformed mesh and maximum deformations (umax= 008 m)
(b) ldquoSafety calculationrdquo Deformed mesh and maximum deformations (umax= 9041103 m)
Figure 27 Deformed mesh and maximum deformations in the end of (a) plastic calculation and
(b) safety calculation
30deg
System description and current design concept
MSc Thesis A Rippi 16
(a) ldquoSafety
calculationrdquo
Mmax = 9392 kNmm Nmax= -4402 kNm
(b) ldquoPlastic
calculationrdquo
Mmax = 6833 kNmm Nmax= -1372 kNm
Figure 28 Bending moments and axial forces exerted on the sheet pile during the (a) safety
calculation and (b) plastic calculation
On the other hand the aforementioned design procedure guaranties a strong structure capable
of probably undertaking more than the expected loads and thus ensuring the safety against
flooding of the landside However the knowledge on fully probabilistic methods that has been
obtained as well as the development of the technology can introduce a reliability analysis of the
system as a preferable way for the future design concept Such a procedure should be
accompanied also by field test that would be able to validate FEM results as well as in-situ
measurements and inspections of the soil properties for calibrating their statistical
characteristics
Due to the conservative results that the aforementioned procedure came up with and because of
the simplifications that this case study was later subjected to in order to be used in this thesis a
new design was carried out in Chapter 6 and the properties of the structural elements were
redefined
The inclination towards probabilistic methods has not only been created due to the incomplete
design regulations for the system under consideration but also the current proposed safety
standards which are discussed in the section below recommends a risk based safety assessment
of the primary flood defenses which in principle implies the evaluation of the failure probability
In the next section an overview of the new recommended safety standards related to primary
flood defenses is given and the required reliability for the stability of the aforementioned case
study is elaborated
System description and current design concept
MSc Thesis A Rippi 17
23 Safety standards An important factor for the evaluation of a structurersquos reliability is the safety standards that
have been set and according which the acceptance or not of the failure probability is made After
the safety assessment5 of 2011 it has been declared that 33 (1225 km) of the primary flood
defenses (3767 km) in the Netherlands does not comply with the safety standards For this
purpose a new session of investigations has commenced in order to assess the reliability of the
existing dikes and the potential amendment of the safety standards for future constructions
Taking into account this latest information and within the framework of the Flood Risk in the
Netherlands 2 project6 (VNK2) a new Delta program was founded whose one of the main
decisions was the ldquoDelta Safety Decisionrdquo This decision contains a proposal for new safety
standards of the primary dikes as they are depicted in Figure 29 which are stricter than the
previous ones and formulated in terms of failure probability
According to this project a new policy regarding the safety assessment of the flood defenses has
been settled and proposed for the future design criteria which is based on the evaluation of the
acceptable flood risk (probability of failure (x) consequences) of each dike ring rather than the
probability of exceedance in order to achieve a level of protection that is in balance with the
societal value (Cost Benefit Analysis and life loss)
VNK has been already using reliability evaluation techniques in order to calibrate partial safety
factors Especially in geotechnical engineering the high inherent uncertainty of the soil
properties renders the evaluation of the structural reliability essential for the safety assessment
of the structure Subsequently this implies the investigation of the response of the different
reliability methods applied in a real case study of a dike The techniques that are used to
evaluate the probability of failure are discussed thoroughly in Chapter 4 and in Appendix F In
case of a complicated structure such as a dike with a retaining wall both the reliability of the
dike body and the reliability of the wall shall be evaluated and eventually a common standard
for the reliability of the system shall be defined
5 A safety assessment is an evaluation of the function of the flood protection structure and it is being conducted every approximately 6 years after the reassessment of the existing hydraulic boundary conditions 6 in Dutch Veiligheid Nederland in Kaart 2 or VNK2 It has been initiated by the Ministry of Infrastructure and the Environment (IampM) the Association of Regional Water Authorities (UvW) and the Association of Provincial Authorities (IPO)
System description and current design concept
MSc Thesis A Rippi 18
Figure 29 Maximum admissible flooding probability for primary defences according to Delta
programme 2015 (Deltaprogramma 2014)
As far as the retaining walls are concerned CUR (2005) design procedure distinguishes the
following three safety classes for retaining walls with corresponding reliability indexes
Table 21 Safety classes and corresponding reliability indices (CUR 2005)
Class I Relatively simple constructions no personal safety risks and relatively minor
damage in the case of overall failure ϐ=25
Class II Considerable damage in the case of overall failure minor personal safety
risks ϐ=34
Class III Major damage in the case of overall failure andor considerable personal
safety risks ϐ=42
System description and current design concept
MSc Thesis A Rippi 19
Additionally as far as a dike section is concerned the required overall reliability index shall be
concluded according to the new norms that are shown in Figure 29 However the reliability
index of the different failure mechanisms of a dike shall be estimated separately
According to Vergouwe et al (2014) that is part of the VNK2 report the major failure
mechanisms in dike ring 43 where the case study under investigation is located (see Figure 26)
is uplift and piping and macro-instability of the inner slope Especially in Figure 210 the
estimated percentage of the contribution of each failure mechanism to the overall probability of
flooding7 is indicated In Vergouwe et al (2014) the overall flooding probability of the specific
dike section that the case study concerns was evaluated to 17800 Therefore the macro-
instability on which this thesis is concentrated with a contribution percentage of 226 has a
probability of about 310-5 (
) and thus a target reliability index β of
approximately 4 This reliability level can be considered as a benchmark for evaluating the one
that will be entailed in the end of this thesis
Figure 210 Percentage contribution of the different failure mechanisms to the probability of flooding (source Vergouwe et al 2014)
It should be mentioned though that the aforementioned safety standards concerning the dike
ring 43 were defined according to the probability of exceedance of a certain water level that the
dike has to retain Therefore this complies with the old norms and not with those appeared in
Figure 29 However for large engineered systems such as flood defence systems with large
potential consequences and substantial investments it is worthwhile to assign target reliability
levels based on a risk assessment of the area surrounded by the specific dike ring There are
three widely used types of criteria for evaluating the risks related to floods and major industrial
hazards (Vrijling et al 2011)
Economic criteria
Individual Risk Criteria
Societal Risk criteria
7 The dependencies among the failure mechanisms were not taken into account Therefore this probability stated in Vergouwe et al (2014) is not exactly equal to the probability of flooding
System description and current design concept
MSc Thesis A Rippi 20
In order to check then if a certain system abides by the target reliability that is required the
failure probability of each failure mechanism shall be calculated based on a LSF The
aggregation of the failure probabilities of all the modes gives the failure probability of the
system which is supposed to be compared with the required in order to reassure the reliability
of the structure Therefore this thesis is dealing with the calculation of this failure probability of
a certain system modelled in FEM analysis and understanding of the systemrsquos behaviour and
finally the investigation of the robustness of coupling FEM with reliability methods for
evaluating the reliability of a structure In the next chapter an overview over the research has
been done until now concerning coupling and reliability methods applications is presented and
a discussion over the most important is made
MSc Thesis A Rippi 21
3 Literature study
In this chapter an overview of some principles related to FEM and the uncertainties in
geotechnical engineering is presented More precisely an introduction into FEM concept and a
discussion over the different types of FEM is made Moreover the uncertainties in geotechnical
engineering are stressed and the general framework according to which they are handled is
described Finally previous studies that are associated with the application of FEM on soil
structures and the reliability analysis are discussed
31 Background
311 Finite Element Modeling
FEM is a numerical method whose essence is to convert a problem described by partial
differential equations over space and time into one by dividing the space-time continuum into a
set of discrete elements and assuming that the unknowns vary over each element FEM solution
process is as follows
1 Divide structure into pieces (elements with nodes) (discretizationmeshing)
2 Connect the elements at the nodes to form an approximate system of equations for the
whole structure (forming element matrices)
3 Solve the system of equations involving unknown quantities at the nodes (eg
displacements)
4 Calculate desired quantities (ie strains and stresses) at selected elements
The properties of each element are set the same as the material properties that have been
defined by the user Then the Deterministic Finite Element Method (the finite element method
that was formulated with deterministic variables) can be used in conjunction with means and
standard deviations of the input variables to obtain reliability estimates
In the simple coupling of FEM with reliability analysis each parameter that is considered as
stochastic is given a particular probability density function estimated either by field tests or by
engineering judgement This approach is also called ldquosingle random variablerdquo and it assumes
that the spatial correlation length is infinite In other words the soil is considered to be
homogeneous and the stochastic property assigned to the soil is taken at random from a certain
probability distribution
Nevertheless a more realistic model should take into account the spatial correlation within
smaller regions where the stochastic property is allowed to vary For that purpose the Random
Finite Element Method (RFEM) was introduced by Fenton and Vanmarcke (1990) in which the
random variables are correlated to one another using auto-correlation functions
The most sophisticated method of FEM is the Stochastic Finite Element Method introduced by
Ghanem and Spanos (1991) which accounts for the uncertainties in the input parameters
Literature study
MSc Thesis A Rippi 22
implicitly within the finite element calculation This aims at representing the complete response
probability distribution function (PDF) in an intrinsic way Two steps are basically applied for
that purpose
Discretization of the problem and definition of the random vector with an unknown
joint PDF and
Expansion of the response in a particular basis of random vectors with a finite variance
called the polynomial chaos
In this thesis the Deterministic Finite Element Method is to be used coupled with reliability
analysis According to Waarts (2000) coupling FEM with reliability methods (FERM) would lead
to the following advantages
In comparison to standard finite element analysis it gives direct insight into the
structural reliability and decisive parameters
Structures designed using FERM will either be safer orand more economically built in
comparison to structures designed using safety factors and classical constitutive models
FERM is expected to be useful for the determination of LSF that cannot be explicitly formulated
and that differ in each case such as soil limit state in different structural schematizations
Moreover it is likely to be valuable in areas where little knowledge exists on the systems
reliability of structures with multiple components (ie a dike with a sheet pile wall)
FEM in practice
The conventional method for stability analysis in a soil body is represented by LEM although
FEM is increasingly used by designersresearchers The latter has been proved to be quite
realistic for the progressive behaviour (ie stress-strain development in different construction
phases) of a soil system under the effect of stress redistribution in comparison with classical
models Especially in their master thesis Johansson amp Sandeman (2014) compared the
deformations and the forces measured at a deep excavation supported by anchored sheet pile
wall in a railway tunnel located in Gothenburg in Sweden with 1D finite element software the
2D finite element software (Plaxis) and hand calculations They proved that Plaxis produces
reliable results for horizontal deformations in the sheet pile wall and anchor forces when
compared to in-situ measurements
Moreover in Gonzaacutelez et al (2013) it was proved that the simplification of reality done by the
classical methods such as Blumrsquos Engelrsquos Kreyrsquos methods etc (for further information for these
methods a reference is made to Gonzaacutelez et al (2013)) allows us to generally understand the
behaviour of the system wall-soil Nonetheless the results that came out of this analysis were
found to be quite conservative whereas FEM managed to give a more realistic interpretation of
the wallrsquos movement
Seed et al (2006) investigated the performance of the New Orleans Flood Protection Systems
during hurricane Katrina in August 2005 by simulating the different levee sections with FEM As
an example in Figure 31 pictures of the failure of the Lower Ninth ward IHNC East Bank in
New Orleans after the hurricane Katrina are presented FEM was proved to be an efficient tool
to realistically interpret the shape and the triggers of the failure mechanism
Literature study
MSc Thesis A Rippi 23
Figure 31 Failure of the Lower Ninth ward IHNC East Bank in New Orleans modelled in FEM with Plaxis (source Seed et al 2006) At the bottom right the shear strain contours are depicted that are developed through the embankment and at the bottom left the deformed mesh is presented illustrating the failure mode of the levee
In recent implementation of design concepts and technical recommendations such as CUR 166
ndash Damwandconstructies (2005) which refers to the design of retaining walls it can be observed
that guidelines for Finite Element Analysis (FEA) have been set parallel to the existing analytical
or empirical calculation methods Therefore FEA tends to become more and more accepted as
an alternative for Serviceability Limit State (SLS) as well as for ULS design
According to Wyllie and Mah (2004) LEM give an estimation of the factor of safety with no
information on deformation of the slope However in numerical analysis the failure surface can
evolve during the calculation in a way that is representative of the natural evolution of the
physical failure plane in the slope In that way a better insight into the evolution of failure
mechanisms can be gained
Cundall (2002) compared the characteristics of numerical solutions and LEM in solving the
factor of safety of slopes and concluded that continuum mechanics-based numerical methods ie
FEM have the following advantages
No pre-defined slip surface is needed
The slip surface can be of any shape (in contrast with Bishop and Kranz stability)
Multiple failure surfaces are possible
No static assumptions are needed
Literature study
MSc Thesis A Rippi 24
Structures (such as footings embankments etc) and structural elements (such as
beams cables sheet piles etc) and interface can be included without concern about
compatibility
It is also important to recognize that LEM only identifies the onset of failure whereas FEM
includes the effect of stress redistribution and progressive failure after its initiation Numerical
models can also be used to determine the factor of safety of a slope in which a number of failure
mechanisms can exist simultaneously or where the mechanism of failure may change as
progressive failure occurs (Hoek et al 2000)
On the other hand the accuracy of FEM is dependent on usersrsquo settings such as the refinement
of the mesh generation the loading steps etc Moreover even if FEM can be functional and easy
to use it is essential that the user has a deep knowledge of the subject under investigation (soil
and structural mechanics) in order to be sceptical and critical with the FEM results able to
interpret the output behaviour of the structure and changecorrect everything that seems
peculiar Moreover the usage of FEM implies also some basic knowledge on numerical methods
and the general background of the FEM so as to solve possible numerical errors
An ideal way of modelling in FEM would be a prior calibration of the program according to data
related with the structurersquos properties and performance (ie stress generation) However this
means that enough field measurements should be carried out and under specific conditions in
order to be considered as a base for FEM calibration Unfortunately such field tests are not
always available and if so their reliability is on doubt For example measurements related with
the developed moments or displacements of a retaining wall inside a dike section are rare to be
found but even if there are some available they cannot represent the situation near the failure
domain where most of the engineers are worried about
312 Uncertainties and Sensitivity analysis
Uncertainty in geotechnical engineering can be categorized into aleatoric epistemic and
decision model uncertainty as they are depicted in Figure 32 (Baecher amp Christian (2003))
Aleatoric uncertainty is associated with the natural variability which is defined as the intrinsic
randomness of natural processes Such variability can be expressed by changes of the soil
properties over time at a certain location (temporal variability) or over space at a single time
(spatial variability) Epistemic uncertainty can be divided into the site characterization model
and parameters uncertainty and it is attributed to lack of information about events or lack of
understanding the physical laws that prohibits the ability of creating a realistic model Finally
the decision uncertainty describes the difficulty of being aware of social objectives defining
social values such as discount rates and predicting the planning horizon
Literature study
MSc Thesis A Rippi 25
Figure 32 Categories of uncertainty in geotechnical engineering (modified from Baecher amp
Christian (2003))
Below the general steps of an uncertainty study are described and an introduction into the
sensitivity analysis concept is made
Global methodology of an uncertainty study
A first step of an uncertainty study8 can be described as ldquothe definition of the problemrdquo Initially
the variables of interest (or else the output variables) of which the uncertainty is to be
quantified shall be specified In sequence given several input variables for which the user may
have data andor expertengineering judgment a model denoted usually by a mathematical
function should be introduced that enables the computation of the set variable of interest
After the general context has been staged we should choose the criteria with which the
uncertainty can be evaluated The most complete measure of uncertainty when dealing with a
random vector is the probability distribution In order to assess the value of this distribution
function the following criteria can be followed
Probability of exceeding a threshold the aim is to assess the probability that the variable
of interest exceeds a threshold important for the goals at stake
Quantiles the aim is to assess the threshold that a variable of interest may exceed with a
probability equal to a given value
Central dispersion the aim is to obtain an ldquoorder of magnituderdquo of the uncertainty by
specifying the average value and the variance of a variable of interest
8 In the engineering world this is also called reliability analysis However in this section the general uncertainty study is described In particular reliability describes the ability of a system or component to function under stated conditions for a specified time period
Literature study
MSc Thesis A Rippi 26
The next step is to define a model to represent and quantify the uncertainties of the input
variables One shall investigate each variable as a singularity and come up with the most
suitable probability density function (mostly depending on available data) Besides it is
essential to assess also the potential correlations among the variables that can be included in a
joint probability density function which is discussed later in this section
Eventually the uncertainties of the input variables shall be translated in terms of uncertainty on
the variables of interest This procedure is called uncertainty propagation and can be carried
out via several reliability methods (approximation methods or sampling methods) that are
extensively elaborated in Chapter 4 In Figure 33 a schematization of the different steps during
an uncertainty study is given
Figure 33 Methodology of uncertainty study
Last but not least a better understanding of uncertaintiesrsquo influence can be achieved by
analyzing the contribution of the different uncertainty sources to the uncertainty of the
variables of interest via a sensitivity analysis Such an analysis aims at identifying important
parameters for the system response besides it gives a better insight into the model used to
quantify the uncertainties Before conducting an uncertainty analysis it is advisable to filter out
parameters of less significance in order to reduce the modeling computational effort Below
different methods for sensitivity analysis are described
Sensitivity analysis
Sensitivity analysis (SA) is the study of how the variation in the output of a model can be
apportioned quantitatively and qualitatively to the variation in the model parameters (Saltelli
et al 2004) Saltelli (2004) proposes one possible way of grouping these methods into three
classes (i) screening methods (ii) global SA methods and (iii) local SA methods
i Screening methods
Screening is a particular instance of a sampling-based method The objective here is to identify
which input variables are contributing significantly to the output uncertainty in high-
dimensionality models rather than exactly quantifying sensitivity (ie in terms of variance)
Screening tends to have a relatively low computational cost when compared to other
Literature study
MSc Thesis A Rippi 27
approaches and can be used in a preliminary analysis to weed out less influential variables
before applying a more informative analysis to the remaining set One of the most commonly
used screening methods is the elementary effect method
ii Global SA methods
Global SA techniques incorporate the whole range of variation and the probability density
function of the input parameters to calculate their influence on the output Many global
sensitivity analysis techniques are now available such as
Variance-based methods
o Sobolrsquos method9
o Fourier Amplitude Sensitivity Test (FAST)10
o Analysis of Covariance (ANCOVA)
Regression-based methods11
Both FAST and Sobolrsquos method rely on the assumption of parameter independence while
ANCOVA can also handle correlated input parameters The main principles of FAST method are
presented in Appendix A1 For more information about the other methods the reader can
betake himself to the related references
iii Local SA methods
Local SA methods provide the slope of the calculated model output in the parameter space at a
given set of values (Turanyi and Rabitz 2004) This basically means that local methods involve
taking the partial derivative of the output Y with respect to an input factor Xi |
|
where the
subscript indicates that the derivative is taken at some point in the space of the input
Examples for these are (Schweckendiek 2006)
the α-values in a FORM calculation (further discussion in section 422)
Local Probabilistic Sensitivity Measure
In reliability analysis the local sensitivities will be more important than the global ones in the
sense that the influence of all variables in specific points such as the design point cannot be
identified by the global methods However local SA can only inspect one point at a time and the
sensitivity index of a specific parameter is dependent on the central values of the other
parameters
In this thesis the influence of the different parameters on the response of the system is
evaluated according to local sensitivity indices However before the main part of the reliability
analysis starts it was necessary to filter out the less influencing variables in order to reduce the
number of the major variables and make the analysis more efficient and computationally
affordable For this purpose the global sensitivity method FAST was used whose results are
presented in Chapter 7
9 created by Sobolrsquo (1993) 10 introduced by Cukier et al (1973) 11 introduced by Helton et al (2005)
Literature study
MSc Thesis A Rippi 28
Probabilities as a measure of uncertainties
Most engineers today use the concept of probabilities as the standard way to talk about
uncertainty One way to quantify a probability of a random variable is the calculation of the
cumulative probability function (CDF) of non-exceedance which can be obtained by
the probability distribution function For a random vector this reads
( ) (11)
From this the joint probability density function of this random vector can be determined as
( )
( )
(12)
This function is depicted in Figure 34 by means of contour levels The joint probability density
function has the shape of a hill as it is showed in the 3D figure and the inclination of the ellipses
reveals a correlation between the two variables and
Figure 34 Joint probability density function in 2D and 3D
The functions that join or couple multivariate distribution functions to their one-dimensional
marginal distribution functions are called copulas These are going to be used in order to define
the joint density probability distribution from which the random variables are taking their
values However no detailed explanation is made in this report regarding copulas and its
characteristics as it is automatically implemented in OT whereas the user has to define only the
type of copula that heshe considers the most appropriate (a reference is made to Nelson (1999)
for further information) According to the type of copula that is selected the order of the values
of the variables can be chosen within an iterative reliability method In this thesis an
independent copula was used considering that the input variables are independent the ones
from the others due to the time limit
Literature study
MSc Thesis A Rippi 29
32 Previous Studies In this section a summary of the most relevant researches on the topic of FEM applied in soil
structures and coupling with reliability methods is given
To begin with an attempt to estimate the reliability of a structure modeled in FEM has been
made by Waarts (2000) by introducing an optimized reliability method in terms of
computational effort and efficiency
In particular in Waarts (2000) two adopted reliability methods are introduced both making
use of a response surface (a detailed explanation of the reliability methods and the response
surface techniques is made in Chapter 4 and Appendix F) These adaptive response surfaces are
used in combination with FORM and DS respectively The accuracy and the effectiveness of
these methods are investigated on the basis of artificial LSFs and a comparison is made with the
existing standard reliability methods The most efficient combinations of response surface
techniques and reliability methods were with FORM (FORM-ARS) and DS (DARS) Comparing
these two methods DARS predominated over FORM-ARS as it can cope with a much wider
range of limit state functions
In a later stage the above best performing reliability method (ie DARS) is further investigated
in terms of its efficiency on the basis of complex structures reliability In Figure 35 a couple of
case studies used to verify the performance of DARS are given
Figure 35 Left Two-story two-bay frame structure Right overview of a sheet pile installation
(source Waarts 2000)
The conclusions of this research showed that DARS serves its purpose and that the benefit from
using it increases with increasing number of random variables In Figure 36 the Limit State
Function Evaluations (LSFE) that are carried out as a function of the number of variables are
presented
Literature study
MSc Thesis A Rippi 30
Figure 36 Number of LSFE as a function of the number of variables n The upper line depicts the
required number of LSFE N=160 (it is supposed that on average 3 LSFE per sample are required
and consequently the number of LSFE equals 3x160n=480n) The lower line shows the
performance of DARS (source Waarts 2000)
More recently a methodology for assessing the reliability of geotechnical structures and in
particular deep excavations has been developed by Schweckendiek (2006) More precisely this
study deals with the application of different reliability methods in combination with FEM which
carries out the LSFE The reliability methods are applied initially on simple examples in order to
be calibrated and eventually the most efficient methods are used for the reliability assessment
of a deep excavation with a sheet pile wall The several failure mechanisms are treated in detail
concerning the structural elements of the excavation as well as the soil medium
Finally the failure probability of the system is estimated according to a fault tree
schematization of a deep excavation with one layer anchored sheet pile wall Uncertainties in
the soil properties the phreatic levels and also the corrosion of the sheet pile wall were taken
into account In Figure 37 a picture of the case study is shown together with a FEA illustration
and the importance factors of different soil properties in terms of the probability of failure after
a FORM analysis
Literature study
MSc Thesis A Rippi 31
Figure 37 Finite Element simulation and contribution of several structural and soil properties in
the failure mechanism of an anchored retaining wall in layered soil (source Schweckendiek 2006)
For a similar case study as in Schweckendiek (2006) a cantilever and an anchored sheet pile
wall the safety of the structure based on LEM is evaluated in a research conducted by Gonzaacutelez
et al (2013) Moreover in order to compare the efficiency of the classical methods with FEM
the finite element software Plaxis has been used This research shows the limitation of the
classical models and the efficiency of FEM as far as an anchored sheet pile wall in concerned
More precisely LEM results found to be conservative as they do not take into account the
confinement around the free length of the anchor rod that increases the passive pressure After
comparing numerical results with the classical methods in anchored walls the failure
mechanism given by the FEM suggests an intermediate failure state where the sheet pile wall
describes a translation movement on the deep zone and at the same time a rotation movement
around the anchor point that is closer to the reality
Furthermore Moumlllmann and Vermeer (2008) conducted a reliability analysis of a dike failure A
case study at river Elbe in Saxony Germany has been used and the failure probabilities of
different dike cross sections involving different failure modes were compared More precisely
overflowwave overtopping upliftpiping slope stability and damage of the revetment on the
waterside were taken into account For each failure mode the particular failure probability was
determined using the software package PC-Ring (Vrouwenvelder et al 1999) while the
reliability method used for the assessment of the failure probability was FORM That was
coupled with a slope stability software where Bishoprsquos approach was performed
However this method is limited to circular slip surfaces and prescribed pore pressure
distributions within the dike In order to overcome this limitation FEM was used for further
analysis In combination with FEM an adaptive response surface technique was used in order to
represent the limit state function of each failure mode In Figure 38 the simulated structure
Literature study
MSc Thesis A Rippi 32
together with the response surface of the slope instability failure mode is showed Note that this
structure is approached with a single soil layer (same soil within and under the dike) and that
no structural element is implemented inside the dike
Figure 38 Dike model for the probabilistic FEA of dike stability and the corresponding response
surface of the limit state function for 3 stochastic input parameters (source Moumlllmann and
Vermeer 2008)
The failure probability of each of the aforementioned failure mechanisms were finally calculated
and compared with those recommended in each case while the performance of the coupling
between FEM and response surface method was assessed
Additionally a more recent investigation in terms of dikes with sheet piles employing FEA was
carried out by Breedeveld (2011) The main objective of this study was to display the
implementation of partial factors for design purposes using FEM and especially the software
Plaxis in simulating dikes with a sheet pile wall The dike was firstly simulated as a single
structure without reinforcement while in sequence the occurring stresses and pore pressures
were calculated with an existing sheet pile wall The results of the distribution of the effective
stresses within and below the dike are illustrated in Figure 39 His case study has been
introduced in Chapter 2 while in Chapter 6 a more detailed explanation of how it was modelled
in FEM is given This case study will be used as an example case in this thesis
(a)
Literature study
MSc Thesis A Rippi 33
(b)
Figure 39 Effective stress distribution before (a) and after (b) the installation of the sheet pile wall (source Breedeveld 2011)
Last but not least Bach (2014) has been occupied with the reliability analysis of bored piles and
the case study that he examined is presented in Figure 310 The major objectives of this
research were (1) Propose models to calibrate resistance factors for the Load and Resistance
Factor Design (LRFD) (see section 423 for further explanation of this method) approach and
find a suitable model aiming to directly determine reliability of a bored pile considering some
types of defect that may occur in the bored pile (2) Select a quality control method and evaluate
its reliability when applied to bored piles
For that purpose he made a coupling calculation between the finite element software Plaxis
(version 90) and the numerical probabilistic toolbox Prob2B in order to design bored pile
foundations in light of the reliability-based design Two failure modes the geotechnical failure
mode and the structural failure mode were proposed in order to comprehensively assess the
reliability of an axially loaded pile The parameter uncertainty was considered through the use
of statistical parameters and probability distributions for material parameters in soil models
The soil parameters were treated as random variables The geometry parameters of pile were
used as deterministic quantities because a change in pile shape in the calculation process
requires establishing a new mesh which is now impossible with regard to the Plaxis software
Figure 310 Nine calculation cases considered with different pile toe levels (source Bach 2014)
Literature study
MSc Thesis A Rippi 34
For the reliability of the pile the ULS of the pile was investigated and finally a displacement
criterion of a percentage of the pile diameter used depending on the soil type and
characteristics that the structure is founded on
33 Overview In this chapter concepts that are related to FEM probabilities uncertainty and sensitivity
analysis were discussed The introduction of FEM applications and the principles of how to
handle uncertainties will help the reader to better understand and follow the process of this
thesis
Furthermore preliminary researches related to several aspects of this thesis are presented that
mainly concern
Reliability methods and their applications [Waarts 2000 Schweckendiek 2006 Bach
2014]
Coupling FEM with reliability methods [Waarts 2000 Schweckendiek 2006 Bach
2014]
(Reliability) Analysis of geotechnical structures and especially dikes reinforced with
retaining walls [Waarts 2000 Schweckendiek 2006 Moumlllmann and Vermeer 2008
Breedeveld 2011 Gonzaacutelez et al 2013 Bach 2014]
As someone might have noticed from the overview of the existing research on the topic both the
performance of different reliability methods and the simulation of a dike with a sheet pile wall
in FEM have been carried out These references were quite helpful in order to get an idea of
coupling FEM with reliability methods as well as FEM and reliability methods individually
Moreover a first dive into the system under investigation and the variables to be handled was
made However noone has analysed the reliability of such a complicated system that is more
and more often seemed to be applied in the Netherlands Thus it would be essential to evaluate
the implementation and the results of the reliability analysis of a dike with sheet piles simulated
in FEM The conclusions of such a research are going to contribute to design and optimization
concepts and hopefully to a better understanding of the system behaviour
MSc Thesis A Rippi 35
4 Structural Reliability Analysis
The aim of this chapter is to introduce the basic theory of a reliability analysis and to summarize
the principles of different reliability methods that are widely used for the uncertainty
evaluation of a system Particularly three of them (MC DS and FORM)12 were used in this thesis
and are described in this chapter Finally a brief evaluation of the selected reliability methods
and the way that they are applied on this thesis by coupling them with FEM are discussed
41 Basics of Reliability Analysis In the consideration of the reliability of an element the determination of the probability of
failure is the central issue The limit between failure and non-failure is defined as a limit state
and the reliability is the probability that this limit state is not exceeded The limit states are
interpreted through the so-called limit state functions (LSF) whose general form is
(41)
in which R is the strength or more general the resistance to failure and S is the load or that
which is conducive to failure (ldquosolicitationrdquo) The basic principle of structural design is that the
resistance needs to be higher than the load or in other words that the LSF is larger than zero
(Zgt0) The main objective of the design is to ensure that this performance criterion is valid
throughout the lifetime of a structure Nevertheless the majority of the quantities that both
resistance and load consists of are uncertain Therefore a probability of satisfying the preceding
criterion is estimated and expresses the reliability of the structure P(Zgt0) The probability of
failure is
(42)
Hence the probability of failure is complement to the reliability to the sense that
(43)
The LSF is plotted in Figure 41 indicating the failure space and the limit state function by Z=0
12 Besides RS techniques were used for proceeding with the analysis but most for solving numerical problems rather than evaluating the failure probability The RS techniques are still under investigation and need to be tested before being applied in real cases However it is highly possible that this would be a potential way of evaluating the failure probability in the future
Structural Reliability Analysis
MSc Thesis A Rippi 36
Figure 41 Limit Sate function
In this case the design concept is based on the evaluation of the so-called design point which is
the point in the failure space with the greatest probability density Generally this point is
located on the border between the safe and the unsafe areas
In the structural domain the Joint committee on structural safety (1981) proposed a level-
classification of the calculation methods This classification includes the following three levels
Level III these methods calculate the probability of failure by considering the probability
density functions of all strength and load variables The reliability of an element is linked
directly to the probability of failure
Level II this level comprises a number of methods for determining the probability of
failure and thus the reliability It entails a linearization of the LSF in a carefully selected
point These methods approximate the probability distribution of each variable by a
standard normal distribution
Level I at this level no failure probabilities are calculated The level I calculation is a
design method according to the standards which consider an element sufficiently reliable
if a certain margin is present between the representative values of the strength and the
loads This margin is created by taking the so-called partial safety factors into account in
the design
In the next section the different reliability methods to be used in this thesis are further
elaborated
42 Overview of Reliability Analysis Methods
421 Level III Methods
The foundation of the Level III failure probability calculation is the mathematical formulation of
the subset of the probability space which involves failure (Zlt0) Level III reliability methods
Structural Reliability Analysis
MSc Thesis A Rippi 37
(also known as fully probabilistic methods) compute the probability of failure based on the
exact probability density function and the exact limit state functions Therefore such methods
are considered to be the most accurate Well-known sampling methods are (Quasi-) MC
simulation Importance Simulation DS and the Latin Hybercube Simulation Below the former is
discussed as it represents the fundamental theory on which the sampling methods are based on
while DS is also further explained as it is continuously used in this thesis For the rest a brief
explanation is provided in Appendix F There are also other sampling methods such as the
Numerical and the Directional integration which will not be described in this study However
for further information about them a reference though is made to Waarts (2000) In Level III
methods errors can only occur by a too low number of simulations
Crude Monte Carlo Simulation
The MC method uses the possibility of drawing random numbers from a uniform probability
density function between zero and one If and are the
vectors for the resistance and the load respectively that consists of n variables then by taking
multiple realizations of the uniform probability distribution a value can be determined for
every and
By repeating this procedure a large number of times the failure probability can be estimated
(44)
where is the total number of simulations (nm draws from the uniform distribution in which
m is the number of base variables) and is the number of simulations for which Zlt0
The simulationrsquos relative error is
(45)
If the expected value of the relative error is zero the standard deviation is
radic
(46)
Based on the central limit theorem the error is normally distributed provided is sufficiently
large The probability that the relative error is smaller than the given value E is then
(47)
Thus for the reliability of the relative error is smaller than For the wanted k and
E the required number of simulations can be determined with
Structural Reliability Analysis
MSc Thesis A Rippi 38
(48)
If for instance a reliability of 95 is required for a maximum relative error E=01 the required
number of simulations amounts to
(49)
The number of simulations is therefore still dependent on the probability of failure
MC simulation is applicable whatever the complexity of the deterministic model is However the
number of samples required to get an accurate estimation of may be dissuasive especially
when the value of is small (10-k) Thus crude MC is not applicable when small values of are
sought andor when the CPU cost of each run of the model is non-negligible
In Appendix F and especially in sections F1 and F2 a further elaboration of the MC simulation
is presented Moreover the methods of Quasi-MC Importance sampling and Latin Hybercube
are described
Directional Sampling
The directional simulation method is an accelerated sampling method It implies a preliminary
iso-probabilistic transformation as for FORM and SORM methods however it remains based on
sampling and is thus not an approximation method Below the method as it is described in Open
TURNS 15 Reference Guide (2015 pp190-193) is presented
Each simulation of the directional simulation algorithm is made of three steps Using the
probability distribution of a random vector we seek to evaluate the following probability
( ) (410)
Here is a deterministic vector and ( ) the limit state function which enables the
definition of the event For the iteration these steps are the
following
Let A point is drawn randomly on S according to a uniform
distribution
In the direction starting from the origin and the passing through solutions of the
equation ( ) (ie limits of ) The set of values of that belong to is
deduced for these solutions it is a subset of
Then one calculates the probability By property of independent
standard variable
is a random variable distributed according to chi-square
distribution which makes the computation effortless
Structural Reliability Analysis
MSc Thesis A Rippi 39
Finally the estimate of the probability after N simulations is the following
sum
(411)
The main idea is that for each direction we go along the direction by step of a specified length
from the origin to the maximum distant point and we check if there is a sign change on each
segment so formed (see Figure 42) We go until the maximum distant point Then for all the
segments where a sign change is detected we research the root λ in the segment (there are
also other strategies available in OT however all of them follow the same principle of
searching) The following figure illustrates the principle of an iteration in dimension 2
Figure 42 DS of a 2-dimensional problem
Similar to MC method the outcome of the estimated probability of failure is a random variable
and the error in the estimate can be made as small as possible by taking a sufficient number of
samples For DS the standard deviation σ of the estimated failure probability can be quantified
as follows (Melchers 2002)
radic
sum
(412)
From this equation relative errors and the confidence intervals can be estimated As can be
seen the error in the estimated probability of failure will decrease with increasing number of
sampled directions Equation (412) can be used to determine the number of sampled directions
that is required for a reliable estimate of the failure probability The convergence criterion in
this method is usually the coefficient of variation (CoV) which is defined as
Structural Reliability Analysis
MSc Thesis A Rippi 40
422 Level II Methods
Level II methods (known also as fully probabilistic methods with approximations) can take all
the probabilistic properties of the random variables into account but they include
approximations of the limit state function and therefore their use and outcomes should be
inspected and evaluated in order to be considered as reliable However experience in that
methods has shown that the computational effort is profoundly decreasing in comparison with
Level III methods and that the application of them can provide important parameters such as
the reliability index (β) and the influence factors (α) Some of the most known Level II methods
are First Order Second Moment (FOSM) Method First and Second Order Reliability Methods
(FORM and SORM) and the Point Estimate method (PEM) In the next paragraphs FORM is
elaborated as it will be mainly used in this thesis while in Appendix F3 the principles of FOSM
are discussed
First Order Reliability Method (FORM)
The principles of FORM are based on the First-Order Second-Moment (FOSM) approach which
first introduced the reliability index concept This method is presented in Appendix F3 in detail
However the reliability index estimated using the FOSM approach is not ldquoinvariantrdquo which
means that the index value depends on the limit state function format (Farrokh 2007) Hasofer
and Lind (1974) proposed an invariant definition of the reliability index that leads to a
geometric interpretation known as first-order reliability method (FORM)13 The first step of this
method is the transformation of the random variables to equivalent standard normally
distributed variables and the whole procedure is carried out in u-space (or else standard space)
For variables that are normally distributed this step is as follows
(413)
For other types of distributions there are procedures available for carrying out this
transformation such as Generalised Nataf and Rosenblatt transformations In this study these
transformations are not discussed however for further explanation a reference is made to the
Reference Guide of OpenTURNS 15 (2015)
In the sequence the limit state function Z is expressed in terms of
The second step is the approximation of the function with the first two terms of the Taylor-
polynomial The approximation reads (CUR 1997)
( ) ( ) sum
(414)
where is the point where the linearization takes place This approximation of Z is linear and
according to the central limit theorem it is normally distributed The expected value of the LSF
can be approximated by the expected value of the linearized function
13 VNK2 uses FORM in view of the calculation time required In case of inaccuracies VNK2 switches to more time-consuming techniques such as DS whereby a large number of draws are made from the probability density functions of the various stochastic variables (uncertain quantities)
Structural Reliability Analysis
MSc Thesis A Rippi 41
( ) sum
(415)
While the standard deviation (sum of contributions of each variable to the variance of Z) is
defined as
radicsum
( )
(416)
Using the definition of FOSM for the reliability index [ ]
where [ ] is the mean value of
the limit state function and substituting Eq 415 and Eq 416 in it the reliability index can be
approximated However linearization in different points can lead to different values of the
reliability index Hasofer and Lind definition of the reliability index overcomes this limitation
and renders it equal to the minimum distance from the origin to the design point
(417)
where radic
Looking for the design point is basically an optimization problem Many analytical and
numerical approaches can be used for that purpose A relatively straightforward method to do
this is by firstly assuming that the design point is the mean value (the starting point can also be
another point according to an engineering judgement in order to accelerate the optimization
procedure) The obtained β-value is used to determine a new point in which the LSF is
linearized In this case the importance factors αi are calculated as
( )
radicsum (
( ) )
(418)
where is the design point and is the number of variables expresses the
contribution of the variance of each variable to the total variance of Z in the design point Figure
43 illustrates the linear approximation of the limit state function and the aforementioned
parameters for a two dimensional problem The probability that Zlt0 can be determined using
the standard normal CDF
(
) (419)
Structural Reliability Analysis
MSc Thesis A Rippi 42
The new calculation point is determined by
(420)
After some iterations the chosen optimization algorithm finds the final design point and the
new reliability index In Appendix A2 an overview of the different optimization algorithms
available in OT is presented together with an evaluation of their performance in the specific
case study
Figure 43 Two-dimensional illustration of u-space LSF and design point
423 Level I Methods (semi-probabilistic)
At the beginning of the probabilistic concepts incorporation in the field of structural
engineering the most notable development was the implementation of LFRD At the past single
factors of safety on the ratio of total resistance to total load were used in order to ensure that
the stresses developed from an applied load on a structure were lower than the allowable ones
LRFD replaces those factors by introducing a set of partial safety factors on the individual
components of load and resistance These partial factors can be selected such as they account
for uncertainties associated with the resistance and the load respectively
The current design philosophy in CUR 166 and Eurocode is characterized by the use of
characteristic values of the parameters (in Appendix E a further explanation of the
characteristic values is given together with the way that they are converted into the mean
values for being used in the reliability analysis later on) The values of the partial safety factors
are additional factors to the characteristic values Therefore the characteristic values are
multiplied with the aforementioned safety factors ( ) and the following criterion has
to be satisfied for a reliable structural performance (see Figure 44)
(421)
Structural Reliability Analysis
MSc Thesis A Rippi 43
The main disadvantage of LRFD is the fact that there is a set of load and resistance factors that
need to be calibrated in order to cover the specific cases that are likely to occur This is because
of the lack of sufficient statistical data that are necessary for this calibration procedure
Consequently it cannot reassure that all the designs to which the set of factors is applied result
in the reliability level that was aimed for As a result in majority of the cases a conservative
calibration that gives the required reliability is chosen which can lead to ldquoover-designrdquo
structures
Figure 44 Design with partial factors for the load and resistance parameters (source
Schweckendiek 2006)
The evaluation of an elementrsquos reliability starts in principle with calculating the probability of
failure and subsequently the reliability for the given strength and load In practise the problem
is often that the strength is unknown but it has to be determined for a given reliability The
determination of the required reliability can be estimated with the help of Level II and III
methods by iteratively adjusting the strength in the calculation until a sufficiently small
probability of failure is found
In the design domain regulations and guidelines follow the standard that the characteristic
value14 of the strength is divided by a factor and that the characteristic value of the load is
multiplied by a factor as it is described in Eq 421
The link between Level I and the estimation of the failure probability has been achieved through
Level II methods The design point that results from a Level II method calculation is the point
with the greatest joint probability density of the strength and the load and it is therefore
possible that when failure occurs the strength and load values will be close to that point The
design values can be specified as
(422)
(423)
14 The characteristic value of a parameter is its main representative value and its value can be specified as a mean value an upper or lower value or a nominal value according to EN 1990 definition (2002)
Structural Reliability Analysis
MSc Thesis A Rippi 44
where are the coefficient of variation for the resistance and the solicitation respectively
15 and are the importance factor and β is the reliability factor These can be also expressed
in terms of characteristic values as follows
(424)
By substituting function 423 with 421 and 422 the partial safety factors are defined as
(425)
(426)
where 16 and are the values for load or resistance respectively to which a probability of
(non-)exceedance of 95 corresponds (non-exceedance for the load and exceedance for the
resistance) (for a standard normal distributed parameter this is 1645)
424 Response Surface Techniques (RS)
In case that the models presented above tend to be time consuming for the limit state function
evaluation or convergence problems of the optimization algorithms occur a better methodology
is recommended known as response surface The RS is mainly used when (unknown) response is
only available from experiments or complex FEM computations (for example large highly-non-
linear FEM models) Indeed once a RS has been built up the various standard methods may be
applied at a negligible cost An analytical limit state function replaces the real response function
The main idea is that the response consisting of a complex function of input variables is
approximated by a simple function of the input variables
A list of possible response surfaces techniques is given below
Linear and quadratic Taylor expansions
Polynomial RS based on least squares
Kriging method
Polynomial chaos expansion
A detailed explanation of the aforementioned methods can be found in OpenTURNS 15 (2015a)
The first method is associated with the approximation of the model response around a specific
set of input parameters while the rest seek a global approximation of the model over its whole
domain of definition The most sophisticated one is the polynomial chaos expansion but one the
15 α values are negative for solicitation (ie RV for which increasing their values lead to adversely effects on the reliability) and positive for resistance (ie RV for which increasing their values implies a beneficial effect on the reliability) 16 is negative and can be negative or positive
Structural Reliability Analysis
MSc Thesis A Rippi 45
other hand its complexity does not make it attractive In this research a polynomial RS is
applied based on least squares
In combination with FEA the standard procedure is as follows
1 Select the most important random variables on the basis of engineering basis
2 A SA is carried out in combination with FEM
3 Reduce the stochastic variables if needed according to the SA
4 A RS is constructed through the response data
5 A reliability calculation is carried out using the RS instead of the real response
A polynomial RS is generally constructed by fitting a quadratic function to the sampling points
whose general expression is
sum
sum
sum sum
(427)
The type of the reliability method to be used is of little importance since the time consuming
LSFE (using FEM) are replaced by analytical expressions However of main importance is now
the accuracy of the RS compared to the real response There are several parameters in order to
evaluate the goodness of fit of the RS to the real model among which is the well-known R-
squared
In Figure 45 an example of a quadratic response surface is showed that is fitted on the data
response indicated with the black circles This was a 6-dimensional problem that means 6
different variables were included in the system (non-visualized dimensions are kept constant at
their mean value) In this thesis the response of the system is to be acquired after the coupling
of FEM with a reliability method and that would give the ldquoblackrdquo circles that are shown in Figure
45 The response surface technique applies a curve fitting on the real response of the system
(ldquoblackrdquo circles) whose analytical formula can later be used for the reliability analysis of similar
type of systems with negligible computation time
Structural Reliability Analysis
MSc Thesis A Rippi 46
Figure 45 Quadratic response surface with 28 terms (ie ) for 6 dimensions ( ie x
=[ x1 hellip x6])
43 Coupling Reliability Analysis with FEM The limit state function evaluations (LSFE) will be carried out with the software Plaxis 2D 2015
which is a special two-dimensional finite element software used to perform deformations and
stability analysis for various types of geotechnical applications Moreover considering the case
study under investigation in this thesis Plaxis in contrast with other FEM such as Abacus
Comsol DIANA etc offers several techniques to realistically simulate structural elements such
as sheet pile walls and anchors and their interaction with soil while the variety of the
constitutive models for the soil body that are available and the ability to include the history of
the construction phases can lead to a better analysis of the systemrsquos behavior in terms of the
stress level and the deformations It is essential at that point to mention that using FEM for this
purpose means that the limit state formulation is implicit and can only be solved numerically
The reliability analysis is carried out through an uncertainty package In this section firstly a
description of the reliability package and its possibilities are given Finally an explanation of the
coupling procedure between the reliability tool and FEM is given together with the calculation
process that was followed for the parameters manipulation
431 The functionality and possibilities of OT
The reliability analysis tool to be utilized in this research is OpenTURNS (OT) which is a
scientific library usable as a Python module dedicated to the treatment of uncertainties and it is
still under development during the work of this thesis Several reliability packages are available
such as OpenEarth (applied in Matlab) Prob2b (applied in Probox software of TNO) and
Probabilistic Toolkit (software of Deltares) Plaxis 2015 features Python connection possibility
and thus coupling was decided to be carried out with an uncertainty package in Python and OT
is the most developed one Moreover until now only OT is an open source package and can be
Structural Reliability Analysis
MSc Thesis A Rippi 47
used easily from anyone as there is a wide community of experienced people supporting with
their knowledge while many related manual reports are already available for starting learning
A list of the current available reliability methods in OT is given in Figure 46 In this figure apart
from the standard methods the alternative method of the Adaptive Response surfaces is
introduced that is mainly used when (unknown) response is only available from complex FE
computations
Figure 46 Reliability methods available in OpenTURNS
In this thesis the methods to be basically applied are the FORM and the DS analysis while MC is
also used mainly for confirming the application of FORM Additionally the RS technique has
been mainly implemented for enhancing the performance of FORM and DS analysis
OT handles 47 types of distributions amongst which the ones used in this research are Normal
Lognormal Uniform and Truncated Normal distribution Regarding the joint distributions that
are available to be selected for random vectors 11 types of copulas are existing in OT amongst
which the most known ones are the Independent the Gumbel and the Normal copula In this
thesis the independent copula has been used during the reliability analysis which means that
the variables were assumed to be independent among each other However it is strongly
advised a further elaboration and research considering correlation matrices for specific soil
parameters
432 Coupling OpenTURNS-Plaxis
The finite element simulation will be carried out in Plaxis 2015 Plaxis is a finite element
software for plane strain and axi-symmetric modelling of soil and rock behaviour Moreover it
supports a fully automatic mesh generation allowing for a virtually infinite number of 6-node
and 15-node elements
Structural Reliability Analysis
MSc Thesis A Rippi 48
The coupling of reliability analysis and FEA requires an interface for the communication
between each other When a reliability tool is coupled with another software program the
reliability program carries out the whole reliability analysis and it uses the other program only
for the evaluation of the limit state function More precisely OT should be able to read and
amend Plaxis output values for important variables such as material parameters pore pressures
generation and stresses development and corresponding deformations inside the dike
Respectively Plaxis has to be also capable of obtaining the new values that has been set by OT
for the variables that are treated as stochastic during an iterative process according to the
reliability assessment In Figure 47 an illustration of the coupling methodology and its function
is shown
Figure 47 Coupling scheme OpenTURNS-Plaxis
In principle an input file is firstly required where the user set the preferable reliability method
to be used the stochastic input parameters and their probability distributions the joint
probability distribution and the corresponding correlation matrix and finally the limit state
function is formed depending on the situation In Appendix D an example of the input files that
were used for soil sheet pile wall and anchor analysis is shown However such input files
should be interpreted so as to be readable by both Plaxis and OT Therefore an input interpreter
was created which is actually a python script that helps OT to start up the reliability analysis
according to the assigned method variables distributions and LSF As it was mentioned before
the evaluation of the limit state function is conducted by Plaxis For that purpose the input
interpreter should be also able to send the next set of input parameters to Plaxis However an
additional means of connecting the interpreter with Plaxis is also needed This is can be
Structural Reliability Analysis
MSc Thesis A Rippi 49
achieved via a Plaxis interface which actually call Plaxis to make the calculations while it also
transfers the required value of the limit state function to input interpreter and this in turn to OT
Eventually the probability of failure is obtained as a model result However it is also essential
that Plaxis simulation procedure converges to the desirable criteria and under the physical
boundary conditions that have been determined Likewise the convergence criteria of the
reliability methods shall be manipulated so as the optimization algorithms to able to converge
efficiently (see Appendix A2 for further explanation)
44 Overview In section 42 a summary of the principles of the main reliability methods was made From
Level III methods MC and DS were presented whereas from Level II the basics of a FORM
analysis were introduced Besides the concept of the Level I method was explained Even
though this method was not used in this research for the evaluation of the probability of failure
it was applied in section 643 for the deterministic analysis of the system under investigation
and the preliminary calculations of the dimensions In Appendix F more reliability methods are
introduced and they are further described In the sequence their implementation in conjunction
with Plaxis was discussed through the coupling of FEM with OT For more information
concerning OT special features (ie SA optimization algorithms and probability distributions) a
reference is made to Appendices A and E
Before any of these reliability methods was applied on the case study of this thesis their
performance and their compatibility with Plaxis were tested with simple examples At first the
probability of failure was evaluated with FORM SORM Importance Sampling (IS) crude MC and
DS for simple artificial linear and non-linear limit state functions
and
respectively where R B and F are the random variables distributed normally
or lognormally The crude MC technique has been applied in order to validate the failure
probability in case the various methods give different results Secondly a flexible circular
footing over a one-layer soil material was used from Plaxis tutorial in order to test the coupling
between Plaxis and the reliability method and reassure the functionality of the interface In this
pilot example the objective was to determine the probability that the settlements of the footing
exceed a certain threshold of settlements and thus The respective
components that were taken into account for the failure mechanism was the cohesion (c) the
friction angle (φ) and the specific weight (γ) of the soil
From a qualitative evaluation of the reliability methods that has been done and according to the
experience gained from the aforementioned applications of some methods valuable conclusions
were drawn in terms of their efficiency More precisely regarding the sampling methods it does
not required previous knowledge of the failure domain apart from the IS method in which a
starting point inside the failure domain shall be provided Moreover the calculation effort
depends on the magnitude of the failure probability and the required accuracy whereas for the
IS it depends also on the selection of the sampling region In Waarts (2000) it has been proved
that DS tends to be more efficient than MC for low dimensional problems as it is indicated in
Figure 48 (number of random variables nlt100)
Structural Reliability Analysis
MSc Thesis A Rippi 50
Figure 48 Required number of samples for MC and DS as a function of the random variables
(source Waarts 2000)
As far as the approximation reliability methods (FORM and SORM) are concerned it is not
necessary to be aware of the failure region in advance Furthermore it has been observed that
the required iterations and the calculation time is quite lower in comparison with the sampling
methods However the accuracy of the method is highly dependent on the shape of the LSF
In Table 41 a summary of the most applied reliability methods evaluation is presented
According to this evaluation it was decided that the methods of FORM and DS as well as the
combination of them with RS techniques are going to be tested and evaluated in terms of their
efficiency and robustness These methods were chosen from both Level II and Level III methods
due to their expected reduced computational time the non-requirement of previous knowledge
about the LSF and their satisfactory performance in similar case studies In Chapter 5 the
different failure mechanisms and the relevant LSF that are going to be utilized in this thesis are
presented whereas in Chapter 7 the results deduced from the reliability analysis are elaborated
In Chapter 7 there is also an indication of which reliability method was used for each systemrsquos
component the selection of which was based on the aforementioned evaluation
Table 41Evaluation of reliability methods
Method Previous knowledge Accuracy Calculation Effort
FORM not required full accuracy for
Gaussian variables and linear LSF
depends on LSF linearity and the number of random
variables
SORM not required exact up to 2nd order
LSF error dependent on the shape of LSF
depends on LSF linearity and the number of random
variables
Structural Reliability Analysis
MSc Thesis A Rippi 51
Crude MC not required can be controlled by convergence criteria
depends on the magnitude of failure probability and the
required accuracy
IS required can be controlled by convergence criteria
depends on the choices made for the sampling region
DS not required can be controlled by convergence criteria
for low dimensional problems (nlt100) DS is more efficient
than MC
Especially RS were created based on a SA sampling in order basically to increase the efficiency
of FROM and DS analysis in case of FEM numerical errors Precisely an analytical formula or a
response surface was needed that can approximately represent the performance of the structure
under investigation and that is called in case of Plaxis errors This formula was essential in
order to keep the reliability analysis running by providing a response value for the LSF when
Plaxis calculation is unsuccessful and thus incapable of returning a result It should be
mentioned that in case a RS is implemented for the reliability analysis Figure 47 can be
reduced to a graph consisting of only the input file OT and the input interpreter as no Plaxis
calculations would be necessary any more The LSFE will then be performed through the
response surface that is provided by the user Due to this limited use of the RS method it will
not be further treated in this thesis
Structural Reliability Analysis
MSc Thesis A Rippi 52
MSc Thesis A Rippi 53
5 Failure Mechanisms and Limit State Functions
The reliability analysis of a structure requires the definition of the different failure modes that
are relevant to the corresponding structural elements In this chapter the possible failure
mechanisms of the systemrsquos elements expressed in LSF are identified and their role in the
system reliability is explained Based on these LSF the reliability of the system components was
evaluated These LSFs focus on the ULS whose exceedance implies failure of the corresponding
component
51 Introduction to the system analysis and the limit states For system reliability analysis all the possible failure mechanisms have to be identified and
summarized in a fault tree In Figures 51 and 52 the fault trees of a dike section without
structural elements and a retaining structure with sheet piles are depicted respectively In this
project the combination of these two separate structures ie a reinforced dike section with an
anchored sheet pile wall is to be simulated and studied in terms of its failure modes
As it is described later in this chapter failure can be expressed in different ways depending on
the structure and stakeholders demands and safety standards This research is mainly focused
on the ULS of failure and Mohr Coulomb was chosen as the soil failure criterion due to its
efficiency on detecting failure Such a failure criterion is actually defining the stress strain
relationship and the gradual weakening of the soil (ie the gradual reduction of the strength
parameters such as the friction angle and the cohesion due to the stresses development until
soil collapses) under a stress condition In Plaxis new version there is the ability of gradually
reducing structural properties such the strength of steel elements However the results
retrieved of such a safety analysis have not been yet investigated and for that reason such
calculations are not included in this project
As far as the dike section is concerned the failure mechanism to be investigated in this project is
the macro instability of the slope rather than the other mechanisms depicted in Figure 51 On
the other hand because of Plaxis limitations to distinguish among the different structural
elements failure mechanisms each element was considered as a singular case for investigating
its failure mode In the next section a distinction is also made between the ULS and the SLS
For a reinforced dike with retaining walls there are basically four classes of structural elements
(see Figures 12 and 21)
∙ Retaining wall (ie sheet piles)
∙ Anchors
∙ Walings
∙ Soil structure (ie dike section)
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 54
For the reliability analysis of each of the elements the probability of failure is to be determined
In the next sections of this chapter the respective LSF that are considered for each class are
further elaborated according to the ULS criterion while in the end the general combined fault
tree is to be schematized
Figure 51 Fault tree of a dike section
Figure 52 Fault tree for retaining structures using anchored sheet piles (source Schweckendiek 2006)
The system failure in this research is considered as a serial system of the anchor sheet pile
wall and soil body LSFs The fault tree of the system failure is presented in Figure 53 together
with the LSF of each component as they are formulated according to sections 522 and 523
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 55
Figure 53 Fault tree of the system failure
It should be mentioned at that point that the probability of failure of the system differs from the
overall probability of flooding which takes into account any potential failure mechanism of the
system combined with the uncertainties of the water elevation More precisely as far as the dike
safety is concerned a probability of failure (where ) under a certain
water level is estimated from the fault tree of Figure 51 which in sequence is multiplied
with the occurrence probability of the corresponding water level in order to estimate the
overall probability of failure for the specific water elevation Then that product is accumulated
over a required range of water levels that can jeopardise the overall stability of the structure in
order to calculate the overall probability of failure or else the probability of flooding This
probability is then compared to the one established from the safety standards so as to reassure
the safety or not of the structure can be estimated as follows
int
sum (51)
In this thesis the probability is to be estimated where is a specific water level
and especially the design water level as it is considered to be the most challenging part of the
procedure described above gives the probability of occurrence of the water level (ie it is
the PDF of ) After setting up the steps and implementing them successfully for the estimation
of the calculation of the overall probability of failure is just a repetition of the same
procedure for more water levels It should be mentioned that in this thesis failure consists only
of the macro-instability failure mode whereas the rest are excluded for the time being A
simplified alternative for the overcoming the integral of Eq51 is to repeat the calculations for
certain water levels and then sum their products instead of integrating the full range of the
them
52 Limit State Functions
521 Serviceability Limit State
The SLS is evaluated in design calculations in order to avoid excessive deformations of a
structure that could lead to the loss of its service requirements and its functionality In some
cases the SLS criteria are used to prevent damage to adjacent structures in the surroundings
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 56
For example concerning the design of a dike section attention should be paid to deformations
caused to buildings that are located in the inland part of the dike in case of an extreme event
In Figure 54 an example of a deformed dike is depicted In this figure some of the potential
locations that excessive deformations might be experienced are illustrated such as
∙ the vertical settlement of top of dike (arrow 1)
∙ the horizontal displacement of the top of the sheet pile (it might also be the middle part
of the sheet pile that ends up with the largest deformation according to the loading
conditions and the specific soil structure different deformations might occur) (arrow 2)
∙ Uplift upstream of the dike (arrow 3) special attention shall be paid to this type of
deformation not only due to the direct effect on inland structures but also as a sign of
developing piping mechanism
(a)
(b)
Figure 54 Figure (a) shows the initial undeformed structure while (b) illustrates the deformed mesh (not to scale) (the plotted length of the anchor does not have any physical meaning) In (b) the possible locations and directions (black arrows) of potentially excessive deformations
According to the current Dutch regulations (Larsen et al 2013) the SLS criteria for designing
reinforced dikes with sheet pile walls require that
∙ the settlement on the top of dike does not exceed 10 cm and
∙ the horizontal displacements of the sheet pile wall does not overcome 10 cm
In this research the SLS criteria and especially the dike settlements were used roughly for the
preliminary determination of the structures characteristics that are discussed in section 643
Dike Anchor
Sheet pile wall
1
2 3
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 57
More precisely displacements at the top of the dike were also taken into account as
complementary to the safety factor in order to determine the required length of the sheet pile
wall As far as the reliability analysis in concerned the LSF were formulated based on the ULS of
the different systemrsquos components that are presented in the next section
522 ULS for Structural Members
In the design process one is most interested in the ULS of a failure mechanism This state
describes the situation wherein the acting extreme loads are just balanced by the strength of the
construction If that limit state is exceeded the construction will lose its functionality and thus
collapse or fail In general most attention is paid to the behaviour of the structure after
completion However during construction there are also periods in which the construction may
fail The different phases of construction are listed in section 642 In the present section the
analytical LSF of the structural elements are given as they are going to be used in the reliability
analysis
Sheet pile wall
The most relevant failure mode for the sheet pile wall is the exceedance of the yield strength
which corresponds to the ultimate steel strength The response of the structure is mainly due to
bending moments and the axial forces (shear forces are considered to be negligible) Where an
axial force is present allowance should be made for its effect on the moment resistance
Figure 55 The water and earth pressures as well as the anchor forces lead to bending moments and axial forces on the sheet pile wall
In Figure 55 an example of the axial forces and the bending moments that can be developed in
the sheet pile wall with one anchor layer are illustrated after an extreme water level loading
Accordingly the maximum stresses on the sheet pile wall are composed of a bending moment
and a normal force component17
[
] (52)
where [kNm] and [kN] are the bending moment and the axial normal force
respectively that depend on the depth level where they are calculated over the sheet pile length
[m3] is the elastic section modulus and [m2] the cross-sectional area of the sheet pile
wall
17 the vertical anchor force component is reducing by its interaction with the soil over depth
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 58
Bending moment and axial force can be variable over the depth and that is why they are
expressed as a function of z-depth FEM has the advantage to take into account second order
effects ie a stiffer structure will experience higher bending moments than a more flexible one
Taking into account the above the limit state function can be formed as the difference between
the maximum developed stress and the yield stress
[
] (53)
where and can be characterized as the load variables while and can be considered
as the resistance variables and are assumed to be constant over depth
Concerning the permissible displacements for the ULS an upper limit for the maximum
horizontal displacements of the top of the sheet pile has been set to 150L where L is the
vertical length of the sheet pile wall (Larsen et al 2013) This requirement was mainly used in
section 643 where the structural properties were determined after the deterministic analysis
Anchors
Anchors are loaded by their reaction to the horizontal loads on the retaining walls The failure of
the anchor element is actually represented by the failure of the steel members of the anchor
(tubes bars cables etc) that are loaded by traction forces
As it is illustrated in Figure 56 the anchor experiences an axial tensile N that is nearly constant
over its length
Figure 56 Axial loading of anchor inside a dike
The elastic behaviour of an anchor involves only a relationship between axial force N and
displacement (elongation) u of the form
[ ] (54)
where EA [kN] is the anchor stiffness consisting of the steel Youngrsquos modulus E [kNm2] and the
anchor cross section A [m2] and L [m] in the length of the anchor
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 59
Similarly to the sheet pile wall the limit state function of the anchor involves the certain yield or
ultimate strength of the steel members and the maximum stress that the anchor experiences
during its loading Consequently the LSF is as following
(55)
where [kN] is the calculated anchor force and [m2] is the cross sectional area of the
anchor (both of them considered to be constant over the depth) It is essential to mention that
the anchor is also subjected to bending moments due to soil settlements (that are implicitly
illustrated via the uniformly distributed load q over the tie rod) that should be taken into
account in order to investigate the displacements of the tie rod itself However in this thesis
only the axial forces on the anchor are considered without taking into account the individual
deformations and its reaction with the surrounding soil
Walings
The waling is the element that transfers the loads from the retaining wall to the anchors (see
Figure 21) The loading of the walings can be schematized as a continuous beam on several
supports as it is depicted in Figure 57
Figure 57 Loading of walings
If Nα [kNm] is the calculated anchor force per m then the developed bending in the support Ms
and in the opening Mo can be approximated as follows
[ ] (56)
(57)
where La is the mutual anchor distance Considering as the design moment of the waling to be
the limit state function can be formulated as follows
Waling Sheet pile Anchor
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 60
(58)
For a conservative design of the waling the limit state function will give the same or lower
failure probability than the anchor itself Therefore it will actually not be necessary to carry out
this analysis The anchor failure will be the determinant mechanism (Schweckendiek 2006)
523 ULS for Soil Failure
In this thesis emphasis is given on the dike global instability which actually consists of several
failure modes Figure 58 indicates the possible failure mechanisms of flood defences The
combination if these can lead to the overall instability of the dike and thus to soil body failure
However Plaxis is not capable of accounting for all of them More precisely Plaxis assumes the
soil to be a continuous body and thus it can model movements in the scale of soil bodies (using a
relative fine mesh in relation to the moving soil body) but not on a soil particle level Plaxis can
simulate the groundwater flow in a soil body and from this someone can deduce input for a
piping analysis (see failure mechanism G in Figure 58) (using other softwaremethods) but as
it has been mentioned above it cannot determine movement of soil particles due to
groundwater flow Moreover Plaxis cannot deal with the flow and waves occurring in ldquoopen
waterrdquo ie water outside the soil in a canal a lake or sea for instance
Figure 58 Possible failure mechanisms of flood defences (source Flood Defences 2015)
Therefore Plaxis can handle the failure modes (C) (D) (E) (J) (K) and (L) of those illustrated in
Figure 57 In this thesis loads due to extreme temperature conditions or ships are not
considered while the settlements are taken into account implicitly though the mechanisms (C)
(D) and (E) These mechanisms are thus investigated in this thesis and they mainly refer to the
macro-instability of the inner and outer slope of the dike or the overall dike body (ie horizontal
sliding (mechanism D))
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 61
For these failure modes soil failure is defined as shear failure In Figure 59 the most relevant
patterns of the macro-instability regarding the shear strength of the soil are illustrated
Applying FEA the most critical failure mode is determined automatically However it is not
always straightforward what the trigger mechanism of the failure was and that is why more
investigation and FEA tests are needed in order to obtain a clear view
(a) Outer slope failure (b) ldquoActive siderdquo failure
(c) ldquoPassive siderdquo failure (d) Overall failure
(e) ldquoKranzrdquo stability
Figure 59 Macro-instability failure modes in a dike with an anchored sheet pile wall
In Figure 59 failure modes (a)-(d) follows Mohr-Coulomb failure criterion In principle the
total stress state inside a dike section consists of the effective stresses and the pore pressures
according to Terzaghi principle
[
] (59)
Mohr-Coulomb failure criterion basically considers as limit for the failure to occur the
maximum shear stress that the soil can withstand This is determined from the friction
angle the cohesion of the soil and the current stress state The drained shear strength
(similarly in for the undrained shear strength) in the soil according to Mohr-Coulomb is defined
as follows
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 62
[
] (510)
where is the effective friction angle In Appendix B1 a detailed explanation of this failure
criterion is attached
Taking into account the above mentioned failure mechanism (a) is mainly triggered by the
increase of the pore pressures in the outer slope which subsequently causes a decrease in the
effective stresses ( ) which at the same time leads to a reduction of the shear strength
(considering Mohr-Coulomb failure criterion) When the shear stresses in the soil exceed the
shear strength a slip plane forms and a soil wedge collapses
Failure modes (b) and (c) are primarily determined from the active and passive effective
stresses respectively As far as the ldquoactive siderdquo failure is concerned displacements of the sheet
pile wall towards the downstream side of the dike lead to development of active stresses
where is the active pressure coefficient During the ldquoactiverdquo failure the
retaining soil is allowed to relax which leads to a decrease of horizontal stresses and
simultaneously to a decrease in shear strength Similarly to failure mode (a) a soil wedge
collapses The result is an increased earth pressure on the wall for which it is not designed
However this type of failure in dikes is not such determinant for the soil and the sheet pile wall
failure because there is always the resistance of the passive side of the dike that keeps the active
side stable enough Such a mechanism is thus more relevant for deep excavations with retaining
walls
In mechanism (c) the passive soil resistance is exceeded by the horizontal loads In this case the
wall moves inland and a wedge of soil compresses The shear strength is larger due to the
deformation of the wall and the horizontal stresses increase stresses where is
the passive pressure coefficient ( ) This failure mechanism usually occurs due to an
underestimation of the sheet pile length or due to the presence of a weak soil layer in the
passive side
Failure mode (d) is actually a combination of (b) and (c) possibly accompanied also with the
development of excess pore pressures as it was discussed in failure mechanism (a)
Moreover too short anchors could cause a soil block to collapse as it is showed in Figure 59(e)
(ldquoKranzrdquo stability)
As it has been already mentioned for mechanism (d) failure modes can be correlated and
combined with each other in order to lead to a final failure state Therefore the common failure
probability would be smaller than the sum of the singular probabilities In this thesis the total
probability of failure of the soil body due to global instability is to be determined This is due to
the fact that the application of FEM is capable of simulating the combination of the failure
modes depicted in Figure 58 and thus the overall failure (related to (C) (D) and (E) modes of
Figure 58) can be evaluated However with the classical engineering approach for the stresses
calculation each one of the failure modes depicted in Figure 59 should be separately evaluated
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 63
Below a description of the available methods to formulate the LSF of the soil failure is given
After the evaluation of these alternatives a selection was made regarding the most suitable LSF
for the current case study
Limit State functions for soil failure
Soil mechanical failure can be analysed in several ways (Schweckendiek 2006)
1 Excessive Deformations
2 φ-c Reduction
3 Relative Shear Resistance
4 Plaxis definition of soil collapse
The possibilities and limitations of the aforementioned methods are briefly discussed below
1 Excessive Deformations
Similarly to the SLS a limit state function can be formulated by deformations that are
unacceptable As it is illustrated in Figure 54(b) there are some points (the top of the dike the
sheet pile top and the inland soil level behind the dike) the displacements of which shall be
limited to the minimum possible during the structurersquos lifetime in order for the structure to
meet the required service standards Therefore the calculated deformations at locations 1 2
and 3 could give a clear indication of failure given the maximum acceptable deformation
In case that there is a limit value for each location then the limit state function can be formed as
follows
[ ] (511)
with this criterion designer can control the developed deformation on the system and improve
the structure so as to limit the displacements to the required level and subsequently increase
the reliability
On the other hand such an approach might exhibit several problems during the reliability
analysis that are listed below
Many failure mechanisms suddenly happen and as a result no significant displacements
are observed before moving very close to failure In sequence this may cause problems
for the iterative procedures of some reliability methods such as FORM and DS
The determination of suitable maximum admissible deformations is not
straightforward They have to be large enough to serve as failure criterion and
simultaneously they shall not be larger than the values that can be calculated within
the limits of equilibrium in FEM-calculations This requires previous knowledge on the
analysed system as well as on the feasibility of FEM-calculations on the specific subject
Last but not least the location of the undesirable displacements in a dike section is also
under investigation as there can be multiple vulnerable spots on the dike that should
be kept under a certain range of deformations in order not to cause a collapse and that
they are not always known in advance
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 64
2 φ-c Reduction
An available type of calculation in Plaxis is the so-called ldquoSafety Calculationrdquo with which global
safety factors can be computed In Appendix B2 a more detailed elaboration of this method is
exhibited This feature can also be used in reliability analysis by formulating the LSF as follows
(512)
where is the Multiplier Safety Factor obtained by the φ-c reduction The main concept is
that if the safety factor is smaller than 1 then it is considered as failure This method can provide
us with the probability of failure for a general soil body collapse However even if this method
seems to be quite simple there are some issues during its implementation in a reliability
analysis
For complex limit state functions there can be convergence problems for some
reliability methods with iterative procedures (ie FORM) This can be explained from the
fact that in a φ-c reduction soil strength properties follows a certain path as it is
described in Appendix B2 independently of the values for the random variables that
have been set from the reliability method
The safety factor is a general safety factor regarding the failure of the system
Therefore sometimes it is not straightforward what the ldquotriggerrdquo factors are that led to a
certain type of failure
A safety calculation in Plaxis is time-consuming and the outcome can be unstable (Plaxis
calculationsrsquo convergence depends on the number of calculation steps)
Plaxis cannot handle safety factors below 1 Therefore a new limit of the safety factor
shall be introduced in this limit state function
3 Relative Shear Resistance
In this method the basic idea is to define soil failure according to the failure criterion that Plaxis
is set to use for the analysis In this research the Mohr Coulomb failure criterion is applied in
combination with a linear ndashelastic perfectly-plastic stress-strain relationship (that is the so-
called Mohr-Coulomb model) For this model the initiation of plasticity is considered as failure
Therefore the maximum shear resistance is defined just before plastic yielding occur for any
given stress state Of course the occurrence of plasticity does not directly indicate the failure of
the soil structure however this model can give a first estimate for the stress state and the
deformations
According to this method the relative shear resistance is defined as the ratio between the
mobilized shear stress and the maximum shear resistance In Figure 510 the Mohr-Coulomb
model is illustrated in a Cartesian stress-plane In the graph the principle stresses σ1 and σ3 are
indicated More precisely the principle stresses are defined as follows
radic
(513)
(514)
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 65
radic
(515)
Figure 510 Mohrs stress circle and Mohr-Coulomb failure criterion
Hence the mobilized shear stress is
radic
(516)
And the maximum shear resistance which is the distance of the mean stress point to the yield
surface is defined as
(517)
Then the relative shear resistance is a measure for load-resistance ratio that can be determined
in any integration point in the soil continuum
(518)
However this criterion requires prior knowledge of the possible relevant failure mechanisms
and their potential locations on the soil structure This is necessary in this method in order to
choose a suitable cluster of integration points where the average value of the relative shear
strength is to be determined Therefore the limit state function to be considered is not
straightforward and it needs problem investigation in advance
4 Plaxis definition of soil collapse
In Plaxis the construction stages are analysed by performing a Load advancement ultimate level
procedure which is controlled by a total multiplier
(519)
where is the load that is applied by Plaxis is the load at the beginning of the
calculation phase (ie the load that has been reached at the end of the previous calculation
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 66
phase) and is the defined load configuration (ie water level soil weight vertical load
on the top of the dike etc) Thus the main idea is that Plaxis applies an incremental load until it
reaches the defined one When has reached an ultimate level (which by default is 1)
the current phase is finished However if a staged construction calculation has not properly
finished the multiplier fails to converge to the desired ultimate level at the end of the
phase and a warning of ldquoSoil collapsesrdquo appears in the Log info box
In other words a collapse load has been reached In this case the total specified load has not
been applied In physical terms this means that the current value of the stiffness parameter CSP
is less than 0015 (Plaxis 2D 2015b) It is a measure for the amount of plasticity that occurs
during the calculation When the solution is fully elastic CSP is equal to unity whereas at failure
it approaches zero Therefore soil reaches its upper limit of plasticity and it collapses which can
be visualized as a settlement of the dike body Such a warning in the analysis is thus assumed to
be a possible failure situation
53 Overview In this chapter an overview of the relevant possible LSF of the structural elements and the dike
body respectively has been given The different LSFs are formulated and discussed whereas the
failure of each component is going to be evaluated in the ULS of each component
The system failure in this research is considered as the combination of the anchor sheet pile
wall and soil body LSF Particularly for the anchor and the sheet pile failure their yield stress
was chosen as a limitation for their failure by adopting the next LSFs
[(
)] sheet pile wall LSF (520)
anchor LSF (521)
As far as the soil body is concerned in this research method 4 is to be used in order to identify
soil failure The limit state function is then formed as a single value in case of a successful and an
unsuccessful computation Therefore the limit state function was chosen to be equal to 1
(actually the value of is retrieved from Plaxis that in case of a successful calculation is
1) in case of a successful Plaxis analysis and -1 in case that soil collapses in any phase (error in
Plaxis Log file ldquoSoil Collapsesrdquo) This formula can be written as follows
(522)
This accounts for the global instability of the dike body in contrast with the other methods that
are related to a certain dike area (methods 1 and 3) or they are cannot handle overall failure
due to numerical issues (method 2) More precisely methods 1 and 3 were excluded due to the
prior knowledge that is required in method 3 while in method 1 a definition of the maximum
admissible deformations shall be first introduced However attention shall be paid to the type
of the warning that Plaxis gives because it might also be that the maximum number of load steps
was not sufficient In that case the phase must be recalculated using a larger number of steps
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 67
Someone could argue that such a failure can be considered as a system failure rather than a soil
failure alone Partially this is true as the incapability of the sheet pile wall or the anchor to keep
the upstream part of the dike stable can lead to non-convergence and thus according to Plaxis
definition to failure This incapability can be translated into various scenarios of failures that
stem from either the structural elements or the soil body weakness and that are illustrated in
Figure 511 This picture shows the different scenarios that can take place due to macro
instability that was discussed in 523 and the failure of the retaining wallrsquos elements
Figure 511 Combined failure scenarios (initial and developed mechanisms) (source Larsen et al 2013)
However this does not mean that the sheet pile wall or the anchor fail because of exceeding
their yield stress but rather due to possible unacceptable deformations This is also an
advantage of the inherent residual strength of the structural elements that allows a large
deformation before they fail However the residual strength of the elements was not taken into
account in this thesis as they were considered as elastic Moreover the message for the ldquosoil
collapsesrdquo can appear in a phase prior to the installation of the wall which means that this type
of failure is exclusively attributed to the deficiency of the soil strength Therefore it is not
always straightforward what is failing first and lead to the system failure Consequently by
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 68
considering such an analysis representative of a system failure it is likely to underestimate the
overall probability of failure whereas if it is considered as a soil failure only it might
overestimate the total probability of failure since it is later combined with the rest LSF of the
anchor and the sheet pile wall In this research it will be considered only for the soil failure that
brings the results to the safe side even if it does not lead to the best optimization In this thesis a
serial system of the anchor the sheet pile and the soil body LSF was considered the fault tree of
which is depicted in Figure 53
Another way to evaluate the reliability of the different elements could be the reach of a
threshold displacement according to the SLS criteria Such a threshold could be also considered
for the system as a whole rather than the individual elements However in that case someone
should be cautious regarding the choice of the value for the maximum deformations as well as
the location of their occurrence
Concluding in this section the possible failure mechanisms of a flood defence (Figure 58) were
shown Then it was clarified which of these mechanisms can be modelled in Plaxis and they
were further explained and schematized in Figure 58 with the presence of the retaining wall as
they would have to be handled in case of hand calculations Plaxis can incorporate these
mechanisms automatically and thus no distinguish among the different patterns of these
specific modes is necessary Last but not least in Figure 511 the possible failure modes for the
system of the dike and the retaining wall that can be simulated in Plaxis are depicted In this
thesis the plasticity of the structural elements is not considered and thus possible plastic hinges
in the anchor and the sheet pile are not taken into account Therefore scenario 3 is excluded
In Chapter 7 the results after the reliability analysis of the aforementioned LSFs are given It
was essential though before proceeding with the reliability analysis to get a better
understanding into the system under investigation For this purpose in the next chapter a
description of the case study as it has been simulated in Plaxis is given while also the different
structural properties are defined for the following reliability analysis The soil variables to be
considered as stochastic are presented and a first insight into of the systemrsquos behaviour is
obtained through the mean values calculation
MSc Thesis A Rippi 69
6 Case Study-Dike with an anchored sheet pile wall
In this chapter a description of the case study is presented together with the relevant soil and
structural parameters In sequence a deterministic analysis follows in order to specify the basic
structural characteristics and obtain a first sense of the stresses magnitude developed on the
structure This was carried out based on the recommendations given in the CUR 166 (2005)
The aim of this deterministic analysis is to redefine the structural properties in order to avoid
having an overdesigned structure understand the current design procedure and see if there are
any possibilities of improvement by applying the proposed reliability analysis
61 Case Description The case study concerns a dike section with an anchored sheet pile wall which was initially used
to showcase the current methodology of designing dikes with sheet piles using partial factors
For more information a reference is made to Breedeveld (2011)
In this research this case study has been modified and simplified for the needs of the thesis and
it is illustrated in Figure 61 while in Table 61 the soil materials are identified Details for the
properties of the soil layers are given in section 62 We assume the structure to be
homogeneous to the third dimension and therefore a plane-strain model is applied The dike
soil materials (layers 3 and 4 in Figure 61) were adopted from the case study used in
Breedeveld (2011) that was presented in Chapter 2 Especially in Figure 25 the original case
study as it was modeled in Plaxis is shown
Table 61 Soil materials that corresponds to Figure 61
1 Aquifer Sand layer 2 Aquitard Clay layer 3 Dike new material Clay layer 4 Dike old material Clay layer
The geometrical properties the groundwater level and the design water level were taken
identical to the original case study as well as the soil properties of the dike materials The soil
layers under the dike were modified for the need of the thesis and were modeled using random
average quantities for which the statistics were chosen arbitrarily but in a realistic range
according to the NEN6740- Table 1 (see Appendix C) included in the Dutch code for
geotechnical structures Moreover the vertical load was removed in order to examine the
impact of the water elevation on the dike stability
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 70
Figure 61 Case study Dike with anchored sheet pile wall Dimensions and Soil materials
62 Soil Parameters The soil parameters for the dike material were obtained from the report of Breedeveld (2011)
where for some parameters the characteristic values were available whereas for some others
the mean values were given For the soil layers below the dike Table 1 of NEN6740 was used as
it was mentioned before which gives characteristic values For those parameters whose value
was a characteristic 18 one a transformation was made according to their probability
distributions in order to acquire the mean values that are required for the reliability analysis
Only the properties required for the use of the Mohr-Coulomb model are presented
The parameters of each soil layer are presented in Tables 63(a) and (b) and the front number of
each layer indicates the corresponding soil layer as showed in Figure 61
The distribution types and the coefficients of variation for each parameter were chosen
according to the knowledge that has been obtained until now about the physically possible
ranges of such parameters and the recommendations that have been given in several researches
up to now (Baecheramp Christian 2003 Schweckendiek 2006 GriffithsampFenton 2007 Phoon
2008 Wolters 2012 Bach 2014) The distributions and the CoV of all parameters are displayed
in Table 62 They were considered to be the same for all soil layers Particularly for the
saturated and unsaturated volumetric weight γsat and γunsat respectively a relationship was
established in order to derive the one from the other In general γunsat varies between the real
dry weight and γsat depending each time on the degree of saturation An estimated maximum
difference between γunsat and γsat is approximately 4 kNm3 (Brinkgreve 2015) In this thesis γsat
was expressed as the summation of γunsat and a variable with uniform distribution in the range
of [0 2]19 Initially a deterministic relationship was defined between the two variables but later
on it was proved the variable was important for the soil failure and thus it was decided to
consider it as random
18 for the definition of the characteristic value see Appendix E 19 The range of [0 4] was not chosen in order to avoid extreme conditions
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 71
Table 62 Soil parameter distributions
Soil parameter Symbol DistributionRelation COV Unit
Unsaturated
Volumetric weight γunsat Normal (microσ) 5 [kNm3]
Saturated Volumetric
weight γsat γsat= γunsat+U(02) 5 [kNm3]
Cohesion c Lognormal (microσ0) 20 [kPa]
Friction angle φ Truncated normal (microσ045) 10 [ ˚]
Youngrsquos Modulus E Lognormal (microσ0) 25 [kPa]
Poissonrsquos ratio ν Truncated normal (microσ005) 10 [-]
Interface strength Rinter Truncated normal (microσ0099) 20 [-]
Table 63 (a) Soil statistical properties under the dike and (b) soil statistical properties in the dike
(a) 1 SAND (very silty)
Soil property Symbol Characteristic
value
Mean
value STD Unit
Saturated
Volumetric
weight
γsat 20 22 11 [kNm3]
Unsaturated
Volumetric
weight
γunsat 19 21 103 [kNm3]
Cohesion c 0 0 0 [kPa]
Friction angle φ 30 36 36 [ ˚]
Dilatancy angle ψ 0 0 0 [ ˚]
Youngrsquos Modulus E 20000 30769 7692 [kPa]
Poissonrsquos ratio ν 025 03 003 [-]
Interface strength Rinter 044 066 013 [-]
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 72
2 CLAY (clean medium)
Soil property Symbol Characteristic
value
Mean
value STD Unit
Saturated
Volumetric
weight
γsat 17 185 093 [kNm3]
Unsaturated
Volumetric
weight
γunsat 17 185 093 [kNm3]
Cohesion c 10 14 282 [kPa]
Friction angle φ 175 21 21 [ ˚]
Dilatancy angle ψ 0 0 0 [ ˚]
Youngrsquos Modulus E 2000 3077 769 [kPa]
Poissonrsquos ratio ν 029 035 0035 [-]
Interface strength Rinter 034 05 01 [-]
(b) 3 DIKE NEW (very sandy clay)
Soil property Symbol Characteristic
value
Mean
value STD Unit
Saturated
Volumetric
weight
γsat 17 185 093 [kNm3]
Unsaturated
Volumetric
weight
γunsat 17 185 093 [kNm3]
Cohesion c 4 564 113 [kPa]
Friction angle φ 29 347 347 [ ˚]
Dilatancy angle ψ 0 0 0 [ ˚]
Youngrsquos Modulus E 1625 2500 625 [kPa]
Poissonrsquos ratio ν 029 035 0035 [-]
Interface strength Rinter 034 05 01 [-]
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 73
4 DIKE OLD (little sandy medium clay)
Soil property Symbol Characteristic
value
Mean
value STD Unit
Saturated
Volumetric
weight
γsat 195 212 106 [kNm3]
Unsaturated
Volumetric
weight
γunsat 19 207 103 [kNm3]
Cohesion c 13 183 367 [kPa]
Friction angle φ 28 335 335 [ ˚]
Dilatancy angle ψ 0 0 0 [ ˚]
Youngrsquos Modulus E 2925 4500 1125 [kPa]
Poissonrsquos ratio ν 029 035 0035 [-]
Interface strength Rinter 034 05 01 [-]
63 Finite Element Model The structure has been modelled with the finite element software Plaxis 2D 2015 In Figure 62
the model as it was created and the automatically generated mesh are presented The ldquoboxrdquo
around the dike section was used to refine the mesh in this area The mesh in the interfaces
between the sheet pile and the soil was also refined for a better representation of the potential
high stresses that can develop in that area
Figure 62 FEM model and generated mesh for case study
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 74
The sheet pile has been modelled with elastic20 plate elements and the free anchor length with
an elastic fixed-end-anchor element The grout body of the anchor was not explicitly modelled
but it is implied in the end of the anchor as its final point is considered to be fixed in the soil A
fixed-end-anchor was chosen because even if it is a simplification of an anchorage system it can
still give reliable results in terms of the anchor axial force which is actually included in the limit
state function However the interaction of the soil and the grout body in that case cannot be
simulated The interaction and the modelling of a grouted body in the soil are still under
investigation due to several difficulties in modelling the real conditions
For all the steel members a Youngrsquos modulus of E=210 GPa was used while the specific
structural parameters are to be specified after the deterministic analysis that is described in the
next section
64 Deterministic Analysis The design concept of dikes with sheet piles modelled in FEM is based on the combination of
several partial factors in order to evaluate the overall required safety factor Additionally the
design recommendations of CUR 166 (2005) are considered for using the design or the
characteristic values of the structural parameters and the loads The design procedure (see
section 22) that has been initially followed in this structure is reported in Breedeveld (2011)
and the following structural elements have been defined accordingly
Cross sectional area of the sheet pile (per meter)
Length of the sheet pile wall
Steel quality of the piles
Cross sectional area of the anchor (per meter)
Free length of the anchor (without the grounded body)
Steel quality of the anchorrsquos reinforcement
Profile of waling
The required overall safety factor was calculated up to 18 according to the current design
procedure described in section 22 according which the above mentioned structural parameters
were defined Moreover apart from the overall safety factor in Breedeveld (2011) additional
partial factors are applied on the deduced forces and bending moments of the anchor and the
sheet pile wall as they are defined in CUR 166 (2005) In particular in Table 64 the different
partial factors that are applied on the several structural developed actions are given
Table 64 Partial factors for the different structural actions according to CUR 166 (2005)
Structural action Partial factor γ
Sheet pile normal force Fn 115
Sheet pile bending moment M 115
Anchor normal force FA 125
Waling bending moment Mw 110
20 The elements were considered as elastic in order to avoid numerical problems in FORM analysis For a better explanation a reference is made to the evaluation of the optimization algorithms in Appendix A2
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 75
However as it is mentioned in section 62 alterations have been made in the soil layers below
the dike while also the vertical load has been removed On the top of that in section 22 it was
shown that the large applied partial factors led to an overestimation of the design values for the
moments of the sheet pile Therefore a redesign of the sheet pile and the anchor is carried out
in this section in order to determine the new required characteristics of the structural elements
according to the new loading situation with less strict partial factors In particular in this thesis
the required overall safety factor was considered 12 according to CUR 166 (2005) calculation
scheme that is described in the next section The reason for using a lower safety factor was to
avoid an overestimation of the design forces and moments of the sheet pile wall that would
render a reliability analysis meaningless as the structure would be quite safe Such an analysis is
also important in order to show the discrepancy between the two different design procedures
and indicate the optimal one
However the reliability analysis was carried out for the mean values of the soil parameters and
for a safety factor of 10 in order to account only for soil properties uncertainties A safety factor
of 10 was not directly used as the design procedure followed in section 643 accounted only for
the macro-instability of the system and thus other failure mechanisms such as piping
overtopping infiltration etc were not taken into account In that way we prevent coming up
with a retaining wall that would be volatile under other crucial failure mechanisms
In section 641 the calculation scheme to be used in this thesis for the deterministic analysis is
further described In the sequence in section 642 the construction phases that were followed
in Plaxis simulation are listed and illustrated and finally in section 643 the results of the
preliminary design are displayed
641 Calculation Scheme and Design Values
In this research only the design philosophy suggested in CUR 166 is applied and the desired
safety factor of the structure is considered to be 12 for structures classified in ldquosafety class IIIrdquo
in order to avoid an overestimation of the structurersquos dimensions as it happened in the original
case study At this point it should be also clarified that the redesign was conducted considering
only the overall stability of the structure and no other failure mechanisms such as piping
internal erosion or overtopping
The prescribed calculation schemes in CUR 166 (2005) for designing structures with retaining
walls are basically the following
Calculation scheme A Calculations with design values The calculations are executed
using the design values for the soil parameters retaining height water levels and
stiffness of the structure It uses two different soil stiffnesses When using a high soil
stiffness the anchor force is generally relatively high The advantage of this scheme is
that it requires relative little effort The disadvantage is that the deformations may be
overestimated due to the use of design values in every stage However a designer is not
interested in deformations when performing a ULS calculation
Calculation scheme B Calculations with characteristic values The calculations are
executed using characteristic values of the soil strength (crsquo and φrsquo) Design values are
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 76
used for the retaining height water levels external loads and stiffness of the soil Here
again two different soil stiffnesses are used In the end of the governing stage a phi-c
reduction should be carried out As the characteristic values for the soil parameters are
used the deformations are most probably smaller than in case of calculation scheme A
The safety margin lies in the fact that phi-c reduction is imposed to obtain a safety factor
of 115 (class II) or 12 (class III)
For the soil stiffness parameters two options are given in CUR 166 a low and a high value
(61)
where is the (low and the high) design value the (low and the high) characteristic value
and is the mean value of the stiffness parameter In Woltersrsquos master thesis (2012) both
have been applied and the appeared to be governing for the bending moment in the wall
and the anchor force
In this research the Calculation scheme B was used and the as the soil stiffness
parameter It should be mentioned that regarding the soil parameters the characteristic values
were available with an exemption for the stiffness for which the mean values were given The
groundwater level was manually schematized according to TAW (2004) because of lack of data
while the maximum water level in the river side was taken equal to the design water level
according the Dutch regulations for the specific dike region
In Tables 65 and 66 the partial factors according to CUR 166 and the corresponding design
values for the different soil properties are presented As someone can notice in these tables
only the strength parameters (phi and c) and the stiffness parameter (E) are exhibited For the
rest parameters the design value is identical to the characteristic value (see Tables 63(a) and
(b)) as a partial factor of 1 is used
Table 65 Partial factors and design values for the soil layers under the dike
1 SAND (very silty)
Soil property Symbol Partial factor γk Design value Unit
Cohesion c 11 0 [kPa]
Friction angle φ 12 2569 [ ˚]
Youngrsquos Modulus E 13 1538462 [kPa]
2 CLAY (medium)
Soil property Symbol Partial factor γk Design value Unit
Cohesion c 11 91 [kPa]
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 77
Friction angle φ 12 1472 [ ˚]
Youngrsquos Modulus E 13 153846 [kPa]
Table 66 Partial factors and design values for the soil materials in the dike
3 DIKE NEW (medium clay)
Soil property Symbol Partial factor γk Design value Unit
Cohesion c 11 364 [kPa]
Friction angle φ 12 2479 [ ˚]
Youngrsquos Modulus E 13 1250 [kPa]
4 DIKE OLD (stiff clay)
Soil property Symbol Partial factor γk Design value Unit
Cohesion c 11 1182 [kPa]
Friction angle φ 12 239 [ ˚]
Youngrsquos Modulus E 13 2250 [kPa]
642 Construction Stages
The construction and the gradual loading of the dike ware modelled as follows (see Figure 63)
1 K0-procedure for the generation of the initial stresses under horizontal groundwater
level
2 Dike self-weight under horizontal groundwater level
3 Rise groundwater level to the phreatic water elevation
4 Interpolation21 of the clay layer under the dike and activation of the sheet pile wall and
the anchor
5 Apply extreme water level conditions on the dike structure
6 (φ-c reduction for the determination of the safety factor)
21 Due to different permeability between the clay and the sand layer an approximation of the developed pore pressures is carried out through interpolation between the two layers
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 78
1
2
3
4
5
Figure 63 Construction and loading stages
For the soil modelling the Mohr-Coulomb model was used with a non-associative flow rule
(φneψ) For phases 1-4 the calculations were conducted under drained conditions whereas in
phases 5 and 6 the undrained behaviour of clay layers was considered
643 Determination of the structural elementsrsquo characteristics
According to the calculation scheme that was described above the design characteristics of the
structural elements were obtained from phase 6 In particular the target overall safety factor is
12 with which initially the length of the sheet pile wall was calculated
+370 NAP
+700 NAP
+1050 NAP
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 79
Firstly a length of 10 m and a profile of AZ 12 with a steel quality of 240 kNm2 yield stress
were used while for the anchor a cross-section of 628 mm2 and S355 were chosen for the first
calculation However even if the safety factor reached 12 it seemed to be quite unstable during
the calculation steps while the developed anchor stress was above the ultimate yield stress
What is more the deformations resulted in that case were more than 30 cm Therefore in order
to stabilize the safety factor and to decrease the displacements the length of the sheet pile was
increased 1 m (so 11 m in total) while also a better steel quality for the sheet pile was chosen
S355 Furthermore the anchor cross section was increased to 933 mm2 with a steel quality of
MW450 Last but not least the waling was designed based on the anchor force
In the following paragraphs the design procedure of each element is described providing also
the results of moments and forces that were retrieved from Plaxis
Sheet Pile wall
As it was partially mentioned in the previous paragraph the length of the sheet piles is
considered to be sufficient when the safety factor of the system22 reaches at least 12 For the
reasons discussed above the sheet pile length was decided to be 11 m and with a profile of AZ
12 S355
With this configuration the safety factor reached is 13 as it is depicted at the right graph of
Figure 64 Someone can notice that the safety factor from 117 in the situation without a sheet
pile structure increased to 13 with the structure
Figure 64 Safety factor deduced from the phi-c reduction calculation without the anchored sheet pile wall at the left (safety factor=117) and with the wall at the right (safety factor=13) The steps on the x-axis represent the calculation steps have been carried out in order to reach an equilibrium after incrementally reducing the strength parameters ie phi and c
The design moment is to be taken for a safety factor of 12 that is reached in step 34 where it
was found to be 1519 kNmm and the corresponding axial force 9133 kNm Therefore
considering Eq 52 for checking the exceedance of the ultimate yield stress and applying the
prescribed partial factor of 115 for both the moment and the axial force we get
22 Plaxis always gives the safety factor of the system which is under investigation
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 80
In Figure 65 the maximum bending moments with a safety factor 10 and 12 are respectively
showed From the values being shown in that figure someone can easily distinguish the
difference between the two moments It must be also mentioned in that point that the required
safety factor defined in Chapter 2 was estimated at 18 (and not 12 that is considered here)
which gives excessive design moments (almost 7 times larger) For that reason in this research
a lower safety factor was assumed avoiding thus the design of a quite conservative structure
Figure 65 Bending moments on the sheet pile for a safety factor safety factor=10 at the left and
safety factor=12 at the right
The explanation behind this discrepancy is that in case of the phi-c reduction procedure the
artificially decreasing strength of the soil and on the other hand the non-reduced strength and
stiffness23 parameters of the structural elements tends to lead in an increased arching
phenomenon that makes the structure attract the most of the load
In this research the reliability of the structure will be evaluated for a safety factor of 1 or else
during phase 5 and by considering the mean values of the soil parameters As someone can
notice from Figure 65 the magnitude of the maximum bending moment for a safety factor =10
is quite low which implies a possible minor contribution to the overall failure
At this point it is essential to mention that Mohr-Coulomb model is capable of simulating the
elastic behaviour and the state at failure Therefore in case that plasticity of the soil or the
structural elements plays an important role a different constitutive model is recommended to
be applied
Anchor
For a safety factor 12 the anchor force was found to be 10387 kNm Considering a mutual
anchor distance of 3 m the total axial force that is undertaken by the rod multiplied also with
the prescribed design partial factor is
23 ldquoStrengthrdquo and ldquostiffnessrdquo are usually confused terms Strength is a measure of the maximum load that can be placed on a material before it permanently deforms or breaks and it is represented by the yield stress σy in metals and by φ and c in soil Stiffness is a measure of the amount of deflection that a load causes in a material and it is represented by the Youngrsquos modulus E and the Poisson ratio ν both in soil and metals
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 81
Therefore choosing an anchor with a cross sectional area of A = 933 mm2 and a steel quality of
MW450 ( ) the design stress level is
In comparison with the sheet piles the anchor does develop large stresses as it was concluded
from the results in the deterministic calculations This is probably attributed to the high
stiffness of the anchor that apparently attracts the load and thus the anchor develops large axial
force
Waling
The type of the waling system is defined according to the anchor force as it was described in
section 522 Therefore by choosing a waling type of 2UPE 200 the developed bending
moment can be estimated as follows
where is the design bending moment of the walling system according to the chosen profile
In Table 67 the characteristics of each structural element that was described above are
displayed in detail
65 Overview In this chapter an introduction of the case study as it has been used in this research is made The
soil materials and their properties are summarized and their mean and design values are
defined The probability distributions of the soil parameters to be considered as stochastic were
proposed according to the literature study and the engineering judgement Moreover the mesh
generation and the modelling features are discussed
Last but not least due the observed overestimation of the retaining wallrsquos design moment and
axial force according to the current design regulation a new deterministic analysis was carried
out in order to redefine the structural properties and come up with a less conservative structure
This makes the reliability analysis of the system meaningful in a sense that an extremely safe
structure would give a failure probability of zero Besides the scope of a reliability analysis is
also the optimization of the construction that corresponds to a financially attractive and
simultaneously safe structure This can be made by conducting a risk analysis whose first step
would be the analysis that is carried out in this research Therefore the aim is to start with the
marginal required structural properties according to the safety standards In a risk analysis that
should conducted for several dimensions of the structure until the most optimized one is
identified However this is out of the scope of this thesis
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 82
In Table 67 the structural properties of the sheet pile wall the anchor and the waling are
summarized as they were found in the deterministic analysis
Table 67 Design parameters of structural elements
SHEET PILES
Property Symbol Value Unit
Profile - AZ 12 [-]
Steel quality - S355 [-]
Length L 11 [m]
Moment of inertia I 21430 [cm4m]
Elastic section modulus Wel 1245 [cm3m]
Mass w 943 [kgm]
Sheet pile thickness d 85 [mm]
Cross sectional area A 1201 [cm2m]
ANCHOR
Property Symbol Value Unit
Steel quality - MW450 [-]
Free length Lafree 104 [m]
angle φ 30 [deg]
Cross sectional area A 933 [mm2]
Mutual anchor distance s 3 [m]
WALING
Property Symbol Value Unit
Profile - 2UPE200 [-]
Steel quality - S355 [-]
Elastic section modulus Wel 191 [cm3m]
In the original case study the sheet pile wall is 162 m long and it has a profile of AZ 28 (S240)
while the anchor is an MW450 with a cross-section of 1134 mm2 As it is noticed both sheet
pilesrsquo and anchorrsquos properties have been reduced after the new deterministic analysis However
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 83
it should be mentioned that in the original case there was a vertical load on the top of the dike
that was displaced while the clay layers under the dike was merged into one in the new case
study These changes might lead to less heavy structure However the difference among the
design properties between the new and the original case study is mainly due to the required
safety factor that was assumed In the new case study the safety factor was set to 12 whereas in
the original cases study a safety factor of 18 was considered as it was calculated according to
the partial safety factors
In the next chapter the results after the reliability analysis of the just designed case study are
presented
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 84
MSc Thesis A Rippi 85
7 Reliability analysis results with stochastic soil properties
In this chapter the probability of failure of the different structural elements that were discussed
in Chapter 4 will be estimated by varying the soil properties of the several soil layers under and
inside the dike At the beginning a better insight into the system behavior is gained by
inspecting the deterministic Plaxis calculations for the mean values of the soil variables After
that a global SA is carried out in order to evaluate the most important soil parameters and
inspect their impact on the different structural elements in combination with each other and
individually Finally the results regarding the probability of failure of the anchor the sheet pile
wall and the soil are presented as they deduced by the reliability analysis (concerning the ULS of
the elements) Finally the systemrsquos reliability is evaluated The methodology that was followed
during the probabilistic analysis is described below and it is mainly dedicated to the DS
rationale that was used in order to obtain the influence factors and to evaluate the results of
Plaxis calculations
71 Method description For the calculation of the failure probability of the systemrsquos different components FORM and DS
were used for the retaining wall and the soil body respectively From these methods the
probability of failure and the important factors are going to be deduced for each component (ie
anchor sheet pile wall and soil body) However before starting with the reliability analysis it is
essential to carry out a sensitivity analysis in order to obtain a first impression of what are the
most influencing variables and thus determine the input random parameters for the reliability
analysis In the figure below the steps for the reliability analysis as they were followed in this
thesis are presented
Figure 71 Steps of the reliability analysis
Below the methodology of both the sensitivity and the reliability analysis are discussed and
explained in detail
Sensitivity analysis
In this thesis the FAST method was utilized for conducting the sensitivity analysis which is
further elaborated in Appendix A1 Here an explanation is given for the quantitative measure of
sensitivity which is represented by Sensitivity Indices The first-order sensitivity index of an
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 86
input variable pi is the measure of the main (direct) effect of pi on the output variance
(where i ne j) the second-order sensitivity indices measures the interaction effect of pi and pj on
the output variance Other higher-order indices are defined in the same manner The total
sensitivity index is the sum of all sensitivity indices involving factor pi (Homma amp Saltelli
1996) For example the total sensitivity index of factor 1 for a model with 3 input factors is
given as
The first-order (Si) and total ( ) indices can be interpreted as following (Pandya et al 2008)
high pi is an influent parameter
and both small pi is not an influent parameter (neither alone nor in interaction
with other parameters)
and nearly the same no interaction of pi with the other parameters
and very different high interaction of pi with the other parameters
Reliability analysis
As far as the FORM analysis is concerned section 422 gives sufficient information about the
methodology that is followed in order to obtain the failure probability and the influence factors
(ie α-values) The main concept and principles of DS are briefly explained in section 421
whereas here a more precise description of the implementation of DS and the how the α2-
values can be deduced is given
In particular DS is a method in which ldquodirectionsrdquo are sampled The concept of this method is
schematically depicted in Figure 72(a) For each sampled direction it is evaluated whereas
along this line the LSF Z equals to zero or not This procedure is repeated for a number of
directions and this method is applied in the standard normal space (u-space see also section
422) using equations to transform the standard normal variables u1 hellip un to their
corresponding ldquorealrdquo variables (in the x-space) X1 hellipXn and vice versa
(a) (b)
Figure 72 (a) Schematic view of the DS method in the u-space The numbers 0 1 2 3 4 show successive LSF evaluations along one direction (b) Schematic view of the line search algorithm β is the distance from the origin in the u-space along one direction The numbers 0 1 2 3 4 correspond to the ones shown in figure (a) (source Hydra-Ring 12 2015)
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 87
The search procedure for the location on the line where Z=0 is often referred to as the ldquoline
search methodrdquo and it is depicted in Figure 72 (a) and (b) In both figures the numbers 0-4 refer
to successive Z-function evaluations along one direction Figure 72(a) shows 3 evaluated
directions in the u-space whereas Figure 72(b) shows the (fictional) value of the Z-function
along one direction (where ||u|| is the distance from the origin along the line) The maximum
length of the direction line as well as the step size along the direction (for example the distance
between 0 and 1 along one line) can be steered in order to increase the efficiency of the method
and presumably decrease the computational time
In the sequence a method was developed in this thesis in order to estimate the α2-values of the
random variables out of a DS analysis as OT does not have an available method so far For that
purpose a transformation of the output samples to u-space was firstly carried out Then the
distance to the origin of all samples (in u-space) that are located on a direction where failure
(Z=0) was detected is calculated as follows
radicsum
i=1hellip (71)
where n is the number of random variables and uij is the ith ldquofailurerdquo sample of the jth random
variable is equivalent with the reliability index of this direction βi The α-values for each
random variable can then be calculated as (for more information about these values see
sections 422- Eq 417 and 423)
(72)
After that three different methods were created in order to evaluate the influence of the
random variables on the response of the LSF ie the α2-values (square of Eq 72)
Shortest distance ldquoβminrdquo
Average 10
Average all
In the Shortest distance method the sample with the smallest distance to the origin is
considered to be the design point and thus the corresponding influence factors are estimated
according to that point The smallest distance is also equivalent to the reliability index βmin This
method is well-known for estimating the design point out of MC techniques (Vrijling amp Van
Gelder 2002) However for a soil failure modeled in FEM this might not be always indicative of
the real influence that the soil properties can have on the system failure Therefore the last two
techniques were used as additional in order to get a better insight into the dominant soil
properties In Figure 73 a general scheme of the methodology that was followed for obtaining
the α2-values is depicted where also the different techniques are presented
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 88
Figure 73 DS methodology for calculating α2-values
The second method Average 10 takes into account the α2-values of the samples in the failure
domain whose distance to the origin lies within the 10 higher than βmin and averages them
Therefore it is firstly required the calculation of βmin as it was described for the Shortest distance
method Such a technique helps to identify the other directions that are close to the one which
gives the shortest distance to the origin and thus inspect if the important variables that were
indicated according to the first method are influential for other failure mechanisms as well
Last but not least the third method Average all averages the α2-values of all the samples
located in the failure domain This shows the overall contribution of the random variables to the
response of the LSF under investigation Concerning the soil failure LSF that is described in
section 73 such a method can be more suitable for distinguishing the soil properties that are
involved in the majority of the failure mechanisms rather than only in the one that the first
method indicates
72 Mean values calculations The results of the deterministic analysis that are presented in the previous chapter included a
degree of conservatism due to the load and material factors (partial safety factors) In order to
overcome this limitation structural reliability calculations were carried out and are discussed in
the next sections by considering relevant soil parameters as random However before starting
the probabilistic calculations by coupling Plaxis with OT that cannot be fully interpreted and
verified to a certain extent it is essential to first carry out some deterministic calculation Such
calculations can help to get a better insight into the behavior of the system and show
qualitatively what kind of response to wait later in the analyses These calculations were taken
place based on the mean values of the soil properties
Deformations
The analysis of this case study focuses on the ULS of the structure For the stresses and
deformation calculations the Mohr-Coulomb model is used that is generally accepted for the
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 89
simulation of the soil behavior until failure and for detecting the failure modes (a more detailed
explanation of this model is provided in Appendix B1) However this model is not so suitable
for analyzing the deformations of soft soils or the settlements around the wall
The deformation pattern that would more likely occur due to primarily the soil body failure can
be illustrated by conducting a ϕ-c reduction calculation In Figure 74 (a)-(c) the displacement
pattern after the ϕ-c reduction of the mean values of the friction angle and the cohesion
respectively As it can be seen a slip surface in the inner side of the dike has been created that
rotates towards the inland
(a) deformed mesh
(b) shadings
(c) arrows
Figure 74 Deformations after the ϕ-c reduction (mean values)
Stresses on the structural elements and the soil body
The anchor and the sheet pile wall are installed in phase 4 In phase 5 the anchor normal force
increases and especially for the mean values it amounts to Na = 7983kNm
As far as the sheet pile wall is concerned the bending moments and the normal forces that are
developed on the wall are indicated in Figure 75 as they were deduced from phase 5 Similarly
to the anchor the maximum stresses occur in phase 5 and are mainly located in the area of the
wall which is surrounded by the clay layer (brownish soil layer in Figure 74(a)) As someone
can observe the bending moments are not significant taking into account that the maximum one
observed reaches 2132 kNm
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 90
Figure 75 Bending moments and normal forces for the mean values calculation
For the interpretation of the stress field of the soil body the effective principle stresses and the
relative shear stress τrel (see Eq 517) were illustrated in Figure 76 (a) and (b) respectively In
Figure 76 (a) the crosses indicate the direction and the magnitude of the principle stresses σ1rsquo
and σ3rsquo and from which an active soil behavior (
) can be noticed on the river side of
the retaining wall whereas a passive behavior (
) is dominant in the inland side
(a) Effective principle stresses
(b) Relative
shear stress τrel
Figure 76 Effective principle stresses and relative shear stress in mean values calculation
In Figure 76 (b) the distribution of the relative shear strength shows the potential developing
plastic area in case of failure that expands from the river side of the wall and around it It can
also betoken the possible shape of formation of a slip surface that starts from the outer side of
the dike and goes around the wall
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 91
73 Sensitivity Analysis Results In section 312 different types of SA are presented and that can be applied through OT In this
research due to the time limit and the amount of parameters FAST method was used for
carrying out the SA as it allows the calculation of the first order sensitivity indices as well as the
total order indices and the computational time required is less than the other available methods
In this section the results from the conducted SA are presented separately for the anchor the
sheet pile and the soil performance Regarding the sheet pile wall and the anchor the sensitivity
of the total developed stresses towards the soil properties was evaluated according to Eq 53
and 55 respectively The influence of the soil parameters on the soil behaviour was evaluated
based on the calculation of the overall safety factor after a safety calculation (ϕ-c reduction
method) The total order indices are considered for distinguishing the most important
parameters while their difference with the first order indices Si indicates the level of interaction
effect amongst the parameters on the output variance
As far as the anchor and the sheet pile wall are concerned the SA was carried out in two steps
(1) SA with all the soil parameters (listed in Table 62) and (2) SA with the most important
parameters found in step 1 In Figure 77(a) and (b) the sensitivity indices from step (1) are
presented as they were found for the anchor and the sheet pile respectively Step 1 mainly
provides an overview of the contribution of the variablesrsquo variance to the total variance of the
limit state response and it helps to identify the most influencing parameters Step 2 was
basically carried out in order to build the response surface based on the SA output that was
later utilized in the FORM analysis as a complementary function in case of Plaxis errors At that
point it should be mentioned that the Dikenew material was not considered neither for the SA
nor for the reliability analysis of the structural elements as its influence was assumed to be
insignificant relative to the other soil layers However Dikenew was considered in case of the
soil body investigation
From Figure 77(a) it can be concluded that as far as the anchor is concerned the shear stiffness
G of the clay layer seems to be the most crucial soil property In particular the anchorrsquos stress
level seems also to be sensitive towards the Poisson ratio of the clay layer ν the friction angle φ
of the sand layer and the strength parameters φ c of the Dikeold material Moreover the
interface strength Rinter between the Dikeold material and the sheet pile wall seems to display
an additional notable influence as well The unit soil weight γ does not appear to affect the
anchor performance significantly
As far as the sheet pile wall is concerned the shear stiffness of the clay layer and the Dikeold
material seem to be dominant according to Figure 77(b) From these two soil layers it appears
that also the strength properties (φ c and Rinter) as well as the unit soil weight γ have a relative
impact on the stress level of the wall Furthermore the sand layer contributes via especially its
friction angle and shear stiffness
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 92
(a)
(b)
Figure 77 Total and First order sensitivity indices for the (a) anchor and the (b) sheet pile in step 1
As it was mentioned before step 2 was used for building the response surfaces for the anchor
and the sheet pile limit states The input random variables in this step are the most influencing
ones that were found in step 1 In Table 71 the soil properties that were considered as random
during step 2 are presented for both anchor and sheet pile It is essential to mention that the RS
was constructed by fitting a quadratic polynomial function to the sampling points of the SA
whose general expression is given in Eq 425 In Figure 78(a) and (b) an illustration of these
response surfaces is made However the figures depict the response of the anchor and the sheet
pile stress level over the fluctuation of only two variables amongst the 14 that are presented in
Table 71
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 93
Table 71 Soil properties considered for the RS construction of the anchor and the sheet pile respectively
Soil layer Anchor Sheet pile
Clay Sand
Dikeold
The linear regression of all the parameters was based on least squares and the R2 parameter
was found 0989 for the sheet pile and 0996 for the anchor Therefore the regression was quite
representative for the domain of the samples over which it was adjusted However there were
many residuals (ie difference between the observed and the predicted value) in the order of
103 and 104 which means that a possible expansion of this RS to an extended domain of the
input variables may lead to an underestimation of the limit state response
(a)
(b)
Figure 78 Polynomial response surface adjusted for the (a) anchor and (b) sheet pile stress level For the anchor the friction angle of the sand layer and the shear stiffness of the clay layer were accounted for this illustration whereas for the sheet pile wall the shear stiffness of the sand and the cohesion of the clay layer were used
The SA for the soil performance was also conducted in two steps However in that case there
was no need of a response surface construction since the LSF for the reliability analysis (Eq
522) takes the value of -1 in case of Plaxis error due to soil collapse Eq 522 though takes only
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 94
two different values -1 and 1 for soil collapse or not respectively which cannot be used for
carrying a SA For that reason the φ-c reduction method was used and the sensitivity of the
value of the safety factor (Msf) was investigated instead The second step of the SA was carried
out in order to divide the soil properties into two different groups of variables and thus
investigate the sensitivity of the soil in a more efficient way rather than including all the
parameters at one SA It should be mentioned that in the soil analysis soil properties of the
Dikenew material were added In Figure 79(a) and (b) the total and the first order indices are
depicted for step 1 and 2 respectively Step 1 includes all the stiffness and strength parameters
of all the soil layers whereas in step 2 the most important variables that were found in step 1
together with the unit soil weight γ are included Subsequently step 1 detects the most
influencing stiffness and strength properties and in step 2 the impact of γ is investigated
As it is observed in Figure 79(a) and (b) the safety factor appeared to be sensitive towards the
sandrsquos friction angle while also the strength properties (c and φ) and the shear stiffness (G) of
the clay layer play an important role Moreover the cohesion of the Dikeold material and the
unit soil weight of the clay sand and Dikenew layer seem to be influencing as well As it has
been mentioned above for this SA the LSF was different than the one considered in the
reliability analysis Therefore even if the SA at that point helps to evaluate in general the most
dominant soil properties the engineering judgement should also be included for the selection of
random variables that are going to be utilized in the reliability analysis
(a)
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 95
(b)
Figure 79 Sensitivity indices for the soil in step 1 and 2
It is also advisable that the difference between the total and the first order sensitivity indices be
investigated in order to have an insight into the most interactive variables The magnitude of the
interaction effect of a variable is crucial for the system behaviour because a variable might not
be important as a singularity but it is possible its combination with another variable to have a
considerable effect on the limit state under investigation Therefore this can be considered as
an additional criterion for choosing the set of the random variables to be used in the reliability
analysis In Figure 710(a) and (b) the difference between the two indices is presented for the
anchor and the sheet pile in (a) and the soil in (b) as it was deduced from the first step of the SA
Regarding the sheet pile wall the most interactive variables are the shear stiffness and the unit
weight of the clay layer the friction angle of the sand and the cohesion shear stiffness interface
strength and unit weight of the Dikeold material Concerning the anchorrsquos limit state the
cohesion of the Dikeold material seems to have the highest interaction with the rest of the
variables Last but not least regarding the soil performance the stiffness of all the soil layers
and the friction angle of the clay and the sand layer have a higher interaction effect amongst the
other parameters The importance of the aforementioned variables can be enhanced
considering both total indices and the difference that they display between their first and total
indices
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 96
(a)
(b)
Figure 710 Interaction effect amongst the parameters in (a) anchor and sheet pile and in (b) soil This effect is expressed as the difference between the total order indices T and the first order indices S as they were deduced from the first step of the SA for both anchorsheet pile and soil respectively
Considering the discussion above some preliminary conclusions can be drawn regarding the
influence of the several variables to the different components of the system
1 both anchor and sheet pile wallrsquos limit states are sensitive mostly towards the shear
stiffness of the soil
2 sheet pile wall stress level is also affected by the unit soil weight more than the anchor
does
3 soil body is mostly influenced by the strength properties and the unit soil weight
4 the interaction effect among the soil properties is more profound for the soil and the
sheet pile limit states rather than the anchorrsquos
It must be mentioned that this SA is a global one that gives an overview of the impact of the soil
parameters in terms of the system behaviour However near the design point a local sensitivity
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 97
can possibly give a better estimation of the variablesrsquo impact Near the design point the
structural elements as well as the soil body usually reaches plasticity that implies a non-linear
behaviour Such a behaviour is unlikely to be initiated during a SA that usually samples around
the mean values This can be counteracted by conducting more iterations so as to have a larger
amount of available samplings but this can make the SA to be quite time consuming and thus to
lose its benefit to give a fast preliminary estimation of the most influencing soil layers and
properties
Besides the purpose of a preliminary global SA is to reduce the number of the stochastic
parameters to a reasonable amount so as the initial reliability analysis to be more efficient and
computationally accessible The parameters that were eventually considered as stochastic in
each LSF are summarized in the table below and they were chosen according to both the SA and
the engineering judgement In the last column the stochastic variables used also for the
systemrsquos analysis are presented
Table 72 Stochastic variables for each LSF
Soil
Failure Anchor Failure
Sheet Pile failure
System failure
Sa
nd
γunsat radic
c
φ radic radic radic radic
G radic radic radic radic
v
radic
Rinter radic
Cla
y
γunsat radic
radic radic
c radic radic radic radic
φ radic radic radic radic
G radic radic radic radic
v
radic
Rinter radic radic radic
Dik
e_n
ew
γunsat radic
c radic
φ radic
G
v
Rinter
Dik
e_o
ld
γunsat
radic radic
c radic radic radic radic
φ radic radic radic radic
G
radic radic radic
v
radic
Rinter radic radic radic
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 98
74 Soil Shear Failure As far as the soil failure is concerned the limit state function that was implemented is given
below (see also sections 52 and 53) It was formulated as a single value depending on the
success of computation in Plaxis
(71)
The reliability method that was chosen in that case is DS due to the formulation of the limit state
function and the convergence issues of FORM There were used 12 input stochastic soil
parameters as it is indicated in Table 72 (see also input file of Soil Failure in Appendix D) The
variables were chosen based on the SA that was presented in the previous section For the soil
failure Poisson ratio was initially included in the analysis but it was noticed that values close to
04 during Plaxis calculations were leading to unrealistic behaviour of the dike More precisely
it has been observed that Poisson ratio equal to 04 can lead to swelling effects on the dike
which consequently lead to a raise of the dike crest instead of collapsing as it would be expected
in a soil failure analysis Therefore Poisson ratio was excluded from the reliability analysis of
the soil failure
The results of the reliability analysis according to DS are presented below In particular 300
iterations were carried out and 3 different failure directions were found Specifically in Table
73 the probability of failure the reliability index the number of iterations and the duration of
the analysis are firstly shown In the sequence the points from each failing direction with the
lowest reliability index are shown
Table 73 Reliability results of DS for the soil failure
DS Pf 1310-8 β 55 Number of LSF calls 1840 Number of iterations (or directions)
300
Elapsed time (hr) 34
Failure points (in real space) Parameter 1st direction 2nd direction 3rd direction Dikenew_phi [deg] 2277 2543 3021 Dikenew_c [kPa] 529 557 444 Dikenew_gammaunsat [kNm3]
1748 1848 1973
Dikenew_gammasat [kNm3]
1948 1984 1997
Sand_phi [deg] 2782 2039 3899 Sand_G [kPa] 808437 715473 715785 Sand_gammaunsat [kNm3]
2183 2122 2097
Sand_gammasat [kNm3] 2352 2145 2283 Clay_G [kPa] 56934 90063 80134 Clay_c [kPa] 704 1593 794
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 99
Clay_phi [deg] 1039 1629 1077 Clay_gammaunsat [kNm3]
1772 1413 1612
Clay_gammasat [kNm3] 1833 1433 1719 Dikeold_c [kPa] 1621 2978 915 Dikeold_phi [deg] 2474 3149 3281
According to the Shortest distance method the failure point from the 2nd direction turned out to
be the design point the importance factors of which are depicted in Figure 714 (first pie chart
at the left) However before concluding for the final design point an assessment of the above
failing directions was made
In particular the combination of the variables from each direction was later implemented in
Plaxis in order to illustrate the failure This is also necessary for verifying and assessing the
failure that Plaxis is giving and inspecting the error in the logfile in order to check if it is a ldquorealrdquo
(ie error=Soil collapses) or a ldquofakerdquo failure (ie error=Load advancement failed or Not enough
load steps) In Figure 711 the ldquorealrdquo failure is defined
Moreover a demonstration of the displacements is also essential in order to certify the failure
shape that normally has the dike crest settled down A different output of the displacements
formation (ie dike crest goes up) would probably imply a ldquofakerdquo failure However it should be
mentioned that a ldquofakerdquo failure in Mohr Coulomb model can be a ldquorealrdquo failure in another model
that can more realistically simulate the deformation patterns of the structure In general even
though Mohr Coulomb can be successfully applied for detecting failure it is not very suitable for
the determination of the displacements and especially for soft soils such that the dike under
investigation is constructed by It should be also mentioned at that point that in this case study
the system is subjected to unloading conditions24 (ie decrease of effective stresses) in soft soil
layers after phase 4 which cannot be modelled successfully by the Mohr-Coulomb model In that
case other constitutive models such as the Soft Soil Model and the Hardening Soil model are
presumably more reliable for simulating such situations For further information for this model
a reference is made to Plaxis (2015) Therefore someone should be always critical and verify
the failure points based on hisher engineering judgement
Figure 711 Prerequisites for real failure in Plaxis
24 From phase 4 to 5 there is an increase in water pressures and thus due to constant total stresses there is a decrease of the effective stresses This can be characterised as an unloading situation
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 100
In order to make it more understandable the different failure points were illustrated in Plaxis
and their deformed mesh and total displacements are presented in Figure 712 (a)-(c) for each
failure direction As one can observe in the first two directions the dike crest settles down
whereas in the 3rd one the dike body seems to swell and move to the right (the blue lines that
surrounds the dike soil layers indicate the initial dike geometry) However even if the 2nd
direction gave a normal pattern of displacements the error message was ldquoNot enough load
stepsrdquo The combination of the variables of the 3rd direction was then tested again with the
Hardening Soil model (HS) in order to illustrate the deformations with a more suitable model
and it was proved that the deformations seem to be realistic as it is depicted in Figure 713 With
Mohr Coulomb only the 1st direction seems to comply with the diagram in Figure 711 and that
would properly be considered as the real design point
(a) Deformed mesh
(b) Total displacements |u| in m (shadings)
(c) Total displacements direction (arrows)
1st direction 2nd direction 3rd direction
Figure 712 Illustration of the design point in soil body failure
(a) Deformed mesh (b) Total displacements |u| in m (shadings)
(c) Total displacements direction (arrows)
Figure 713 Displacements pattern of the 3rd failure direction according to HS
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 101
However taking into account that the other directions showed characteristics of failing behavior
they should also be accounted for the determination of the importance factors Excluding them
from the analysis would not be wise as the error message or the deformation schematization can
be a limitation of FEM or Mohr Coulomb model respectively In any case the points that lead to
failure shall be treated critically in order to decide if they should be included or not In this case
the points from the three directions seem to be realistic as both strength and stiffness
parameters are low enough in order to lead to a potential failure Therefore they were finally
considered for the analysis and the importance factors were averaged over all the points as it is
depicted in Figure 714 (last graph at the right)
Figure 714 Importance factors α2 for soil failure
Another option is also to average over the 10 range of close to the lowest reliability index that
was calculated with the shortest distance method (ie Average 10 method) in order to come
up with the α2 values However such methodology can lead to averaging over points of the same
direction that are not very different with each other Therefore the rest of the failure points are
automatically omitted from the consideration This can be also verified from Figure 714 where
the pie chart of the Shortest Distance and the Average 10 are quite similar whereas the Average
all chart introduces the importance of other soil variables as well
From a general perspective the soil properties of the clay layer seems to be determinant
whereas also sand and Dikeold material play an important role to the soil failure More precisely
according to the Average all method the unit weight and the friction angle of the clay layer
turned out to be the most influencing while also the cohesion of the Dikeold material and
friction angle of the sand layer contribute to the failure domain From a qualitative point of view
the weakening and consequently the settlement of the Dikeold material pushes the subsoil
creating an additional surcharge for the underlying clay and sand layer which act like the
foundation soil the incapability then of the clay layer to withstand the overlying load due to the
low unit weight in combination with the low friction angle of the clay as well as the low stiffness
and strength of the sand layer can lead to the creation of an inner slip surface and thus failure
Since the clay layer can be considered as a foundation soil the importance of the friction angle
and the unit weight can be also explained from the Terzaghirsquos bearing capacity theory (Terzaghi
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 102
1943) where these two factors are crucial for the determination of the ultimate bearing capacity
of the soil However these can be better verified by conducting large scale experiments in dikes
and inspecting the sensitivity of the dikersquos stability towards the soil properties
75 Anchor Failure The reliability of the support system is estimated based on the anchor and the waling probability
of failure However the load on the waling is proportional to the anchor force as it appeared in Eq
57 Therefore after the estimation of the anchor reliability by calculating the probability of
exceedance of a certain admissible anchor force the waling can be designed deterministically
Besides the failure probability of the waling must be lower than the anchorrsquos (given that no
uncertainties of the strength and the structural properties are considered) For the anchor failure
the LSF was formulated as follows
(72)
where [kN] is the anchor force is the yield stress and the cross
section area The reliability method to be utilized in this analysis is FORM and the selection of the
random variables to be used was made according to the sensitivity analysis results and the
engineering judgment and they are listed in Table 72 In particular the variables from the second
step of the sensitivity analysis were considered which will also help at a later stage to compare
FORM sensitivity method and FAST In Table 74 the results of the reliability analysis with FORM
are listed Additionally the design point together with the importance factors of each variable is
presented It should be mentioned at that point that the sensitivity analysis helped to define a
starting point closer to the design point and thus reduce the computational time required for the
analysis More precisely the analysis lasted approximately 16 hours whereas by starting from
mean values from which by default FORM starts the iterations it would take almost a day
Table 74 Reliability results for the anchor failure with FORM
FORM Pf 28410-6 β 45 Number of LSF calls 119 Maximum number of iterations
100
Elapsed time (hr) 16
Design point Parameter value α-value Clay_phi [deg] 1814 -0059 Clay_nu [-] 033 0086 Clay_c [kPa] 1487 -0089 Clay_G [kPa] 41739 0911 Clay_Rinter [-] 051 -0021 Sand_Rinter [-] 069 -0051 Sand_nu [-] 030 0007 Sand_phi [deg] 3259 -0174 Sand_G [kPa] 1130770 0048 Dikeold_phi [deg] 2842 -0020 Dikeold_nu [-] 036 -0045
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 103
Dikeold_c [kPa] 1873 -0045 Dikeold_G [kPa] 243183 -0326 Dikeold_Rinter [-] 053 -0072
Figure 715 Importance factors α2 in anchor failure from FORM analysis
The influence coefficients in Figure 715 indicate that this limit state is governed by the shear
stiffness of the clay layer G Furthermore the shear stiffness of the Dikeold material seems to
contribute significantly to the anchor yield stress exceedance It can be concluded that the
problem is still in the elastic domain as far as the soil behaviour is concerned In case of
predominately plastic behaviour the strength properties of the soil (φ and c) become more
important In Figure 716 a demonstration of the design point as it was acquired by the FORM
analysis is presented The figures indicate that presumably due to the clay layer withdrawal in the
passive side as it is depicted in Figure 716(b) with the red shadings sheet pile moves to the right
and subsequently anchor is tensioned enough to reach its yield stress25
(a) Deformed mesh (b) Total displacements |u| in m (shadings)
(c) Total displacements direction (arrows)
Figure 716 Design point illustration from the anchor failure
25 A better representation of the soil-anchor interaction could be possibly obtained if the anchor was modelled as a node-to-node element and the anchorage body as an embedded element
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 104
The calculations were repeated with the 6 most influencing parameters both with FORM and DS
in order to investigate the effect of reducing the number of random variables and in order to
validate the results of FORM analysis In Table 75 the reliability results of respectively FORM and
DS are presented while in Figure 717 the importance factors of both methods are juxtaposed The
design point and the importance factors of DS were estimated based on the Average 10 method
(see section 71)
Table 75 Reliability results of FORM and DS with reduced variables in anchor failure
FORM Pf 8510-6 DS Pf 1510-5 β 43 β 42 Number of LSF calls
28 Number of LSF calls
776
Maximum number of iterations
100 Number of iterations
100
Elapsed time (hr) 034 (starting point close to the design point)
Elapsed time (hr)
115
Design point Design point Parameter value α-value value α-value Clay_nu [-] 033 0119 033 0242 Clay_c [kPa] 1404 -0026 1535 -0212 Clay_G [kPa] 43236 0928 44548 0807 Sand_phi [deg] 3163 -0110 2954 0203 Dikeold_G [kPa] 239465 -0330 251723 -0366 Dikeold_Rinter [-] 053 -0064 060 -0263
Figure 717 Importance factors α2 with FORM and DS analysis respectively
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 105
As it can be noticed comparing FORM results from Table 74 and 75 the reliability index changes
from β=45 with 14 parameters to β=43 with the 6 most influencing parameters This verifies the
importance of the most influencing parameters that it was found in the analysis with the 14
parameters The small discrepancy though between the two reliability indexes might be
attributed to the interaction effect of the soil variables on the output performance
The estimation of failure probability with DS is almost the same with FORM As far as the
importance factors are concerned DS also came up with the conclusion that the shear stiffness of
the clay and the Dikeold material are the most crucial soil properties for the anchor stress level
However according to DS the stiffness of clay ν and the interface strength Rinter between the
Dikeold material and the sheet pile wall seem to also contribute to the limit state of the anchor
The reason why the strength of the clay layer under the dike and the Dikeold material (the part of
the dike that comes in contact with the anchor and the sheet pile wall) are the most important
variables can be explained by the principle of soil arching According to this principle the soil
columns on both sides of the rigid sheet pile wall are more compressive than the soil columns on
the top of the wall (ie overlying dike body) because of the higher stiffness of the wall when
compared with soils As such soil columns on both sides tend to settle more than the soils on top
of the rigid wall and this differential settlement causes a downward shear force acting along the
sides of soil columns on top of the wall As such the vertical load on the wall becomes larger than
the sole weight of soil columns on its top and the anchor that actually receives the most of this
vertical load reaches eventually its yielding stress
Last but not least it should be mentioned that the response surface which was constructed based
on the SA worked efficiently in most of the Plaxis errors cases in a sense that it gave reasonable
results of the anchor stress level
76 Sheet pile wall failure For the reliability analysis of the sheet pile wall the strength property ie the yield stress was
considered constant and equal to 355 kPa Therefore the LSF that was applied for this case was
formulated as follows
[
] (73)
where [kNm] and [kN] are the maximum bending moment and the axial force
that are exerted on the wall respectively is the elastic section modulus ( =1245 cm3m)
and is the cross sectional area ( =1201 cm2m for a AZ12 profile)
In sections 643 and 72 the maximum bending moments of the sheet pile wall were calculated
with the design values of the soil properties and the mean values respectively It has been
observed that the bending moments were not so significant and determinant in comparison with
the anchor force and the safety factor of the system that were also considered (in section 643) in
order to determine the structural properties
In a later stage a SA was carried out for the sheet pile wall that has been further discussed in
section 73 Figure 718 shows the stress level of the wall that was developed after 3150 iterations
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 106
by considering as random variables the ones that are defined in Table 72 As someone can
observe the stress level is generally quite low and even an order lower than the yield stress (ie
355105 Pa)
That small magnitude of bending moments can be probably attributed to the counteraction of the
active earth pressures from the passive ones that due to the homogeneity of the soil in both sides
of the wall are quite similar
Figure 718 Stress level [in Pa] of the sheet pile wall after 3150 iterations of the SA
The reliability analysis has been carried out with FORM method and considering as random
variables those listed in Table 72 However due to the low level of stresses that are developed
and the simultaneously high yield stress the method converged after almost 35 days of
calculations As it can be noticed in Table 76 the analysis came up with an extremely low
probability of failure It should be mentioned at that point that the convergence of the analysis
was successful thanks to the RS that was fitted This means that a Plaxis error of soil collapse
preceded and it activated the RS which implies that the soil failure is more likely to occur before
the sheet pile failure
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 107
Table 76 Reliability results for the sheet pile wall failure with FORM
FORM Pf 7210-38 β 13 Number of LSF calls 2619 Maximum number of iterations
100
Elapsed time (hr) 84 (35 days)
Design point Parameter value α-value Clay_phi [deg] 1164 0261 Clay_c [kPa] 586 0336 Clay_G [kPa] 14275 0663 Clay_Rinter [-] 052 -0016 Clay_gammaunsat [kNm3]
1848 0000
Clay_gammasat [kNm3] 1948 0000 Sand_phi [deg] 3020 0000 Sand_G [kPa] 1193650 0000 Dikeold_phi [deg] 1360 0406 Dikeold_c [kPa] 592 0438 Dikeold_G [kPa] 168798 0000 Dikeold_Rinter [-] 053 -0024 Dikeold_gammaunsat [kNm3]
2275 -0150
Dikeold_gammasat [kNm3]
2375 0000
This can be also verified from the importance factors that are depicted in Figure 719 where apart
from the shear stiffness of the clay layer which seems to be the predominant soil property for the
sheet pile LSF the strength parameters (φ and c) of the clay and the Dikeold materials also play
an important role This implies that the soil behaviour has already passed into its plastic domain
where the governing properties are the strength parameters before the sheet pile exceeds its
yield stress
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 108
Figure 719 Importance factors α2 of the sheet pile stress level in the plastic domain of soil
Nevertheless in order to investigate the potential most influencing soil properties for the sheet
pile LSF in the elastic domain as well the available output data of the FORM analysis have been
assessed The highest observed value of the sheet pile stress before its failure was around
235000 kPa After tracing the results it seems that when FORM tries to reduce further the
stiffness and strength properties of the soil layers soil collapses before the sheet pile failure The
vectors of the random variable that led to a stress level on the sheet pile wall in the range of
230000-235000 kPa were chosen in order to evaluate the importance factors The method that
was followed in that case is the Shortest Distance and the results are shown in Figure 720
Figure 720 Importance factors α2 of the sheet pile stress level in the elastic domain of soil
As it can be noticed from Figure 720 the results in the elastic domain regarding the influence of
the stochastic parameters are comparable with anchorrsquos analysis The stiffness parameter of the
clay layer (69) is the most dominant variable for the system behaviour whereas the stiffness of
the sand (15) and the Dikeold (4) layers seems also to have a significant influence This is not
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 109
surprising as the nature of the loads that generate high stresses on the anchor as well as the sheet
pile is the same namely the horizontal earth and water forces
Concluding in this case study the sheet pile wall is not expected to show a decisive behaviour to
the systemrsquos reliability according to the reliability analysis and the deterministic analysis (see
section 643) It should be stressed the fact that the fitted RS in a lot of calculations seemed not to
perform suitably for the sheet pile and this can be attributed to the fact that the it was created
based on the SA which was far from the plastic domain In that case a better fit of a RS is
recommended However such an action is still under investigation that for future considerations
in the design concept will be contemplated as a valuable technique for complex structures
modelled in FEM whose reliability analysis with conventional methods is time consuming and the
coupling may lead to numerical problems In the next section the analysis of the systemrsquos
reliability is presented
77 System Failure The estimation of the systemrsquos reliability was based on the systemrsquos fault tree of Figure 53
according to which the general LSF was formulated as the minimum of the three LSF that were
presented above for the soil the sheet pile and the anchor and it is given below
[ ]
[(
)]
(74)
According to this LSF the failure that occurs first is counted as a system failure The analysis was
carried out with DS reliability method and the variables that were taken as stochastic are listed in
Table 72 As it was discussed in section 51 the systemrsquos probability of failure differs from the
probability of flooding for which a combination between the systems reliability and the water
level uncertainty shall be made
In Table 77 the reliability analysis results are presented as they were deduced from DS Similarly
to the soil failure in Figure 721 the importance factors according to the Shortest Distance the 10
Average and the Average all methods are illustrated
Table 77 Reliability analysis results from the system failure
DS Pf 2410-3 β 30 Number of LSF calls 1807 Number of iterations (or directions)
300
Elapsed time (hr) 39
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 110
Figure 721 Importance factors α2 of the system failure
The analysis took almost one and a half days and it came with indicative valuable though results
about the system behaviour According to the averaged importance factors in Figure 721
(deduced by the Average all method) we can notice features from both soil and anchor failure
For example the unit weight (14) the friction angle (7) and the cohesion (8) of the clay
layer as well as the friction angle (8) and the stiffness (7) of the sand layer that played an
essential role in the soil failure analysis Moreover the stiffness of the clay (11) and the Dikeold
(9) layers that appeared to be important for the anchor and sheet pile LSF They seem to be also
essential in the system analysis
However the other two methods (Shortest Distance and Average 10) came up with the cohesion
and the interface strength of the Dikeold material to be the most influencing soil properties for
the system For that reason an investigation was conducted regarding the output results of the
system analysis in order to obtain an insight into the different reasons of the system failure More
precisely the failure points were divided into those that come from Plaxis Error and those that
come from the yield stress exceedance of the structural elements It was then observed that many
failure points originate from Plaxis errors that are related to ldquoNot enough load stepsrdquo (Error codes
102 and 112) rather than ldquoSoil seems to collapserdquo (Error code 101)
Therefore in order to estimate the actual important soil properties the vectors of the random
variables that led to errors of ldquoSoil seems to collapserdquo and to anchorsheet pile (SP) failure were
studied separately The importance factors are depicted in Figure 722 which actually verifies the
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 111
importance of the variables that were also deduced from the Averaged all method in Figure 721
(which are basically the unit weight and the stiffness of the clay layer and the friction angle of the
sand layer) It should be mentioned that the failure of the structural elements contributes 65
whereas the soil failure contributes 35 on the total probability of failure The latter would be
possibly lower than 2410-3 that was initially found as several failure points are not attributed to
real failure but to Plaxis numerical errors
Figure 722 Influence factors α2 for the structural elements (left) and the soil LSF (right) respectively
The system analysis can lead to valuable conclusions for the system behaviour as a whole and
individually for the different elements given though that an inspection of the results is followed
Moreover such an analysis is taking into account correlations among the three sub-failure
mechanisms implicitly For example the different LSFs may be affected by the same soil variable
such as the shear stiffness of the clay layer for the anchor and the sheet pile LSF In that case the
most dominant failure mechanism will occur first and the rest are excluded26 However during
the system analysis someone is not aware of what failed or not and therefore heshe is not
capable of distinguishing the weaknesses of the system that need further improvement Besides
with this technique is not also easy to inspect the failure mechanisms that occur in order to get
some knowledge about the systemrsquos behaviour Therefore the reliability analysis of the
components is recommended to be carried out separately if time is available
78 Comparison between Global Sensitivity Analysis and Local Sensitivity
Analysis (FORM)
The aim of this section is a comparison between the importance factors that were deduced in
section 75 according to a FORM analysis and the results of the global SA that were presented in
section 73 for the anchor failure
The output of the global SA is actually the total indices that show the influence of the random
variables The importance factors of FORM and the total indices of global SA cannot be directly
26 This method is not capable though of accounting for the simultaneous occurrence of different failure mechanisms as the system was considered a serial system
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 112
compared as they are different values More precisely the total index of a random variable is the
ratio between the total variance and the portion of the total variance that stems from the
uncertainty of the specific variable and its interactions with the rest (see Eq A10 in Appendix A)
On the other hand importance factors of FORM express the ratio between the normalized value of
the variable in the design point and the reliability index (see Eq 72) Therefore the comparison
between these values is carried out qualitatively in terms of the variables that are distinguished
as the most influencing
In particular according to global SA the shear stiffness G of the clay layer and the cohesion c of
the Dikeold material seem to be the most determinant soil properties Moreover the strength
reduction parameters Rinter of the Dikeold and the clay layers and the friction angle φ of the
Dikeold and the sand layer appear to have a relatively notable contribution to the anchors limit
state
On the other hand FORM analysis has concluded to the same results as the global SA with the
difference that Dikeold material affects the anchorrsquos stress level more with its shear stiffness
rather than its cohesion In Table 78 the results of both FORM and global SA are presented
separately for each soil layer and property (in blue for the global SA and in black for the FORM
analysis)
Table 78 Important soil properties according to global SA and FORM analysis for the anchor failure (radicradic=very important radic=less important but still considerable)
FORM global SA Sand Clay Dikeold
ν
radicradic
ϕ radicradic
radic
c radic radic radicradic
G radicradicradicradic radicradic
Rinter
radic radicradic
The validation of the most important parameters was carried out with the repetition of the
analysis with DS as it was showed in section 75 The small difference of the probability of failure
as well as the same tendency of the important factors of both DS and FORM verified the reliability
of FORM analysis Therefore the similarity of the results deduced according to global SA with the
ones given by FORM indicates the fact that in this case study the global SA should be considered
as a trustable way of SA for obtaining a first overview of the dominant variables that affect the
limit state under investigation Furthermore it is capable of giving a general insight into the
outputrsquos (in this case anchor stress level) behavior and the soil layers that can potentially stand as
predominant during the reliability analysis For example in order to obtain a first sense of the
reliability and the magnitude of the expected probability of failure someone can inspect the
output samples of the LSF deduced from the global SA In Figure 723 the samples of the anchor
stress level are presented as they were deduced after 3150 iterations during the global SA
(computation time =75 hr = 3 days) The fact that after 3150 iterations and variables
combinations the anchor did not reach its ultimate stress level (ie 450 kPa) implies that the
probability of failure might be small This can also help for qualitative decision making concerning
a systemrsquos behavior without conducting a reliability analysis Moreover global SA proved the high
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 113
interaction (large difference between first and total order indices) among the parameters that can
affect the reliability analysis and can help to decide the amount of the random variables
Figure 723 Anchor stress level during global SA (step 1)
However as someone can observe from Table 78 there are also some discrepancies concerning
the dominance of some variables This can be attributed to the fact that the influence of all
parameters in specific points such as the design point cannot be identified by the global SA
Moreover it is possible that the samples used in the SA could not reach the field where the
system has a plastic behavior and where presumably other variables may also affect the limit
state
As general conclusions for global SA some possibilities and limitations are listed below
Possibilities
bull Indicates the level of interaction effect on the output variance
bull Indicate the amount of the random variables that should be considered in the reliability
analysis
bull Gives a general insight into the output behavior
bull Valuable for future use in the construction of response surfaces
Limitations
bull Cannot rely on global SA for the evaluation of the most important parameters near the
design point
bull Difficult to reach plastic zone (large amount of samples are probably needed)
79 Discussion
In this chapter the reliability analysis results were presented for the soil the anchor the sheet
pile and eventually the system failure taking soil properties of the several soil layers as stochastic
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 114
Initially a SA was carried out for each system component in order to identify the major soil
properties and reduce the number of the stochastic parameters The reliability analysis was
performed either with FORM or DS and in both cases the design point and the importance factors
were estimated Eventually the list of subquestions that is proposed in section 11 was answered
The way they were treated and the corresponding conclusions are elaborated below
Which reliability methods are computationally tractable in combination with FEM
Before starting analyzing the reliability of the system it was essential to evaluate and
assess the performance of the available reliability methods in order to be aware of their
performance in terms of computational time and flexibility in different kind of LSFs After
a literature review and the implementation of some of them in simple LSF problems a
general overview has been obtained for the most applied reliability methods concerning
the previous knowledge required for the system behavior in order to provide the right
input parameters its accuracy and the computational effort Eventually FORM and DS
were chosen to be employed for this case study considering the no preliminary knowledge
of the LSF is required their reliable accuracy and the limited computational effort that
they have in comparison with other methods
The anchorrsquos and sheet pilersquos reliability has been evaluated with FORM method whereas
soil and systemrsquos reliability have been assessed by DS Carrying out a FORM analysis and
steering the algorithmrsquos parameters in order to converge proved to be more difficult than
it was expected However an efficient performance was finally achieved and the failure
probability of the anchor failure was successfully estimated together with the importance
factors (or else the α values) The analysis was repeated with the most important variables
and they were validated with DS The FORM analysis took approximately one and half
hour to converge (with 14 random variables) after providing a starting point that it was
estimated to be close to the design point whereas for the validation it took almost half an
hour On the other hand in case of the sheet pile wall FORM analysis took almost 4 days
The computational time also depends on the LSF to be approximated the number of
variable and the complexity of the problem that each time is analyzed As far as the soil
analysis is concerned DS took approximately one and a half day to converge for 300
iterations and 14 random variables
Which limit states are relevant for the envisaged application of retaining walls in dikes and
how can they be formulated using FEM analysis outcomes
The relevant LSFs of a dike with an anchored sheet pile wall were investigated For each
element ie the sheet pile the anchor and the soil body a separate LSF has been set up
The focus of this research was on the ULS that in case of the structural elements it was
represented by the ultimate yield stress of the anchor and the sheet pile respectively as
they were expressed in Eq 72 and 73 The determination of the soil failure has been also
elaborated and the available alternatives that can be considered in that case were
elaborated After a short evaluation the Plaxis definition of soil failure was selected that it
was formulated as it is shown in Eq 71 Finally the systemrsquos LSF was formed as a
combination of the aforementioned limit states More precisely the system was
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 115
considered as a serial and thus system failure occurs when one of the elements fails first
(see Eq 74)
How robust (convergence) are the tractable methods
The robustness of the different reliability methods can be controlled by the convergence
criteria that are defined by the user and depend on the required accuracy In case of FORM
the convergence criteria consist of a set of errors that has mostly to do with the
approximation of the reliability index and the limit state threshold whereas in DS the
convergence is controlled through the coefficient of variation (CoV)
In order to reassure the robustness of FORM the method has been validated with DS The
validation showed that the FORM method complies with the result of DS with a small
difference though in the design point and the importance factors FORM analysis seemed
to converge sufficiently and faster than DS However the choice of the optimization
algorithm plays a key role to the methodrsquos efficiency In this research the convergence
errors were manipulated for increasing the efficiency of the method More precisely it has
been observed that by relaxing the relative error and the constraint error (the explanation
of these errors and an evaluation of the different optimization algorithms are available in
Appendix A2) FORM convergence can be accelerated
The reliability analysis of the soil body and the system was performed with DS Similarly
to FORM analysis in order to improve the performance of the searching algorithm some
of the methodrsquos parameters should be stipulated for improving its efficiency and
reliability Therefore the maximum step size of the algorithms was tripled and it was
proved that for the same time period double iterations were feasible to be carried out
This does not mean that the larger the step size the better it is The choice of the step size
must be made wisely depending on the reliability index that it is expected and
respectively it is allowed to change for achieving a better performance In this case study
it was proved that after 300 iterations the CoV of DS tends to be high and it can therefore
still be improved by carrying out more iterations In Figure 724 an example of a DS
convergence graph concerning the system analysis is shown The middle red line shows
the probability estimation whereas the green lines show the boundaries At that example
the CoV was 040
DS seems to perform successfully both in soil failure with one LSF and in system failure
where 3 different LSF were included In such cases performing an approximation method
such as FORM would not be wise as the output from each iteration does not provide
information for going to the next that a FORM analysis normally needs In contrast it just
gives an indication of failure or not This means that only a random sampling method can
handle it and converge successfully as it was proved
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 116
Figure 724 Directional Sampling convergence graph (at confidence level 095) of the system LSF
What is the contribution of different uncertainties in the failure mechanisms of the system
According to the global SA and the reliability analysis of FORM and DS the friction angle
the cohesion and the shear stiffness properties appeared to have an influence on the
different LSFs This is not surprising as they all have large CoV of 10 20 and 25
respectively Thus their uncertainty generally showed an impact on the systemrsquos
behaviour It is also essential to mention that the global SA has shown that the uncertainty
of the interface strength Rinter (CoV=20) has also a notable contribution to the structural
elements LSFs However the FORM analysis which is more representative close to the
design point came up with a less important impact of that variable
The reliability analysis showed that as far as the structural elements (anchor and sheet
pile) are concerned the shear stiffness of the clay and the Dikeold material is determinant
and especially for the sheet pile the strength parameters of the soft soils seem also to play
a crucial role In contrast with the anchor where the soil body seems still to behave in an
elastic manner during the sheet pile failure soil appears to reach plasticity in more
locations where the strength properties (ϕ and c) are also influential The most dominant
failure mode appeared to be the anchor failure whereas the soil failure is always preceded
the sheet pile failure The dominance of the anchor failure can be explained from the fact
that the most influent parameter in this LSF is the stiffness that is also the variable with
the highest uncertainty (CoV=25) Therefore the plastic limit of the anchor is more
likely to be reached
Regarding the soil body reliability the analysis detected 3 different failure directions in
which the clay and the sand layer below the dike seem to be the most influent More
specifically the friction angle φ the cohesion c the unit weight γ of the clay layer the
friction angle of the sand layer and the cohesion of the Dikeold material turned out to play
an important role The decrease of the clay layerrsquos unit weight lead to the incapability of
the layer to withstand the overlying load of the dike body and at the same time the
weakening of the sand layer contribute to the creation of an inner sliding surface Even if
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 117
the uncertainty of the unit weight is small (CoV=5) it is still important for the stability
of the system as the maximum deformations appear in that layer (see Figure 712) Last
but not least the illustration of the several failure points in Plaxis allows someone to
inspect the failure mode and realize if it is a real failure of the structure or if it is a
weakness of the modelling
Finally during the system analysis multiple failure directions were detected whereas the
pie charts of the importance factors include soil properties that affect both soil and
structural elements as it was expected However due to Plaxis numerical errors several
failure points were found to be not realistic and they should be excluded in order to come
up with a more reliable failure probability which will be presumably lower The most
suitable method for estimating the importance factors though seems to be the Average all
method which is not distorted from the Plaxis numerical errors However it is
recommended that a further research be done on how realistic are the failure points
coming from Plaxis warnings and to verify if they comply with Figure 711 failure
definition
In Table 79 a synopsis of the most important soil properties for the different LSFs is
presented It is also apparent the common contribution that some of them have in
different LSFs In the last two columns of the table the results of the influencing soil
variables for the system LSF are shown with the Average 10 and Average all method
respectively For the reasons that were explained in the previous paragraph the Average
all method is considered to be the most suitable in that case and as it can be noticed from
the table it gives the most representative picture of the governing soil variables of each
LSF
Table 79 Important soil properties for the different LSFs (radicradic=very important radic=less important but still considerable the most important parameters for each LSF is painted in orange)
Soil material
Soil property
Anchor Sheet pile
Soil body
System Average
10
System Average
all
Clay
G radicradic radicradic
radic radicradic
φ
radic radicradic
c
radic radic radic
Rinter γ
radicradic
radicradic
Sand
G
radic radic
φ
radicradic radicradic
c
Rinter γ
Dikeold
G radicradic
radic
φ
radic
c
radicradic radic radicradic radic
Rinter radic
radicradic radic
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 118
γ
radic radic
Rinter was taken from the validation of FORM with DS Figure 717 (right graph)
taken from the sensitivity of the sheet pile in the soil elastic domain Figure 720
Can response surface techniques help to increase the efficiency and robustness of the
reliability model
In this research RS techniques were used only as a source for providing a reasonable
value to the FORM analysis in order to continue running in case of Plaxis errors These
response surfaces were constructed based on the SA results for both the anchor and the
sheet pile wall Apparently the domain of the SA is limited and most of the times it is
incapable of approaching the design point However RS worked efficiently in many cases
where Plaxis failed to give an output result such as in case of the sheet pile LSF Moreover
the RS could be used individually for the estimation of the failure probability (instead of
Plaxis calculations) but a better fit is advised to be firstly achieved that would be
representative in a wider plastic domain of the structural elements Therefore the RS
technique can be used to enhance the efficiency of the reliability analysis of a complex
system in the sense that it can be considered as solution for FEM errors What is more RS
techniques are a quite promising method for the reliability analysis of a system as it can
also be used alone without coupling with FEM and eliminate the computational effort
However this requires more investigation of such a technique
How can the current design approach be improved
In section 22 the current design concept of the case study that is under investigation in
this research is described At that section the design values of the moments and the forces
on the structural elements are presented as well as the required total safety factor FEM
=18 is given as it was initially calculated by Breedeveld (2011) It has been observed that
a significant overestimation is made in terms of the design values that might later lead to
heavy and cost inefficient structures For that purpose alterations was made on the
original case study in terms of the structural elements and some soil properties and a new
deterministic design was made in order to come up with a less heavy structure This
would also render it possible to investigate the reliability of the system rather than taking
zero probabilities or non-convergence Moreover it would finally help to see if there are
any potentials for optimizing the design by applying a full probabilistic analysis
According to the soil reliability analysis the probability of failure was found 1310-8 and
the reliability index β=55 As far as the anchor failure is concerned the probability of
failure found to be 28410-6 with a reliability index β=45 according to the FORM analysis
The sheet pile failure seem not be significant as it will more likely occur after the soil
failure Eventually the system analysis was performed the probability of failure was
found to be 2410-3 and the reliability index β=30 which is high enough to consider the
system safe
The systemrsquos reliability index that came out of this analysis cannot be directly compared
with the required one (ie β =40) because they stem from different probabilities The
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 119
former refers to the probability of failure due to the global instability given a design water
level whereas the latter refers to the probability of flooding that incorporates also the
water level uncertainty It must be mentioned that in reality there is no sheet pile wall
inside the dike section in the dike ring 43 but this section is under investigation for
reinforcement purposes as it does not comply with the safety standards after a recent
inspection that has been carried out Moreover the original case study taken from
Breedeveld (2011) was subjected to multiple alterations for the needs of the current
research Therefore a comparison of the reliability index with the required one from the
regulations would not be meaningful to make However what it can be compared in that
case is the order of magnitude of the reliability index After the analysis a β of 30 was
estimated for the probability of failure whereas for the overall probability of flooding a β
of 40 is required whose magnitude is the comparable with the former This can firstly be
considered a sort of the probability ldquovalidationrdquo in the sense that the analysis gave
realistic results and secondly it can imply that the structure that has been investigated in
this research is likely to be safe towards macro-instability as its coupling with the water
level uncertainty would certainly give a reliability index above 40 Furthermore a
reliability index of 30 indicates that there is room for optimizing the design concept of
such a system if someone considers also the fact that the soil strength of the subsoil was
further reduced for the needs of the current study
Due to the limitations of the modeling and the alterations of the original case study these
reliability indices should not be taken into account as an indicative number for the safety
of this dike section but rather as an order of magnitude in order to realize what the
potentials for improving the design procedure are This research can also contribute to
the way of designing not only the system but also each structural element as the reliability
of each can be calculated in advance According to the methodology developed in this
thesis a comparison can later on be made for a real case situation between the results of a
full reliability analysis and the current design procedure In that sense a tangible
improvement can be suggested for the design concepts and even to introduce a new way
for the future design of dikes with retaining walls
It should be stressed that the conclusions are based on the results of this case study with its
geometry and set of material parameters and the variation coefficients Changes in the geometry
the material parameters or the statistical properties could lead to different results The
generalization of these conclusions should thus be treated carefully
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 120
MSc Thesis A Rippi 121
8 Conclusions and Recommendations
81 Conclusions The advances in structural reliability analysis have made a large variety of methods available The
applicability and efficiency of these methods depends on the problem that is analyzed and on the
models that are used One the other hand the necessity for optimizing the design procedure of
several structures is becoming more and more intense A solution to this demand can be the
reliability analysis of a system for which rare information about its application on real-world
problems is available This work can contribute to making probabilistic analysis more accessible
for dikes reinforced with a sheet pile wall and at the same time to analyzing such a complex
system in terms of the factors that mainly influence the system behavior In particular the main
research question of this research was
How can the probability of failure due to global instability of a dike with a sheet pile wall modeled
by a Finite Element Model be analyzed
For that purpose in this research a reliability analysis was conducted by making use of a soil-
structure model that was simulated in FEM The output of the FEM calculations was utilized as an
input in the probabilistic model that was used in this thesis in order to assess the reliability of a
dike with an anchored sheet pile wall In that case the load was basically the earth and water
pressure whose impact was investigated on the anchor (LSF 1) the sheet pile (LSF 2) and the soil
body (LSF 3) respectively as well as on the system as a whole
It was demonstrated that the calculation effort of this kind of analysis does not necessarily have to
be extremely high Especially when previous knowledge about the system behavior and the
reliability models used is available This can be achieved by carrying out a sensitivity analysis In
this thesis a global SA was conducted that was proved capable of giving a representative
estimation of the most influencing variables and give a general overview of the systemrsquos
performance However close to the design point local SA such as FORM becomes more suitable
As far as the reliability of the structural elements is concerned their limit state was evaluated in
terms of the exceedance of their yielding stress The reliability assessment was carried out with
FORM from which the probability of failure and the influence factors were concluded More
precisely the anchor failure mode seemed to be the most determinant one with the shear
stiffness of the soft soils to be the most important soil properties The sheet pile failure is quite
rare and it is unlikely to occur before the soil failure since the soil body has already entered its
plastic domain
The soil and the system failure were challenging to assess as there are multiple criteria that
someone can choose In this thesis regarding the soil failure the Plaxis definition of soil collapse
was used and it was evaluated with DS In that case the unit weight and the strength parameters
of the soft soil layers seemed to be influencing The detected failure directions were investigated
Conclusions amp Recommendations
MSc Thesis A Rippi 122
in terms of their validity More precisely the failure points were illustrated in Plaxis in order to
visualize the failure and see if it is real failure or not (real if dike crest settles) That procedure
revealed the possibilities and limitations of FEM modeling The advantage is that a potential
failure can be detected and its visualization is possible On the other hand one should be aware of
the limitations related with the constitutive model that is used In this study the Mohr Coulomb
method was applied that although it can efficiently identify possible failure mechanisms the
representation of the deformations is not so realistic
The failure of the system was considered as a serial system of LSF 1 2 and 3 and it was also
assessed with DS The importance factors stemmed from the reliability analysis indicated soil
features that were important in all the LSFs However it should be stressed that the frequency of
the Plaxis numerical errors was such that quite a few failure points seem to be unrealistic In this
thesis the identification of those errors and their partially exclusion from the analysis was
achieved The latter was attained by the construction of the so-called Response Surfaces (RS)
based on a preliminary sensitivity analysis Such a technique found to be valuable in order to
overcome such a problem however a study for fitting more suitable RS in each problem is
recommended Nevertheless the further elimination of those numerical errors is strongly
recommended to be investigated in future research given the results from the current study
Retrieving the importance factors from a DS analysis was also a challenging part In this study
additional methods were developed in order to evaluate the importance factors and define the
design point after a DS analysis It revealed that alternative possible approaches of the
importance factors (Average 10 and Average all methods) can presumably provide a better
understanding of the systemrsquos behaviour Such methods seemed to be valuable for excluding the
ldquonoiserdquo of the possible fake failure points that originated from FEM numerical errors However
due to the inherent randomness of DS in terms of the selection of the directions the design point
shall be inspected and compared with the other failure points that were found Then someone can
contemplate if it is necessary to proceed with an averaging of the importance factors in order to
get a clearer view the variables influence Of course such a procedure takes time that is not always
available but according to the requirements of each research the most preferable assumptions
should be made
Last but not least the reliability analysis came up with relatively low probabilities of failure which
implies that there is still room for optimizing the design concept of a dike with retaining wall
More precisely the original case study of Breedeveld (2011) has been altered by reducing the
strength properties of the subsoil and installing a less heavy sheet pile wall after following a less
strict deterministic design It was then observed that even then the reliability analysis came up
with a respectively high reliability index It should be stressed that the failure probability of the
system should be also coupled with the uncertainty of the water level in order to obtain an overall
picture of the total probability of failure Nevertheless this research proves the potentials to
optimize the design of such a system which presumably will lead to a less expensive construction
However additional failure mechanisms such as piping internal erosion etc should be also
investigated and the design properties of the system (such as the length and cross section area of
the sheet pile wall) shall be based on them
Conclusions amp Recommendations
MSc Thesis A Rippi 123
82 Recommendations The following topics are recommended for further research
Additional research is required on FEM modeling both for the structural elements and the
soil body More in particular in this research it has been proved that Mohr Coulomb
model cannot realistically estimate the deformations in a soft soil In that case there are
other more advanced models such Hardening Soil model and Soft Soil model that are
capable of simulating the soil displacements and the stress-strain relationship closer to
the actual ones However these methods require more parameters to be specified and
their coupling with reliability packages might not be so robust and functional
As far as the structural elements modeling is concerned there are two major topics that
future research should focus on Firstly the anchor rod can be simulated in Plaxis as a
node-to-node element and at the bottom an embedded beam can be installed in order to
represent the anchorage body This would provide a better reaction with the surrounding
soil body and it could probably simulate the anchor-soil interaction more realistically
Another alternative would be to simulate it as a plate element in order to be able to test
the moments that are also developed on the anchor separately Secondly it is
recommended that the elements be investigated close to their plastic area and a reliability
analysis to be carried out with plastic elements instead In this research an attempt was
made for such an analysis but it appeared not to be functional in cooperation with FORM
analysis
Another issue that is recommended to be investigated is the inspection and the validation
of the failure that Plaxis is giving During the soil failure analysis different types of failure
were observed from which some might not be realistic but due to the modeling limitations
Therefore an attempt should be made in order to further eliminate the effect of Plaxis
numerical errors on the reliability analysis that can lead to fake failure modes A check of
the failure points can be also conducted by using a more suitable constitutive model for
the deformations rather than Mohr Coulomb However this is not always feasible for each
individual case Therefore the illustration of the failure mechanisms in large scale
experiments can give a better insight into the systemrsquos behavior and valuable knowledge
can be acquired of the expected failure modes In this way also 3D effects can be
investigated see what is their effect to the dike instability and at what extend they shall be
considered
Apart from the soil body the stresses and the displacements of the structural elements
shall be also verified It is thus recommended that field tests and measurements take place
in embedded elements like sheet piles and anchor in order to validate Plaxis results and
learn what is the most suitable manner for the their simulation (fixed-end-anchor node-
to-node element plate etc)
Concerning the reliability analysis of the specific case study a research on the potential
relevant LSFs is recommended for the different elements Specifically in this thesis the
LSF were chosen in order to represent the ULS of the element There is the possibility that
the SLS is used and certain acceptable deformations can be settled as a threshold for the
Conclusions amp Recommendations
MSc Thesis A Rippi 124
limit states However the choice of this displacement limit is not straightforward and a
qualitative research shall also be conducted in order to identify the impacts of different
values Besides it is not only the value that should be selected but also the location on the
structure that is going to be checked
It is also advised that the potential correlations among the soil parameters to be included
In Wolters (2012) and Bach (2014) correlation matrices are introduced for specific soil
properties that can be included in a future research and reveal what is their effect on the
reliability analysis The influence of correlations depends on what kind of variables are
correlated and what is the influence of those variables on the system behavior However a
preliminary analysis with independent variables is advisable to be carried out at the
beginning so as for someone to be able to analyze the effect of the variables individually
and then evaluate the correlation effect more wisely
It would be also advisable a comparison to be made between the design of a real case
according to a full probabilistic analysis and partial safety factors In that way a direct
insight into the advantages and disadvantages of each method can be gained and
fundamental improvements of the design concept can be made for dikes with retaining
walls
At this moment the most reliable way of conducting a reliability analysis is with sampling
methods like MC or DS However such methods are usually unattractive due to their large
computational time Thus the tolerance of sampling methods accuracy is lowered in order
also to reduce the number of iterations or approximating methods are used like FORM
analysis However even if approximating methods are used they are usually preferred to
be validated with sampling methods in order for their results to be acceptable A real
validation of the failure probability would mean that a structure should be experimentally
tested in different positions and under the same conditions in order to see if the
probability of failure that is calculated in each case is the same Therefore a real
validation of the failure probability cannot yet be initiated but the computational effort of
the sampling methods can be eliminated with the development of technology and the
improvement of the computer science
MSc Thesis A Rippi 125
References
Abdo T amp Rackwitz R (1990) A new beta-point algorithm for large time-invariant and time
variant reliability problems Proceedings of the 3rd IFIP WG 75 Conference Berkeley pp1-12
California USA
Ang AHS and Tang W (1984) Probability Concepts in Engineering Planning and Design Volume
IBasic Principles John Wiley and Sons New York USA
Baecher GB and Christian JT (2003) Reliability and Statistics in Geotechnical Engineering
Chichester West Sussex John Wiley amp Sons Ltd
Behaviour of Soils and Rocks GEO-ENGINEERING - CIE4361 (Vol 2015) Retrieved from
httpsblackboardtudelftnlbbcswebdavpid-2386735-dt-content-rid-
7926659_2courses32281-14150221-possibilities26limitationspdf
Breedeveld J (2011) Technisch Rapport Analyse Macrostabiliteit Dijken met de Eindige Elementen
Methode Bijlage Voorbeeldcase Deltares Delft the Netherlands
Brinkgreve RBJ (2015) Personal discussion TU Delft the Netherlands
Brinkgreve RBJ and Bakker HL (1991) Non-linear finite element analysis of safety factors In
G Beer JR Booker and J P Carter (Eds) Proceedings of the Seventh International Conference on
computer methods and advances in Geomechanics Rotterdam A A Balkema Brookfield
Cukier RI Fortuin M Shuler KE Petschek AG and Schaibly JH (1973) A study of the
sensitivity of coupled reaction systems to uncertainties in rate coefficients I Theory California
University USA
Cundall PA (2002) A discontinuous future for numerical modeling in soil and rock Published in
Proceedings of the third international conference Discrete Element Methods Numerical Modeling
of Discontinua Edited by Cook and Jensen Santa Fe pp 3-4
CUR (2005) Handbook Quay Walls Gouda the Netherlands
CUR-publication 190 (1997) Probabilities in civil engineering Part 1 Probabilistic design in theory
Stichting CUR Gouda
CUR 166 ndash Damwandconstructies (2005) Technical Recommendation 4e herziene uitgave 2005
The Netherlands
Deltaprogramma (2014) Synthesedocument Veiligheid Achtergronddocument B1 Ministrie van
Infrastructuur en Milieu
EN 1990 2002 Basis of Structural Design European Committee for Standardization
References
MSc Thesis A Rippi 126
Farrokh N (2007) Tools and Strategies for Dealing with Uncertainty in Geotechnics In DV
Griffiths and GA Fenton (Ed) Probabilistic methods in Geotechnical Engineering (pp 71-95) New
York NY Springer Vienna
Fenton GA and Vanmarcke EH (1990) Simulation of random fields via local average
subdivision J Eng Mech ASCE 116(8)1733ndash1749
Flood Defences HYDRAULIC ENGINEERING ndash CIE5314 (Vol 2015) Retrieved from
httpsblackboardtudelftnlbbcswebdavpid-2125084-dt-content-rid-
7621002_2courses29759-
131404CIE5314_FloodDefences_05_failure_mechanisms_stability_Schweckendiekpdf
Gemeentewerken-Rotterdam (2003) Alle triaxiaal proeven rotterdam Technical report Gemeentewerken Rotterdam Rotterdam the Netherlands
Ghanem R and Spanos P (1991) Stochastic Finite Elements ndashA Spectral Approach
Springer New York
Gonzaacutelez A Cuadrado A Josa A and Olivella S (2013) Safety assessment of limit equilibrium
methods for the design of cantilever and anchored sheet pile walls Escola de Camins Barcelona
Griffiths DV and Fenton GA (Eds) (2007) Probabilistic Methods in Geotechnical Engineering Lecture notes monographs edited works and proceedings in the field of Mechanics Engineering Computer Science and Applied Mathematics Italy Hasofer AM and Lind NC (1974) An exact and invariant first-order reliability format Journal of
Engineering Mechanics Division ASCE 100(1) 111-121
Hoek E Read J Karzulovic A and Chen ZY (2000) Rock slopes in civil and mining engineering
Published in Proceedings of the International Conference on Geotechnical and Geological
Engineering GeoEng2000 19-24 November Melbourne
Homma T and Saltelli A (1996) Importance measures in in global sensitivity analysis of
nonlinear models Reliab Eng Syst Saf 52(1) pp1-17
Hydra-Ring 12 (2015) Probabilistic toolbox for the WTI2017 Technical Reference Manual Version
12 Deltares Delft
Jaky J (1944) ldquoThe coefficient of earth pressure at rest In Hungarian (A nyugalmi nyomas tenyezoje)rdquo J Soc Hung Eng Arch (Magyar Mernok es Epitesz-Egylet Kozlonye) 355ndash358
Johansson E amp Sandeman E (2014) Modelling of Deep Excavation in Soft Clay A comparison of different calculation methods to in-situ measurements Chalmers University of Technology Gothenburg Sweden
Joint Committee on Structural Safety (1981) General principles on reliability for structural design
International Association for Bridge and Structural Engineering
References
MSc Thesis A Rippi 127
Larsen H Lubking P and Breedeveld J (2013) Ontwerp stabiliteitsschermen (type II) in primaire
waterkeringen (groene versie) Deltares Delft
Melchers R E (2002) Safety and risk in structural engineering Prog Struct Engng Mater
4 193ndash202
Moumlllman A amp Vermeer P (2008) Reliability analysis of a dike failure case study Elbe river
University of Stuttgart Germany
Nelson RB (1999) Lecture Notes in Statistics An Introduction to Copulas Springer New York
OpenTURNS 15 (2015a) Reference Guide OpenTURNSrsquo methods for Step C uncertainty
propagation pp155-161
OpenTURNS 15 (2015b) Use Cases Guide for the Textual User Interface pp190-192
Pandya N Marsdea E Floquet P and Gabas N (2008) Sensitivity analysis of a model for
atmospheric dispersion of toxic gases In B Braunschweig and X Joulia (Ed) 18th European
Symposium on Computer Aided Process Engineering-ESCAPE18 (pp1143-1148) Elsevier BV
Phoon K (Ed) (2008) Reliability-based Design in Geotechnical Engineering Computations and Applications Taylor amp Francis Britain
Plaxis 2D (2015a) Tutorial manual Delft the Netherlands
Plaxis 2D (2015b) Reference Manual The Netherlands
Plaxis (2015) Material Models Manual The Netherlands
Powell MJD (1994) A direct search optimization method that models the objective and
constraint functions by linear interpolation In S Gomez amp JP Hennart (Ed) Advances in
Optimization and Numerical Analysis (pp51-67) Dordrecht Kluwer Academic
Rankine W (1857) On the stability of loose earth Philosophical Transactions of the Royal Society
of London Vol 147
Saltelli A (2004) What is Sensitivity Analysis In A Saltelli K Chan and EM Scott (Ed)
Sensitivity Analysis (pp 3-13) John Wiley amp Sons Publication
Saltelli A S Tarantola F Campolongo and M Ratto (2004) Sensitivity Analysis in Practice
A Guide to Assessing Scientific Models John Wiley amp Sons Ltd
Schittkowski K (1985) NLPQL A Fortran subroutine solving constrained nonlinear
programming problems Annals of Operations Research 5 485-500
Schweckendiek T (2006) Structural Reliability applied to Deep Excavations Coupling Reliability
Methods with Finite Elements TU Delft Delft
References
MSc Thesis A Rippi 128
Schweckendiek T Vrouwenvelder ACWM Calle EOF Jongejan RB and Kanning W (2013)
Partial Factors for Flood Defenses in the Netherlands Modern Geotechnical Codes of Practice ndash
Development and Calibration Fenton GA et al (eds) Special Geotechnical Publication Taylor amp
Francis
Seed RB Bea R G Abdelmalak R I Athanasopoulos A G Boutwell G P Bray J D Briaud J-L Cheung C Cobos-Roa D Cohen-Waeber J Collins B D Ehrensing L Farber D Hanemann M Harder L F Inkabi K S Kammerer A M Karadeniz D Kayen RE Moss R E S Nicks J Nimmala S Pestana J M Porter J Rhee K Riemer M F Roberts K Rogers J D Storesund R Govindasamy A V Vera-Grunauer X Wartman J E Watkins C M Wenk Jr E and Yim SC (2006) Investigation of the Performance of the New Orleans Flood Protection Systems in Hurricane Katrina on August 29 2005 Volume I Main Text and Executive Summary New Orleans USA
Steenbergen RDJM (2013) Lecture notes Probabilistic Design Delft the Netherlands TU Delft
TAW (Technische Adviescommissie voor de Waterkeringen) (2004)Technisch Rapport Waterspanningen bij dijken The Netherlands
Terzaghi K (1942) Theoretical Soil Mechanics New York Wiley ISBN 978-0-471-85305-3 Teunissen JAM (2009) SBW Analyse macrostabiliteit van dijken met Eindige Elementen
Modellen Achtergronden bij Activiteit EEM 04a Opstellen stappenplan Deltares the Netherlands
Terzaghi K (1943) Theoretical Soil Mechanics John Wiley and Sons New York
Turaacutengi T amp Rabitz H (2004) Local Methods In A Saltelli K Chan and EM Scott (Ed)
Sensitivity Analysis (pp 81-99) John Wiley amp Sons Publication
Vergouwe R Huting RJM and van der Scheer P (2014) Veiligheid Nederland in Kaart 2 Overstromingsrisico dijkring 43 Betuwe Tieler- en Culemborgerwaarden The Netherlands
Vrijling JK Schweckendiek T and Kanning W (2011) Safety Standards of Flood Defenses Keynote at ISGSR 2011 (International Symposium on Geotechnical Safety and Risk) Munich Vogt Schuppener Straub amp Braumlu (eds)
Vrijling JK and Gelder PHAJM van (2002) Probabilistic Design in Hydraulic Engineering Lecture notes TUDelft Faculty of Civil Engineering and Geosciences Delft Vrouwenvelder ACWM Steenbergen HMGM and Slijkhuis KAH (1999) Theoriehandleiding PC-Ring Delft the Netherlands TNO
Waarts PH (2000) Structural reliability using Finite Element Analysis An appraisal of DARS
(Directional Adaptive Response surface Sampling) Delft University of Technology Delft the
Netherlands
Wyllie DC amp May CW (2004) Rock Slope Engineering (4th edition) Spon Press London
MSc Thesis A Rippi 129
Appendix A OpenTURNS features
In this Appendix an overview of the most relevant of OT is given Firstly in section 112 an
introduction of the applicable methods in SA is made while in section 52 the results from such an
analysis are presented In Appendix A1 a more detailed elaboration is made of how the FAST
sensitivity method that was applied in this project works
In Appendix A2 an explanation of the different optimization algorithms available for FORM
analysis is given The theory behind these algorithms and their convergence criteria are discussed
while finally an evaluation of their performance after applying them in case of the anchor failure
is presented This evaluation has been conducted in order to decide about the most relatively
reliable algorithm that should be adopted for the reliability analysis
Finally in Appendix A3 the types of probability distributions that have been used in this study
for the random variables and which are available in OT are displayed
A1 Fourier Amplitude Sensitivity Test (FAST) In this section the extended FAST sensitivity method is discussed as it is also described in the
Reference Guide of OpenTURNS 15 (2015) FAST is based on the Fourier decomposition of the
variance of the model response ( ) the latter being represented by its Fourier expansion
is an input random vector of independent components Its key idea is to
recast this representation as a function of a scalar parameter by defining
exploring the support of the input random vector
For each input the same procedure is realized in three steps
1) Sampling
Deterministic space-filling paths with random starting points are defined ie each input Xi is
transformed as follows
( ( )) (A1)
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 130
Figure A1 Search curves in the input space as they are defined from the transformation (the blue
dot indicates the direction of the current search path)
where is the number of input variables N is the length of the discretization of the s-space with
s varying in (-π π) by step 2πN is a random phase-shift chosen uniformly in [0 2π]
is a set of integer frequencies assigned to each input Xi The frequency
associated with the input of interest is set to the maximum admissible frequency satisfying the
Nyquist criterion (which ensures to avoid aliasing effects)
(A2)
with M the interference factor usually equal to 4 or higher It corresponds to the truncation level
of the Fourier series ie the number of harmonics that are retained in the decomposition realised
in the third step of the procedure And the maximum frequency of the complementary set of
frequencies is
(A3)
with the index ldquo-irdquo which meaning ldquoall but irdquo
2) Simulations
Output is computed such as
Then is expanded onto a Fourier series
sum [ ] (A4)
where and are Fourier coefficients defined as follows
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 131
int
(A5)
int
(A6)
These coefficients are estimated thanks to the following discrete formulation
sum
(A7)
sum
(A8)
3) Estimations by frequency analysis
The quantitative measure of sensitivity is represented by Sensitivity Indices The first-order
sensitivity index of an input pi is the measure of the main (direct) effect of pi on the output
variance (where i ne j) the second-order sensitivity indices measures the interaction effect of
pi and pj on the output variance Other higher-order indices are defined in the same manner The
total sensitivity index is the sum of all sensitivity indices involving factor pi (Homma amp Saltelli
1996) For example the total sensitivity index of factor 1 for a model with 3 input factors is
given as
The first-order (Si) and total ( ) indices can be interpreted as following (Pandya et al 2008)
high pi is an influent parameter
and both small pi is not an influent parameter (neither alone nor in interaction with
other parameters)
and nearly the same no interaction of pi with the other parameters
and very different high interaction of pi with the other parameters
Total indices are especially suited to apportion the model output variation to the input factors in a
comprehensive manner The FAST method calculates the first-order and the total sensitivity
indices whereas Sobolrsquo method in addition to these also provides all higher-order sensitivity
indices to determine quantitatively the interaction between parameters However the
computational cost and calculation time of Sobolrsquos method tends to be higher than that of the
FAST method
Combining equations A1-A8 the first-order indices are estimated as follows
sum (
)
sum (
)
(A9)
where is the total variance the portion of D arising from the uncertainty of the ith input and N
the size of the sample using to compute the Fourier series
Subsequently the total-order indices are estimated as follows
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 132
sum (
)
sum (
)
(A10)
where is the part of the variance due to all inputs except the ith input
A2 Optimization Algorithms in FORM
Principles of optimization algorithms
The possible optimization algorithms in OT are
Abdo-Rackwitz (ARF)
Cobyla
Sequential Quadratic Programming (SQP)
Below a description of the main principles of the optimization methods mentioned above is
quoted ARF and SQP methods are elaborated according to Abdo and Rackwitz (1990) while
Cobylarsquos principles according to Powell (1994) Finally an evaluation of the performance of the
different methods follows after applying them in case of the anchor failure
The general optimization problem (objective function) can be written in the classical form
subject to inequality constraints (constraint function) (A11)
The Lagrangian function of the general problem is defined by
sum (A12)
where are the so-called Lagrangian multipliers The SQP method (developed by Schittkowski
(1985)) replaces the original problem by a sequence of quadratic programming problems which
are exactly solvable and which approximate the original one This is done by approximating the
Lagrangian function by its second order Taylor expansion in an initial point u0
(A13)
where
sum
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 133
sum
sum
in which is the gradient operator and represents the Hessian matrix27 of the function f in
the point u0 The optimality conditions for any iteration point k of the sequence of quadratic
expansions are
sum [ sum
] (A14)
(A15)
The exact calculation of the Hessian matrix is generally too expensive and cannot be efficiently
implemented for a general case Therefore the gradient information obtained in each point
during iteration is used to build up an approximation of this matrix using one of the known
update formulas
The new iteration point is defined by
(A16)
where is the step length and is a direction in which a line search is performed The process
stops when the optimality conditions of the original problem are satisfied
The most time consuming part in this algorithm is the updating of the Hessian matrix and the
solution of the system of equations A fair approximation of the Hessian of non-quadratic
functions is also obtained with about n updates of the matrix This means that the approximation
used in the few (say ten) iterations to reach convergence cannot be very good when the problem
has large number of variables The rounding errors during the updating process in large problems
can make the approximate Hessian to become singular Close to singularity the search direction
can be significantly distorted In this case the algorithm has to restart the iteration with a unit
Hessian matrix in the point where singularity occurred
The posterior algorithm ARF developed by Abdo and Rackwitz (1990) overcomes this problem
by obtaining a constant approximation of the true Hessian matrix Only the contribution of the
objective function to the Hessian is considered The numerical decomposition of the matrix
contains the scalar product of the gradients of the constraints as elements in each iteration
27 Hessian is a square matrix of second-order partial derivatives of a scalar-valued function or scalar field It describes the local curvature of a function of many variables
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 134
Cobyla algorithm has been developed by Powell (1994) and it constructs linear polynomial
approximation to the objective and constraint functions by interpolation at the vertices of
simplices (a simplex in n dimensions is the convex hull of n+1 points n being the number of
variables) It generates the next vector of variables from function values at the vertices
of a nondegenerate simplex in In this case there are unique linear
functions and say that interpolate f and at the vertices
and the optimization problem A15 by the linear programming problem
( ) (A18)
( )
The iterative use of expression A18 puts this method in the class of ldquosequential linear
programming algorithmsrdquo It is also essential to notice that in this method the gradients are
derived by interpolation at the vertices of simplices and not analytically as it happens in SQP and
ARF
Additionally Cobyla algorithm makes use of a trust region bound ρ so that the trust region
condition on the new vector of variables is
(A19)
Such a region gives the user some control over the steps that are taken automatically and which
respond satisfactorily to the fact that there may be no finite solution to the linear programming
problem The algorithm also employs a merit function of the form
( ) ( ) [ ( ) ] (A20)
in order to compare the goodness of two different vectors of variables Here is a parameter that
is adjusted automatically depending on the how close to optimizing the objective function the
analysis is and the subscript ldquo+rdquo means that the expression in square brackets is replaced by zero
if and only if its value is negative so we have ( ) ( ) whenever is feasible Parameters
and are changing ( only reduces) automatically according to the improvement of the
optimization problem that sequential vectors might cause Therefore if the change in the merit
function does not happen to improve the optimization then these values are changing Especially
regarding the parameter user gives the initial and the final values of it namely and (it
is recommended that be a reasonable change to make the variables for a coarse exploration
of the calculation while should be approximately the required distance from the final vector
of variables to the solution of the optimization problem)
Convergence criteria
The convergence of the aforementioned algorithms is controlled by the evaluation of the
following errors expressed in the standard space (for the relationship between the real space (x-
space) and the standard space (u-space) see section 422) (OpenTURNS 15 (b) 2015 pp 190-
192)
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 135
The absolute error which is the distance between two successive iterates
(A21)
The constraint error which is the absolute value of the limit state function minus the
threshold
(A22)
The relative error which is the relative distance between two successive iterates (with
regards to the second iterate)
(A23)
The residual error which is the orthogonality error (lack of orthogonality between the
vector linking the center and the iterate and the limit state function)
( ) (A24)
The algorithm converges if all the final error values are less than the maximum value specified by
the user The convergence can be also controlled by altering the maximum number of iterations
which should be higher in case of stricter required errors What is more there is the possibility to
manipulate the starting point that the algorithm uses for its first iteration Usually the default
starting point in a FORM analysis consists of the mean values of the random variables However if
there is already some knowledge about the potential design point it can be set as the starting
point of the algorithm in order to accelerate the convergence and save a lot of time
The aforementioned errors the maximum number of iterations and the starting point were
manipulated during the analysis in order to understand the functionality of the different
algorithms and to improve their efficiency
Evaluation of the algorithms performance
An evaluation of the above optimization algorithms has been made in order to make a decision of
the most suitable for the current project In principle a literature overview regarding the
performance of the algorithms is presented Then the 3 algorithms were tested considering the
anchor failure limit state function (see Eq 55) with a stress limit of 450000 Pa and their results
are discussed below
According to Abdo and Rackwitz (1990) the SQP method seems to have better convergence
behavior for highly curved constraint functions which is especially true when very expensive
structural state functions are involved such as finite element structural analyses However the
mentioned algorithm fails to reach convergence at a problem dimension (amount of random
variables) of around n=50 Therefore for increased dimensionality SQP is less efficient in terms
of storage and CPU time compared with other methods due to singularity of the updated Hessians
matrices In OT a warning is introduced since an analysis has started that a default
implementation of Hessian is being used and that the computation can be severely wrong
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 136
Abdo and Rackwitz (1990) also proved that ARF algorithm can handle problems of up to 2000
variables while the storage requirements and the CPU time are much smaller than with SQP
Therefore optimization problems of high dimensionality can be evaluated
In OT a warning appears at the beginning of a FORM calculation saying that a default
implementation of the gradient is being used and that the computation can be wrong A similar
message also appears for the SQP algorithm saying that a default implementation of the Hessian
matrix is used and that the computationrsquos results might be unreliable Moreover in both
algorithms it is advisable to check the values of the random variables that were used during the
analysis in order to ensure that they fluctuate within acceptable boundaries (Figure A4 is an
example of a variable that surpassed the minimum boundary at some stage)
Firstly the Cobyla algorithm was applied on the case study in combination with a FORM analysis
Cobyla does not require a gradient evaluation of the limit state function and does not use a default
implementation of the Hessians matrix as ARF and SQP does while the computational time is
quite small comparable to the aforementioned methods
The Cobyla algorithm was tested with (i) 3 (ii) 10 and (iii) 19 parameters respectively in order to
check its efficiency28 As it has been mentioned in the previous section the convergence criteria to
be steered are the errors (absolute relative residual and constraint) the maximum number of
iterations and the starting point At first the default errors were used with a value of 10-3 but it
was realised that such an order of errors lead to non-convergence of the algorithm Therefore the
algorithm was then checked after relaxing the maximum errors to values of 01 05 10 50 and
100 Finally the combination of the errors that led to convergence was 01 05 01 02 for the
maximum absolute relative residual and constraint error respectively After the analysis the
number of LSF calls in case (i) (ii) and (iii) was 25 even if the maximum one was set to 100 and
the probability of failure was around 035 The probability of failure was quite high and that is
why an inspection of the output file was carried out In Figures A2 and A3 the values of the
anchor limit state function are presented in the right graph during the iterations It was then
observed that the threshold of 450 kPa was never surpassed during the iterations
28 All of the algorithms were first checked with very simple LSF (linear and quadratic) and their results were verified by MC and DS
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 137
Figure A2 Friction angle and cohesion progress during a FORM analysis with Cobyla and with 3 random variables The figure at the right illustrates the limit state function evolution
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 138
Figure A3 Friction angle and cohesion progress during a FORM analysis with Cobyla and with 19 input parameters The last figure at the right illustrates the limit state function evolution
An additional action was then taken by relaxing the maximum number of iterations from 100 to
200 but no improvement was noticed (the LSF calls were again 25) It was also given extreme
values to the errors and especially to the absolute constraint and relative error in order to check
the response of the algorithm but that did not lead to better results Furthermore the
parameter (see theory part for explanation) was manipulated but even then no further
improvement was noticed in terms of the results In Figures A2 and A3 the performance of the
algorithm during the analysis can be seen through the values of the variables that were assigned
during the iterations
As someone can notice from the graphs above it is obvious that Cobyla does not perform very
well in high dimensionality problems as the variables do not seem to vary a lot More specifically
in Figure A3 with the 19 parameters it can be noticed that the random variable does not change
at all apart from two peaks at the beginning and at the end of the analysis respectively The
incapability of the algorithm to change the parameters in order to reach the limit of the anchor
stress ie 450 kPa (while here a maximum of approximately 267 kPa is reached) leads to a poor
approximation of both the design point and the probability of failure In Figure A2 with only 3
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 139
parameters the variability of the parameter is higher during the analysis which leads to a better
approach of the limit (around 370 kPa which is larger than with the 19 parmaters) but even then
a failure probability of 035 was not a realistic value On the other hand Cobyla is easy to use only
for small number of variables (ie less than 9) with linear LSF otherwise the linear
approximations can be highly inefficient Apparently the limit state function of the anchor stress
level as a function of the soil parameters is much more complex in order to be approximated by
this algorithm
Therefore the other available optimization algorithms ie ARF and SQP were examined as well
Because of the similarity of the two algorithms only the ARF performance is presented below and
a link with SQP is made when it is needed
The ARF algorithm was tested initially with 19 parameters The settings for the errors that were
used were 01 for the absolute relative residual and constraint error respectively and the
maximum number of iterations was set to 100
Figure A4 Friction angle and cohesion progress during a FORM analysis with ARF algorithm and with 19 input parameters The figure at the right illustrates the limit state function evolution while the red spots indicate the points where the 450 kPa threshold was surpassed
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 140
The algorithm worked efficiently which means that the algorithm managed to converge towards
the threshold of 450 kPa of the limit state function as it is depicted in Figure A4 (see red dots) It
was also noticed that the random variables fluctuated a lot during the iteration steps even if the
amount was significant (19 parameters) The capability of the algorithm to vary all the 19
variables evenly and in a wide range makes it possible to find the real design point and thus to
come up with a more reliable failure probability The resulted probability of failure was 3810-9
and the reliability index was ϐ=58
However the analysis took almost 4 days to finish which is a long duration and the number of LSF
calls counted at 4151 Moreover even if the results were promising after tracing all the 19
parameters in order to inspect the algorithmrsquos performance it was found out that some variables
took unacceptable values29 during the analysis As an example in Figure A5 Poissonrsquos ratio
received negative values (see red circle) at the beginning of the iterations There were also other
such examples in the same analysis such as the friction angle which took sky-high values and the
interface strength which took negative values as well as values above 1 which is unacceptable It
should be mentioned that the design point of the algorithm was completely reasonable but even
then such a performance can prevent the algorithm from converging to the right design point
quickly and increase the computational effort
Figure A5 Poisson ratio progress during the FORM analysis with ARF algorithm
Taking into account the above implications there are two main concerns at that stage (a) the
high computational time and (b) the unreasonable values that the variables took Firstly in order
to reduce the computational time the maximum errors were relaxed More precisely it was
noticed that after relaxing the maximum relative error a lot which is more related to the
29 In some cases a value of ldquoinfrdquo was given to the variables that led to termination of the analysis without results
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 141
relaxation factor (in this case an extreme value of 1000030 was tried) and increasing the absolute
error to 05 the number of LSF calls was reduced to 724 and the analysis was finished after
almost 10 hours In Figure A6 the values that the cohesion and the friction angle took during the
iterative procedure while the red spots in the right figure indicate the points that are included in
the failure space Figure A4 shows the progress of friction angle after relaxing the errors
Figure A6 Progress of the friction angle during the FORM analysis with ARF algorithm with errors
05 10000 01 01 for the absolute relative residual and constraint error respectively
As a general conclusion after the test of different combination of maximum errors by relaxing the
constraint absolute and relative error too much successive iterations of a certain variablersquos value
seem to have a wide difference which can detain the algorithm from converging Therefore the
wisest in this case study is to keep the maximum errors below or around 10 except for the
relative and the constraint error that can relax more than 1 (depending on the case) in order to
reduce the computational time to the minimum possible
To prevent having unacceptable values of the parameters and in order to increase the efficiency of
the algorithm the random variables were reduced to 11 after a SA and some of them were
truncated in a certain range according to observations and experience until now However the
algorithm then started to give an error due to zero gradient of the objective function at a specific
point (similarly SQP gave an error due to zero hessian matrix at a specific point) while the upper
limit of the LSFs could not be reached As a countermeasure the errors were relaxed (05 10 05
10 for the absolute relative residual and constraint error respectively) in order for the algorithm
not to be trapped in a specific point while also the step of the algorithm was manipulated It
was finally proved that this kind of error stemmed from the fact that the structural elements were
considered as elastoplastic and thus a limit of the stress level was considered that misled the
algorithm The elements were eventually considered as elastic and the numerical problem was
solved
30 Practically such a value has the meaning that there is no limitation of the relative error during the optimization In the standard space (where these errors are expressed) a value also greater than 10 can be also quite high However here extreme values were given due to the time limit to test multiple values
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 142
Before figuring out that the elastoplastic behaviour was giving the problem Cobyla as a gradient
free algorithm was tested again as a potential solution but with changing the starting point close
to the design one that was found with ARF It is worthwhile to take a look into the performance of
the algorithm with 19 random variables in Figure A7 In that way we boost Cobyla to converge
and it eventually performed quite differently than in Figure A3 and without giving convergence
problems The computational time was estimated to be 9 hours and the probability of failure was
very small in the order of 10-120 However the validation of the failure probability with DS
showed that the result of Cobyla was far from correct and thus ARF was preferred for continuing
with the rest of the analysis For educational reasons a FORM analysis with ARF was again
performed after changing the starting point and it was found that the computational time was
reduced to 4 hours Therefore it is still worthy to be aware of the fact that by changing the
starting point the performance of the algorithm can improve a lot However that requires a prior
knowledge of the potential design point that usually does not exist
Figure A 7 Cobyla performance after changing the starting point closer to the failure point
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 143
It must be also mentioned that by reducing the number of the variables to 11 as it was mentioned
before the algorithms could not converge to the errors that has been set and the computational
time increased compared with the case of 19 variables Taking into account the high interaction
among the variables that was proved in section 73 through the SA it can be possible that the
combination of the soil variables could lead to failure rather than the variables individually
Therefore an analysis was tried again with the complete set of the parameters and it was realised
that the algorithm did converge to the limit value
A3 Distribution Types
Uniform Distribution
If the probability of a result is between a lower limit a and an upper limit b and if every result
between these limits is equally probable a uniform distribution should be used This distribution
is defined by the two limit parameters and the probability density function is written
(A25)
and the CDF is
(A26)
Figure A8 Probability density function of a uniform distribution with a=-2 and b=2
The mean and the variance of the distribution are also defined by the limit parameters
(A27)
(A28)
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 144
Normal Distribution
One of the most commonly used distribution types is the normal distribution or Gauss-
distribution and it is defined by its first two central moments the mean micro and the standard
deviation σ The PDF is given by
radic
(A29)
and its CDF is
int
radic
(A30)
Figure A9 Probability density function of a normal distribution with different mean and standard
deviation values (notice that the curve with micro=0 and σ2=1 belongs to the standard normal distribution)
The normal distribution with mean and standard deviation is called the standard
normal distribution In the most reliability methods such as FORM SORM DS etc the analysis is
carried out in the standard normal space which actually means that the stochastic parameters are
transformed into standard normal variables
Lognormal Distribution
If a variable X has a lognormal distribution the variable Y=ln(X) has a normal distribution The
probability density function is defined as
radic (
) (A31)
where and the mean and the standard deviation of the normally distributed random
variable Y Sometimes it is useful to work with the Yrsquos parameters rather than Xrsquos These
parameters can be expressed as follows
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 145
radic
(A32)
(A33)
where and are the mean and the standard deviation of Y In Figure A10 an example is given
of a log-normally distributed variable with different values of mean and standard deviation
Figure A10 Probability density function of a log-normal distribution with different mean and standard deviation values
Truncated Normal Distribution
The truncated normal distribution is the probability distribution of a normally distributed
random variable whose value is either bounded below or above or both
Suppose has a normal distribution and lies within the interval
Then conditional on has a truncated normal distribution whose probability
density function is
(
) (
)
(A34)
and otherwise and is the probability density function and the CDF of the standard
normal distribution similar to Eq (1) and (2) If
and
then the CDF is
defined as
(A35)
Then the mean value and the variance of the conditional will be respectively
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 146
(A36)
[
(
) ] (A37)
In the next figure the PDF histograms of a variable with a mean of 0 and a standard deviation 1
were plotted for two cases (a) variable is normally distributed (b) variable is truncated-normally
distributed with =-2 and b=2
Figure A 11 Truncated normal distribution in comparison with normal distribution
MSc Thesis A Rippi 147
Appendix B Plaxis 2D (2015) features
In this Appendix the basic features of Plaxis 2D that were treated in this thesis are discussed First
of all an introduction over the Mohr Coulomb model is made as the soil failure criteria stem from
that model Then the initial stress generation in Plaxis is discussed and how that was applied in
the current master thesis while also the φ-c reduction technique is described which was used in
the SA of the soil failure Last but not least a more detailed explanation of the interface strength
Rinter is given which represents the modeling of the soil-structure interaction in Plaxis and it has
been considered as a random variable in the current project
B1 Mohr Coulomb failure criterion Soil tends to behave in a highly non-linear way under load This non-linear stress-strain behaviour
can be modelled at several levels of sophistication There are different constitutive models
available for simulating the behaviour of the soil whose parameters can change depending on the
level of sophistication that is required in each case
As someone can notice in Table B1 different models can be applied in different cases of soil
behaviour Hardening Soil model is becoming more and more interesting in soil structures as it
has been proved to approach the plastic behaviour of the soil and the non-linear stress-strain
relationship in a reliable and promising level Moreover UBCSAND31 and the Hypoplastic model
are more sophisticated models capable of representing several aspects of soil behaviour
however the amount of the model parameters is large and for that reason they are recommended
only in case that enough soil data are available for determining these parameters
Table B 1 Features of material models (source Behaviour of Soils and Rocks 2015)
ModelFeature Elasticity Failure Hardening Softening Small strain stiffness
Cyclic loading
Liquefaction Creep
Mohr-Coulomb x x
Duncan-Chang x x
Hardening Soil x x x
HSsmall x x x
x
Modified Cam-Clay
x x x x
Soft Soil x x x
Soft Soil Creep x x x
x
UBCSAND x x x
x x
Hypoplastic (x) x x x
x (x)
31 UBC are the initials for University of British Columbia in Vancouver in Canada
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 148
In this master thesis the Mohr-Coulomb model is going to be applied which is simple but an
efficient model to describe the soilrsquos failures condition Below a description of that model is made
The Mohr-Coulomb model is a linear elastic perfectly plastic model (see Figure B1) The linear
elastic part is based on Hookersquos law while the perfectly plastic part is based on the Mohr-Coulomb
failure criterion formulated in a non-associated plasticity framework
The main principle of elastoplasticity is that total strains and strain rates consist of elastic and
plastic strain components
(B1)
(B2)
in which the elastic strain rate component abides by the Hookersquos law and thus it can be expressed
as follows
(B3)
where is the elastic stress-strain matrix and the effective stress rate The plastic strain
component is given by
(B4)
where is a scalar defining the magnitude of the plastic strains called plastic multiplier and
is
a vector describing their direction with to be the so-called plastic potential function
Figure B1 Basic principle of elastoplasticity
The occurrence of (elasto)plasticity can be indicated by the yield function which is
illustrated in Figure B2 in principle stress space and it is derived according to the Mohr-Coulomb
failure criterion expressed in the next formula
(B5)
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 149
Considering the sign of this yield function the different soil states can be recognised as follows
if flt0 (inside yield contour) pure elastic behaviour
if f=0 and dflt0 unloading from aplastic state (=elastic behaviour)
if f=0 and df=0 elastoplastic behaviour
if fgt0 (outside yield contour) impossible stress state
Figure B2 Mohr-Coulomb yield surface in principle stress space
Finally in Mohr-Coulomb model the calculation of the plastic strains is achieved by the so-called
non-associated flow rule where with the dilatancy angle
This makes a difference between the friction and the dilatancy angle which prevents the model
from overestimating potential uplift
Concluding this model uses 5 input parameters
Stiffness parameters E (Youngrsquos modulus) and v (Poissonrsquos ratio) as they are included in
the elastic stress-strain matrix according to Hookersquos law
Strength parameters c (cohesion) and (friction angle)
Dilative behaviour ψ (dilatancy angle)
The author is aware of the possibilities and limitations of this model In particular the model is
capable of
a first order approach of the soil behaviour in general
describing quite well the (drained) failure behaviour
It should be mentioned at that point that the Youngrsquos modulus E cannot be steered via the
Python interface (as it appears to be a readonly parameter) and thus the shear stiffness G was
considered instead that is defined as follows
[kPa]
f=0 (yield contour)
flt0
fgt0
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 150
On the other hand the model has also limitations that are referring to
the linear elastic behaviour until failure (no strain- or stress path-dependent stiffness
behaviour)
overestimation of the shear strength in case of an undrained behaviour
However this model is considered reliable for detecting failure and giving realistic stresses
magnitudes in the soil and structural elements
Moreover in this project a non-associative behaviour was assumed by considering a zero
dilatancy angle However it has been proved by Teunissen (2009) that in case of deep failure
mechanisms in a dike structure with soil materials of a friction angle above 30deg problems arise in
the numerical analysis in finite element calculations Particularly the strength in the sliding
surface decreases due to rotation and the ground stresses are relieving around it by releasing
energy which might lead to calculation implications These phenomena are not treated in this
thesis but it is advisable to check and review the results after such an analysis
B2 φ-c Reduction Technique In this section the ldquoSafety Calculationrdquo is applied as it is described in Plaxis (Plaxis 2D (a) 2015)
In the Safety Calculation the φ-c reduction technique (Brinkgreve amp Bakker 1991) is adopted
where the shear strength parameters tanφ and c of the soil as well as the tensile strength are
successively reduced until failure of the structure occurs The dilatancy angle ψ is in principle not
affected by the phic reduction procedure However the dilatancy angle can never be larger than
the friction angle When the friction angle φ has reduced so much that it becomes equal to the
(given) dilatancy angle any further reduction of the friction angle will lead to the same reduction
of the dilatancy angle The strength of interfaces if used is reduced in the same way
The total multiplier is used to define the value of the soil strength parameters at a given
stage in the analysis
(B6)
where the strength parameters with the subscript ldquoinputrdquo refer to the properties entered in the
material sets and the parameters with the subscript ldquoreducedrdquo refer to the reduced values used in
the analysis is set to 10 at the beginning of a calculation to set all material strengths to
their input values
The incremental multiplier is used to specify the increment of the strength reduction of the
first calculation step (this increment is by default 01 but user can change it) The strength
parameters are successively reduced automatically until all required steps have been performed
(by default these steps are set to 100 but this can be subjected to changes depending on the
situation) It must always be checked whether the final step has resulted in a fully developed
failure mechanism In other words if has reached an almost constant value while the
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 151
deformation is continuing that indicates an equilibrium state In that case the factor of safety is
given by
(B7)
B3 Initial Stress Generation The first construction phase in Plaxis was the initial stress generation before the dike has been
constructed In that phase for horizontal soil surfaces with any soil layers parallel to this surface
and horizontal phreatic level32 a K0-procedure can be applied where K0 is the ratio between
horizontal and vertical stresses
(B8)
In practice the value of for a normally consolidated soil is often assumed to be related to the
friction angle by Jakyrsquos empirical expression (Jaky 1944)
Using very low or very high K0-values in the K0 procedure may lead to stresses that violate the
Mohr-Coulomb failure criterion In this case Plaxis automatically reduces the lateral stresses such
that the failure condition is obeyed Hence these stresses points are in a plastic state and are thus
indicated as plastic points Although the corrected stress state obeys the failure condition it may
result in a stress field which is not in equilibrium
B4 Interface Strength In FEM for a proper modelling of soil-structure interaction interface elements can be added to
plates in order to simulate the thin zone of intensely shearing material at the contact between the
plate and the soil
The strength properties of the interface are linked to the strength properties of the adjacent soil
layer and they are reduced by a factor of as follows
(B9)
(B10)
(B11)
where index ldquoirdquo refers to the interfacersquos parameters whereas index ldquosoilrdquo refers to the adjacent soil
parameters The behaviour of the interfaces is defined according to the Coulomb criterion which
distinguishes between elastic behaviour where small displacements can occur within the
interface and plastic interface behaviour when permanent slip may occur
32 For non-horizontal surfaces a Gravity loading is recommended
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 152
When the interface is elastic then both slipping (relative movement parallel to the interface) and
gapping (relative displacements perpendicular to the interface) can occur The magnitudes of the
interface displacements are
(B12)
(B13)
where is the shear modulus of the interface is the 1-D compression modulus of the
interface is the virtual thickness of the interface (generated automatically from Plaxis) is
the elastic interface normal stiffness and is the elastic interface shear stiffness
The shear and compression moduli are related as follows
(B14)
(B15)
Taking into account the above relations the displacements can be highly influenced by the elastic
parameters and thus the strength reduction factor can have a considerable impact on the
systemrsquos behaviour In the Reference manual of Plaxis 2D (2015) a value of 23 is generally
recommended while a value of 1 would not normally be used It should be stated that in the
framework of this thesis there are no available measurements in order to define the sensitivity to
this parameter and hence it is considered uncertain variable following a truncated normal
distribution
Appendix C ndash NEN 6740 ndash Table 1
MSc Thesis A Rippi 153
Appendix C NEN 6740 - Table 1
MSc Thesis A Rippi 154
MSc Thesis A Rippi 155
Appendix D Input Files for the Reliability Analysis
Below an example of the input files that were used for the anchor the sheet pile the soil body and
the system failure is presented in Figures D1 D2 D3 and D4 respectively In particular
regarding the input file of the anchor failure the meaning and the scope of each line is briefly
explained in Table D1
Figure D 1 Anchor analysis input file (with FORM method)
APPENIDX D ndash Input Files for the Reliability Analysis
MSc Thesis A Rippi 156
Table D1 Explanation of input components
A Reliability method and its parameters eg FORM with a maximum number of iterations 100 a maximum absolute error of 05 a relative error of 1000 a residual error of 05 and a constraint error of 05
B Number of random variables
C Soil material as it is specified in Plaxis eg AKleiVenigOnder is the clay layer under the dike
D Soil parameter eg Rinter is the interface strength of the clay layer
E Probability distribution and its parameters eg Truncated normal distribution with mean 066 standard deviation 013 and bounds from 00 to 10
F Copula type
G I Limit state function I is the threshold of the function
H Response surface function that is activated in case of Plaxis errors
Figure D 2 Sheet pile wall analysis input file (with FORM method)
APPENIDX D ndash Input Files for the Reliability Analysis
MSc Thesis A Rippi 157
Figure D 3 Soil analysis input file (with DS method)
APPENIDX D ndash Input Files for the Reliability Analysis
MSc Thesis A Rippi 158
Figure D 4 System analysis input file (with DS method)
MSc Thesis A Rippi 159
Appendix E Characteristic and mean values
The characteristic value of a parameter implies that there is 5 probability that the real value is
higher (load) or lower (resistance) than the characteristic In Figure 22 the probability density
function of a standard normal distributed parameter (mean value equals to zero and standard
deviation 1) is depicted In case of such a distribution the characteristic value is 164σ higher than
the mean value
Figure E1 Characteristic value illustration for a normal distribution (source Wolters 2012)
In this study full probabilistic calculations are taking place in order to estimate the corresponding
structural reliability According to the probabilistic design concept the mean values and the
standard deviations are used instead of characteristic values In that case characteristic values
and shall be converted to their mean values and as follows
where
(E1)
(E2)
and are the values for load or resistance respectively to which a probability of (non-
)exceedance of 95 corresponds (non-exceedance for the load and exceedance for the resistance)
(for a standard normal distributed parameter this is 1645) The definition of the rest parameters
is given in section 423
MSc Thesis A Rippi 160
MSc Thesis A Rippi 161
Appendix F Reliability methods
In this appendix an overview of different reliability methods is given More precisely in F1 an
explanation is given concerning the generation of random samples that is applied in MC as well as
in other sampling methods In Appendix F2 more advanced sampling methods than crude MC are
presented that can be used as accelerating reliability methods in the sense that the computational
time can be reduced Finally in Appendix F3 some basic principles of the FOSM are elaborated
that can help to better understand the concept of FORM as well
F1 Generation of random samples in Monte Carlo
The non-exceedance probability of an arbitrary random variable is uniformly distributed between
zero and one regardless of the distribution of the variable In formula (CUR ndashpublication 190
1997)
(F1)
is the uniformly distributed variable between zero and one and is the non-exceedance
probability Thus for the variable X
(F2)
in which is the inverse of the PDF of X Using this formula a random number X can be
generated from an arbitrary distribution by drawing a number of from the uniform
distribution between zero and one
More or less the same way base variables of a statistical vector can be drawn from a known joint
probability distribution function However the joint PDF must then be formulated as the product
of the conditional probability distributions of the base variables of the vector In formula this is
( )
By taking m realizations of the uniform probability distribution between zero and one a value can
be determined for every
(F3)
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 162
If the variables are statistically independent this can be simplified to
(F4)
By inserting the values for the LSF one can check whether the obtained vector ( ) is
located in the safe area
F2 Other Sampling Methods
Quasi-Monte Carlo Simulation
As an alternative to pseudo-random number generation (used in Crude MC) the quasi-random
number sequences are introduced from where the method Quasi-MC took its name Quasi-MC is a
technique which approximates the probability of failure using low discrepancy sequences33
x1hellipxN instead of randomly generated sequences
If the domain of failure is the goal is to estimate the following
probability
int ( )
(F5)
The main principle of this method is the integral of the function on [ ] can be
approximated by using some low discrepancy sequence x1hellipxN as follows
int ( )
sum
(F6)
The low discrepancy sequence is generated in Δ according to Lebesgue measure and then may be
transformed to any measure micro thanks to the inverse CDF technique in order to approximate the
integral
int ( )
sum
(F7)
In Figure 4 a comparison is made between quasi-random sample and pseudo-random sample in
order to illustrate the difference in sampling for the two methods
33 The discrepancy of a sequence is low if the number of points in the sequence falling into an arbitrary set B is close to proportional to the measure of B as would happen on average (but not for particular samples) in the case of a uniform distribution
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 163
Figure F1 Example of Quasi random Sample (left) and Pseudo Random Sample (right) (source
Waarts 2000)
This method a priori provides an asymptotically higher convergence rate than traditional MC but
no general rule can guarantee a better efficiency of the Quasi-MC sampling than the classical MC
sampling The advantages of Quasi-MC sampling tend to disappear with the increase of the
number of dimensions of the input variables It is recommended to use the Quasi-MC technique
with high sampling sizes or with very low dimensionality (in OT this method is valid only if the
input random variables are independent) (OpenTURNS 15 2015)
Importance Simulation
The main idea of the Importance Sampling method is to replace the initial probability distribution
of the input variables by a more efficient one Efficient means that more events will be counted
in the failure domain and thus reduce the variance of the estimator of the probability of
exceeding a threshold In other words importance sampling starts with the observation that if we
are going to sample randomly we should distribute the points to put the most of them in the
region that contains information (ie the failure region) and to waste as few as possible
In this method a sampling density is used instead of the actual probability density function
For a MC simulation the probability of failure is calculated by
sum ( ( )) ( )
( )
(F8)
where ( ( )) ( )
( ( )) ( )
The problem is that prior knowledge is needed of the failure area which in most cases is not
available In Figure 5 examples of the different sampling methods are shown
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 164
(a) (b)
(c)
Figure F2 Comparison of Importance Sampling (a) MC (b) and Numerical integration (c) (source
Latin Hypercube Simulation
Latin Hypercube Sampling is a sampling method enabling to better cover the domain of variations
of the input variables thanks to a stratified sampling strategy (in OT this method is applicable in
the case of independent input variables) Stratified sampling tries to place the sampling points so
that more of them will be found in regions where the variance of the limit state function g(x) is
largest In Figure F3 a comparison is made between the LHS and MC in terms of their sampling
technique The sampling procedure according to Baecher and Christian (2003) is based on the
next steps
1 Divide the range of each variable into several intervals of equal probability
2 If there are D variables to be sampled each in k regions then for each variable the k
sampling points are placed in independent random order
3 K samplings are made by choosing the first randomized value for each variable the second
randomized value for each variable and so on until k sets of randomized variables have
been chosen
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 165
Figure F 3 Comparison between Latin Hypercube sampling and MC (source OpenTurns 15 2015a)
F3 First Order Second Moment (FOSM) Method The First-Order Second-Moment (FOSM) approach (Ang amp tang 1984) provides analytical
approximation for the mean and the standard deviation of a parameter of interest as a function of
the mean and standard deviations of the various input factors and their correlations
Consider Z to be a function of random variables x1 x2 hellip xn that is
(F9)
In general case are correlated with covariance matrix [C] ie [C]=[σ][R][σ] where [σ]
is a diagonal matrix of standard deviations and [R] is the (positive-definite and symmetric)
correlation matrix with diagonal elements Rii=1 and non-diagonal elements Rij=ρij (ρij is the
correlation coefficient between variables i and j) In scalar notation Cij= σiσjRij (Farrokh 2007)
The first order approximation to the mean variance and standard deviation of the Z function is
based on the first terms of a Taylor series expansion of Z The following first-order estimates of
mean and variance are obtained
[ ]
(F10)
[ ] (F11)
Where the vector b denotes evaluated at the mean values of ie
(F12)
If there is no correlation among the variables Eq 223 can be written as
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 166
sum
(F13)
Then the reliability index can be calculated in 6 steps (Baecher amp Christian 2003)
1 Identify all variables that affect the mechanism that is researched
2 Determine the best estimate (usually the mean value) of each variable E[ ] and use these
to calculate the best estimate of the function E[Z]
3 Estimate the uncertainty in each variable and in particular its variance Var[ ]
4 Perform sensitivity analyses by calculating the partial derivatives of Z with respect to each
of the uncertain variables or by approximating each derivative by the divided difference
5 Use the equation of Var[Z] to obtain the variance of the function Z
6 Calculate the reliability index
[ ]
(F14)
MSc Thesis A Rippi 167
Structural reliability analysis of a dike
with a sheet pile wall
Coupling Reliability methods with Finite Elements
by
A RIPPI
in partial fulfilment of the requirements for the degree of
Master of Science
in Civil Engineering
at the Delft University of Technology
to be defended publicly on Wednesday November 25 2015 at 1100
Graduate Aikaterini Rippi Student ID 4325583 E k-rippihotmailcom
Thesis committee Prof dr ir S N Jonkman TU Delft
Dr ir R B J Brinkgreve TU Delft and Plaxis bv
Dr ir T Schweckendiek TU Delft and Deltares
Dr A Teixeira Deltares
An electronic version of this thesis is available at httprepositorytudelftnl
MSc Thesis A Rippi i
MSc Thesis A Rippi ii
Preface
This thesis is the final challenge in the master Hydraulic Engineering at Delft University of
Technology The report ldquoStructural reliability analysis of a dike with a sheet pile wall Coupling
Reliability methods with Finite Elementsrdquo was completed at Deltares as a part of a larger
research project namely TO2 in collaboration with Toegepast Natuurwetenschappelijk
Onderzoek (TNO) I chose that subject first of all because it combines two things that I enjoyed
a lot at TU Delft probabilities and flood defences systems Secondly it was an opportunity for
me to get acquainted with FEM and geotechnics that I was always interested in
Different people have contributed to the successful completion of this thesis First of all I would
like to express my appreciation to my graduation committee for their guidance and particularly
my daily supervisor Timo Schweckendiek Together we had many fruitful and interesting
discussions on the subject that triggered and motivated me for keep searching Especially I
want to express my gratitude and my thanks to Ana Teixeira and Jonathan Nuttall employees of
Deltares as they were also some of the main contributors to this research They stood by me not
only as colleagues and thesis mentors but also as friends I would like also to thank the rest of
Deltares employees whom ensured a friendly and easy going working environment Last but not
least I want to thank my friends and especially Panagiotis Apostolidis and my family for their
love support and advice throughout all my studies
Katerina Rippi
Delft November 2015
MSc Thesis A Rippi iii
MSc Thesis A Rippi iv
Abstract
Some dike sections in the Netherlands failed to comply with the safety standards and one of the
most promising countermeasures is the construction of retaining walls inside the dike The
Dutch design codes for dikes with retaining walls rely on Finite Element Analysis (FEM) in
combination with partial safety factors However this can lead to conservative designs For this
reason in this research a reliability analysis is carried out with FEM calculations aiming to
demonstrate the feasibility of reliability analysis for such a soil-structure interaction problem
The case study concerns a (river) dike with an anchored sheet pile wall modelled in PLAXIS The
sensitivity and reliability analyses were enabled by coupling the uncertainty software package
OpenTURNS and PLAXIS through a Python interface The most relevant (ultimate) limit states
concern the anchor the sheet pile wall and global instability (soil body failure) The case was
used to investigate the applicability of the First Order Reliability Method (FORM) and
Directional Sampling (DS) to analysing these limit states Finally also the system reliability was
evaluated using sampling-based methods (DS)
Due to the considerable number of random variables before starting the reliability analysis a
sensitivity analysis was conducted for each limit state This indicated the most important soil
layers to be accounted for and the variables to be considered as stochastic The sensitivity
analysis and later on the reliability analysis were based on analytical formulations of the limit
state functions The anchor and the sheet pile limit states were formulated in terms of their
yield stress for global instability loss of equilibrium in the FEM analysis was used to define
failure Moreover the systemrsquos reliability was evaluated by taking into account all the three limit
states that were mentioned previously
The goal is to implement the coupling between FEM and reliability methods in order to analyse
the components of such a system (ie anchor sheet pile wall and dikersquos soil body) estimate the
probability of failure and identify the most important soil properties that affect the behaviour of
each component and the system as a whole The results of this research can be used to assess
and optimize the current design procedure for dikes with retaining walls
MSc Thesis A Rippi v
MSc Thesis A Rippi vi
Table of Contents
Preface ii
Abstract iv
List of abbreviations x
1 Introduction 1
11 Project objective and main research questions 2
12 Research approach and outline 3
2 System description and current design concept 7
21 System description and forces configuration 7
22 Current design concept 12
23 Safety standards 17
3 Literature study 21
31 Background 21
311 Finite Element Modeling 21
312 Uncertainties and Sensitivity analysis 24
32 Previous Studies 29
33 Overview 34
4 Structural Reliability Analysis 35
41 Basics of Reliability Analysis 35
42 Overview of Reliability Analysis Methods 36
421 Level III Methods 36
422 Level II Methods 40
423 Level I Methods (semi-probabilistic) 42
424 Response Surface Techniques (RS) 44
43 Coupling Reliability Analysis with FEM 46
431 The functionality and possibilities of OT 46
432 Coupling OpenTURNS-Plaxis 47
44 Overview 49
5 Failure Mechanisms and Limit State Functions 53
51 Introduction to the system analysis and the limit states 53
52 Limit State Functions 55
521 Serviceability Limit State 55
522 ULS for Structural Members 57
MSc Thesis A Rippi vii
523 ULS for Soil Failure 60
53 Overview 66
6 Case Study-Dike with an anchored sheet pile wall 69
61 Case Description 69
62 Soil Parameters 70
63 Finite Element Model 73
64 Deterministic Analysis 74
641 Calculation Scheme and Design Values 75
642 Construction Stages 77
643 Determination of the structural elementsrsquo characteristics 78
65 Overview 81
7 Reliability analysis results with stochastic soil properties 85
71 Method description 85
72 Mean values calculations 88
73 Sensitivity Analysis Results 91
74 Soil Shear Failure 98
75 Anchor Failure 102
76 Sheet pile wall failure 105
8 Conclusions and Recommendations 121
81 Conclusions 121
82 Recommendations 123
References 125
Appendix A 129
OpenTURNS features 129
A1 Fourier Amplitude Sensitivity Test (FAST) 129
A2 Optimization Algorithms in FORM 132
Principles of optimization algorithms 132
Convergence criteria 134
Evaluation of the algorithms performance 135
A3 Distribution Types 143
Uniform Distribution 143
Normal Distribution 144
Lognormal Distribution 144
Truncated Normal Distribution 145
Appendix B 147
MSc Thesis A Rippi viii
Plaxis 2D (2015) features 147
B1 Mohr Coulomb failure criterion 147
B2 φ-c Reduction Technique 150
B3 Initial Stress Generation 151
B4 Interface Strength 151
Appendix C NEN 6740 - Table 1 153
Appendix D 155
Input Files for the Reliability Analysis 155
Appendix E 159
Characteristic and mean values 159
Appendix F 161
Reliability methods 161
F1 Generation of random samples in Monte Carlo 161
F2 Other Sampling Methods 162
F3 First Order Second Moment (FOSM) Method 165
MSc Thesis A Rippi ix
MSc Thesis A Rippi x
List of abbreviations
CDF Cumulative Distribution Function
CoV Coefficient of Variation
CUR Civieltechnisch Centrum Uitvoering Research en Regelgeving
DS Directional Sampling
FAST Fourier Amplitude Sensitivity Analysis
FEA Finite Element Analysis
FEM Finite Element Model
FERM Finite Element Reliability Method
FORM First Order Reliability Method
FOSM First Order Second Moment
LEM Limit Equilibrium Method
LRFD Load and Resistance Factor Design
LSF Limit State Function
MC Monte Carlo
OT OpenTURNS
PDF Probability Distribution Function
RFEM Random Finite Element Method
RS Response Surface
SA Sensitivity Analysis
SLS Serviceability Limit State
SORM Second Order Reliability Method
ULS Ultimate Limit State
VNK Veiligheid Nederland in Kaart
MSc Thesis A Rippi xi
The roots of education are bitter but the fruit is sweet
Aristotle
MSc Thesis A Rippi 1
1 Introduction
In the Netherlands according to the Flood Protection Program (Hoogwater
Beschermingsprogramma1) and the Room for the River program (Ruimte voor de Rivier2)
alternative structural techniques for the reinforcement of existing dikes or for future dike
constructions additional to the conventional ones have been introduced and tend to be
attractive The heightening of the dike crest and the construction of a stability berm are some of
the most common current actions for dike strengthening Additionally filter layers geotextiles
and drainage systems can be applied in advance in order to prevent several failure mechanisms
Recently cantilever or anchored sheet pile walls and diaphragm walls tend to be used as an
alternative procedure for embankment reinforcing (see Figure 11) Such an alternative can be
chosen for an improvement of the macro stability of the dikersquos inner slope and additionally for
saving space in the land area that needs to be habited
Figure 11 Design options for dike reinforcement (source Flood Defences 2015)
For that purpose engineers need a concrete and unequivocal design methodology for such
combined structures In the meantime Deltares in cooperation with the Water Board of
Rivierenland is working on a design guideline which will be applicable to design the so-called
Type II stability fences3 A draft of such a guideline is elaborated in the report of Larsen et al
(2013) In this report suggestions and recommendations are outlined for the design of such
structures with Finite Element Modeling (FEM) and partial factors Remarks are also made for
the advantages and the limitations of such models and how they should be handled so that
reasonable and trustworthy results can be retrieved from the analysis
1 httpwwwhoogwaterbeschermingsprogrammanldefaultaspx 2 httpalbertawatercomhow-is-water-governedwhat-is-room-for-the-river 3 Type II structures are structures which in combination with a soil construction fulfil water retaining functionalities
Introduction
MSc Thesis A Rippi 1
The concept behind the recommended design criteria (Larsen et al 2013) is the definition and
the evaluation of several partial factors with the view to determine overall safety factors These
factors are further discussed in section 22 However validation of these factors shall be carried
out before being used in any case otherwise either the safety of the structure is jeopardized or
the structure will be overdesigned and cost inefficient
For improving the design criteria researchers for a long time focused on enhancing structural
models (beams shells etc) and constitutive laws (elasticity plasticity damage theories etc)
With the development of computer science a great amount of work has been devoted to
numerically evaluate approximated solutions of the boundary value problems describing the
mechanical system FEM is probably nowadays the most applied approach for the solution of
these problems
However the increasing level of detail of the constitutive models and the constant enhancement
of the computational tools do not solve the problem of identification of the model parameters
and the inherent physical and modelling uncertainties Moreover in most civil engineering
applications the intrinsic randomness of materials (soil rock concrete etc) or loads (water
elevation wind earthquake motion etc) is such that deterministic models are using average or
later on characteristic values of the properties at best lead to rough representations of the
reality
As a counteraction a semi-probabilistic methodology has been developed that was based on the
application of characteristic and design values by using partial factors Current design codes
such as Eurocode provide target reliabilities for different types of structures and structural
elements according to the potential consequences of failure However these partial factors are
not always equally suitable and efficient for all types of structural applications since they have
been calibrated under specific conditions Besides the consequences of failure of flood defences
such as dikes can be comparable to the investments in increasing the reliability of such systems
are For that reason it would be advisable for these systems to define target reliabilities based
on a risk assessment (ie tailor-made solution) rather than using the standard partial factors
coming from general geotechnical design codes which may be either too low or too high for a
given flood defence system
One step of such a risk assessment is accounting for randomness and spatial variability of the
mechanical properties of materials is one of the tasks of stochastic or probabilistic mechanics
which has developed fast in the last decade In this master thesis project the uncertainty of soil
properties is going to be treated in terms of its contribution to failure For that purpose
probabilistic methods are going to be implemented for a dike with an anchored sheet pile wall
(see Figure 12) simulated in a FEM software that is specialized in soil mechanics The successful
implementation and in future research the verification of such methods can be considered as
the most preferable and cost efficient way to design structures with high safety requirements
and not only for the validation of the partial factors Of course such a procedure tends to be
time consuming However the gradual improvement of the current probabilistic methods in
combination with the state-of-the-art computer capabilities as well as the scientific knowledge
gained in terms of different systems behaviour and failure modes can introduce a more
optimized way of designing structures with considerable investments
Introduction
MSc Thesis A Rippi 2
Figure 12 Reinforced dike section with an anchored sheet pile wall
11 Project objective and main research questions At the beginning of this research it has been observed that the design procedure being followed
for dikes with retaining walls modeled in FEM can lead to an overestimation of the critical loads
and thus to a potential cost inefficient final structure An example of the magnitude of the design
values in such a case study is given in the next chapter where also a more detailed explanation
of the current design concept is described
The objective of this study is to implement a full probabilistic analysis for evaluating the
reliability of a dike reinforced with an anchored sheet pile wall Such a research analysis can
subsequently come up with valuable recommendations for the improvement of the present
design approach Soilrsquos and structural elementsrsquo behavior should be taken into account both
separately and as a system Some of the reliability methods can deal with system reliability
problems such as Monte Carlo (MC) and Directional Sampling (DS) whereas for others like
FORM and SORM additional methods should be applied that use reliability information for each
individual limit state function to obtain the systemrsquos reliability
In principle the probability of failure of different limit states is to be computed individually
while in the sequence the system reliability is going to be estimated As far as the soil failure
mechanisms are concerned this research will focus mainly on the global stability of the dike
slope while for the sheet pile wall and the anchorrsquos failure the exceedance of the steel yield
stress is going to be considered
The scope of this thesis is not only to test different reliability models robustnessrsquo in conjunction
with FEM simulations but also to get a better insight into the specific system behavior (ie of a
dike with a retaining wall) analyze each component separately and investigate its response
under certain load conditions Furthermore the minimization of the computational effort and
time could also be carried out meaningfully under the constraint of sufficient accuracy The
accuracy should be such that the probability of failure is acceptable for the ultimate limit state
(ULS) in a normal design process and according to the current safety standards Finally the
Sheet pile
wall
Anchor
Dike section
Introduction
MSc Thesis A Rippi 3
robustness of the coupling between the reliability model and the FEM will be tested via their
capability of adapting to new input parameters without encountering convergence errors
during execution
The main research question of this master thesis project is thus formed as follows
How can the probability of failure of a dike with a sheet pile wall due to global instability modeled
by a Finite Element Model be analyzed
Essential questions regarding the soil models and failure criteria as well as the reliability
methods and the systems behavior are generated which are listed below
Subquestion 1 Which reliability methods are computationally tractable in
combination with FEM
Subquestion 2 How robust (convergence) are the tractable methods
Subquestion 3 Which limit states are relevant for the envisaged application of
retaining walls in dikes and how can they be formulated using FEM
analysis outcomes
Subquestion 4 What is the contribution of different uncertainties in the failure
mechanisms of the system
Subquestion 5 Can response surface techniques help to increase the efficiency and
robustness of the reliability model
Subquestion 6 How can the current design approach for dikes with sheet piles be
improved
In the next chapters the above research question and the related subquestions are going to be
answered by following the methodology that is described in the next section
12 Research approach and outline In this section an overview is given of how the aforementioned objective and sub-questions are
approached The thesis is mainly divided into 8 chapters and 6 appendices In Figure 13 the
thesis outline is illustrated and an indication of which subquestions (Sq1-6) are treated to
which chapter is given In Chapter 1 an introduction into this research content is made and a
first illustration of the system under investigation is presented together with the main research
questions
Chapter 2 has been devoted to the description of the system (ie a dike with an anchored sheet
pile wall) and to the case study that has been adopted for being analyzed in terms of its
reliability The forces configuration is also illustrated and the results of the case study according
to the current design concept are presented and evaluated Last but not least the new
recommended safety standards are included and the current required reliability of the specific
dike section is presented
Introduction
MSc Thesis A Rippi 4
In Chapter 3 a literature study related to FEM uncertainties and sensitivity analysis is
presented Moreover previous studies that are associated with the objective of this research are
mentioned whereas some of their results were also taken into account for proceeding with this
research
The main scope of this thesis is the implementation of different reliability methods on a specific
case study with the view to investigate and analyze its behavior Some of these methods are
continuously mentioned through the test and thus in Chapter 4 an overview is given of the
most well-known reliability methods Eventually an evaluation of these methods is made based
on literature and preliminary testing with simple case studies and the procedure of their
coupling with FEM is discussed The evaluation helped to qualitatively answer subquestion 1
while also an introduction of how special reliability methods such Response Surfaces were used
in this thesis that partially covers subquestion 5 In this chapter an introduction into the Limit
State Functions (LSF) concept is made and how they are considered for the reliability analysis of
a system An LSF actually represents a failure mode that can be detected in a structural or soil
element and it is expressed as a function of several variables In Chapter 5 the failure
mechanisms and the corresponding LSF that are related to the specific case study are identified
and formulated which answers subquestion 3
In Chapter 6 the case study whose reliability is to be evaluated is presented as it was modeled
in FEM The boundary conditions are specified and the soil and structural properties are
indicated and illustrated Moreover a deterministic design is also taking place in order to
roughly estimate the structural elementsrsquo properties This will help to make a qualitative
comparison between the current design procedure and the design according to a fully
probabilistic approach by referring thus to subquestion 6
Eventually the results of the aforementioned analysis are presented in Chapter 7 The results
mainly include the estimated probability of failure of each of the system components as well as
of the system as a whole by considering the soil properties as stochastic Moreover an
interpretation of the failure points is made and an assessment of the level of impact of the
random variables on the systemrsquos behavior is carried out At that chapter subquestions 2 4 and
5 are mainly treated
Last but not least in Chapter 8 the general conclusions are presented together with some
valuable recommendations for future consideration and research Furthermore a reflection on
the methodology and how the different research questions were approached is made
Introduction
MSc Thesis A Rippi 5
Figure 13 Thesis outline
Introduction
MSc Thesis A Rippi 6
MSc Thesis A Rippi 7
2 System description and current design concept
In this section the system to be analyzed is described more in detail and the individual
components are identified Moreover the forcesrsquo configuration is explained and illustrated as
they would have been calculated with the conventional way in parallel with a qualitative
evaluation of the potential expected deformations Finally the current design concept and the
safety standards that are referred to such a system are introduced
21 System description and forces configuration The system to be investigated in this thesis consists of a dike section reinforced with a one-layer
anchored sheet pile wall The system has been simulated by FEM and later on it was coupled
with a reliability package for carrying out a reliability analysis In the figure below an
illustration of the system is depicted and the several elements of an anchored sheet pile wall are
showed
Figure 21 System layout and different components of the anchored sheet pile wall
In particular an anchored sheet pile wall consists of the sheet piles which are anchored in the
soil via a tie rod The rod is grouted most of the times in a sandy layer with a cemented grout
body and it both alleviates the sheet piles from the axial forces exerted by the upper structure
and keeps the wall stable in case of excessive developed moments due to the lateral earth
pressure Last but not least a waling system is applied mainly downstream of the sheet piles in
order to transfer the loads from the piles to the anchors in such a manner so as to avoid
excessive local stresses on the intersection between the sheet piles and the tie rod
Cross-Section
Top View
System description and current design concept
MSc Thesis A Rippi 8
The main load configuration acting on a dike section as well as a sheet pile wall is depicted in
Figures 22-24 together with the possible displacements As far as the soil body stability is
concerned most of the slope stability analysis computer programs are based on the limit
equilibrium concept according which a soil mass tends to slide down under the influence of
gravity The transitional or rotational movement is considered on an assumed or known slip
surface below the soil There an equilibrium should be achieved among the driving and the
resisting forces In that case the driving moments consist of the soil weight the water pressures
and the loads around the center of the slip surface such as a possible vertical load on the top of
the dike crest (ie traffic load) The magnitude of the water pressures is controlled by the water
elevation on the river side
In Figure 22(a) the external and internal forces acting on a slice of the slip surface are
illustrated The driving forces are the soil weight W the water pressure U and any additional
load that can contribute to the rotation of the slip surface which in turn activate the lateral
active earth pressure The resisting forces consist of the lateral passive earth pressure
(Rankine 1857) and the shear strength (Terzaghi 1943) Their expressions are written as
follows
[kN] (21)
[kN] (22)
[kPa] (23)
where is the active lateral earth pressure coefficient and the passive lateral earth
pressure [kNm3] is the unit weight of the corresponding soil layer and H [m] is the thickness
of the soil layer that the lateral earth pressure comes from (in case of multiple soil layers there
should be a distinction among the different forces) and [kPa] is the effective cohesion of the
soil layer The safety factor that is written in the Figure 22(a) as FS is defined by the ratio of the
shear strength (excluding the pore water pressure) divided by the weight of the earth body
(including the pore water pressure)
The loss of equilibrium between the driving and the resisting moments lead to the rotation and
instability of a slip surface as it is illustrated in Figure 22(b) A slide surface can take various
shapes according to the Limit Equilibrium Method (LEM) that is followed According to Bishop
method this surface tends to be circular in Uplift-Van method a horizontally compressed zone
can be also considered whereas in Spencer method the shape of the slide body can be arbitrary
The slip surface can be located either in the landside or in the river side of the dike depending
on the load the soil characteristics and the design of the dike (ie inner or outer berm
reinforcing revetment on the outer slope etc) Most of the times the landside area of the dike is
jeopardized due to different failure mechanisms that are described in section 523 The
instability of the outer slope is mainly relevant in case of sea dikes and a potential damage can
be usually reconstructed until the next flooding
System description and current design concept
MSc Thesis A Rippi 9
(a)
(b) Figure 22 (a) External and internal forces acting on a slip surface of a dike section and (b) possible deformation pattern
As a counteraction for the slope instability a sheet pile wall can be installed inside the dike The
special thing about these structures is that they make possible a greater freedom in form and
functionality than a traditional dike design This structure derives its strength from the
materials used such as steel which are able to withstand higher pressures than clay for instance
The general stability is due to friction and wedging in the bottom
After the reinforcement of the inner side of the dike with a retaining wall the strength capacity
of the wall is also important for the global stability of the system The forces to be taken into
account for the sheet pile wall stability are the active and the passive earth pressures (effective
pressures) and the water pressure as they are illustrated in Figure 23(a) In that figure the
System description and current design concept
MSc Thesis A Rippi 10
forces were simplified in a singular triangular shape as in a homogeneous soil body in order to
explain and depict the overall picture of the acting forces However the stresses distribution
over depth can be more complex depending on the variety of the soil layers that are present In
Figure 23(a) it can be noticed that a sheet pile wall which is installed close to the inner berm
might not have significant instability issues as there is the passive side that contributes to
resisting forces However in the passive side the soil can differ and be weaker than this of the
active side Therefore the passive force in that case might not be very supportive and thus the
wall shall be designed cautiously
Additionally the anchor resistance shall be reassured for the sake of the stability of both the
sheet pile wall and the overlying soil body In Figure 23(a) the forces acting on the anchorage
are displayed The distributed load qs comes from the earth pressure of the upper soil layer (it
might not be uniformly distributed as it is depicted in the figure but more in a trapezoidally type
of load) whereas the tension force Ft4 stems both from the earth pressure and the displacement
of the sheet pile wall The bending and the tension capacity of the anchor are determinant for its
stability In Figure 23(b) a possible displacement pattern of the retaining wall and the
anchorage is illustrated
Furthermore corrosion is an additional weakening impact on the wall that depends on the
water level and the pore water pressures near the structure The thickness of the wall and the
material properties are playing a key role to the resistance towards corrosion This will not be a
subject of this thesis however it should be taken into account in case of design purposes
(a)
4 where O is the average circumference of the pile P [kPa] is the shaft friction at depth z and L is the length of the pile
System description and current design concept
MSc Thesis A Rippi 11
(b) Figure 23 (a) Mobilizing forces acting on the sheet pile wall and the anchorage and (b) possible deformation pattern
It should be mentioned that as far as the pore pressures are concerned in Figure 23(a) only the
hydrostatic pore pressures are depicted Generally the pore pressures (or active pore pressures)
are the sum of the steady state pore pressures and the excess pore pressures
[kNm2] (24)
Steady state pore pressures can be hydrostatic (based on a horizontal phreatic line) or non-
hydrostatic (based on steady state groundwater flow) Excess pore pressures are based on
loading of undrained soils In FEM these various pore pressures are taken into account
automatically In Figure 24 an example of a potential distribution of the active pressures
nearby the sheet pile wall is depicted as it was deduced from Plaxis Therefore as it can be
noticed from the figure the distribution can indeed sometimes approximated as triangular
System description and current design concept
MSc Thesis A Rippi 12
Figure 24 Example distribution of the active pore pressures acting nearby the sheet pile wall
The knowledge of the type of forces that are exerted on both the structural components and the
soil body as well as of the most expected form of displacements is valuable in order for
someone to be able to evaluate the results deduced from FEM and detect potential modelling
errors Therefore this validation was necessary before starting with the reliability analysis
In the next sections the current design approach of the system described above is elaborated
Additionally the Dutch recommended safety standards are presented for this type of structures
which show the need of carrying out a reliability analysis For that purpose a case study was
adopted that it is presented first as it was designed according to the current regulations while
finally a comparison is made between the original and the new case study as it was found to be
according to the reliability analysis
22 Current design concept In the Netherlands dike reinforcement is becoming a major issue concerning the design of flood
defenses Special reinforcing structures such as cofferdams anchored sheet piles and diaphragm
walls are used for strengthening the dike Especially in the context of the Flood Protection
Program and the Room for the River program many dike reinforcement projects have been
suggested
In the report of Larsen et al (2013) the design procedure of a dike reinforced with sheet pile
wall is prescribed using FEM The safety philosophy that is followed in this technical report
actually composes the current procedure for designing dikes with sheet pile walls and it is
linked with the usage of partial safety factors Particularly the required overall safety factor
FEM that should be compared with the one from FEM calculations is determined as follows
SFEM
b d m n
(25)
where
System description and current design concept
MSc Thesis A Rippi 13
Partial safety factor indicating the uncertainties of the soil composition and
the water pressures (also called schematization factor)
Partial safety factor which is related to the calculation model and the way the
calculations have been carried out (also called model factor)
Partial safety factor which is related to the material parameters (also called
material factor)
Partial safety factor associated with damage caused during the soil tests (also
called loss factor)
Safety factor of load
Each of these partial factors shall be at first calibrated by conducting a full probabilistic analysis
of the system under consideration In most of the cases these partial factors have been already
calibrated on previous similar projects and then they are reused for any similar case In
Schweckendiek et al (2013) a new approach of using partial factors for flood defences is
proposed whose application is still under consideration At the moment the aforementioned
partial factors for dike reinforced with sheet pile walls modelled in FEM are estimated
according to mostly engineering judgement and political negotiations
In sequence the FEM design procedure that is applied for sheet pile walls inside soil structures
is based on a so-called ldquosafety calculationrdquo (for information for this type of calculation see
Appendix B2) In this calculation an artificial reduction to the soil strength parameters (friction
angle and cohesion) is applied in order to evaluate the overall safety factor and confirm if this is
lower or higher than the required one that was estimated according to Eq 25 Furthermore the
resulting moments and forces developed on the structural elements during the specific
calculation are used for their design
This is a quite helpful and indicative calculation type available in Plaxis that can give a sense of
the structurersquos safety factor and the possible failure mechanisms that can occur under the
predefined load configuration However such a method of calculating the developed stresses
can be misleading in terms of the moments and forces acting on the structural elements and the
total deformations of the system
In this thesis the reliability of a dike with an anchored sheet pile wall is going to be evaluated
and in particular the case study of Breedeveld (2011) is to be used for that purpose In Figure
26 the location of the existing dike section is presented while in Figure 25 the structure as it
was modelled in FEM is illustrated It is essential to mention that the retaining wall does not
exist in reality but it was placed so as to implement and demonstrate the current design
regulations in the report of Breedeveld (2011) The anchorrsquos angle has been taken equal to 30deg
from the vertical which is generally a steep anchor inclination That is expected to reduce the
part of the horizontal force that is undertaken by the anchor and thus the most of it exerts on
the sheet pile wall The angle is usually taken 30deg-45deg from the horizontal axis (CUR 2005) but
this is of course depended on the load conditions and the construction requirements for each
System description and current design concept
MSc Thesis A Rippi 14
case The influence of the anchor orientation and the construction methods that are generally
followed for an anchored sheet pile wall are not considered in this thesis The case study was
taken as it was modelled in the aforementioned project and the reader should be aware that this
serves the proceedings of a reliability analysis and not of a perfectly constructed retaining wall
Figure 25 Original dike section in Varik-Heesselt as it is presented in Breedeveld (2011)
In particular based on this case study and trying to understand the design criteria a
comparison was carried out between the ldquosafety calculationrdquo and a normal ldquoplastic calculationrdquo
(ie elastoplastic drained or undrained analysis) In both cases the boundary conditions the
external loads and the input parameters are the same The difference between them is the
calculation procedure followed in order to come up with final stresses which in the case of the
ldquoplastic calculationrdquo does not include a reduction of the soil strength properties In Figures 27
and 28 the results from both calculations are presented In Figure 27 the total deformations
are exhibited while in Figure 28 the moments and the axial forces on the sheet pile are shown
Figure 26 Location of the Varik-Heeselt dike section (section number TG-078) in the Dike ring 43
BetuweTieler-en Culemborgerwaarden west (source Breedeveld 2011)
System description and current design concept
MSc Thesis A Rippi 15
As it is realised from this figure the discrepancy between the two results is quite high while as
far as the deformations are concerned no realistic conclusions can be drawn from a ldquosafety
calculationrdquo Moreover taking as design moments the moments that are deduced from the
ldquosafety calculationrdquo will lead to a much heavier and more expensive sheet pile cross section than
the ldquoplastic calculationrdquo
It is also essential to mention that the required safety factor was estimated up to FEM =18
according to the partial-factor procedure that is defined in Eq 25 Such a safety factor as
prerequisite for coming up with the design values of the moments and forces of the sheet pile
and the procedure of the ldquosafety calculationrdquo in FEM can lead to a conservative and financially
inaccessible design
(a) ldquoPlastic calculationrdquo Deformed mesh and maximum deformations (umax= 008 m)
(b) ldquoSafety calculationrdquo Deformed mesh and maximum deformations (umax= 9041103 m)
Figure 27 Deformed mesh and maximum deformations in the end of (a) plastic calculation and
(b) safety calculation
30deg
System description and current design concept
MSc Thesis A Rippi 16
(a) ldquoSafety
calculationrdquo
Mmax = 9392 kNmm Nmax= -4402 kNm
(b) ldquoPlastic
calculationrdquo
Mmax = 6833 kNmm Nmax= -1372 kNm
Figure 28 Bending moments and axial forces exerted on the sheet pile during the (a) safety
calculation and (b) plastic calculation
On the other hand the aforementioned design procedure guaranties a strong structure capable
of probably undertaking more than the expected loads and thus ensuring the safety against
flooding of the landside However the knowledge on fully probabilistic methods that has been
obtained as well as the development of the technology can introduce a reliability analysis of the
system as a preferable way for the future design concept Such a procedure should be
accompanied also by field test that would be able to validate FEM results as well as in-situ
measurements and inspections of the soil properties for calibrating their statistical
characteristics
Due to the conservative results that the aforementioned procedure came up with and because of
the simplifications that this case study was later subjected to in order to be used in this thesis a
new design was carried out in Chapter 6 and the properties of the structural elements were
redefined
The inclination towards probabilistic methods has not only been created due to the incomplete
design regulations for the system under consideration but also the current proposed safety
standards which are discussed in the section below recommends a risk based safety assessment
of the primary flood defenses which in principle implies the evaluation of the failure probability
In the next section an overview of the new recommended safety standards related to primary
flood defenses is given and the required reliability for the stability of the aforementioned case
study is elaborated
System description and current design concept
MSc Thesis A Rippi 17
23 Safety standards An important factor for the evaluation of a structurersquos reliability is the safety standards that
have been set and according which the acceptance or not of the failure probability is made After
the safety assessment5 of 2011 it has been declared that 33 (1225 km) of the primary flood
defenses (3767 km) in the Netherlands does not comply with the safety standards For this
purpose a new session of investigations has commenced in order to assess the reliability of the
existing dikes and the potential amendment of the safety standards for future constructions
Taking into account this latest information and within the framework of the Flood Risk in the
Netherlands 2 project6 (VNK2) a new Delta program was founded whose one of the main
decisions was the ldquoDelta Safety Decisionrdquo This decision contains a proposal for new safety
standards of the primary dikes as they are depicted in Figure 29 which are stricter than the
previous ones and formulated in terms of failure probability
According to this project a new policy regarding the safety assessment of the flood defenses has
been settled and proposed for the future design criteria which is based on the evaluation of the
acceptable flood risk (probability of failure (x) consequences) of each dike ring rather than the
probability of exceedance in order to achieve a level of protection that is in balance with the
societal value (Cost Benefit Analysis and life loss)
VNK has been already using reliability evaluation techniques in order to calibrate partial safety
factors Especially in geotechnical engineering the high inherent uncertainty of the soil
properties renders the evaluation of the structural reliability essential for the safety assessment
of the structure Subsequently this implies the investigation of the response of the different
reliability methods applied in a real case study of a dike The techniques that are used to
evaluate the probability of failure are discussed thoroughly in Chapter 4 and in Appendix F In
case of a complicated structure such as a dike with a retaining wall both the reliability of the
dike body and the reliability of the wall shall be evaluated and eventually a common standard
for the reliability of the system shall be defined
5 A safety assessment is an evaluation of the function of the flood protection structure and it is being conducted every approximately 6 years after the reassessment of the existing hydraulic boundary conditions 6 in Dutch Veiligheid Nederland in Kaart 2 or VNK2 It has been initiated by the Ministry of Infrastructure and the Environment (IampM) the Association of Regional Water Authorities (UvW) and the Association of Provincial Authorities (IPO)
System description and current design concept
MSc Thesis A Rippi 18
Figure 29 Maximum admissible flooding probability for primary defences according to Delta
programme 2015 (Deltaprogramma 2014)
As far as the retaining walls are concerned CUR (2005) design procedure distinguishes the
following three safety classes for retaining walls with corresponding reliability indexes
Table 21 Safety classes and corresponding reliability indices (CUR 2005)
Class I Relatively simple constructions no personal safety risks and relatively minor
damage in the case of overall failure ϐ=25
Class II Considerable damage in the case of overall failure minor personal safety
risks ϐ=34
Class III Major damage in the case of overall failure andor considerable personal
safety risks ϐ=42
System description and current design concept
MSc Thesis A Rippi 19
Additionally as far as a dike section is concerned the required overall reliability index shall be
concluded according to the new norms that are shown in Figure 29 However the reliability
index of the different failure mechanisms of a dike shall be estimated separately
According to Vergouwe et al (2014) that is part of the VNK2 report the major failure
mechanisms in dike ring 43 where the case study under investigation is located (see Figure 26)
is uplift and piping and macro-instability of the inner slope Especially in Figure 210 the
estimated percentage of the contribution of each failure mechanism to the overall probability of
flooding7 is indicated In Vergouwe et al (2014) the overall flooding probability of the specific
dike section that the case study concerns was evaluated to 17800 Therefore the macro-
instability on which this thesis is concentrated with a contribution percentage of 226 has a
probability of about 310-5 (
) and thus a target reliability index β of
approximately 4 This reliability level can be considered as a benchmark for evaluating the one
that will be entailed in the end of this thesis
Figure 210 Percentage contribution of the different failure mechanisms to the probability of flooding (source Vergouwe et al 2014)
It should be mentioned though that the aforementioned safety standards concerning the dike
ring 43 were defined according to the probability of exceedance of a certain water level that the
dike has to retain Therefore this complies with the old norms and not with those appeared in
Figure 29 However for large engineered systems such as flood defence systems with large
potential consequences and substantial investments it is worthwhile to assign target reliability
levels based on a risk assessment of the area surrounded by the specific dike ring There are
three widely used types of criteria for evaluating the risks related to floods and major industrial
hazards (Vrijling et al 2011)
Economic criteria
Individual Risk Criteria
Societal Risk criteria
7 The dependencies among the failure mechanisms were not taken into account Therefore this probability stated in Vergouwe et al (2014) is not exactly equal to the probability of flooding
System description and current design concept
MSc Thesis A Rippi 20
In order to check then if a certain system abides by the target reliability that is required the
failure probability of each failure mechanism shall be calculated based on a LSF The
aggregation of the failure probabilities of all the modes gives the failure probability of the
system which is supposed to be compared with the required in order to reassure the reliability
of the structure Therefore this thesis is dealing with the calculation of this failure probability of
a certain system modelled in FEM analysis and understanding of the systemrsquos behaviour and
finally the investigation of the robustness of coupling FEM with reliability methods for
evaluating the reliability of a structure In the next chapter an overview over the research has
been done until now concerning coupling and reliability methods applications is presented and
a discussion over the most important is made
MSc Thesis A Rippi 21
3 Literature study
In this chapter an overview of some principles related to FEM and the uncertainties in
geotechnical engineering is presented More precisely an introduction into FEM concept and a
discussion over the different types of FEM is made Moreover the uncertainties in geotechnical
engineering are stressed and the general framework according to which they are handled is
described Finally previous studies that are associated with the application of FEM on soil
structures and the reliability analysis are discussed
31 Background
311 Finite Element Modeling
FEM is a numerical method whose essence is to convert a problem described by partial
differential equations over space and time into one by dividing the space-time continuum into a
set of discrete elements and assuming that the unknowns vary over each element FEM solution
process is as follows
1 Divide structure into pieces (elements with nodes) (discretizationmeshing)
2 Connect the elements at the nodes to form an approximate system of equations for the
whole structure (forming element matrices)
3 Solve the system of equations involving unknown quantities at the nodes (eg
displacements)
4 Calculate desired quantities (ie strains and stresses) at selected elements
The properties of each element are set the same as the material properties that have been
defined by the user Then the Deterministic Finite Element Method (the finite element method
that was formulated with deterministic variables) can be used in conjunction with means and
standard deviations of the input variables to obtain reliability estimates
In the simple coupling of FEM with reliability analysis each parameter that is considered as
stochastic is given a particular probability density function estimated either by field tests or by
engineering judgement This approach is also called ldquosingle random variablerdquo and it assumes
that the spatial correlation length is infinite In other words the soil is considered to be
homogeneous and the stochastic property assigned to the soil is taken at random from a certain
probability distribution
Nevertheless a more realistic model should take into account the spatial correlation within
smaller regions where the stochastic property is allowed to vary For that purpose the Random
Finite Element Method (RFEM) was introduced by Fenton and Vanmarcke (1990) in which the
random variables are correlated to one another using auto-correlation functions
The most sophisticated method of FEM is the Stochastic Finite Element Method introduced by
Ghanem and Spanos (1991) which accounts for the uncertainties in the input parameters
Literature study
MSc Thesis A Rippi 22
implicitly within the finite element calculation This aims at representing the complete response
probability distribution function (PDF) in an intrinsic way Two steps are basically applied for
that purpose
Discretization of the problem and definition of the random vector with an unknown
joint PDF and
Expansion of the response in a particular basis of random vectors with a finite variance
called the polynomial chaos
In this thesis the Deterministic Finite Element Method is to be used coupled with reliability
analysis According to Waarts (2000) coupling FEM with reliability methods (FERM) would lead
to the following advantages
In comparison to standard finite element analysis it gives direct insight into the
structural reliability and decisive parameters
Structures designed using FERM will either be safer orand more economically built in
comparison to structures designed using safety factors and classical constitutive models
FERM is expected to be useful for the determination of LSF that cannot be explicitly formulated
and that differ in each case such as soil limit state in different structural schematizations
Moreover it is likely to be valuable in areas where little knowledge exists on the systems
reliability of structures with multiple components (ie a dike with a sheet pile wall)
FEM in practice
The conventional method for stability analysis in a soil body is represented by LEM although
FEM is increasingly used by designersresearchers The latter has been proved to be quite
realistic for the progressive behaviour (ie stress-strain development in different construction
phases) of a soil system under the effect of stress redistribution in comparison with classical
models Especially in their master thesis Johansson amp Sandeman (2014) compared the
deformations and the forces measured at a deep excavation supported by anchored sheet pile
wall in a railway tunnel located in Gothenburg in Sweden with 1D finite element software the
2D finite element software (Plaxis) and hand calculations They proved that Plaxis produces
reliable results for horizontal deformations in the sheet pile wall and anchor forces when
compared to in-situ measurements
Moreover in Gonzaacutelez et al (2013) it was proved that the simplification of reality done by the
classical methods such as Blumrsquos Engelrsquos Kreyrsquos methods etc (for further information for these
methods a reference is made to Gonzaacutelez et al (2013)) allows us to generally understand the
behaviour of the system wall-soil Nonetheless the results that came out of this analysis were
found to be quite conservative whereas FEM managed to give a more realistic interpretation of
the wallrsquos movement
Seed et al (2006) investigated the performance of the New Orleans Flood Protection Systems
during hurricane Katrina in August 2005 by simulating the different levee sections with FEM As
an example in Figure 31 pictures of the failure of the Lower Ninth ward IHNC East Bank in
New Orleans after the hurricane Katrina are presented FEM was proved to be an efficient tool
to realistically interpret the shape and the triggers of the failure mechanism
Literature study
MSc Thesis A Rippi 23
Figure 31 Failure of the Lower Ninth ward IHNC East Bank in New Orleans modelled in FEM with Plaxis (source Seed et al 2006) At the bottom right the shear strain contours are depicted that are developed through the embankment and at the bottom left the deformed mesh is presented illustrating the failure mode of the levee
In recent implementation of design concepts and technical recommendations such as CUR 166
ndash Damwandconstructies (2005) which refers to the design of retaining walls it can be observed
that guidelines for Finite Element Analysis (FEA) have been set parallel to the existing analytical
or empirical calculation methods Therefore FEA tends to become more and more accepted as
an alternative for Serviceability Limit State (SLS) as well as for ULS design
According to Wyllie and Mah (2004) LEM give an estimation of the factor of safety with no
information on deformation of the slope However in numerical analysis the failure surface can
evolve during the calculation in a way that is representative of the natural evolution of the
physical failure plane in the slope In that way a better insight into the evolution of failure
mechanisms can be gained
Cundall (2002) compared the characteristics of numerical solutions and LEM in solving the
factor of safety of slopes and concluded that continuum mechanics-based numerical methods ie
FEM have the following advantages
No pre-defined slip surface is needed
The slip surface can be of any shape (in contrast with Bishop and Kranz stability)
Multiple failure surfaces are possible
No static assumptions are needed
Literature study
MSc Thesis A Rippi 24
Structures (such as footings embankments etc) and structural elements (such as
beams cables sheet piles etc) and interface can be included without concern about
compatibility
It is also important to recognize that LEM only identifies the onset of failure whereas FEM
includes the effect of stress redistribution and progressive failure after its initiation Numerical
models can also be used to determine the factor of safety of a slope in which a number of failure
mechanisms can exist simultaneously or where the mechanism of failure may change as
progressive failure occurs (Hoek et al 2000)
On the other hand the accuracy of FEM is dependent on usersrsquo settings such as the refinement
of the mesh generation the loading steps etc Moreover even if FEM can be functional and easy
to use it is essential that the user has a deep knowledge of the subject under investigation (soil
and structural mechanics) in order to be sceptical and critical with the FEM results able to
interpret the output behaviour of the structure and changecorrect everything that seems
peculiar Moreover the usage of FEM implies also some basic knowledge on numerical methods
and the general background of the FEM so as to solve possible numerical errors
An ideal way of modelling in FEM would be a prior calibration of the program according to data
related with the structurersquos properties and performance (ie stress generation) However this
means that enough field measurements should be carried out and under specific conditions in
order to be considered as a base for FEM calibration Unfortunately such field tests are not
always available and if so their reliability is on doubt For example measurements related with
the developed moments or displacements of a retaining wall inside a dike section are rare to be
found but even if there are some available they cannot represent the situation near the failure
domain where most of the engineers are worried about
312 Uncertainties and Sensitivity analysis
Uncertainty in geotechnical engineering can be categorized into aleatoric epistemic and
decision model uncertainty as they are depicted in Figure 32 (Baecher amp Christian (2003))
Aleatoric uncertainty is associated with the natural variability which is defined as the intrinsic
randomness of natural processes Such variability can be expressed by changes of the soil
properties over time at a certain location (temporal variability) or over space at a single time
(spatial variability) Epistemic uncertainty can be divided into the site characterization model
and parameters uncertainty and it is attributed to lack of information about events or lack of
understanding the physical laws that prohibits the ability of creating a realistic model Finally
the decision uncertainty describes the difficulty of being aware of social objectives defining
social values such as discount rates and predicting the planning horizon
Literature study
MSc Thesis A Rippi 25
Figure 32 Categories of uncertainty in geotechnical engineering (modified from Baecher amp
Christian (2003))
Below the general steps of an uncertainty study are described and an introduction into the
sensitivity analysis concept is made
Global methodology of an uncertainty study
A first step of an uncertainty study8 can be described as ldquothe definition of the problemrdquo Initially
the variables of interest (or else the output variables) of which the uncertainty is to be
quantified shall be specified In sequence given several input variables for which the user may
have data andor expertengineering judgment a model denoted usually by a mathematical
function should be introduced that enables the computation of the set variable of interest
After the general context has been staged we should choose the criteria with which the
uncertainty can be evaluated The most complete measure of uncertainty when dealing with a
random vector is the probability distribution In order to assess the value of this distribution
function the following criteria can be followed
Probability of exceeding a threshold the aim is to assess the probability that the variable
of interest exceeds a threshold important for the goals at stake
Quantiles the aim is to assess the threshold that a variable of interest may exceed with a
probability equal to a given value
Central dispersion the aim is to obtain an ldquoorder of magnituderdquo of the uncertainty by
specifying the average value and the variance of a variable of interest
8 In the engineering world this is also called reliability analysis However in this section the general uncertainty study is described In particular reliability describes the ability of a system or component to function under stated conditions for a specified time period
Literature study
MSc Thesis A Rippi 26
The next step is to define a model to represent and quantify the uncertainties of the input
variables One shall investigate each variable as a singularity and come up with the most
suitable probability density function (mostly depending on available data) Besides it is
essential to assess also the potential correlations among the variables that can be included in a
joint probability density function which is discussed later in this section
Eventually the uncertainties of the input variables shall be translated in terms of uncertainty on
the variables of interest This procedure is called uncertainty propagation and can be carried
out via several reliability methods (approximation methods or sampling methods) that are
extensively elaborated in Chapter 4 In Figure 33 a schematization of the different steps during
an uncertainty study is given
Figure 33 Methodology of uncertainty study
Last but not least a better understanding of uncertaintiesrsquo influence can be achieved by
analyzing the contribution of the different uncertainty sources to the uncertainty of the
variables of interest via a sensitivity analysis Such an analysis aims at identifying important
parameters for the system response besides it gives a better insight into the model used to
quantify the uncertainties Before conducting an uncertainty analysis it is advisable to filter out
parameters of less significance in order to reduce the modeling computational effort Below
different methods for sensitivity analysis are described
Sensitivity analysis
Sensitivity analysis (SA) is the study of how the variation in the output of a model can be
apportioned quantitatively and qualitatively to the variation in the model parameters (Saltelli
et al 2004) Saltelli (2004) proposes one possible way of grouping these methods into three
classes (i) screening methods (ii) global SA methods and (iii) local SA methods
i Screening methods
Screening is a particular instance of a sampling-based method The objective here is to identify
which input variables are contributing significantly to the output uncertainty in high-
dimensionality models rather than exactly quantifying sensitivity (ie in terms of variance)
Screening tends to have a relatively low computational cost when compared to other
Literature study
MSc Thesis A Rippi 27
approaches and can be used in a preliminary analysis to weed out less influential variables
before applying a more informative analysis to the remaining set One of the most commonly
used screening methods is the elementary effect method
ii Global SA methods
Global SA techniques incorporate the whole range of variation and the probability density
function of the input parameters to calculate their influence on the output Many global
sensitivity analysis techniques are now available such as
Variance-based methods
o Sobolrsquos method9
o Fourier Amplitude Sensitivity Test (FAST)10
o Analysis of Covariance (ANCOVA)
Regression-based methods11
Both FAST and Sobolrsquos method rely on the assumption of parameter independence while
ANCOVA can also handle correlated input parameters The main principles of FAST method are
presented in Appendix A1 For more information about the other methods the reader can
betake himself to the related references
iii Local SA methods
Local SA methods provide the slope of the calculated model output in the parameter space at a
given set of values (Turanyi and Rabitz 2004) This basically means that local methods involve
taking the partial derivative of the output Y with respect to an input factor Xi |
|
where the
subscript indicates that the derivative is taken at some point in the space of the input
Examples for these are (Schweckendiek 2006)
the α-values in a FORM calculation (further discussion in section 422)
Local Probabilistic Sensitivity Measure
In reliability analysis the local sensitivities will be more important than the global ones in the
sense that the influence of all variables in specific points such as the design point cannot be
identified by the global methods However local SA can only inspect one point at a time and the
sensitivity index of a specific parameter is dependent on the central values of the other
parameters
In this thesis the influence of the different parameters on the response of the system is
evaluated according to local sensitivity indices However before the main part of the reliability
analysis starts it was necessary to filter out the less influencing variables in order to reduce the
number of the major variables and make the analysis more efficient and computationally
affordable For this purpose the global sensitivity method FAST was used whose results are
presented in Chapter 7
9 created by Sobolrsquo (1993) 10 introduced by Cukier et al (1973) 11 introduced by Helton et al (2005)
Literature study
MSc Thesis A Rippi 28
Probabilities as a measure of uncertainties
Most engineers today use the concept of probabilities as the standard way to talk about
uncertainty One way to quantify a probability of a random variable is the calculation of the
cumulative probability function (CDF) of non-exceedance which can be obtained by
the probability distribution function For a random vector this reads
( ) (11)
From this the joint probability density function of this random vector can be determined as
( )
( )
(12)
This function is depicted in Figure 34 by means of contour levels The joint probability density
function has the shape of a hill as it is showed in the 3D figure and the inclination of the ellipses
reveals a correlation between the two variables and
Figure 34 Joint probability density function in 2D and 3D
The functions that join or couple multivariate distribution functions to their one-dimensional
marginal distribution functions are called copulas These are going to be used in order to define
the joint density probability distribution from which the random variables are taking their
values However no detailed explanation is made in this report regarding copulas and its
characteristics as it is automatically implemented in OT whereas the user has to define only the
type of copula that heshe considers the most appropriate (a reference is made to Nelson (1999)
for further information) According to the type of copula that is selected the order of the values
of the variables can be chosen within an iterative reliability method In this thesis an
independent copula was used considering that the input variables are independent the ones
from the others due to the time limit
Literature study
MSc Thesis A Rippi 29
32 Previous Studies In this section a summary of the most relevant researches on the topic of FEM applied in soil
structures and coupling with reliability methods is given
To begin with an attempt to estimate the reliability of a structure modeled in FEM has been
made by Waarts (2000) by introducing an optimized reliability method in terms of
computational effort and efficiency
In particular in Waarts (2000) two adopted reliability methods are introduced both making
use of a response surface (a detailed explanation of the reliability methods and the response
surface techniques is made in Chapter 4 and Appendix F) These adaptive response surfaces are
used in combination with FORM and DS respectively The accuracy and the effectiveness of
these methods are investigated on the basis of artificial LSFs and a comparison is made with the
existing standard reliability methods The most efficient combinations of response surface
techniques and reliability methods were with FORM (FORM-ARS) and DS (DARS) Comparing
these two methods DARS predominated over FORM-ARS as it can cope with a much wider
range of limit state functions
In a later stage the above best performing reliability method (ie DARS) is further investigated
in terms of its efficiency on the basis of complex structures reliability In Figure 35 a couple of
case studies used to verify the performance of DARS are given
Figure 35 Left Two-story two-bay frame structure Right overview of a sheet pile installation
(source Waarts 2000)
The conclusions of this research showed that DARS serves its purpose and that the benefit from
using it increases with increasing number of random variables In Figure 36 the Limit State
Function Evaluations (LSFE) that are carried out as a function of the number of variables are
presented
Literature study
MSc Thesis A Rippi 30
Figure 36 Number of LSFE as a function of the number of variables n The upper line depicts the
required number of LSFE N=160 (it is supposed that on average 3 LSFE per sample are required
and consequently the number of LSFE equals 3x160n=480n) The lower line shows the
performance of DARS (source Waarts 2000)
More recently a methodology for assessing the reliability of geotechnical structures and in
particular deep excavations has been developed by Schweckendiek (2006) More precisely this
study deals with the application of different reliability methods in combination with FEM which
carries out the LSFE The reliability methods are applied initially on simple examples in order to
be calibrated and eventually the most efficient methods are used for the reliability assessment
of a deep excavation with a sheet pile wall The several failure mechanisms are treated in detail
concerning the structural elements of the excavation as well as the soil medium
Finally the failure probability of the system is estimated according to a fault tree
schematization of a deep excavation with one layer anchored sheet pile wall Uncertainties in
the soil properties the phreatic levels and also the corrosion of the sheet pile wall were taken
into account In Figure 37 a picture of the case study is shown together with a FEA illustration
and the importance factors of different soil properties in terms of the probability of failure after
a FORM analysis
Literature study
MSc Thesis A Rippi 31
Figure 37 Finite Element simulation and contribution of several structural and soil properties in
the failure mechanism of an anchored retaining wall in layered soil (source Schweckendiek 2006)
For a similar case study as in Schweckendiek (2006) a cantilever and an anchored sheet pile
wall the safety of the structure based on LEM is evaluated in a research conducted by Gonzaacutelez
et al (2013) Moreover in order to compare the efficiency of the classical methods with FEM
the finite element software Plaxis has been used This research shows the limitation of the
classical models and the efficiency of FEM as far as an anchored sheet pile wall in concerned
More precisely LEM results found to be conservative as they do not take into account the
confinement around the free length of the anchor rod that increases the passive pressure After
comparing numerical results with the classical methods in anchored walls the failure
mechanism given by the FEM suggests an intermediate failure state where the sheet pile wall
describes a translation movement on the deep zone and at the same time a rotation movement
around the anchor point that is closer to the reality
Furthermore Moumlllmann and Vermeer (2008) conducted a reliability analysis of a dike failure A
case study at river Elbe in Saxony Germany has been used and the failure probabilities of
different dike cross sections involving different failure modes were compared More precisely
overflowwave overtopping upliftpiping slope stability and damage of the revetment on the
waterside were taken into account For each failure mode the particular failure probability was
determined using the software package PC-Ring (Vrouwenvelder et al 1999) while the
reliability method used for the assessment of the failure probability was FORM That was
coupled with a slope stability software where Bishoprsquos approach was performed
However this method is limited to circular slip surfaces and prescribed pore pressure
distributions within the dike In order to overcome this limitation FEM was used for further
analysis In combination with FEM an adaptive response surface technique was used in order to
represent the limit state function of each failure mode In Figure 38 the simulated structure
Literature study
MSc Thesis A Rippi 32
together with the response surface of the slope instability failure mode is showed Note that this
structure is approached with a single soil layer (same soil within and under the dike) and that
no structural element is implemented inside the dike
Figure 38 Dike model for the probabilistic FEA of dike stability and the corresponding response
surface of the limit state function for 3 stochastic input parameters (source Moumlllmann and
Vermeer 2008)
The failure probability of each of the aforementioned failure mechanisms were finally calculated
and compared with those recommended in each case while the performance of the coupling
between FEM and response surface method was assessed
Additionally a more recent investigation in terms of dikes with sheet piles employing FEA was
carried out by Breedeveld (2011) The main objective of this study was to display the
implementation of partial factors for design purposes using FEM and especially the software
Plaxis in simulating dikes with a sheet pile wall The dike was firstly simulated as a single
structure without reinforcement while in sequence the occurring stresses and pore pressures
were calculated with an existing sheet pile wall The results of the distribution of the effective
stresses within and below the dike are illustrated in Figure 39 His case study has been
introduced in Chapter 2 while in Chapter 6 a more detailed explanation of how it was modelled
in FEM is given This case study will be used as an example case in this thesis
(a)
Literature study
MSc Thesis A Rippi 33
(b)
Figure 39 Effective stress distribution before (a) and after (b) the installation of the sheet pile wall (source Breedeveld 2011)
Last but not least Bach (2014) has been occupied with the reliability analysis of bored piles and
the case study that he examined is presented in Figure 310 The major objectives of this
research were (1) Propose models to calibrate resistance factors for the Load and Resistance
Factor Design (LRFD) (see section 423 for further explanation of this method) approach and
find a suitable model aiming to directly determine reliability of a bored pile considering some
types of defect that may occur in the bored pile (2) Select a quality control method and evaluate
its reliability when applied to bored piles
For that purpose he made a coupling calculation between the finite element software Plaxis
(version 90) and the numerical probabilistic toolbox Prob2B in order to design bored pile
foundations in light of the reliability-based design Two failure modes the geotechnical failure
mode and the structural failure mode were proposed in order to comprehensively assess the
reliability of an axially loaded pile The parameter uncertainty was considered through the use
of statistical parameters and probability distributions for material parameters in soil models
The soil parameters were treated as random variables The geometry parameters of pile were
used as deterministic quantities because a change in pile shape in the calculation process
requires establishing a new mesh which is now impossible with regard to the Plaxis software
Figure 310 Nine calculation cases considered with different pile toe levels (source Bach 2014)
Literature study
MSc Thesis A Rippi 34
For the reliability of the pile the ULS of the pile was investigated and finally a displacement
criterion of a percentage of the pile diameter used depending on the soil type and
characteristics that the structure is founded on
33 Overview In this chapter concepts that are related to FEM probabilities uncertainty and sensitivity
analysis were discussed The introduction of FEM applications and the principles of how to
handle uncertainties will help the reader to better understand and follow the process of this
thesis
Furthermore preliminary researches related to several aspects of this thesis are presented that
mainly concern
Reliability methods and their applications [Waarts 2000 Schweckendiek 2006 Bach
2014]
Coupling FEM with reliability methods [Waarts 2000 Schweckendiek 2006 Bach
2014]
(Reliability) Analysis of geotechnical structures and especially dikes reinforced with
retaining walls [Waarts 2000 Schweckendiek 2006 Moumlllmann and Vermeer 2008
Breedeveld 2011 Gonzaacutelez et al 2013 Bach 2014]
As someone might have noticed from the overview of the existing research on the topic both the
performance of different reliability methods and the simulation of a dike with a sheet pile wall
in FEM have been carried out These references were quite helpful in order to get an idea of
coupling FEM with reliability methods as well as FEM and reliability methods individually
Moreover a first dive into the system under investigation and the variables to be handled was
made However noone has analysed the reliability of such a complicated system that is more
and more often seemed to be applied in the Netherlands Thus it would be essential to evaluate
the implementation and the results of the reliability analysis of a dike with sheet piles simulated
in FEM The conclusions of such a research are going to contribute to design and optimization
concepts and hopefully to a better understanding of the system behaviour
MSc Thesis A Rippi 35
4 Structural Reliability Analysis
The aim of this chapter is to introduce the basic theory of a reliability analysis and to summarize
the principles of different reliability methods that are widely used for the uncertainty
evaluation of a system Particularly three of them (MC DS and FORM)12 were used in this thesis
and are described in this chapter Finally a brief evaluation of the selected reliability methods
and the way that they are applied on this thesis by coupling them with FEM are discussed
41 Basics of Reliability Analysis In the consideration of the reliability of an element the determination of the probability of
failure is the central issue The limit between failure and non-failure is defined as a limit state
and the reliability is the probability that this limit state is not exceeded The limit states are
interpreted through the so-called limit state functions (LSF) whose general form is
(41)
in which R is the strength or more general the resistance to failure and S is the load or that
which is conducive to failure (ldquosolicitationrdquo) The basic principle of structural design is that the
resistance needs to be higher than the load or in other words that the LSF is larger than zero
(Zgt0) The main objective of the design is to ensure that this performance criterion is valid
throughout the lifetime of a structure Nevertheless the majority of the quantities that both
resistance and load consists of are uncertain Therefore a probability of satisfying the preceding
criterion is estimated and expresses the reliability of the structure P(Zgt0) The probability of
failure is
(42)
Hence the probability of failure is complement to the reliability to the sense that
(43)
The LSF is plotted in Figure 41 indicating the failure space and the limit state function by Z=0
12 Besides RS techniques were used for proceeding with the analysis but most for solving numerical problems rather than evaluating the failure probability The RS techniques are still under investigation and need to be tested before being applied in real cases However it is highly possible that this would be a potential way of evaluating the failure probability in the future
Structural Reliability Analysis
MSc Thesis A Rippi 36
Figure 41 Limit Sate function
In this case the design concept is based on the evaluation of the so-called design point which is
the point in the failure space with the greatest probability density Generally this point is
located on the border between the safe and the unsafe areas
In the structural domain the Joint committee on structural safety (1981) proposed a level-
classification of the calculation methods This classification includes the following three levels
Level III these methods calculate the probability of failure by considering the probability
density functions of all strength and load variables The reliability of an element is linked
directly to the probability of failure
Level II this level comprises a number of methods for determining the probability of
failure and thus the reliability It entails a linearization of the LSF in a carefully selected
point These methods approximate the probability distribution of each variable by a
standard normal distribution
Level I at this level no failure probabilities are calculated The level I calculation is a
design method according to the standards which consider an element sufficiently reliable
if a certain margin is present between the representative values of the strength and the
loads This margin is created by taking the so-called partial safety factors into account in
the design
In the next section the different reliability methods to be used in this thesis are further
elaborated
42 Overview of Reliability Analysis Methods
421 Level III Methods
The foundation of the Level III failure probability calculation is the mathematical formulation of
the subset of the probability space which involves failure (Zlt0) Level III reliability methods
Structural Reliability Analysis
MSc Thesis A Rippi 37
(also known as fully probabilistic methods) compute the probability of failure based on the
exact probability density function and the exact limit state functions Therefore such methods
are considered to be the most accurate Well-known sampling methods are (Quasi-) MC
simulation Importance Simulation DS and the Latin Hybercube Simulation Below the former is
discussed as it represents the fundamental theory on which the sampling methods are based on
while DS is also further explained as it is continuously used in this thesis For the rest a brief
explanation is provided in Appendix F There are also other sampling methods such as the
Numerical and the Directional integration which will not be described in this study However
for further information about them a reference though is made to Waarts (2000) In Level III
methods errors can only occur by a too low number of simulations
Crude Monte Carlo Simulation
The MC method uses the possibility of drawing random numbers from a uniform probability
density function between zero and one If and are the
vectors for the resistance and the load respectively that consists of n variables then by taking
multiple realizations of the uniform probability distribution a value can be determined for
every and
By repeating this procedure a large number of times the failure probability can be estimated
(44)
where is the total number of simulations (nm draws from the uniform distribution in which
m is the number of base variables) and is the number of simulations for which Zlt0
The simulationrsquos relative error is
(45)
If the expected value of the relative error is zero the standard deviation is
radic
(46)
Based on the central limit theorem the error is normally distributed provided is sufficiently
large The probability that the relative error is smaller than the given value E is then
(47)
Thus for the reliability of the relative error is smaller than For the wanted k and
E the required number of simulations can be determined with
Structural Reliability Analysis
MSc Thesis A Rippi 38
(48)
If for instance a reliability of 95 is required for a maximum relative error E=01 the required
number of simulations amounts to
(49)
The number of simulations is therefore still dependent on the probability of failure
MC simulation is applicable whatever the complexity of the deterministic model is However the
number of samples required to get an accurate estimation of may be dissuasive especially
when the value of is small (10-k) Thus crude MC is not applicable when small values of are
sought andor when the CPU cost of each run of the model is non-negligible
In Appendix F and especially in sections F1 and F2 a further elaboration of the MC simulation
is presented Moreover the methods of Quasi-MC Importance sampling and Latin Hybercube
are described
Directional Sampling
The directional simulation method is an accelerated sampling method It implies a preliminary
iso-probabilistic transformation as for FORM and SORM methods however it remains based on
sampling and is thus not an approximation method Below the method as it is described in Open
TURNS 15 Reference Guide (2015 pp190-193) is presented
Each simulation of the directional simulation algorithm is made of three steps Using the
probability distribution of a random vector we seek to evaluate the following probability
( ) (410)
Here is a deterministic vector and ( ) the limit state function which enables the
definition of the event For the iteration these steps are the
following
Let A point is drawn randomly on S according to a uniform
distribution
In the direction starting from the origin and the passing through solutions of the
equation ( ) (ie limits of ) The set of values of that belong to is
deduced for these solutions it is a subset of
Then one calculates the probability By property of independent
standard variable
is a random variable distributed according to chi-square
distribution which makes the computation effortless
Structural Reliability Analysis
MSc Thesis A Rippi 39
Finally the estimate of the probability after N simulations is the following
sum
(411)
The main idea is that for each direction we go along the direction by step of a specified length
from the origin to the maximum distant point and we check if there is a sign change on each
segment so formed (see Figure 42) We go until the maximum distant point Then for all the
segments where a sign change is detected we research the root λ in the segment (there are
also other strategies available in OT however all of them follow the same principle of
searching) The following figure illustrates the principle of an iteration in dimension 2
Figure 42 DS of a 2-dimensional problem
Similar to MC method the outcome of the estimated probability of failure is a random variable
and the error in the estimate can be made as small as possible by taking a sufficient number of
samples For DS the standard deviation σ of the estimated failure probability can be quantified
as follows (Melchers 2002)
radic
sum
(412)
From this equation relative errors and the confidence intervals can be estimated As can be
seen the error in the estimated probability of failure will decrease with increasing number of
sampled directions Equation (412) can be used to determine the number of sampled directions
that is required for a reliable estimate of the failure probability The convergence criterion in
this method is usually the coefficient of variation (CoV) which is defined as
Structural Reliability Analysis
MSc Thesis A Rippi 40
422 Level II Methods
Level II methods (known also as fully probabilistic methods with approximations) can take all
the probabilistic properties of the random variables into account but they include
approximations of the limit state function and therefore their use and outcomes should be
inspected and evaluated in order to be considered as reliable However experience in that
methods has shown that the computational effort is profoundly decreasing in comparison with
Level III methods and that the application of them can provide important parameters such as
the reliability index (β) and the influence factors (α) Some of the most known Level II methods
are First Order Second Moment (FOSM) Method First and Second Order Reliability Methods
(FORM and SORM) and the Point Estimate method (PEM) In the next paragraphs FORM is
elaborated as it will be mainly used in this thesis while in Appendix F3 the principles of FOSM
are discussed
First Order Reliability Method (FORM)
The principles of FORM are based on the First-Order Second-Moment (FOSM) approach which
first introduced the reliability index concept This method is presented in Appendix F3 in detail
However the reliability index estimated using the FOSM approach is not ldquoinvariantrdquo which
means that the index value depends on the limit state function format (Farrokh 2007) Hasofer
and Lind (1974) proposed an invariant definition of the reliability index that leads to a
geometric interpretation known as first-order reliability method (FORM)13 The first step of this
method is the transformation of the random variables to equivalent standard normally
distributed variables and the whole procedure is carried out in u-space (or else standard space)
For variables that are normally distributed this step is as follows
(413)
For other types of distributions there are procedures available for carrying out this
transformation such as Generalised Nataf and Rosenblatt transformations In this study these
transformations are not discussed however for further explanation a reference is made to the
Reference Guide of OpenTURNS 15 (2015)
In the sequence the limit state function Z is expressed in terms of
The second step is the approximation of the function with the first two terms of the Taylor-
polynomial The approximation reads (CUR 1997)
( ) ( ) sum
(414)
where is the point where the linearization takes place This approximation of Z is linear and
according to the central limit theorem it is normally distributed The expected value of the LSF
can be approximated by the expected value of the linearized function
13 VNK2 uses FORM in view of the calculation time required In case of inaccuracies VNK2 switches to more time-consuming techniques such as DS whereby a large number of draws are made from the probability density functions of the various stochastic variables (uncertain quantities)
Structural Reliability Analysis
MSc Thesis A Rippi 41
( ) sum
(415)
While the standard deviation (sum of contributions of each variable to the variance of Z) is
defined as
radicsum
( )
(416)
Using the definition of FOSM for the reliability index [ ]
where [ ] is the mean value of
the limit state function and substituting Eq 415 and Eq 416 in it the reliability index can be
approximated However linearization in different points can lead to different values of the
reliability index Hasofer and Lind definition of the reliability index overcomes this limitation
and renders it equal to the minimum distance from the origin to the design point
(417)
where radic
Looking for the design point is basically an optimization problem Many analytical and
numerical approaches can be used for that purpose A relatively straightforward method to do
this is by firstly assuming that the design point is the mean value (the starting point can also be
another point according to an engineering judgement in order to accelerate the optimization
procedure) The obtained β-value is used to determine a new point in which the LSF is
linearized In this case the importance factors αi are calculated as
( )
radicsum (
( ) )
(418)
where is the design point and is the number of variables expresses the
contribution of the variance of each variable to the total variance of Z in the design point Figure
43 illustrates the linear approximation of the limit state function and the aforementioned
parameters for a two dimensional problem The probability that Zlt0 can be determined using
the standard normal CDF
(
) (419)
Structural Reliability Analysis
MSc Thesis A Rippi 42
The new calculation point is determined by
(420)
After some iterations the chosen optimization algorithm finds the final design point and the
new reliability index In Appendix A2 an overview of the different optimization algorithms
available in OT is presented together with an evaluation of their performance in the specific
case study
Figure 43 Two-dimensional illustration of u-space LSF and design point
423 Level I Methods (semi-probabilistic)
At the beginning of the probabilistic concepts incorporation in the field of structural
engineering the most notable development was the implementation of LFRD At the past single
factors of safety on the ratio of total resistance to total load were used in order to ensure that
the stresses developed from an applied load on a structure were lower than the allowable ones
LRFD replaces those factors by introducing a set of partial safety factors on the individual
components of load and resistance These partial factors can be selected such as they account
for uncertainties associated with the resistance and the load respectively
The current design philosophy in CUR 166 and Eurocode is characterized by the use of
characteristic values of the parameters (in Appendix E a further explanation of the
characteristic values is given together with the way that they are converted into the mean
values for being used in the reliability analysis later on) The values of the partial safety factors
are additional factors to the characteristic values Therefore the characteristic values are
multiplied with the aforementioned safety factors ( ) and the following criterion has
to be satisfied for a reliable structural performance (see Figure 44)
(421)
Structural Reliability Analysis
MSc Thesis A Rippi 43
The main disadvantage of LRFD is the fact that there is a set of load and resistance factors that
need to be calibrated in order to cover the specific cases that are likely to occur This is because
of the lack of sufficient statistical data that are necessary for this calibration procedure
Consequently it cannot reassure that all the designs to which the set of factors is applied result
in the reliability level that was aimed for As a result in majority of the cases a conservative
calibration that gives the required reliability is chosen which can lead to ldquoover-designrdquo
structures
Figure 44 Design with partial factors for the load and resistance parameters (source
Schweckendiek 2006)
The evaluation of an elementrsquos reliability starts in principle with calculating the probability of
failure and subsequently the reliability for the given strength and load In practise the problem
is often that the strength is unknown but it has to be determined for a given reliability The
determination of the required reliability can be estimated with the help of Level II and III
methods by iteratively adjusting the strength in the calculation until a sufficiently small
probability of failure is found
In the design domain regulations and guidelines follow the standard that the characteristic
value14 of the strength is divided by a factor and that the characteristic value of the load is
multiplied by a factor as it is described in Eq 421
The link between Level I and the estimation of the failure probability has been achieved through
Level II methods The design point that results from a Level II method calculation is the point
with the greatest joint probability density of the strength and the load and it is therefore
possible that when failure occurs the strength and load values will be close to that point The
design values can be specified as
(422)
(423)
14 The characteristic value of a parameter is its main representative value and its value can be specified as a mean value an upper or lower value or a nominal value according to EN 1990 definition (2002)
Structural Reliability Analysis
MSc Thesis A Rippi 44
where are the coefficient of variation for the resistance and the solicitation respectively
15 and are the importance factor and β is the reliability factor These can be also expressed
in terms of characteristic values as follows
(424)
By substituting function 423 with 421 and 422 the partial safety factors are defined as
(425)
(426)
where 16 and are the values for load or resistance respectively to which a probability of
(non-)exceedance of 95 corresponds (non-exceedance for the load and exceedance for the
resistance) (for a standard normal distributed parameter this is 1645)
424 Response Surface Techniques (RS)
In case that the models presented above tend to be time consuming for the limit state function
evaluation or convergence problems of the optimization algorithms occur a better methodology
is recommended known as response surface The RS is mainly used when (unknown) response is
only available from experiments or complex FEM computations (for example large highly-non-
linear FEM models) Indeed once a RS has been built up the various standard methods may be
applied at a negligible cost An analytical limit state function replaces the real response function
The main idea is that the response consisting of a complex function of input variables is
approximated by a simple function of the input variables
A list of possible response surfaces techniques is given below
Linear and quadratic Taylor expansions
Polynomial RS based on least squares
Kriging method
Polynomial chaos expansion
A detailed explanation of the aforementioned methods can be found in OpenTURNS 15 (2015a)
The first method is associated with the approximation of the model response around a specific
set of input parameters while the rest seek a global approximation of the model over its whole
domain of definition The most sophisticated one is the polynomial chaos expansion but one the
15 α values are negative for solicitation (ie RV for which increasing their values lead to adversely effects on the reliability) and positive for resistance (ie RV for which increasing their values implies a beneficial effect on the reliability) 16 is negative and can be negative or positive
Structural Reliability Analysis
MSc Thesis A Rippi 45
other hand its complexity does not make it attractive In this research a polynomial RS is
applied based on least squares
In combination with FEA the standard procedure is as follows
1 Select the most important random variables on the basis of engineering basis
2 A SA is carried out in combination with FEM
3 Reduce the stochastic variables if needed according to the SA
4 A RS is constructed through the response data
5 A reliability calculation is carried out using the RS instead of the real response
A polynomial RS is generally constructed by fitting a quadratic function to the sampling points
whose general expression is
sum
sum
sum sum
(427)
The type of the reliability method to be used is of little importance since the time consuming
LSFE (using FEM) are replaced by analytical expressions However of main importance is now
the accuracy of the RS compared to the real response There are several parameters in order to
evaluate the goodness of fit of the RS to the real model among which is the well-known R-
squared
In Figure 45 an example of a quadratic response surface is showed that is fitted on the data
response indicated with the black circles This was a 6-dimensional problem that means 6
different variables were included in the system (non-visualized dimensions are kept constant at
their mean value) In this thesis the response of the system is to be acquired after the coupling
of FEM with a reliability method and that would give the ldquoblackrdquo circles that are shown in Figure
45 The response surface technique applies a curve fitting on the real response of the system
(ldquoblackrdquo circles) whose analytical formula can later be used for the reliability analysis of similar
type of systems with negligible computation time
Structural Reliability Analysis
MSc Thesis A Rippi 46
Figure 45 Quadratic response surface with 28 terms (ie ) for 6 dimensions ( ie x
=[ x1 hellip x6])
43 Coupling Reliability Analysis with FEM The limit state function evaluations (LSFE) will be carried out with the software Plaxis 2D 2015
which is a special two-dimensional finite element software used to perform deformations and
stability analysis for various types of geotechnical applications Moreover considering the case
study under investigation in this thesis Plaxis in contrast with other FEM such as Abacus
Comsol DIANA etc offers several techniques to realistically simulate structural elements such
as sheet pile walls and anchors and their interaction with soil while the variety of the
constitutive models for the soil body that are available and the ability to include the history of
the construction phases can lead to a better analysis of the systemrsquos behavior in terms of the
stress level and the deformations It is essential at that point to mention that using FEM for this
purpose means that the limit state formulation is implicit and can only be solved numerically
The reliability analysis is carried out through an uncertainty package In this section firstly a
description of the reliability package and its possibilities are given Finally an explanation of the
coupling procedure between the reliability tool and FEM is given together with the calculation
process that was followed for the parameters manipulation
431 The functionality and possibilities of OT
The reliability analysis tool to be utilized in this research is OpenTURNS (OT) which is a
scientific library usable as a Python module dedicated to the treatment of uncertainties and it is
still under development during the work of this thesis Several reliability packages are available
such as OpenEarth (applied in Matlab) Prob2b (applied in Probox software of TNO) and
Probabilistic Toolkit (software of Deltares) Plaxis 2015 features Python connection possibility
and thus coupling was decided to be carried out with an uncertainty package in Python and OT
is the most developed one Moreover until now only OT is an open source package and can be
Structural Reliability Analysis
MSc Thesis A Rippi 47
used easily from anyone as there is a wide community of experienced people supporting with
their knowledge while many related manual reports are already available for starting learning
A list of the current available reliability methods in OT is given in Figure 46 In this figure apart
from the standard methods the alternative method of the Adaptive Response surfaces is
introduced that is mainly used when (unknown) response is only available from complex FE
computations
Figure 46 Reliability methods available in OpenTURNS
In this thesis the methods to be basically applied are the FORM and the DS analysis while MC is
also used mainly for confirming the application of FORM Additionally the RS technique has
been mainly implemented for enhancing the performance of FORM and DS analysis
OT handles 47 types of distributions amongst which the ones used in this research are Normal
Lognormal Uniform and Truncated Normal distribution Regarding the joint distributions that
are available to be selected for random vectors 11 types of copulas are existing in OT amongst
which the most known ones are the Independent the Gumbel and the Normal copula In this
thesis the independent copula has been used during the reliability analysis which means that
the variables were assumed to be independent among each other However it is strongly
advised a further elaboration and research considering correlation matrices for specific soil
parameters
432 Coupling OpenTURNS-Plaxis
The finite element simulation will be carried out in Plaxis 2015 Plaxis is a finite element
software for plane strain and axi-symmetric modelling of soil and rock behaviour Moreover it
supports a fully automatic mesh generation allowing for a virtually infinite number of 6-node
and 15-node elements
Structural Reliability Analysis
MSc Thesis A Rippi 48
The coupling of reliability analysis and FEA requires an interface for the communication
between each other When a reliability tool is coupled with another software program the
reliability program carries out the whole reliability analysis and it uses the other program only
for the evaluation of the limit state function More precisely OT should be able to read and
amend Plaxis output values for important variables such as material parameters pore pressures
generation and stresses development and corresponding deformations inside the dike
Respectively Plaxis has to be also capable of obtaining the new values that has been set by OT
for the variables that are treated as stochastic during an iterative process according to the
reliability assessment In Figure 47 an illustration of the coupling methodology and its function
is shown
Figure 47 Coupling scheme OpenTURNS-Plaxis
In principle an input file is firstly required where the user set the preferable reliability method
to be used the stochastic input parameters and their probability distributions the joint
probability distribution and the corresponding correlation matrix and finally the limit state
function is formed depending on the situation In Appendix D an example of the input files that
were used for soil sheet pile wall and anchor analysis is shown However such input files
should be interpreted so as to be readable by both Plaxis and OT Therefore an input interpreter
was created which is actually a python script that helps OT to start up the reliability analysis
according to the assigned method variables distributions and LSF As it was mentioned before
the evaluation of the limit state function is conducted by Plaxis For that purpose the input
interpreter should be also able to send the next set of input parameters to Plaxis However an
additional means of connecting the interpreter with Plaxis is also needed This is can be
Structural Reliability Analysis
MSc Thesis A Rippi 49
achieved via a Plaxis interface which actually call Plaxis to make the calculations while it also
transfers the required value of the limit state function to input interpreter and this in turn to OT
Eventually the probability of failure is obtained as a model result However it is also essential
that Plaxis simulation procedure converges to the desirable criteria and under the physical
boundary conditions that have been determined Likewise the convergence criteria of the
reliability methods shall be manipulated so as the optimization algorithms to able to converge
efficiently (see Appendix A2 for further explanation)
44 Overview In section 42 a summary of the principles of the main reliability methods was made From
Level III methods MC and DS were presented whereas from Level II the basics of a FORM
analysis were introduced Besides the concept of the Level I method was explained Even
though this method was not used in this research for the evaluation of the probability of failure
it was applied in section 643 for the deterministic analysis of the system under investigation
and the preliminary calculations of the dimensions In Appendix F more reliability methods are
introduced and they are further described In the sequence their implementation in conjunction
with Plaxis was discussed through the coupling of FEM with OT For more information
concerning OT special features (ie SA optimization algorithms and probability distributions) a
reference is made to Appendices A and E
Before any of these reliability methods was applied on the case study of this thesis their
performance and their compatibility with Plaxis were tested with simple examples At first the
probability of failure was evaluated with FORM SORM Importance Sampling (IS) crude MC and
DS for simple artificial linear and non-linear limit state functions
and
respectively where R B and F are the random variables distributed normally
or lognormally The crude MC technique has been applied in order to validate the failure
probability in case the various methods give different results Secondly a flexible circular
footing over a one-layer soil material was used from Plaxis tutorial in order to test the coupling
between Plaxis and the reliability method and reassure the functionality of the interface In this
pilot example the objective was to determine the probability that the settlements of the footing
exceed a certain threshold of settlements and thus The respective
components that were taken into account for the failure mechanism was the cohesion (c) the
friction angle (φ) and the specific weight (γ) of the soil
From a qualitative evaluation of the reliability methods that has been done and according to the
experience gained from the aforementioned applications of some methods valuable conclusions
were drawn in terms of their efficiency More precisely regarding the sampling methods it does
not required previous knowledge of the failure domain apart from the IS method in which a
starting point inside the failure domain shall be provided Moreover the calculation effort
depends on the magnitude of the failure probability and the required accuracy whereas for the
IS it depends also on the selection of the sampling region In Waarts (2000) it has been proved
that DS tends to be more efficient than MC for low dimensional problems as it is indicated in
Figure 48 (number of random variables nlt100)
Structural Reliability Analysis
MSc Thesis A Rippi 50
Figure 48 Required number of samples for MC and DS as a function of the random variables
(source Waarts 2000)
As far as the approximation reliability methods (FORM and SORM) are concerned it is not
necessary to be aware of the failure region in advance Furthermore it has been observed that
the required iterations and the calculation time is quite lower in comparison with the sampling
methods However the accuracy of the method is highly dependent on the shape of the LSF
In Table 41 a summary of the most applied reliability methods evaluation is presented
According to this evaluation it was decided that the methods of FORM and DS as well as the
combination of them with RS techniques are going to be tested and evaluated in terms of their
efficiency and robustness These methods were chosen from both Level II and Level III methods
due to their expected reduced computational time the non-requirement of previous knowledge
about the LSF and their satisfactory performance in similar case studies In Chapter 5 the
different failure mechanisms and the relevant LSF that are going to be utilized in this thesis are
presented whereas in Chapter 7 the results deduced from the reliability analysis are elaborated
In Chapter 7 there is also an indication of which reliability method was used for each systemrsquos
component the selection of which was based on the aforementioned evaluation
Table 41Evaluation of reliability methods
Method Previous knowledge Accuracy Calculation Effort
FORM not required full accuracy for
Gaussian variables and linear LSF
depends on LSF linearity and the number of random
variables
SORM not required exact up to 2nd order
LSF error dependent on the shape of LSF
depends on LSF linearity and the number of random
variables
Structural Reliability Analysis
MSc Thesis A Rippi 51
Crude MC not required can be controlled by convergence criteria
depends on the magnitude of failure probability and the
required accuracy
IS required can be controlled by convergence criteria
depends on the choices made for the sampling region
DS not required can be controlled by convergence criteria
for low dimensional problems (nlt100) DS is more efficient
than MC
Especially RS were created based on a SA sampling in order basically to increase the efficiency
of FROM and DS analysis in case of FEM numerical errors Precisely an analytical formula or a
response surface was needed that can approximately represent the performance of the structure
under investigation and that is called in case of Plaxis errors This formula was essential in
order to keep the reliability analysis running by providing a response value for the LSF when
Plaxis calculation is unsuccessful and thus incapable of returning a result It should be
mentioned that in case a RS is implemented for the reliability analysis Figure 47 can be
reduced to a graph consisting of only the input file OT and the input interpreter as no Plaxis
calculations would be necessary any more The LSFE will then be performed through the
response surface that is provided by the user Due to this limited use of the RS method it will
not be further treated in this thesis
Structural Reliability Analysis
MSc Thesis A Rippi 52
MSc Thesis A Rippi 53
5 Failure Mechanisms and Limit State Functions
The reliability analysis of a structure requires the definition of the different failure modes that
are relevant to the corresponding structural elements In this chapter the possible failure
mechanisms of the systemrsquos elements expressed in LSF are identified and their role in the
system reliability is explained Based on these LSF the reliability of the system components was
evaluated These LSFs focus on the ULS whose exceedance implies failure of the corresponding
component
51 Introduction to the system analysis and the limit states For system reliability analysis all the possible failure mechanisms have to be identified and
summarized in a fault tree In Figures 51 and 52 the fault trees of a dike section without
structural elements and a retaining structure with sheet piles are depicted respectively In this
project the combination of these two separate structures ie a reinforced dike section with an
anchored sheet pile wall is to be simulated and studied in terms of its failure modes
As it is described later in this chapter failure can be expressed in different ways depending on
the structure and stakeholders demands and safety standards This research is mainly focused
on the ULS of failure and Mohr Coulomb was chosen as the soil failure criterion due to its
efficiency on detecting failure Such a failure criterion is actually defining the stress strain
relationship and the gradual weakening of the soil (ie the gradual reduction of the strength
parameters such as the friction angle and the cohesion due to the stresses development until
soil collapses) under a stress condition In Plaxis new version there is the ability of gradually
reducing structural properties such the strength of steel elements However the results
retrieved of such a safety analysis have not been yet investigated and for that reason such
calculations are not included in this project
As far as the dike section is concerned the failure mechanism to be investigated in this project is
the macro instability of the slope rather than the other mechanisms depicted in Figure 51 On
the other hand because of Plaxis limitations to distinguish among the different structural
elements failure mechanisms each element was considered as a singular case for investigating
its failure mode In the next section a distinction is also made between the ULS and the SLS
For a reinforced dike with retaining walls there are basically four classes of structural elements
(see Figures 12 and 21)
∙ Retaining wall (ie sheet piles)
∙ Anchors
∙ Walings
∙ Soil structure (ie dike section)
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 54
For the reliability analysis of each of the elements the probability of failure is to be determined
In the next sections of this chapter the respective LSF that are considered for each class are
further elaborated according to the ULS criterion while in the end the general combined fault
tree is to be schematized
Figure 51 Fault tree of a dike section
Figure 52 Fault tree for retaining structures using anchored sheet piles (source Schweckendiek 2006)
The system failure in this research is considered as a serial system of the anchor sheet pile
wall and soil body LSFs The fault tree of the system failure is presented in Figure 53 together
with the LSF of each component as they are formulated according to sections 522 and 523
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 55
Figure 53 Fault tree of the system failure
It should be mentioned at that point that the probability of failure of the system differs from the
overall probability of flooding which takes into account any potential failure mechanism of the
system combined with the uncertainties of the water elevation More precisely as far as the dike
safety is concerned a probability of failure (where ) under a certain
water level is estimated from the fault tree of Figure 51 which in sequence is multiplied
with the occurrence probability of the corresponding water level in order to estimate the
overall probability of failure for the specific water elevation Then that product is accumulated
over a required range of water levels that can jeopardise the overall stability of the structure in
order to calculate the overall probability of failure or else the probability of flooding This
probability is then compared to the one established from the safety standards so as to reassure
the safety or not of the structure can be estimated as follows
int
sum (51)
In this thesis the probability is to be estimated where is a specific water level
and especially the design water level as it is considered to be the most challenging part of the
procedure described above gives the probability of occurrence of the water level (ie it is
the PDF of ) After setting up the steps and implementing them successfully for the estimation
of the calculation of the overall probability of failure is just a repetition of the same
procedure for more water levels It should be mentioned that in this thesis failure consists only
of the macro-instability failure mode whereas the rest are excluded for the time being A
simplified alternative for the overcoming the integral of Eq51 is to repeat the calculations for
certain water levels and then sum their products instead of integrating the full range of the
them
52 Limit State Functions
521 Serviceability Limit State
The SLS is evaluated in design calculations in order to avoid excessive deformations of a
structure that could lead to the loss of its service requirements and its functionality In some
cases the SLS criteria are used to prevent damage to adjacent structures in the surroundings
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 56
For example concerning the design of a dike section attention should be paid to deformations
caused to buildings that are located in the inland part of the dike in case of an extreme event
In Figure 54 an example of a deformed dike is depicted In this figure some of the potential
locations that excessive deformations might be experienced are illustrated such as
∙ the vertical settlement of top of dike (arrow 1)
∙ the horizontal displacement of the top of the sheet pile (it might also be the middle part
of the sheet pile that ends up with the largest deformation according to the loading
conditions and the specific soil structure different deformations might occur) (arrow 2)
∙ Uplift upstream of the dike (arrow 3) special attention shall be paid to this type of
deformation not only due to the direct effect on inland structures but also as a sign of
developing piping mechanism
(a)
(b)
Figure 54 Figure (a) shows the initial undeformed structure while (b) illustrates the deformed mesh (not to scale) (the plotted length of the anchor does not have any physical meaning) In (b) the possible locations and directions (black arrows) of potentially excessive deformations
According to the current Dutch regulations (Larsen et al 2013) the SLS criteria for designing
reinforced dikes with sheet pile walls require that
∙ the settlement on the top of dike does not exceed 10 cm and
∙ the horizontal displacements of the sheet pile wall does not overcome 10 cm
In this research the SLS criteria and especially the dike settlements were used roughly for the
preliminary determination of the structures characteristics that are discussed in section 643
Dike Anchor
Sheet pile wall
1
2 3
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 57
More precisely displacements at the top of the dike were also taken into account as
complementary to the safety factor in order to determine the required length of the sheet pile
wall As far as the reliability analysis in concerned the LSF were formulated based on the ULS of
the different systemrsquos components that are presented in the next section
522 ULS for Structural Members
In the design process one is most interested in the ULS of a failure mechanism This state
describes the situation wherein the acting extreme loads are just balanced by the strength of the
construction If that limit state is exceeded the construction will lose its functionality and thus
collapse or fail In general most attention is paid to the behaviour of the structure after
completion However during construction there are also periods in which the construction may
fail The different phases of construction are listed in section 642 In the present section the
analytical LSF of the structural elements are given as they are going to be used in the reliability
analysis
Sheet pile wall
The most relevant failure mode for the sheet pile wall is the exceedance of the yield strength
which corresponds to the ultimate steel strength The response of the structure is mainly due to
bending moments and the axial forces (shear forces are considered to be negligible) Where an
axial force is present allowance should be made for its effect on the moment resistance
Figure 55 The water and earth pressures as well as the anchor forces lead to bending moments and axial forces on the sheet pile wall
In Figure 55 an example of the axial forces and the bending moments that can be developed in
the sheet pile wall with one anchor layer are illustrated after an extreme water level loading
Accordingly the maximum stresses on the sheet pile wall are composed of a bending moment
and a normal force component17
[
] (52)
where [kNm] and [kN] are the bending moment and the axial normal force
respectively that depend on the depth level where they are calculated over the sheet pile length
[m3] is the elastic section modulus and [m2] the cross-sectional area of the sheet pile
wall
17 the vertical anchor force component is reducing by its interaction with the soil over depth
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 58
Bending moment and axial force can be variable over the depth and that is why they are
expressed as a function of z-depth FEM has the advantage to take into account second order
effects ie a stiffer structure will experience higher bending moments than a more flexible one
Taking into account the above the limit state function can be formed as the difference between
the maximum developed stress and the yield stress
[
] (53)
where and can be characterized as the load variables while and can be considered
as the resistance variables and are assumed to be constant over depth
Concerning the permissible displacements for the ULS an upper limit for the maximum
horizontal displacements of the top of the sheet pile has been set to 150L where L is the
vertical length of the sheet pile wall (Larsen et al 2013) This requirement was mainly used in
section 643 where the structural properties were determined after the deterministic analysis
Anchors
Anchors are loaded by their reaction to the horizontal loads on the retaining walls The failure of
the anchor element is actually represented by the failure of the steel members of the anchor
(tubes bars cables etc) that are loaded by traction forces
As it is illustrated in Figure 56 the anchor experiences an axial tensile N that is nearly constant
over its length
Figure 56 Axial loading of anchor inside a dike
The elastic behaviour of an anchor involves only a relationship between axial force N and
displacement (elongation) u of the form
[ ] (54)
where EA [kN] is the anchor stiffness consisting of the steel Youngrsquos modulus E [kNm2] and the
anchor cross section A [m2] and L [m] in the length of the anchor
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 59
Similarly to the sheet pile wall the limit state function of the anchor involves the certain yield or
ultimate strength of the steel members and the maximum stress that the anchor experiences
during its loading Consequently the LSF is as following
(55)
where [kN] is the calculated anchor force and [m2] is the cross sectional area of the
anchor (both of them considered to be constant over the depth) It is essential to mention that
the anchor is also subjected to bending moments due to soil settlements (that are implicitly
illustrated via the uniformly distributed load q over the tie rod) that should be taken into
account in order to investigate the displacements of the tie rod itself However in this thesis
only the axial forces on the anchor are considered without taking into account the individual
deformations and its reaction with the surrounding soil
Walings
The waling is the element that transfers the loads from the retaining wall to the anchors (see
Figure 21) The loading of the walings can be schematized as a continuous beam on several
supports as it is depicted in Figure 57
Figure 57 Loading of walings
If Nα [kNm] is the calculated anchor force per m then the developed bending in the support Ms
and in the opening Mo can be approximated as follows
[ ] (56)
(57)
where La is the mutual anchor distance Considering as the design moment of the waling to be
the limit state function can be formulated as follows
Waling Sheet pile Anchor
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 60
(58)
For a conservative design of the waling the limit state function will give the same or lower
failure probability than the anchor itself Therefore it will actually not be necessary to carry out
this analysis The anchor failure will be the determinant mechanism (Schweckendiek 2006)
523 ULS for Soil Failure
In this thesis emphasis is given on the dike global instability which actually consists of several
failure modes Figure 58 indicates the possible failure mechanisms of flood defences The
combination if these can lead to the overall instability of the dike and thus to soil body failure
However Plaxis is not capable of accounting for all of them More precisely Plaxis assumes the
soil to be a continuous body and thus it can model movements in the scale of soil bodies (using a
relative fine mesh in relation to the moving soil body) but not on a soil particle level Plaxis can
simulate the groundwater flow in a soil body and from this someone can deduce input for a
piping analysis (see failure mechanism G in Figure 58) (using other softwaremethods) but as
it has been mentioned above it cannot determine movement of soil particles due to
groundwater flow Moreover Plaxis cannot deal with the flow and waves occurring in ldquoopen
waterrdquo ie water outside the soil in a canal a lake or sea for instance
Figure 58 Possible failure mechanisms of flood defences (source Flood Defences 2015)
Therefore Plaxis can handle the failure modes (C) (D) (E) (J) (K) and (L) of those illustrated in
Figure 57 In this thesis loads due to extreme temperature conditions or ships are not
considered while the settlements are taken into account implicitly though the mechanisms (C)
(D) and (E) These mechanisms are thus investigated in this thesis and they mainly refer to the
macro-instability of the inner and outer slope of the dike or the overall dike body (ie horizontal
sliding (mechanism D))
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 61
For these failure modes soil failure is defined as shear failure In Figure 59 the most relevant
patterns of the macro-instability regarding the shear strength of the soil are illustrated
Applying FEA the most critical failure mode is determined automatically However it is not
always straightforward what the trigger mechanism of the failure was and that is why more
investigation and FEA tests are needed in order to obtain a clear view
(a) Outer slope failure (b) ldquoActive siderdquo failure
(c) ldquoPassive siderdquo failure (d) Overall failure
(e) ldquoKranzrdquo stability
Figure 59 Macro-instability failure modes in a dike with an anchored sheet pile wall
In Figure 59 failure modes (a)-(d) follows Mohr-Coulomb failure criterion In principle the
total stress state inside a dike section consists of the effective stresses and the pore pressures
according to Terzaghi principle
[
] (59)
Mohr-Coulomb failure criterion basically considers as limit for the failure to occur the
maximum shear stress that the soil can withstand This is determined from the friction
angle the cohesion of the soil and the current stress state The drained shear strength
(similarly in for the undrained shear strength) in the soil according to Mohr-Coulomb is defined
as follows
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 62
[
] (510)
where is the effective friction angle In Appendix B1 a detailed explanation of this failure
criterion is attached
Taking into account the above mentioned failure mechanism (a) is mainly triggered by the
increase of the pore pressures in the outer slope which subsequently causes a decrease in the
effective stresses ( ) which at the same time leads to a reduction of the shear strength
(considering Mohr-Coulomb failure criterion) When the shear stresses in the soil exceed the
shear strength a slip plane forms and a soil wedge collapses
Failure modes (b) and (c) are primarily determined from the active and passive effective
stresses respectively As far as the ldquoactive siderdquo failure is concerned displacements of the sheet
pile wall towards the downstream side of the dike lead to development of active stresses
where is the active pressure coefficient During the ldquoactiverdquo failure the
retaining soil is allowed to relax which leads to a decrease of horizontal stresses and
simultaneously to a decrease in shear strength Similarly to failure mode (a) a soil wedge
collapses The result is an increased earth pressure on the wall for which it is not designed
However this type of failure in dikes is not such determinant for the soil and the sheet pile wall
failure because there is always the resistance of the passive side of the dike that keeps the active
side stable enough Such a mechanism is thus more relevant for deep excavations with retaining
walls
In mechanism (c) the passive soil resistance is exceeded by the horizontal loads In this case the
wall moves inland and a wedge of soil compresses The shear strength is larger due to the
deformation of the wall and the horizontal stresses increase stresses where is
the passive pressure coefficient ( ) This failure mechanism usually occurs due to an
underestimation of the sheet pile length or due to the presence of a weak soil layer in the
passive side
Failure mode (d) is actually a combination of (b) and (c) possibly accompanied also with the
development of excess pore pressures as it was discussed in failure mechanism (a)
Moreover too short anchors could cause a soil block to collapse as it is showed in Figure 59(e)
(ldquoKranzrdquo stability)
As it has been already mentioned for mechanism (d) failure modes can be correlated and
combined with each other in order to lead to a final failure state Therefore the common failure
probability would be smaller than the sum of the singular probabilities In this thesis the total
probability of failure of the soil body due to global instability is to be determined This is due to
the fact that the application of FEM is capable of simulating the combination of the failure
modes depicted in Figure 58 and thus the overall failure (related to (C) (D) and (E) modes of
Figure 58) can be evaluated However with the classical engineering approach for the stresses
calculation each one of the failure modes depicted in Figure 59 should be separately evaluated
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 63
Below a description of the available methods to formulate the LSF of the soil failure is given
After the evaluation of these alternatives a selection was made regarding the most suitable LSF
for the current case study
Limit State functions for soil failure
Soil mechanical failure can be analysed in several ways (Schweckendiek 2006)
1 Excessive Deformations
2 φ-c Reduction
3 Relative Shear Resistance
4 Plaxis definition of soil collapse
The possibilities and limitations of the aforementioned methods are briefly discussed below
1 Excessive Deformations
Similarly to the SLS a limit state function can be formulated by deformations that are
unacceptable As it is illustrated in Figure 54(b) there are some points (the top of the dike the
sheet pile top and the inland soil level behind the dike) the displacements of which shall be
limited to the minimum possible during the structurersquos lifetime in order for the structure to
meet the required service standards Therefore the calculated deformations at locations 1 2
and 3 could give a clear indication of failure given the maximum acceptable deformation
In case that there is a limit value for each location then the limit state function can be formed as
follows
[ ] (511)
with this criterion designer can control the developed deformation on the system and improve
the structure so as to limit the displacements to the required level and subsequently increase
the reliability
On the other hand such an approach might exhibit several problems during the reliability
analysis that are listed below
Many failure mechanisms suddenly happen and as a result no significant displacements
are observed before moving very close to failure In sequence this may cause problems
for the iterative procedures of some reliability methods such as FORM and DS
The determination of suitable maximum admissible deformations is not
straightforward They have to be large enough to serve as failure criterion and
simultaneously they shall not be larger than the values that can be calculated within
the limits of equilibrium in FEM-calculations This requires previous knowledge on the
analysed system as well as on the feasibility of FEM-calculations on the specific subject
Last but not least the location of the undesirable displacements in a dike section is also
under investigation as there can be multiple vulnerable spots on the dike that should
be kept under a certain range of deformations in order not to cause a collapse and that
they are not always known in advance
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 64
2 φ-c Reduction
An available type of calculation in Plaxis is the so-called ldquoSafety Calculationrdquo with which global
safety factors can be computed In Appendix B2 a more detailed elaboration of this method is
exhibited This feature can also be used in reliability analysis by formulating the LSF as follows
(512)
where is the Multiplier Safety Factor obtained by the φ-c reduction The main concept is
that if the safety factor is smaller than 1 then it is considered as failure This method can provide
us with the probability of failure for a general soil body collapse However even if this method
seems to be quite simple there are some issues during its implementation in a reliability
analysis
For complex limit state functions there can be convergence problems for some
reliability methods with iterative procedures (ie FORM) This can be explained from the
fact that in a φ-c reduction soil strength properties follows a certain path as it is
described in Appendix B2 independently of the values for the random variables that
have been set from the reliability method
The safety factor is a general safety factor regarding the failure of the system
Therefore sometimes it is not straightforward what the ldquotriggerrdquo factors are that led to a
certain type of failure
A safety calculation in Plaxis is time-consuming and the outcome can be unstable (Plaxis
calculationsrsquo convergence depends on the number of calculation steps)
Plaxis cannot handle safety factors below 1 Therefore a new limit of the safety factor
shall be introduced in this limit state function
3 Relative Shear Resistance
In this method the basic idea is to define soil failure according to the failure criterion that Plaxis
is set to use for the analysis In this research the Mohr Coulomb failure criterion is applied in
combination with a linear ndashelastic perfectly-plastic stress-strain relationship (that is the so-
called Mohr-Coulomb model) For this model the initiation of plasticity is considered as failure
Therefore the maximum shear resistance is defined just before plastic yielding occur for any
given stress state Of course the occurrence of plasticity does not directly indicate the failure of
the soil structure however this model can give a first estimate for the stress state and the
deformations
According to this method the relative shear resistance is defined as the ratio between the
mobilized shear stress and the maximum shear resistance In Figure 510 the Mohr-Coulomb
model is illustrated in a Cartesian stress-plane In the graph the principle stresses σ1 and σ3 are
indicated More precisely the principle stresses are defined as follows
radic
(513)
(514)
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 65
radic
(515)
Figure 510 Mohrs stress circle and Mohr-Coulomb failure criterion
Hence the mobilized shear stress is
radic
(516)
And the maximum shear resistance which is the distance of the mean stress point to the yield
surface is defined as
(517)
Then the relative shear resistance is a measure for load-resistance ratio that can be determined
in any integration point in the soil continuum
(518)
However this criterion requires prior knowledge of the possible relevant failure mechanisms
and their potential locations on the soil structure This is necessary in this method in order to
choose a suitable cluster of integration points where the average value of the relative shear
strength is to be determined Therefore the limit state function to be considered is not
straightforward and it needs problem investigation in advance
4 Plaxis definition of soil collapse
In Plaxis the construction stages are analysed by performing a Load advancement ultimate level
procedure which is controlled by a total multiplier
(519)
where is the load that is applied by Plaxis is the load at the beginning of the
calculation phase (ie the load that has been reached at the end of the previous calculation
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 66
phase) and is the defined load configuration (ie water level soil weight vertical load
on the top of the dike etc) Thus the main idea is that Plaxis applies an incremental load until it
reaches the defined one When has reached an ultimate level (which by default is 1)
the current phase is finished However if a staged construction calculation has not properly
finished the multiplier fails to converge to the desired ultimate level at the end of the
phase and a warning of ldquoSoil collapsesrdquo appears in the Log info box
In other words a collapse load has been reached In this case the total specified load has not
been applied In physical terms this means that the current value of the stiffness parameter CSP
is less than 0015 (Plaxis 2D 2015b) It is a measure for the amount of plasticity that occurs
during the calculation When the solution is fully elastic CSP is equal to unity whereas at failure
it approaches zero Therefore soil reaches its upper limit of plasticity and it collapses which can
be visualized as a settlement of the dike body Such a warning in the analysis is thus assumed to
be a possible failure situation
53 Overview In this chapter an overview of the relevant possible LSF of the structural elements and the dike
body respectively has been given The different LSFs are formulated and discussed whereas the
failure of each component is going to be evaluated in the ULS of each component
The system failure in this research is considered as the combination of the anchor sheet pile
wall and soil body LSF Particularly for the anchor and the sheet pile failure their yield stress
was chosen as a limitation for their failure by adopting the next LSFs
[(
)] sheet pile wall LSF (520)
anchor LSF (521)
As far as the soil body is concerned in this research method 4 is to be used in order to identify
soil failure The limit state function is then formed as a single value in case of a successful and an
unsuccessful computation Therefore the limit state function was chosen to be equal to 1
(actually the value of is retrieved from Plaxis that in case of a successful calculation is
1) in case of a successful Plaxis analysis and -1 in case that soil collapses in any phase (error in
Plaxis Log file ldquoSoil Collapsesrdquo) This formula can be written as follows
(522)
This accounts for the global instability of the dike body in contrast with the other methods that
are related to a certain dike area (methods 1 and 3) or they are cannot handle overall failure
due to numerical issues (method 2) More precisely methods 1 and 3 were excluded due to the
prior knowledge that is required in method 3 while in method 1 a definition of the maximum
admissible deformations shall be first introduced However attention shall be paid to the type
of the warning that Plaxis gives because it might also be that the maximum number of load steps
was not sufficient In that case the phase must be recalculated using a larger number of steps
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 67
Someone could argue that such a failure can be considered as a system failure rather than a soil
failure alone Partially this is true as the incapability of the sheet pile wall or the anchor to keep
the upstream part of the dike stable can lead to non-convergence and thus according to Plaxis
definition to failure This incapability can be translated into various scenarios of failures that
stem from either the structural elements or the soil body weakness and that are illustrated in
Figure 511 This picture shows the different scenarios that can take place due to macro
instability that was discussed in 523 and the failure of the retaining wallrsquos elements
Figure 511 Combined failure scenarios (initial and developed mechanisms) (source Larsen et al 2013)
However this does not mean that the sheet pile wall or the anchor fail because of exceeding
their yield stress but rather due to possible unacceptable deformations This is also an
advantage of the inherent residual strength of the structural elements that allows a large
deformation before they fail However the residual strength of the elements was not taken into
account in this thesis as they were considered as elastic Moreover the message for the ldquosoil
collapsesrdquo can appear in a phase prior to the installation of the wall which means that this type
of failure is exclusively attributed to the deficiency of the soil strength Therefore it is not
always straightforward what is failing first and lead to the system failure Consequently by
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 68
considering such an analysis representative of a system failure it is likely to underestimate the
overall probability of failure whereas if it is considered as a soil failure only it might
overestimate the total probability of failure since it is later combined with the rest LSF of the
anchor and the sheet pile wall In this research it will be considered only for the soil failure that
brings the results to the safe side even if it does not lead to the best optimization In this thesis a
serial system of the anchor the sheet pile and the soil body LSF was considered the fault tree of
which is depicted in Figure 53
Another way to evaluate the reliability of the different elements could be the reach of a
threshold displacement according to the SLS criteria Such a threshold could be also considered
for the system as a whole rather than the individual elements However in that case someone
should be cautious regarding the choice of the value for the maximum deformations as well as
the location of their occurrence
Concluding in this section the possible failure mechanisms of a flood defence (Figure 58) were
shown Then it was clarified which of these mechanisms can be modelled in Plaxis and they
were further explained and schematized in Figure 58 with the presence of the retaining wall as
they would have to be handled in case of hand calculations Plaxis can incorporate these
mechanisms automatically and thus no distinguish among the different patterns of these
specific modes is necessary Last but not least in Figure 511 the possible failure modes for the
system of the dike and the retaining wall that can be simulated in Plaxis are depicted In this
thesis the plasticity of the structural elements is not considered and thus possible plastic hinges
in the anchor and the sheet pile are not taken into account Therefore scenario 3 is excluded
In Chapter 7 the results after the reliability analysis of the aforementioned LSFs are given It
was essential though before proceeding with the reliability analysis to get a better
understanding into the system under investigation For this purpose in the next chapter a
description of the case study as it has been simulated in Plaxis is given while also the different
structural properties are defined for the following reliability analysis The soil variables to be
considered as stochastic are presented and a first insight into of the systemrsquos behaviour is
obtained through the mean values calculation
MSc Thesis A Rippi 69
6 Case Study-Dike with an anchored sheet pile wall
In this chapter a description of the case study is presented together with the relevant soil and
structural parameters In sequence a deterministic analysis follows in order to specify the basic
structural characteristics and obtain a first sense of the stresses magnitude developed on the
structure This was carried out based on the recommendations given in the CUR 166 (2005)
The aim of this deterministic analysis is to redefine the structural properties in order to avoid
having an overdesigned structure understand the current design procedure and see if there are
any possibilities of improvement by applying the proposed reliability analysis
61 Case Description The case study concerns a dike section with an anchored sheet pile wall which was initially used
to showcase the current methodology of designing dikes with sheet piles using partial factors
For more information a reference is made to Breedeveld (2011)
In this research this case study has been modified and simplified for the needs of the thesis and
it is illustrated in Figure 61 while in Table 61 the soil materials are identified Details for the
properties of the soil layers are given in section 62 We assume the structure to be
homogeneous to the third dimension and therefore a plane-strain model is applied The dike
soil materials (layers 3 and 4 in Figure 61) were adopted from the case study used in
Breedeveld (2011) that was presented in Chapter 2 Especially in Figure 25 the original case
study as it was modeled in Plaxis is shown
Table 61 Soil materials that corresponds to Figure 61
1 Aquifer Sand layer 2 Aquitard Clay layer 3 Dike new material Clay layer 4 Dike old material Clay layer
The geometrical properties the groundwater level and the design water level were taken
identical to the original case study as well as the soil properties of the dike materials The soil
layers under the dike were modified for the need of the thesis and were modeled using random
average quantities for which the statistics were chosen arbitrarily but in a realistic range
according to the NEN6740- Table 1 (see Appendix C) included in the Dutch code for
geotechnical structures Moreover the vertical load was removed in order to examine the
impact of the water elevation on the dike stability
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 70
Figure 61 Case study Dike with anchored sheet pile wall Dimensions and Soil materials
62 Soil Parameters The soil parameters for the dike material were obtained from the report of Breedeveld (2011)
where for some parameters the characteristic values were available whereas for some others
the mean values were given For the soil layers below the dike Table 1 of NEN6740 was used as
it was mentioned before which gives characteristic values For those parameters whose value
was a characteristic 18 one a transformation was made according to their probability
distributions in order to acquire the mean values that are required for the reliability analysis
Only the properties required for the use of the Mohr-Coulomb model are presented
The parameters of each soil layer are presented in Tables 63(a) and (b) and the front number of
each layer indicates the corresponding soil layer as showed in Figure 61
The distribution types and the coefficients of variation for each parameter were chosen
according to the knowledge that has been obtained until now about the physically possible
ranges of such parameters and the recommendations that have been given in several researches
up to now (Baecheramp Christian 2003 Schweckendiek 2006 GriffithsampFenton 2007 Phoon
2008 Wolters 2012 Bach 2014) The distributions and the CoV of all parameters are displayed
in Table 62 They were considered to be the same for all soil layers Particularly for the
saturated and unsaturated volumetric weight γsat and γunsat respectively a relationship was
established in order to derive the one from the other In general γunsat varies between the real
dry weight and γsat depending each time on the degree of saturation An estimated maximum
difference between γunsat and γsat is approximately 4 kNm3 (Brinkgreve 2015) In this thesis γsat
was expressed as the summation of γunsat and a variable with uniform distribution in the range
of [0 2]19 Initially a deterministic relationship was defined between the two variables but later
on it was proved the variable was important for the soil failure and thus it was decided to
consider it as random
18 for the definition of the characteristic value see Appendix E 19 The range of [0 4] was not chosen in order to avoid extreme conditions
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 71
Table 62 Soil parameter distributions
Soil parameter Symbol DistributionRelation COV Unit
Unsaturated
Volumetric weight γunsat Normal (microσ) 5 [kNm3]
Saturated Volumetric
weight γsat γsat= γunsat+U(02) 5 [kNm3]
Cohesion c Lognormal (microσ0) 20 [kPa]
Friction angle φ Truncated normal (microσ045) 10 [ ˚]
Youngrsquos Modulus E Lognormal (microσ0) 25 [kPa]
Poissonrsquos ratio ν Truncated normal (microσ005) 10 [-]
Interface strength Rinter Truncated normal (microσ0099) 20 [-]
Table 63 (a) Soil statistical properties under the dike and (b) soil statistical properties in the dike
(a) 1 SAND (very silty)
Soil property Symbol Characteristic
value
Mean
value STD Unit
Saturated
Volumetric
weight
γsat 20 22 11 [kNm3]
Unsaturated
Volumetric
weight
γunsat 19 21 103 [kNm3]
Cohesion c 0 0 0 [kPa]
Friction angle φ 30 36 36 [ ˚]
Dilatancy angle ψ 0 0 0 [ ˚]
Youngrsquos Modulus E 20000 30769 7692 [kPa]
Poissonrsquos ratio ν 025 03 003 [-]
Interface strength Rinter 044 066 013 [-]
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 72
2 CLAY (clean medium)
Soil property Symbol Characteristic
value
Mean
value STD Unit
Saturated
Volumetric
weight
γsat 17 185 093 [kNm3]
Unsaturated
Volumetric
weight
γunsat 17 185 093 [kNm3]
Cohesion c 10 14 282 [kPa]
Friction angle φ 175 21 21 [ ˚]
Dilatancy angle ψ 0 0 0 [ ˚]
Youngrsquos Modulus E 2000 3077 769 [kPa]
Poissonrsquos ratio ν 029 035 0035 [-]
Interface strength Rinter 034 05 01 [-]
(b) 3 DIKE NEW (very sandy clay)
Soil property Symbol Characteristic
value
Mean
value STD Unit
Saturated
Volumetric
weight
γsat 17 185 093 [kNm3]
Unsaturated
Volumetric
weight
γunsat 17 185 093 [kNm3]
Cohesion c 4 564 113 [kPa]
Friction angle φ 29 347 347 [ ˚]
Dilatancy angle ψ 0 0 0 [ ˚]
Youngrsquos Modulus E 1625 2500 625 [kPa]
Poissonrsquos ratio ν 029 035 0035 [-]
Interface strength Rinter 034 05 01 [-]
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 73
4 DIKE OLD (little sandy medium clay)
Soil property Symbol Characteristic
value
Mean
value STD Unit
Saturated
Volumetric
weight
γsat 195 212 106 [kNm3]
Unsaturated
Volumetric
weight
γunsat 19 207 103 [kNm3]
Cohesion c 13 183 367 [kPa]
Friction angle φ 28 335 335 [ ˚]
Dilatancy angle ψ 0 0 0 [ ˚]
Youngrsquos Modulus E 2925 4500 1125 [kPa]
Poissonrsquos ratio ν 029 035 0035 [-]
Interface strength Rinter 034 05 01 [-]
63 Finite Element Model The structure has been modelled with the finite element software Plaxis 2D 2015 In Figure 62
the model as it was created and the automatically generated mesh are presented The ldquoboxrdquo
around the dike section was used to refine the mesh in this area The mesh in the interfaces
between the sheet pile and the soil was also refined for a better representation of the potential
high stresses that can develop in that area
Figure 62 FEM model and generated mesh for case study
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 74
The sheet pile has been modelled with elastic20 plate elements and the free anchor length with
an elastic fixed-end-anchor element The grout body of the anchor was not explicitly modelled
but it is implied in the end of the anchor as its final point is considered to be fixed in the soil A
fixed-end-anchor was chosen because even if it is a simplification of an anchorage system it can
still give reliable results in terms of the anchor axial force which is actually included in the limit
state function However the interaction of the soil and the grout body in that case cannot be
simulated The interaction and the modelling of a grouted body in the soil are still under
investigation due to several difficulties in modelling the real conditions
For all the steel members a Youngrsquos modulus of E=210 GPa was used while the specific
structural parameters are to be specified after the deterministic analysis that is described in the
next section
64 Deterministic Analysis The design concept of dikes with sheet piles modelled in FEM is based on the combination of
several partial factors in order to evaluate the overall required safety factor Additionally the
design recommendations of CUR 166 (2005) are considered for using the design or the
characteristic values of the structural parameters and the loads The design procedure (see
section 22) that has been initially followed in this structure is reported in Breedeveld (2011)
and the following structural elements have been defined accordingly
Cross sectional area of the sheet pile (per meter)
Length of the sheet pile wall
Steel quality of the piles
Cross sectional area of the anchor (per meter)
Free length of the anchor (without the grounded body)
Steel quality of the anchorrsquos reinforcement
Profile of waling
The required overall safety factor was calculated up to 18 according to the current design
procedure described in section 22 according which the above mentioned structural parameters
were defined Moreover apart from the overall safety factor in Breedeveld (2011) additional
partial factors are applied on the deduced forces and bending moments of the anchor and the
sheet pile wall as they are defined in CUR 166 (2005) In particular in Table 64 the different
partial factors that are applied on the several structural developed actions are given
Table 64 Partial factors for the different structural actions according to CUR 166 (2005)
Structural action Partial factor γ
Sheet pile normal force Fn 115
Sheet pile bending moment M 115
Anchor normal force FA 125
Waling bending moment Mw 110
20 The elements were considered as elastic in order to avoid numerical problems in FORM analysis For a better explanation a reference is made to the evaluation of the optimization algorithms in Appendix A2
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 75
However as it is mentioned in section 62 alterations have been made in the soil layers below
the dike while also the vertical load has been removed On the top of that in section 22 it was
shown that the large applied partial factors led to an overestimation of the design values for the
moments of the sheet pile Therefore a redesign of the sheet pile and the anchor is carried out
in this section in order to determine the new required characteristics of the structural elements
according to the new loading situation with less strict partial factors In particular in this thesis
the required overall safety factor was considered 12 according to CUR 166 (2005) calculation
scheme that is described in the next section The reason for using a lower safety factor was to
avoid an overestimation of the design forces and moments of the sheet pile wall that would
render a reliability analysis meaningless as the structure would be quite safe Such an analysis is
also important in order to show the discrepancy between the two different design procedures
and indicate the optimal one
However the reliability analysis was carried out for the mean values of the soil parameters and
for a safety factor of 10 in order to account only for soil properties uncertainties A safety factor
of 10 was not directly used as the design procedure followed in section 643 accounted only for
the macro-instability of the system and thus other failure mechanisms such as piping
overtopping infiltration etc were not taken into account In that way we prevent coming up
with a retaining wall that would be volatile under other crucial failure mechanisms
In section 641 the calculation scheme to be used in this thesis for the deterministic analysis is
further described In the sequence in section 642 the construction phases that were followed
in Plaxis simulation are listed and illustrated and finally in section 643 the results of the
preliminary design are displayed
641 Calculation Scheme and Design Values
In this research only the design philosophy suggested in CUR 166 is applied and the desired
safety factor of the structure is considered to be 12 for structures classified in ldquosafety class IIIrdquo
in order to avoid an overestimation of the structurersquos dimensions as it happened in the original
case study At this point it should be also clarified that the redesign was conducted considering
only the overall stability of the structure and no other failure mechanisms such as piping
internal erosion or overtopping
The prescribed calculation schemes in CUR 166 (2005) for designing structures with retaining
walls are basically the following
Calculation scheme A Calculations with design values The calculations are executed
using the design values for the soil parameters retaining height water levels and
stiffness of the structure It uses two different soil stiffnesses When using a high soil
stiffness the anchor force is generally relatively high The advantage of this scheme is
that it requires relative little effort The disadvantage is that the deformations may be
overestimated due to the use of design values in every stage However a designer is not
interested in deformations when performing a ULS calculation
Calculation scheme B Calculations with characteristic values The calculations are
executed using characteristic values of the soil strength (crsquo and φrsquo) Design values are
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 76
used for the retaining height water levels external loads and stiffness of the soil Here
again two different soil stiffnesses are used In the end of the governing stage a phi-c
reduction should be carried out As the characteristic values for the soil parameters are
used the deformations are most probably smaller than in case of calculation scheme A
The safety margin lies in the fact that phi-c reduction is imposed to obtain a safety factor
of 115 (class II) or 12 (class III)
For the soil stiffness parameters two options are given in CUR 166 a low and a high value
(61)
where is the (low and the high) design value the (low and the high) characteristic value
and is the mean value of the stiffness parameter In Woltersrsquos master thesis (2012) both
have been applied and the appeared to be governing for the bending moment in the wall
and the anchor force
In this research the Calculation scheme B was used and the as the soil stiffness
parameter It should be mentioned that regarding the soil parameters the characteristic values
were available with an exemption for the stiffness for which the mean values were given The
groundwater level was manually schematized according to TAW (2004) because of lack of data
while the maximum water level in the river side was taken equal to the design water level
according the Dutch regulations for the specific dike region
In Tables 65 and 66 the partial factors according to CUR 166 and the corresponding design
values for the different soil properties are presented As someone can notice in these tables
only the strength parameters (phi and c) and the stiffness parameter (E) are exhibited For the
rest parameters the design value is identical to the characteristic value (see Tables 63(a) and
(b)) as a partial factor of 1 is used
Table 65 Partial factors and design values for the soil layers under the dike
1 SAND (very silty)
Soil property Symbol Partial factor γk Design value Unit
Cohesion c 11 0 [kPa]
Friction angle φ 12 2569 [ ˚]
Youngrsquos Modulus E 13 1538462 [kPa]
2 CLAY (medium)
Soil property Symbol Partial factor γk Design value Unit
Cohesion c 11 91 [kPa]
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 77
Friction angle φ 12 1472 [ ˚]
Youngrsquos Modulus E 13 153846 [kPa]
Table 66 Partial factors and design values for the soil materials in the dike
3 DIKE NEW (medium clay)
Soil property Symbol Partial factor γk Design value Unit
Cohesion c 11 364 [kPa]
Friction angle φ 12 2479 [ ˚]
Youngrsquos Modulus E 13 1250 [kPa]
4 DIKE OLD (stiff clay)
Soil property Symbol Partial factor γk Design value Unit
Cohesion c 11 1182 [kPa]
Friction angle φ 12 239 [ ˚]
Youngrsquos Modulus E 13 2250 [kPa]
642 Construction Stages
The construction and the gradual loading of the dike ware modelled as follows (see Figure 63)
1 K0-procedure for the generation of the initial stresses under horizontal groundwater
level
2 Dike self-weight under horizontal groundwater level
3 Rise groundwater level to the phreatic water elevation
4 Interpolation21 of the clay layer under the dike and activation of the sheet pile wall and
the anchor
5 Apply extreme water level conditions on the dike structure
6 (φ-c reduction for the determination of the safety factor)
21 Due to different permeability between the clay and the sand layer an approximation of the developed pore pressures is carried out through interpolation between the two layers
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 78
1
2
3
4
5
Figure 63 Construction and loading stages
For the soil modelling the Mohr-Coulomb model was used with a non-associative flow rule
(φneψ) For phases 1-4 the calculations were conducted under drained conditions whereas in
phases 5 and 6 the undrained behaviour of clay layers was considered
643 Determination of the structural elementsrsquo characteristics
According to the calculation scheme that was described above the design characteristics of the
structural elements were obtained from phase 6 In particular the target overall safety factor is
12 with which initially the length of the sheet pile wall was calculated
+370 NAP
+700 NAP
+1050 NAP
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 79
Firstly a length of 10 m and a profile of AZ 12 with a steel quality of 240 kNm2 yield stress
were used while for the anchor a cross-section of 628 mm2 and S355 were chosen for the first
calculation However even if the safety factor reached 12 it seemed to be quite unstable during
the calculation steps while the developed anchor stress was above the ultimate yield stress
What is more the deformations resulted in that case were more than 30 cm Therefore in order
to stabilize the safety factor and to decrease the displacements the length of the sheet pile was
increased 1 m (so 11 m in total) while also a better steel quality for the sheet pile was chosen
S355 Furthermore the anchor cross section was increased to 933 mm2 with a steel quality of
MW450 Last but not least the waling was designed based on the anchor force
In the following paragraphs the design procedure of each element is described providing also
the results of moments and forces that were retrieved from Plaxis
Sheet Pile wall
As it was partially mentioned in the previous paragraph the length of the sheet piles is
considered to be sufficient when the safety factor of the system22 reaches at least 12 For the
reasons discussed above the sheet pile length was decided to be 11 m and with a profile of AZ
12 S355
With this configuration the safety factor reached is 13 as it is depicted at the right graph of
Figure 64 Someone can notice that the safety factor from 117 in the situation without a sheet
pile structure increased to 13 with the structure
Figure 64 Safety factor deduced from the phi-c reduction calculation without the anchored sheet pile wall at the left (safety factor=117) and with the wall at the right (safety factor=13) The steps on the x-axis represent the calculation steps have been carried out in order to reach an equilibrium after incrementally reducing the strength parameters ie phi and c
The design moment is to be taken for a safety factor of 12 that is reached in step 34 where it
was found to be 1519 kNmm and the corresponding axial force 9133 kNm Therefore
considering Eq 52 for checking the exceedance of the ultimate yield stress and applying the
prescribed partial factor of 115 for both the moment and the axial force we get
22 Plaxis always gives the safety factor of the system which is under investigation
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 80
In Figure 65 the maximum bending moments with a safety factor 10 and 12 are respectively
showed From the values being shown in that figure someone can easily distinguish the
difference between the two moments It must be also mentioned in that point that the required
safety factor defined in Chapter 2 was estimated at 18 (and not 12 that is considered here)
which gives excessive design moments (almost 7 times larger) For that reason in this research
a lower safety factor was assumed avoiding thus the design of a quite conservative structure
Figure 65 Bending moments on the sheet pile for a safety factor safety factor=10 at the left and
safety factor=12 at the right
The explanation behind this discrepancy is that in case of the phi-c reduction procedure the
artificially decreasing strength of the soil and on the other hand the non-reduced strength and
stiffness23 parameters of the structural elements tends to lead in an increased arching
phenomenon that makes the structure attract the most of the load
In this research the reliability of the structure will be evaluated for a safety factor of 1 or else
during phase 5 and by considering the mean values of the soil parameters As someone can
notice from Figure 65 the magnitude of the maximum bending moment for a safety factor =10
is quite low which implies a possible minor contribution to the overall failure
At this point it is essential to mention that Mohr-Coulomb model is capable of simulating the
elastic behaviour and the state at failure Therefore in case that plasticity of the soil or the
structural elements plays an important role a different constitutive model is recommended to
be applied
Anchor
For a safety factor 12 the anchor force was found to be 10387 kNm Considering a mutual
anchor distance of 3 m the total axial force that is undertaken by the rod multiplied also with
the prescribed design partial factor is
23 ldquoStrengthrdquo and ldquostiffnessrdquo are usually confused terms Strength is a measure of the maximum load that can be placed on a material before it permanently deforms or breaks and it is represented by the yield stress σy in metals and by φ and c in soil Stiffness is a measure of the amount of deflection that a load causes in a material and it is represented by the Youngrsquos modulus E and the Poisson ratio ν both in soil and metals
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 81
Therefore choosing an anchor with a cross sectional area of A = 933 mm2 and a steel quality of
MW450 ( ) the design stress level is
In comparison with the sheet piles the anchor does develop large stresses as it was concluded
from the results in the deterministic calculations This is probably attributed to the high
stiffness of the anchor that apparently attracts the load and thus the anchor develops large axial
force
Waling
The type of the waling system is defined according to the anchor force as it was described in
section 522 Therefore by choosing a waling type of 2UPE 200 the developed bending
moment can be estimated as follows
where is the design bending moment of the walling system according to the chosen profile
In Table 67 the characteristics of each structural element that was described above are
displayed in detail
65 Overview In this chapter an introduction of the case study as it has been used in this research is made The
soil materials and their properties are summarized and their mean and design values are
defined The probability distributions of the soil parameters to be considered as stochastic were
proposed according to the literature study and the engineering judgement Moreover the mesh
generation and the modelling features are discussed
Last but not least due the observed overestimation of the retaining wallrsquos design moment and
axial force according to the current design regulation a new deterministic analysis was carried
out in order to redefine the structural properties and come up with a less conservative structure
This makes the reliability analysis of the system meaningful in a sense that an extremely safe
structure would give a failure probability of zero Besides the scope of a reliability analysis is
also the optimization of the construction that corresponds to a financially attractive and
simultaneously safe structure This can be made by conducting a risk analysis whose first step
would be the analysis that is carried out in this research Therefore the aim is to start with the
marginal required structural properties according to the safety standards In a risk analysis that
should conducted for several dimensions of the structure until the most optimized one is
identified However this is out of the scope of this thesis
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 82
In Table 67 the structural properties of the sheet pile wall the anchor and the waling are
summarized as they were found in the deterministic analysis
Table 67 Design parameters of structural elements
SHEET PILES
Property Symbol Value Unit
Profile - AZ 12 [-]
Steel quality - S355 [-]
Length L 11 [m]
Moment of inertia I 21430 [cm4m]
Elastic section modulus Wel 1245 [cm3m]
Mass w 943 [kgm]
Sheet pile thickness d 85 [mm]
Cross sectional area A 1201 [cm2m]
ANCHOR
Property Symbol Value Unit
Steel quality - MW450 [-]
Free length Lafree 104 [m]
angle φ 30 [deg]
Cross sectional area A 933 [mm2]
Mutual anchor distance s 3 [m]
WALING
Property Symbol Value Unit
Profile - 2UPE200 [-]
Steel quality - S355 [-]
Elastic section modulus Wel 191 [cm3m]
In the original case study the sheet pile wall is 162 m long and it has a profile of AZ 28 (S240)
while the anchor is an MW450 with a cross-section of 1134 mm2 As it is noticed both sheet
pilesrsquo and anchorrsquos properties have been reduced after the new deterministic analysis However
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 83
it should be mentioned that in the original case there was a vertical load on the top of the dike
that was displaced while the clay layers under the dike was merged into one in the new case
study These changes might lead to less heavy structure However the difference among the
design properties between the new and the original case study is mainly due to the required
safety factor that was assumed In the new case study the safety factor was set to 12 whereas in
the original cases study a safety factor of 18 was considered as it was calculated according to
the partial safety factors
In the next chapter the results after the reliability analysis of the just designed case study are
presented
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 84
MSc Thesis A Rippi 85
7 Reliability analysis results with stochastic soil properties
In this chapter the probability of failure of the different structural elements that were discussed
in Chapter 4 will be estimated by varying the soil properties of the several soil layers under and
inside the dike At the beginning a better insight into the system behavior is gained by
inspecting the deterministic Plaxis calculations for the mean values of the soil variables After
that a global SA is carried out in order to evaluate the most important soil parameters and
inspect their impact on the different structural elements in combination with each other and
individually Finally the results regarding the probability of failure of the anchor the sheet pile
wall and the soil are presented as they deduced by the reliability analysis (concerning the ULS of
the elements) Finally the systemrsquos reliability is evaluated The methodology that was followed
during the probabilistic analysis is described below and it is mainly dedicated to the DS
rationale that was used in order to obtain the influence factors and to evaluate the results of
Plaxis calculations
71 Method description For the calculation of the failure probability of the systemrsquos different components FORM and DS
were used for the retaining wall and the soil body respectively From these methods the
probability of failure and the important factors are going to be deduced for each component (ie
anchor sheet pile wall and soil body) However before starting with the reliability analysis it is
essential to carry out a sensitivity analysis in order to obtain a first impression of what are the
most influencing variables and thus determine the input random parameters for the reliability
analysis In the figure below the steps for the reliability analysis as they were followed in this
thesis are presented
Figure 71 Steps of the reliability analysis
Below the methodology of both the sensitivity and the reliability analysis are discussed and
explained in detail
Sensitivity analysis
In this thesis the FAST method was utilized for conducting the sensitivity analysis which is
further elaborated in Appendix A1 Here an explanation is given for the quantitative measure of
sensitivity which is represented by Sensitivity Indices The first-order sensitivity index of an
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 86
input variable pi is the measure of the main (direct) effect of pi on the output variance
(where i ne j) the second-order sensitivity indices measures the interaction effect of pi and pj on
the output variance Other higher-order indices are defined in the same manner The total
sensitivity index is the sum of all sensitivity indices involving factor pi (Homma amp Saltelli
1996) For example the total sensitivity index of factor 1 for a model with 3 input factors is
given as
The first-order (Si) and total ( ) indices can be interpreted as following (Pandya et al 2008)
high pi is an influent parameter
and both small pi is not an influent parameter (neither alone nor in interaction
with other parameters)
and nearly the same no interaction of pi with the other parameters
and very different high interaction of pi with the other parameters
Reliability analysis
As far as the FORM analysis is concerned section 422 gives sufficient information about the
methodology that is followed in order to obtain the failure probability and the influence factors
(ie α-values) The main concept and principles of DS are briefly explained in section 421
whereas here a more precise description of the implementation of DS and the how the α2-
values can be deduced is given
In particular DS is a method in which ldquodirectionsrdquo are sampled The concept of this method is
schematically depicted in Figure 72(a) For each sampled direction it is evaluated whereas
along this line the LSF Z equals to zero or not This procedure is repeated for a number of
directions and this method is applied in the standard normal space (u-space see also section
422) using equations to transform the standard normal variables u1 hellip un to their
corresponding ldquorealrdquo variables (in the x-space) X1 hellipXn and vice versa
(a) (b)
Figure 72 (a) Schematic view of the DS method in the u-space The numbers 0 1 2 3 4 show successive LSF evaluations along one direction (b) Schematic view of the line search algorithm β is the distance from the origin in the u-space along one direction The numbers 0 1 2 3 4 correspond to the ones shown in figure (a) (source Hydra-Ring 12 2015)
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 87
The search procedure for the location on the line where Z=0 is often referred to as the ldquoline
search methodrdquo and it is depicted in Figure 72 (a) and (b) In both figures the numbers 0-4 refer
to successive Z-function evaluations along one direction Figure 72(a) shows 3 evaluated
directions in the u-space whereas Figure 72(b) shows the (fictional) value of the Z-function
along one direction (where ||u|| is the distance from the origin along the line) The maximum
length of the direction line as well as the step size along the direction (for example the distance
between 0 and 1 along one line) can be steered in order to increase the efficiency of the method
and presumably decrease the computational time
In the sequence a method was developed in this thesis in order to estimate the α2-values of the
random variables out of a DS analysis as OT does not have an available method so far For that
purpose a transformation of the output samples to u-space was firstly carried out Then the
distance to the origin of all samples (in u-space) that are located on a direction where failure
(Z=0) was detected is calculated as follows
radicsum
i=1hellip (71)
where n is the number of random variables and uij is the ith ldquofailurerdquo sample of the jth random
variable is equivalent with the reliability index of this direction βi The α-values for each
random variable can then be calculated as (for more information about these values see
sections 422- Eq 417 and 423)
(72)
After that three different methods were created in order to evaluate the influence of the
random variables on the response of the LSF ie the α2-values (square of Eq 72)
Shortest distance ldquoβminrdquo
Average 10
Average all
In the Shortest distance method the sample with the smallest distance to the origin is
considered to be the design point and thus the corresponding influence factors are estimated
according to that point The smallest distance is also equivalent to the reliability index βmin This
method is well-known for estimating the design point out of MC techniques (Vrijling amp Van
Gelder 2002) However for a soil failure modeled in FEM this might not be always indicative of
the real influence that the soil properties can have on the system failure Therefore the last two
techniques were used as additional in order to get a better insight into the dominant soil
properties In Figure 73 a general scheme of the methodology that was followed for obtaining
the α2-values is depicted where also the different techniques are presented
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 88
Figure 73 DS methodology for calculating α2-values
The second method Average 10 takes into account the α2-values of the samples in the failure
domain whose distance to the origin lies within the 10 higher than βmin and averages them
Therefore it is firstly required the calculation of βmin as it was described for the Shortest distance
method Such a technique helps to identify the other directions that are close to the one which
gives the shortest distance to the origin and thus inspect if the important variables that were
indicated according to the first method are influential for other failure mechanisms as well
Last but not least the third method Average all averages the α2-values of all the samples
located in the failure domain This shows the overall contribution of the random variables to the
response of the LSF under investigation Concerning the soil failure LSF that is described in
section 73 such a method can be more suitable for distinguishing the soil properties that are
involved in the majority of the failure mechanisms rather than only in the one that the first
method indicates
72 Mean values calculations The results of the deterministic analysis that are presented in the previous chapter included a
degree of conservatism due to the load and material factors (partial safety factors) In order to
overcome this limitation structural reliability calculations were carried out and are discussed in
the next sections by considering relevant soil parameters as random However before starting
the probabilistic calculations by coupling Plaxis with OT that cannot be fully interpreted and
verified to a certain extent it is essential to first carry out some deterministic calculation Such
calculations can help to get a better insight into the behavior of the system and show
qualitatively what kind of response to wait later in the analyses These calculations were taken
place based on the mean values of the soil properties
Deformations
The analysis of this case study focuses on the ULS of the structure For the stresses and
deformation calculations the Mohr-Coulomb model is used that is generally accepted for the
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 89
simulation of the soil behavior until failure and for detecting the failure modes (a more detailed
explanation of this model is provided in Appendix B1) However this model is not so suitable
for analyzing the deformations of soft soils or the settlements around the wall
The deformation pattern that would more likely occur due to primarily the soil body failure can
be illustrated by conducting a ϕ-c reduction calculation In Figure 74 (a)-(c) the displacement
pattern after the ϕ-c reduction of the mean values of the friction angle and the cohesion
respectively As it can be seen a slip surface in the inner side of the dike has been created that
rotates towards the inland
(a) deformed mesh
(b) shadings
(c) arrows
Figure 74 Deformations after the ϕ-c reduction (mean values)
Stresses on the structural elements and the soil body
The anchor and the sheet pile wall are installed in phase 4 In phase 5 the anchor normal force
increases and especially for the mean values it amounts to Na = 7983kNm
As far as the sheet pile wall is concerned the bending moments and the normal forces that are
developed on the wall are indicated in Figure 75 as they were deduced from phase 5 Similarly
to the anchor the maximum stresses occur in phase 5 and are mainly located in the area of the
wall which is surrounded by the clay layer (brownish soil layer in Figure 74(a)) As someone
can observe the bending moments are not significant taking into account that the maximum one
observed reaches 2132 kNm
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 90
Figure 75 Bending moments and normal forces for the mean values calculation
For the interpretation of the stress field of the soil body the effective principle stresses and the
relative shear stress τrel (see Eq 517) were illustrated in Figure 76 (a) and (b) respectively In
Figure 76 (a) the crosses indicate the direction and the magnitude of the principle stresses σ1rsquo
and σ3rsquo and from which an active soil behavior (
) can be noticed on the river side of
the retaining wall whereas a passive behavior (
) is dominant in the inland side
(a) Effective principle stresses
(b) Relative
shear stress τrel
Figure 76 Effective principle stresses and relative shear stress in mean values calculation
In Figure 76 (b) the distribution of the relative shear strength shows the potential developing
plastic area in case of failure that expands from the river side of the wall and around it It can
also betoken the possible shape of formation of a slip surface that starts from the outer side of
the dike and goes around the wall
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 91
73 Sensitivity Analysis Results In section 312 different types of SA are presented and that can be applied through OT In this
research due to the time limit and the amount of parameters FAST method was used for
carrying out the SA as it allows the calculation of the first order sensitivity indices as well as the
total order indices and the computational time required is less than the other available methods
In this section the results from the conducted SA are presented separately for the anchor the
sheet pile and the soil performance Regarding the sheet pile wall and the anchor the sensitivity
of the total developed stresses towards the soil properties was evaluated according to Eq 53
and 55 respectively The influence of the soil parameters on the soil behaviour was evaluated
based on the calculation of the overall safety factor after a safety calculation (ϕ-c reduction
method) The total order indices are considered for distinguishing the most important
parameters while their difference with the first order indices Si indicates the level of interaction
effect amongst the parameters on the output variance
As far as the anchor and the sheet pile wall are concerned the SA was carried out in two steps
(1) SA with all the soil parameters (listed in Table 62) and (2) SA with the most important
parameters found in step 1 In Figure 77(a) and (b) the sensitivity indices from step (1) are
presented as they were found for the anchor and the sheet pile respectively Step 1 mainly
provides an overview of the contribution of the variablesrsquo variance to the total variance of the
limit state response and it helps to identify the most influencing parameters Step 2 was
basically carried out in order to build the response surface based on the SA output that was
later utilized in the FORM analysis as a complementary function in case of Plaxis errors At that
point it should be mentioned that the Dikenew material was not considered neither for the SA
nor for the reliability analysis of the structural elements as its influence was assumed to be
insignificant relative to the other soil layers However Dikenew was considered in case of the
soil body investigation
From Figure 77(a) it can be concluded that as far as the anchor is concerned the shear stiffness
G of the clay layer seems to be the most crucial soil property In particular the anchorrsquos stress
level seems also to be sensitive towards the Poisson ratio of the clay layer ν the friction angle φ
of the sand layer and the strength parameters φ c of the Dikeold material Moreover the
interface strength Rinter between the Dikeold material and the sheet pile wall seems to display
an additional notable influence as well The unit soil weight γ does not appear to affect the
anchor performance significantly
As far as the sheet pile wall is concerned the shear stiffness of the clay layer and the Dikeold
material seem to be dominant according to Figure 77(b) From these two soil layers it appears
that also the strength properties (φ c and Rinter) as well as the unit soil weight γ have a relative
impact on the stress level of the wall Furthermore the sand layer contributes via especially its
friction angle and shear stiffness
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 92
(a)
(b)
Figure 77 Total and First order sensitivity indices for the (a) anchor and the (b) sheet pile in step 1
As it was mentioned before step 2 was used for building the response surfaces for the anchor
and the sheet pile limit states The input random variables in this step are the most influencing
ones that were found in step 1 In Table 71 the soil properties that were considered as random
during step 2 are presented for both anchor and sheet pile It is essential to mention that the RS
was constructed by fitting a quadratic polynomial function to the sampling points of the SA
whose general expression is given in Eq 425 In Figure 78(a) and (b) an illustration of these
response surfaces is made However the figures depict the response of the anchor and the sheet
pile stress level over the fluctuation of only two variables amongst the 14 that are presented in
Table 71
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 93
Table 71 Soil properties considered for the RS construction of the anchor and the sheet pile respectively
Soil layer Anchor Sheet pile
Clay Sand
Dikeold
The linear regression of all the parameters was based on least squares and the R2 parameter
was found 0989 for the sheet pile and 0996 for the anchor Therefore the regression was quite
representative for the domain of the samples over which it was adjusted However there were
many residuals (ie difference between the observed and the predicted value) in the order of
103 and 104 which means that a possible expansion of this RS to an extended domain of the
input variables may lead to an underestimation of the limit state response
(a)
(b)
Figure 78 Polynomial response surface adjusted for the (a) anchor and (b) sheet pile stress level For the anchor the friction angle of the sand layer and the shear stiffness of the clay layer were accounted for this illustration whereas for the sheet pile wall the shear stiffness of the sand and the cohesion of the clay layer were used
The SA for the soil performance was also conducted in two steps However in that case there
was no need of a response surface construction since the LSF for the reliability analysis (Eq
522) takes the value of -1 in case of Plaxis error due to soil collapse Eq 522 though takes only
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 94
two different values -1 and 1 for soil collapse or not respectively which cannot be used for
carrying a SA For that reason the φ-c reduction method was used and the sensitivity of the
value of the safety factor (Msf) was investigated instead The second step of the SA was carried
out in order to divide the soil properties into two different groups of variables and thus
investigate the sensitivity of the soil in a more efficient way rather than including all the
parameters at one SA It should be mentioned that in the soil analysis soil properties of the
Dikenew material were added In Figure 79(a) and (b) the total and the first order indices are
depicted for step 1 and 2 respectively Step 1 includes all the stiffness and strength parameters
of all the soil layers whereas in step 2 the most important variables that were found in step 1
together with the unit soil weight γ are included Subsequently step 1 detects the most
influencing stiffness and strength properties and in step 2 the impact of γ is investigated
As it is observed in Figure 79(a) and (b) the safety factor appeared to be sensitive towards the
sandrsquos friction angle while also the strength properties (c and φ) and the shear stiffness (G) of
the clay layer play an important role Moreover the cohesion of the Dikeold material and the
unit soil weight of the clay sand and Dikenew layer seem to be influencing as well As it has
been mentioned above for this SA the LSF was different than the one considered in the
reliability analysis Therefore even if the SA at that point helps to evaluate in general the most
dominant soil properties the engineering judgement should also be included for the selection of
random variables that are going to be utilized in the reliability analysis
(a)
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 95
(b)
Figure 79 Sensitivity indices for the soil in step 1 and 2
It is also advisable that the difference between the total and the first order sensitivity indices be
investigated in order to have an insight into the most interactive variables The magnitude of the
interaction effect of a variable is crucial for the system behaviour because a variable might not
be important as a singularity but it is possible its combination with another variable to have a
considerable effect on the limit state under investigation Therefore this can be considered as
an additional criterion for choosing the set of the random variables to be used in the reliability
analysis In Figure 710(a) and (b) the difference between the two indices is presented for the
anchor and the sheet pile in (a) and the soil in (b) as it was deduced from the first step of the SA
Regarding the sheet pile wall the most interactive variables are the shear stiffness and the unit
weight of the clay layer the friction angle of the sand and the cohesion shear stiffness interface
strength and unit weight of the Dikeold material Concerning the anchorrsquos limit state the
cohesion of the Dikeold material seems to have the highest interaction with the rest of the
variables Last but not least regarding the soil performance the stiffness of all the soil layers
and the friction angle of the clay and the sand layer have a higher interaction effect amongst the
other parameters The importance of the aforementioned variables can be enhanced
considering both total indices and the difference that they display between their first and total
indices
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 96
(a)
(b)
Figure 710 Interaction effect amongst the parameters in (a) anchor and sheet pile and in (b) soil This effect is expressed as the difference between the total order indices T and the first order indices S as they were deduced from the first step of the SA for both anchorsheet pile and soil respectively
Considering the discussion above some preliminary conclusions can be drawn regarding the
influence of the several variables to the different components of the system
1 both anchor and sheet pile wallrsquos limit states are sensitive mostly towards the shear
stiffness of the soil
2 sheet pile wall stress level is also affected by the unit soil weight more than the anchor
does
3 soil body is mostly influenced by the strength properties and the unit soil weight
4 the interaction effect among the soil properties is more profound for the soil and the
sheet pile limit states rather than the anchorrsquos
It must be mentioned that this SA is a global one that gives an overview of the impact of the soil
parameters in terms of the system behaviour However near the design point a local sensitivity
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 97
can possibly give a better estimation of the variablesrsquo impact Near the design point the
structural elements as well as the soil body usually reaches plasticity that implies a non-linear
behaviour Such a behaviour is unlikely to be initiated during a SA that usually samples around
the mean values This can be counteracted by conducting more iterations so as to have a larger
amount of available samplings but this can make the SA to be quite time consuming and thus to
lose its benefit to give a fast preliminary estimation of the most influencing soil layers and
properties
Besides the purpose of a preliminary global SA is to reduce the number of the stochastic
parameters to a reasonable amount so as the initial reliability analysis to be more efficient and
computationally accessible The parameters that were eventually considered as stochastic in
each LSF are summarized in the table below and they were chosen according to both the SA and
the engineering judgement In the last column the stochastic variables used also for the
systemrsquos analysis are presented
Table 72 Stochastic variables for each LSF
Soil
Failure Anchor Failure
Sheet Pile failure
System failure
Sa
nd
γunsat radic
c
φ radic radic radic radic
G radic radic radic radic
v
radic
Rinter radic
Cla
y
γunsat radic
radic radic
c radic radic radic radic
φ radic radic radic radic
G radic radic radic radic
v
radic
Rinter radic radic radic
Dik
e_n
ew
γunsat radic
c radic
φ radic
G
v
Rinter
Dik
e_o
ld
γunsat
radic radic
c radic radic radic radic
φ radic radic radic radic
G
radic radic radic
v
radic
Rinter radic radic radic
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 98
74 Soil Shear Failure As far as the soil failure is concerned the limit state function that was implemented is given
below (see also sections 52 and 53) It was formulated as a single value depending on the
success of computation in Plaxis
(71)
The reliability method that was chosen in that case is DS due to the formulation of the limit state
function and the convergence issues of FORM There were used 12 input stochastic soil
parameters as it is indicated in Table 72 (see also input file of Soil Failure in Appendix D) The
variables were chosen based on the SA that was presented in the previous section For the soil
failure Poisson ratio was initially included in the analysis but it was noticed that values close to
04 during Plaxis calculations were leading to unrealistic behaviour of the dike More precisely
it has been observed that Poisson ratio equal to 04 can lead to swelling effects on the dike
which consequently lead to a raise of the dike crest instead of collapsing as it would be expected
in a soil failure analysis Therefore Poisson ratio was excluded from the reliability analysis of
the soil failure
The results of the reliability analysis according to DS are presented below In particular 300
iterations were carried out and 3 different failure directions were found Specifically in Table
73 the probability of failure the reliability index the number of iterations and the duration of
the analysis are firstly shown In the sequence the points from each failing direction with the
lowest reliability index are shown
Table 73 Reliability results of DS for the soil failure
DS Pf 1310-8 β 55 Number of LSF calls 1840 Number of iterations (or directions)
300
Elapsed time (hr) 34
Failure points (in real space) Parameter 1st direction 2nd direction 3rd direction Dikenew_phi [deg] 2277 2543 3021 Dikenew_c [kPa] 529 557 444 Dikenew_gammaunsat [kNm3]
1748 1848 1973
Dikenew_gammasat [kNm3]
1948 1984 1997
Sand_phi [deg] 2782 2039 3899 Sand_G [kPa] 808437 715473 715785 Sand_gammaunsat [kNm3]
2183 2122 2097
Sand_gammasat [kNm3] 2352 2145 2283 Clay_G [kPa] 56934 90063 80134 Clay_c [kPa] 704 1593 794
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 99
Clay_phi [deg] 1039 1629 1077 Clay_gammaunsat [kNm3]
1772 1413 1612
Clay_gammasat [kNm3] 1833 1433 1719 Dikeold_c [kPa] 1621 2978 915 Dikeold_phi [deg] 2474 3149 3281
According to the Shortest distance method the failure point from the 2nd direction turned out to
be the design point the importance factors of which are depicted in Figure 714 (first pie chart
at the left) However before concluding for the final design point an assessment of the above
failing directions was made
In particular the combination of the variables from each direction was later implemented in
Plaxis in order to illustrate the failure This is also necessary for verifying and assessing the
failure that Plaxis is giving and inspecting the error in the logfile in order to check if it is a ldquorealrdquo
(ie error=Soil collapses) or a ldquofakerdquo failure (ie error=Load advancement failed or Not enough
load steps) In Figure 711 the ldquorealrdquo failure is defined
Moreover a demonstration of the displacements is also essential in order to certify the failure
shape that normally has the dike crest settled down A different output of the displacements
formation (ie dike crest goes up) would probably imply a ldquofakerdquo failure However it should be
mentioned that a ldquofakerdquo failure in Mohr Coulomb model can be a ldquorealrdquo failure in another model
that can more realistically simulate the deformation patterns of the structure In general even
though Mohr Coulomb can be successfully applied for detecting failure it is not very suitable for
the determination of the displacements and especially for soft soils such that the dike under
investigation is constructed by It should be also mentioned at that point that in this case study
the system is subjected to unloading conditions24 (ie decrease of effective stresses) in soft soil
layers after phase 4 which cannot be modelled successfully by the Mohr-Coulomb model In that
case other constitutive models such as the Soft Soil Model and the Hardening Soil model are
presumably more reliable for simulating such situations For further information for this model
a reference is made to Plaxis (2015) Therefore someone should be always critical and verify
the failure points based on hisher engineering judgement
Figure 711 Prerequisites for real failure in Plaxis
24 From phase 4 to 5 there is an increase in water pressures and thus due to constant total stresses there is a decrease of the effective stresses This can be characterised as an unloading situation
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 100
In order to make it more understandable the different failure points were illustrated in Plaxis
and their deformed mesh and total displacements are presented in Figure 712 (a)-(c) for each
failure direction As one can observe in the first two directions the dike crest settles down
whereas in the 3rd one the dike body seems to swell and move to the right (the blue lines that
surrounds the dike soil layers indicate the initial dike geometry) However even if the 2nd
direction gave a normal pattern of displacements the error message was ldquoNot enough load
stepsrdquo The combination of the variables of the 3rd direction was then tested again with the
Hardening Soil model (HS) in order to illustrate the deformations with a more suitable model
and it was proved that the deformations seem to be realistic as it is depicted in Figure 713 With
Mohr Coulomb only the 1st direction seems to comply with the diagram in Figure 711 and that
would properly be considered as the real design point
(a) Deformed mesh
(b) Total displacements |u| in m (shadings)
(c) Total displacements direction (arrows)
1st direction 2nd direction 3rd direction
Figure 712 Illustration of the design point in soil body failure
(a) Deformed mesh (b) Total displacements |u| in m (shadings)
(c) Total displacements direction (arrows)
Figure 713 Displacements pattern of the 3rd failure direction according to HS
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 101
However taking into account that the other directions showed characteristics of failing behavior
they should also be accounted for the determination of the importance factors Excluding them
from the analysis would not be wise as the error message or the deformation schematization can
be a limitation of FEM or Mohr Coulomb model respectively In any case the points that lead to
failure shall be treated critically in order to decide if they should be included or not In this case
the points from the three directions seem to be realistic as both strength and stiffness
parameters are low enough in order to lead to a potential failure Therefore they were finally
considered for the analysis and the importance factors were averaged over all the points as it is
depicted in Figure 714 (last graph at the right)
Figure 714 Importance factors α2 for soil failure
Another option is also to average over the 10 range of close to the lowest reliability index that
was calculated with the shortest distance method (ie Average 10 method) in order to come
up with the α2 values However such methodology can lead to averaging over points of the same
direction that are not very different with each other Therefore the rest of the failure points are
automatically omitted from the consideration This can be also verified from Figure 714 where
the pie chart of the Shortest Distance and the Average 10 are quite similar whereas the Average
all chart introduces the importance of other soil variables as well
From a general perspective the soil properties of the clay layer seems to be determinant
whereas also sand and Dikeold material play an important role to the soil failure More precisely
according to the Average all method the unit weight and the friction angle of the clay layer
turned out to be the most influencing while also the cohesion of the Dikeold material and
friction angle of the sand layer contribute to the failure domain From a qualitative point of view
the weakening and consequently the settlement of the Dikeold material pushes the subsoil
creating an additional surcharge for the underlying clay and sand layer which act like the
foundation soil the incapability then of the clay layer to withstand the overlying load due to the
low unit weight in combination with the low friction angle of the clay as well as the low stiffness
and strength of the sand layer can lead to the creation of an inner slip surface and thus failure
Since the clay layer can be considered as a foundation soil the importance of the friction angle
and the unit weight can be also explained from the Terzaghirsquos bearing capacity theory (Terzaghi
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 102
1943) where these two factors are crucial for the determination of the ultimate bearing capacity
of the soil However these can be better verified by conducting large scale experiments in dikes
and inspecting the sensitivity of the dikersquos stability towards the soil properties
75 Anchor Failure The reliability of the support system is estimated based on the anchor and the waling probability
of failure However the load on the waling is proportional to the anchor force as it appeared in Eq
57 Therefore after the estimation of the anchor reliability by calculating the probability of
exceedance of a certain admissible anchor force the waling can be designed deterministically
Besides the failure probability of the waling must be lower than the anchorrsquos (given that no
uncertainties of the strength and the structural properties are considered) For the anchor failure
the LSF was formulated as follows
(72)
where [kN] is the anchor force is the yield stress and the cross
section area The reliability method to be utilized in this analysis is FORM and the selection of the
random variables to be used was made according to the sensitivity analysis results and the
engineering judgment and they are listed in Table 72 In particular the variables from the second
step of the sensitivity analysis were considered which will also help at a later stage to compare
FORM sensitivity method and FAST In Table 74 the results of the reliability analysis with FORM
are listed Additionally the design point together with the importance factors of each variable is
presented It should be mentioned at that point that the sensitivity analysis helped to define a
starting point closer to the design point and thus reduce the computational time required for the
analysis More precisely the analysis lasted approximately 16 hours whereas by starting from
mean values from which by default FORM starts the iterations it would take almost a day
Table 74 Reliability results for the anchor failure with FORM
FORM Pf 28410-6 β 45 Number of LSF calls 119 Maximum number of iterations
100
Elapsed time (hr) 16
Design point Parameter value α-value Clay_phi [deg] 1814 -0059 Clay_nu [-] 033 0086 Clay_c [kPa] 1487 -0089 Clay_G [kPa] 41739 0911 Clay_Rinter [-] 051 -0021 Sand_Rinter [-] 069 -0051 Sand_nu [-] 030 0007 Sand_phi [deg] 3259 -0174 Sand_G [kPa] 1130770 0048 Dikeold_phi [deg] 2842 -0020 Dikeold_nu [-] 036 -0045
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 103
Dikeold_c [kPa] 1873 -0045 Dikeold_G [kPa] 243183 -0326 Dikeold_Rinter [-] 053 -0072
Figure 715 Importance factors α2 in anchor failure from FORM analysis
The influence coefficients in Figure 715 indicate that this limit state is governed by the shear
stiffness of the clay layer G Furthermore the shear stiffness of the Dikeold material seems to
contribute significantly to the anchor yield stress exceedance It can be concluded that the
problem is still in the elastic domain as far as the soil behaviour is concerned In case of
predominately plastic behaviour the strength properties of the soil (φ and c) become more
important In Figure 716 a demonstration of the design point as it was acquired by the FORM
analysis is presented The figures indicate that presumably due to the clay layer withdrawal in the
passive side as it is depicted in Figure 716(b) with the red shadings sheet pile moves to the right
and subsequently anchor is tensioned enough to reach its yield stress25
(a) Deformed mesh (b) Total displacements |u| in m (shadings)
(c) Total displacements direction (arrows)
Figure 716 Design point illustration from the anchor failure
25 A better representation of the soil-anchor interaction could be possibly obtained if the anchor was modelled as a node-to-node element and the anchorage body as an embedded element
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 104
The calculations were repeated with the 6 most influencing parameters both with FORM and DS
in order to investigate the effect of reducing the number of random variables and in order to
validate the results of FORM analysis In Table 75 the reliability results of respectively FORM and
DS are presented while in Figure 717 the importance factors of both methods are juxtaposed The
design point and the importance factors of DS were estimated based on the Average 10 method
(see section 71)
Table 75 Reliability results of FORM and DS with reduced variables in anchor failure
FORM Pf 8510-6 DS Pf 1510-5 β 43 β 42 Number of LSF calls
28 Number of LSF calls
776
Maximum number of iterations
100 Number of iterations
100
Elapsed time (hr) 034 (starting point close to the design point)
Elapsed time (hr)
115
Design point Design point Parameter value α-value value α-value Clay_nu [-] 033 0119 033 0242 Clay_c [kPa] 1404 -0026 1535 -0212 Clay_G [kPa] 43236 0928 44548 0807 Sand_phi [deg] 3163 -0110 2954 0203 Dikeold_G [kPa] 239465 -0330 251723 -0366 Dikeold_Rinter [-] 053 -0064 060 -0263
Figure 717 Importance factors α2 with FORM and DS analysis respectively
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 105
As it can be noticed comparing FORM results from Table 74 and 75 the reliability index changes
from β=45 with 14 parameters to β=43 with the 6 most influencing parameters This verifies the
importance of the most influencing parameters that it was found in the analysis with the 14
parameters The small discrepancy though between the two reliability indexes might be
attributed to the interaction effect of the soil variables on the output performance
The estimation of failure probability with DS is almost the same with FORM As far as the
importance factors are concerned DS also came up with the conclusion that the shear stiffness of
the clay and the Dikeold material are the most crucial soil properties for the anchor stress level
However according to DS the stiffness of clay ν and the interface strength Rinter between the
Dikeold material and the sheet pile wall seem to also contribute to the limit state of the anchor
The reason why the strength of the clay layer under the dike and the Dikeold material (the part of
the dike that comes in contact with the anchor and the sheet pile wall) are the most important
variables can be explained by the principle of soil arching According to this principle the soil
columns on both sides of the rigid sheet pile wall are more compressive than the soil columns on
the top of the wall (ie overlying dike body) because of the higher stiffness of the wall when
compared with soils As such soil columns on both sides tend to settle more than the soils on top
of the rigid wall and this differential settlement causes a downward shear force acting along the
sides of soil columns on top of the wall As such the vertical load on the wall becomes larger than
the sole weight of soil columns on its top and the anchor that actually receives the most of this
vertical load reaches eventually its yielding stress
Last but not least it should be mentioned that the response surface which was constructed based
on the SA worked efficiently in most of the Plaxis errors cases in a sense that it gave reasonable
results of the anchor stress level
76 Sheet pile wall failure For the reliability analysis of the sheet pile wall the strength property ie the yield stress was
considered constant and equal to 355 kPa Therefore the LSF that was applied for this case was
formulated as follows
[
] (73)
where [kNm] and [kN] are the maximum bending moment and the axial force
that are exerted on the wall respectively is the elastic section modulus ( =1245 cm3m)
and is the cross sectional area ( =1201 cm2m for a AZ12 profile)
In sections 643 and 72 the maximum bending moments of the sheet pile wall were calculated
with the design values of the soil properties and the mean values respectively It has been
observed that the bending moments were not so significant and determinant in comparison with
the anchor force and the safety factor of the system that were also considered (in section 643) in
order to determine the structural properties
In a later stage a SA was carried out for the sheet pile wall that has been further discussed in
section 73 Figure 718 shows the stress level of the wall that was developed after 3150 iterations
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 106
by considering as random variables the ones that are defined in Table 72 As someone can
observe the stress level is generally quite low and even an order lower than the yield stress (ie
355105 Pa)
That small magnitude of bending moments can be probably attributed to the counteraction of the
active earth pressures from the passive ones that due to the homogeneity of the soil in both sides
of the wall are quite similar
Figure 718 Stress level [in Pa] of the sheet pile wall after 3150 iterations of the SA
The reliability analysis has been carried out with FORM method and considering as random
variables those listed in Table 72 However due to the low level of stresses that are developed
and the simultaneously high yield stress the method converged after almost 35 days of
calculations As it can be noticed in Table 76 the analysis came up with an extremely low
probability of failure It should be mentioned at that point that the convergence of the analysis
was successful thanks to the RS that was fitted This means that a Plaxis error of soil collapse
preceded and it activated the RS which implies that the soil failure is more likely to occur before
the sheet pile failure
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 107
Table 76 Reliability results for the sheet pile wall failure with FORM
FORM Pf 7210-38 β 13 Number of LSF calls 2619 Maximum number of iterations
100
Elapsed time (hr) 84 (35 days)
Design point Parameter value α-value Clay_phi [deg] 1164 0261 Clay_c [kPa] 586 0336 Clay_G [kPa] 14275 0663 Clay_Rinter [-] 052 -0016 Clay_gammaunsat [kNm3]
1848 0000
Clay_gammasat [kNm3] 1948 0000 Sand_phi [deg] 3020 0000 Sand_G [kPa] 1193650 0000 Dikeold_phi [deg] 1360 0406 Dikeold_c [kPa] 592 0438 Dikeold_G [kPa] 168798 0000 Dikeold_Rinter [-] 053 -0024 Dikeold_gammaunsat [kNm3]
2275 -0150
Dikeold_gammasat [kNm3]
2375 0000
This can be also verified from the importance factors that are depicted in Figure 719 where apart
from the shear stiffness of the clay layer which seems to be the predominant soil property for the
sheet pile LSF the strength parameters (φ and c) of the clay and the Dikeold materials also play
an important role This implies that the soil behaviour has already passed into its plastic domain
where the governing properties are the strength parameters before the sheet pile exceeds its
yield stress
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 108
Figure 719 Importance factors α2 of the sheet pile stress level in the plastic domain of soil
Nevertheless in order to investigate the potential most influencing soil properties for the sheet
pile LSF in the elastic domain as well the available output data of the FORM analysis have been
assessed The highest observed value of the sheet pile stress before its failure was around
235000 kPa After tracing the results it seems that when FORM tries to reduce further the
stiffness and strength properties of the soil layers soil collapses before the sheet pile failure The
vectors of the random variable that led to a stress level on the sheet pile wall in the range of
230000-235000 kPa were chosen in order to evaluate the importance factors The method that
was followed in that case is the Shortest Distance and the results are shown in Figure 720
Figure 720 Importance factors α2 of the sheet pile stress level in the elastic domain of soil
As it can be noticed from Figure 720 the results in the elastic domain regarding the influence of
the stochastic parameters are comparable with anchorrsquos analysis The stiffness parameter of the
clay layer (69) is the most dominant variable for the system behaviour whereas the stiffness of
the sand (15) and the Dikeold (4) layers seems also to have a significant influence This is not
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 109
surprising as the nature of the loads that generate high stresses on the anchor as well as the sheet
pile is the same namely the horizontal earth and water forces
Concluding in this case study the sheet pile wall is not expected to show a decisive behaviour to
the systemrsquos reliability according to the reliability analysis and the deterministic analysis (see
section 643) It should be stressed the fact that the fitted RS in a lot of calculations seemed not to
perform suitably for the sheet pile and this can be attributed to the fact that the it was created
based on the SA which was far from the plastic domain In that case a better fit of a RS is
recommended However such an action is still under investigation that for future considerations
in the design concept will be contemplated as a valuable technique for complex structures
modelled in FEM whose reliability analysis with conventional methods is time consuming and the
coupling may lead to numerical problems In the next section the analysis of the systemrsquos
reliability is presented
77 System Failure The estimation of the systemrsquos reliability was based on the systemrsquos fault tree of Figure 53
according to which the general LSF was formulated as the minimum of the three LSF that were
presented above for the soil the sheet pile and the anchor and it is given below
[ ]
[(
)]
(74)
According to this LSF the failure that occurs first is counted as a system failure The analysis was
carried out with DS reliability method and the variables that were taken as stochastic are listed in
Table 72 As it was discussed in section 51 the systemrsquos probability of failure differs from the
probability of flooding for which a combination between the systems reliability and the water
level uncertainty shall be made
In Table 77 the reliability analysis results are presented as they were deduced from DS Similarly
to the soil failure in Figure 721 the importance factors according to the Shortest Distance the 10
Average and the Average all methods are illustrated
Table 77 Reliability analysis results from the system failure
DS Pf 2410-3 β 30 Number of LSF calls 1807 Number of iterations (or directions)
300
Elapsed time (hr) 39
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 110
Figure 721 Importance factors α2 of the system failure
The analysis took almost one and a half days and it came with indicative valuable though results
about the system behaviour According to the averaged importance factors in Figure 721
(deduced by the Average all method) we can notice features from both soil and anchor failure
For example the unit weight (14) the friction angle (7) and the cohesion (8) of the clay
layer as well as the friction angle (8) and the stiffness (7) of the sand layer that played an
essential role in the soil failure analysis Moreover the stiffness of the clay (11) and the Dikeold
(9) layers that appeared to be important for the anchor and sheet pile LSF They seem to be also
essential in the system analysis
However the other two methods (Shortest Distance and Average 10) came up with the cohesion
and the interface strength of the Dikeold material to be the most influencing soil properties for
the system For that reason an investigation was conducted regarding the output results of the
system analysis in order to obtain an insight into the different reasons of the system failure More
precisely the failure points were divided into those that come from Plaxis Error and those that
come from the yield stress exceedance of the structural elements It was then observed that many
failure points originate from Plaxis errors that are related to ldquoNot enough load stepsrdquo (Error codes
102 and 112) rather than ldquoSoil seems to collapserdquo (Error code 101)
Therefore in order to estimate the actual important soil properties the vectors of the random
variables that led to errors of ldquoSoil seems to collapserdquo and to anchorsheet pile (SP) failure were
studied separately The importance factors are depicted in Figure 722 which actually verifies the
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 111
importance of the variables that were also deduced from the Averaged all method in Figure 721
(which are basically the unit weight and the stiffness of the clay layer and the friction angle of the
sand layer) It should be mentioned that the failure of the structural elements contributes 65
whereas the soil failure contributes 35 on the total probability of failure The latter would be
possibly lower than 2410-3 that was initially found as several failure points are not attributed to
real failure but to Plaxis numerical errors
Figure 722 Influence factors α2 for the structural elements (left) and the soil LSF (right) respectively
The system analysis can lead to valuable conclusions for the system behaviour as a whole and
individually for the different elements given though that an inspection of the results is followed
Moreover such an analysis is taking into account correlations among the three sub-failure
mechanisms implicitly For example the different LSFs may be affected by the same soil variable
such as the shear stiffness of the clay layer for the anchor and the sheet pile LSF In that case the
most dominant failure mechanism will occur first and the rest are excluded26 However during
the system analysis someone is not aware of what failed or not and therefore heshe is not
capable of distinguishing the weaknesses of the system that need further improvement Besides
with this technique is not also easy to inspect the failure mechanisms that occur in order to get
some knowledge about the systemrsquos behaviour Therefore the reliability analysis of the
components is recommended to be carried out separately if time is available
78 Comparison between Global Sensitivity Analysis and Local Sensitivity
Analysis (FORM)
The aim of this section is a comparison between the importance factors that were deduced in
section 75 according to a FORM analysis and the results of the global SA that were presented in
section 73 for the anchor failure
The output of the global SA is actually the total indices that show the influence of the random
variables The importance factors of FORM and the total indices of global SA cannot be directly
26 This method is not capable though of accounting for the simultaneous occurrence of different failure mechanisms as the system was considered a serial system
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 112
compared as they are different values More precisely the total index of a random variable is the
ratio between the total variance and the portion of the total variance that stems from the
uncertainty of the specific variable and its interactions with the rest (see Eq A10 in Appendix A)
On the other hand importance factors of FORM express the ratio between the normalized value of
the variable in the design point and the reliability index (see Eq 72) Therefore the comparison
between these values is carried out qualitatively in terms of the variables that are distinguished
as the most influencing
In particular according to global SA the shear stiffness G of the clay layer and the cohesion c of
the Dikeold material seem to be the most determinant soil properties Moreover the strength
reduction parameters Rinter of the Dikeold and the clay layers and the friction angle φ of the
Dikeold and the sand layer appear to have a relatively notable contribution to the anchors limit
state
On the other hand FORM analysis has concluded to the same results as the global SA with the
difference that Dikeold material affects the anchorrsquos stress level more with its shear stiffness
rather than its cohesion In Table 78 the results of both FORM and global SA are presented
separately for each soil layer and property (in blue for the global SA and in black for the FORM
analysis)
Table 78 Important soil properties according to global SA and FORM analysis for the anchor failure (radicradic=very important radic=less important but still considerable)
FORM global SA Sand Clay Dikeold
ν
radicradic
ϕ radicradic
radic
c radic radic radicradic
G radicradicradicradic radicradic
Rinter
radic radicradic
The validation of the most important parameters was carried out with the repetition of the
analysis with DS as it was showed in section 75 The small difference of the probability of failure
as well as the same tendency of the important factors of both DS and FORM verified the reliability
of FORM analysis Therefore the similarity of the results deduced according to global SA with the
ones given by FORM indicates the fact that in this case study the global SA should be considered
as a trustable way of SA for obtaining a first overview of the dominant variables that affect the
limit state under investigation Furthermore it is capable of giving a general insight into the
outputrsquos (in this case anchor stress level) behavior and the soil layers that can potentially stand as
predominant during the reliability analysis For example in order to obtain a first sense of the
reliability and the magnitude of the expected probability of failure someone can inspect the
output samples of the LSF deduced from the global SA In Figure 723 the samples of the anchor
stress level are presented as they were deduced after 3150 iterations during the global SA
(computation time =75 hr = 3 days) The fact that after 3150 iterations and variables
combinations the anchor did not reach its ultimate stress level (ie 450 kPa) implies that the
probability of failure might be small This can also help for qualitative decision making concerning
a systemrsquos behavior without conducting a reliability analysis Moreover global SA proved the high
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 113
interaction (large difference between first and total order indices) among the parameters that can
affect the reliability analysis and can help to decide the amount of the random variables
Figure 723 Anchor stress level during global SA (step 1)
However as someone can observe from Table 78 there are also some discrepancies concerning
the dominance of some variables This can be attributed to the fact that the influence of all
parameters in specific points such as the design point cannot be identified by the global SA
Moreover it is possible that the samples used in the SA could not reach the field where the
system has a plastic behavior and where presumably other variables may also affect the limit
state
As general conclusions for global SA some possibilities and limitations are listed below
Possibilities
bull Indicates the level of interaction effect on the output variance
bull Indicate the amount of the random variables that should be considered in the reliability
analysis
bull Gives a general insight into the output behavior
bull Valuable for future use in the construction of response surfaces
Limitations
bull Cannot rely on global SA for the evaluation of the most important parameters near the
design point
bull Difficult to reach plastic zone (large amount of samples are probably needed)
79 Discussion
In this chapter the reliability analysis results were presented for the soil the anchor the sheet
pile and eventually the system failure taking soil properties of the several soil layers as stochastic
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 114
Initially a SA was carried out for each system component in order to identify the major soil
properties and reduce the number of the stochastic parameters The reliability analysis was
performed either with FORM or DS and in both cases the design point and the importance factors
were estimated Eventually the list of subquestions that is proposed in section 11 was answered
The way they were treated and the corresponding conclusions are elaborated below
Which reliability methods are computationally tractable in combination with FEM
Before starting analyzing the reliability of the system it was essential to evaluate and
assess the performance of the available reliability methods in order to be aware of their
performance in terms of computational time and flexibility in different kind of LSFs After
a literature review and the implementation of some of them in simple LSF problems a
general overview has been obtained for the most applied reliability methods concerning
the previous knowledge required for the system behavior in order to provide the right
input parameters its accuracy and the computational effort Eventually FORM and DS
were chosen to be employed for this case study considering the no preliminary knowledge
of the LSF is required their reliable accuracy and the limited computational effort that
they have in comparison with other methods
The anchorrsquos and sheet pilersquos reliability has been evaluated with FORM method whereas
soil and systemrsquos reliability have been assessed by DS Carrying out a FORM analysis and
steering the algorithmrsquos parameters in order to converge proved to be more difficult than
it was expected However an efficient performance was finally achieved and the failure
probability of the anchor failure was successfully estimated together with the importance
factors (or else the α values) The analysis was repeated with the most important variables
and they were validated with DS The FORM analysis took approximately one and half
hour to converge (with 14 random variables) after providing a starting point that it was
estimated to be close to the design point whereas for the validation it took almost half an
hour On the other hand in case of the sheet pile wall FORM analysis took almost 4 days
The computational time also depends on the LSF to be approximated the number of
variable and the complexity of the problem that each time is analyzed As far as the soil
analysis is concerned DS took approximately one and a half day to converge for 300
iterations and 14 random variables
Which limit states are relevant for the envisaged application of retaining walls in dikes and
how can they be formulated using FEM analysis outcomes
The relevant LSFs of a dike with an anchored sheet pile wall were investigated For each
element ie the sheet pile the anchor and the soil body a separate LSF has been set up
The focus of this research was on the ULS that in case of the structural elements it was
represented by the ultimate yield stress of the anchor and the sheet pile respectively as
they were expressed in Eq 72 and 73 The determination of the soil failure has been also
elaborated and the available alternatives that can be considered in that case were
elaborated After a short evaluation the Plaxis definition of soil failure was selected that it
was formulated as it is shown in Eq 71 Finally the systemrsquos LSF was formed as a
combination of the aforementioned limit states More precisely the system was
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 115
considered as a serial and thus system failure occurs when one of the elements fails first
(see Eq 74)
How robust (convergence) are the tractable methods
The robustness of the different reliability methods can be controlled by the convergence
criteria that are defined by the user and depend on the required accuracy In case of FORM
the convergence criteria consist of a set of errors that has mostly to do with the
approximation of the reliability index and the limit state threshold whereas in DS the
convergence is controlled through the coefficient of variation (CoV)
In order to reassure the robustness of FORM the method has been validated with DS The
validation showed that the FORM method complies with the result of DS with a small
difference though in the design point and the importance factors FORM analysis seemed
to converge sufficiently and faster than DS However the choice of the optimization
algorithm plays a key role to the methodrsquos efficiency In this research the convergence
errors were manipulated for increasing the efficiency of the method More precisely it has
been observed that by relaxing the relative error and the constraint error (the explanation
of these errors and an evaluation of the different optimization algorithms are available in
Appendix A2) FORM convergence can be accelerated
The reliability analysis of the soil body and the system was performed with DS Similarly
to FORM analysis in order to improve the performance of the searching algorithm some
of the methodrsquos parameters should be stipulated for improving its efficiency and
reliability Therefore the maximum step size of the algorithms was tripled and it was
proved that for the same time period double iterations were feasible to be carried out
This does not mean that the larger the step size the better it is The choice of the step size
must be made wisely depending on the reliability index that it is expected and
respectively it is allowed to change for achieving a better performance In this case study
it was proved that after 300 iterations the CoV of DS tends to be high and it can therefore
still be improved by carrying out more iterations In Figure 724 an example of a DS
convergence graph concerning the system analysis is shown The middle red line shows
the probability estimation whereas the green lines show the boundaries At that example
the CoV was 040
DS seems to perform successfully both in soil failure with one LSF and in system failure
where 3 different LSF were included In such cases performing an approximation method
such as FORM would not be wise as the output from each iteration does not provide
information for going to the next that a FORM analysis normally needs In contrast it just
gives an indication of failure or not This means that only a random sampling method can
handle it and converge successfully as it was proved
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 116
Figure 724 Directional Sampling convergence graph (at confidence level 095) of the system LSF
What is the contribution of different uncertainties in the failure mechanisms of the system
According to the global SA and the reliability analysis of FORM and DS the friction angle
the cohesion and the shear stiffness properties appeared to have an influence on the
different LSFs This is not surprising as they all have large CoV of 10 20 and 25
respectively Thus their uncertainty generally showed an impact on the systemrsquos
behaviour It is also essential to mention that the global SA has shown that the uncertainty
of the interface strength Rinter (CoV=20) has also a notable contribution to the structural
elements LSFs However the FORM analysis which is more representative close to the
design point came up with a less important impact of that variable
The reliability analysis showed that as far as the structural elements (anchor and sheet
pile) are concerned the shear stiffness of the clay and the Dikeold material is determinant
and especially for the sheet pile the strength parameters of the soft soils seem also to play
a crucial role In contrast with the anchor where the soil body seems still to behave in an
elastic manner during the sheet pile failure soil appears to reach plasticity in more
locations where the strength properties (ϕ and c) are also influential The most dominant
failure mode appeared to be the anchor failure whereas the soil failure is always preceded
the sheet pile failure The dominance of the anchor failure can be explained from the fact
that the most influent parameter in this LSF is the stiffness that is also the variable with
the highest uncertainty (CoV=25) Therefore the plastic limit of the anchor is more
likely to be reached
Regarding the soil body reliability the analysis detected 3 different failure directions in
which the clay and the sand layer below the dike seem to be the most influent More
specifically the friction angle φ the cohesion c the unit weight γ of the clay layer the
friction angle of the sand layer and the cohesion of the Dikeold material turned out to play
an important role The decrease of the clay layerrsquos unit weight lead to the incapability of
the layer to withstand the overlying load of the dike body and at the same time the
weakening of the sand layer contribute to the creation of an inner sliding surface Even if
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 117
the uncertainty of the unit weight is small (CoV=5) it is still important for the stability
of the system as the maximum deformations appear in that layer (see Figure 712) Last
but not least the illustration of the several failure points in Plaxis allows someone to
inspect the failure mode and realize if it is a real failure of the structure or if it is a
weakness of the modelling
Finally during the system analysis multiple failure directions were detected whereas the
pie charts of the importance factors include soil properties that affect both soil and
structural elements as it was expected However due to Plaxis numerical errors several
failure points were found to be not realistic and they should be excluded in order to come
up with a more reliable failure probability which will be presumably lower The most
suitable method for estimating the importance factors though seems to be the Average all
method which is not distorted from the Plaxis numerical errors However it is
recommended that a further research be done on how realistic are the failure points
coming from Plaxis warnings and to verify if they comply with Figure 711 failure
definition
In Table 79 a synopsis of the most important soil properties for the different LSFs is
presented It is also apparent the common contribution that some of them have in
different LSFs In the last two columns of the table the results of the influencing soil
variables for the system LSF are shown with the Average 10 and Average all method
respectively For the reasons that were explained in the previous paragraph the Average
all method is considered to be the most suitable in that case and as it can be noticed from
the table it gives the most representative picture of the governing soil variables of each
LSF
Table 79 Important soil properties for the different LSFs (radicradic=very important radic=less important but still considerable the most important parameters for each LSF is painted in orange)
Soil material
Soil property
Anchor Sheet pile
Soil body
System Average
10
System Average
all
Clay
G radicradic radicradic
radic radicradic
φ
radic radicradic
c
radic radic radic
Rinter γ
radicradic
radicradic
Sand
G
radic radic
φ
radicradic radicradic
c
Rinter γ
Dikeold
G radicradic
radic
φ
radic
c
radicradic radic radicradic radic
Rinter radic
radicradic radic
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 118
γ
radic radic
Rinter was taken from the validation of FORM with DS Figure 717 (right graph)
taken from the sensitivity of the sheet pile in the soil elastic domain Figure 720
Can response surface techniques help to increase the efficiency and robustness of the
reliability model
In this research RS techniques were used only as a source for providing a reasonable
value to the FORM analysis in order to continue running in case of Plaxis errors These
response surfaces were constructed based on the SA results for both the anchor and the
sheet pile wall Apparently the domain of the SA is limited and most of the times it is
incapable of approaching the design point However RS worked efficiently in many cases
where Plaxis failed to give an output result such as in case of the sheet pile LSF Moreover
the RS could be used individually for the estimation of the failure probability (instead of
Plaxis calculations) but a better fit is advised to be firstly achieved that would be
representative in a wider plastic domain of the structural elements Therefore the RS
technique can be used to enhance the efficiency of the reliability analysis of a complex
system in the sense that it can be considered as solution for FEM errors What is more RS
techniques are a quite promising method for the reliability analysis of a system as it can
also be used alone without coupling with FEM and eliminate the computational effort
However this requires more investigation of such a technique
How can the current design approach be improved
In section 22 the current design concept of the case study that is under investigation in
this research is described At that section the design values of the moments and the forces
on the structural elements are presented as well as the required total safety factor FEM
=18 is given as it was initially calculated by Breedeveld (2011) It has been observed that
a significant overestimation is made in terms of the design values that might later lead to
heavy and cost inefficient structures For that purpose alterations was made on the
original case study in terms of the structural elements and some soil properties and a new
deterministic design was made in order to come up with a less heavy structure This
would also render it possible to investigate the reliability of the system rather than taking
zero probabilities or non-convergence Moreover it would finally help to see if there are
any potentials for optimizing the design by applying a full probabilistic analysis
According to the soil reliability analysis the probability of failure was found 1310-8 and
the reliability index β=55 As far as the anchor failure is concerned the probability of
failure found to be 28410-6 with a reliability index β=45 according to the FORM analysis
The sheet pile failure seem not be significant as it will more likely occur after the soil
failure Eventually the system analysis was performed the probability of failure was
found to be 2410-3 and the reliability index β=30 which is high enough to consider the
system safe
The systemrsquos reliability index that came out of this analysis cannot be directly compared
with the required one (ie β =40) because they stem from different probabilities The
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 119
former refers to the probability of failure due to the global instability given a design water
level whereas the latter refers to the probability of flooding that incorporates also the
water level uncertainty It must be mentioned that in reality there is no sheet pile wall
inside the dike section in the dike ring 43 but this section is under investigation for
reinforcement purposes as it does not comply with the safety standards after a recent
inspection that has been carried out Moreover the original case study taken from
Breedeveld (2011) was subjected to multiple alterations for the needs of the current
research Therefore a comparison of the reliability index with the required one from the
regulations would not be meaningful to make However what it can be compared in that
case is the order of magnitude of the reliability index After the analysis a β of 30 was
estimated for the probability of failure whereas for the overall probability of flooding a β
of 40 is required whose magnitude is the comparable with the former This can firstly be
considered a sort of the probability ldquovalidationrdquo in the sense that the analysis gave
realistic results and secondly it can imply that the structure that has been investigated in
this research is likely to be safe towards macro-instability as its coupling with the water
level uncertainty would certainly give a reliability index above 40 Furthermore a
reliability index of 30 indicates that there is room for optimizing the design concept of
such a system if someone considers also the fact that the soil strength of the subsoil was
further reduced for the needs of the current study
Due to the limitations of the modeling and the alterations of the original case study these
reliability indices should not be taken into account as an indicative number for the safety
of this dike section but rather as an order of magnitude in order to realize what the
potentials for improving the design procedure are This research can also contribute to
the way of designing not only the system but also each structural element as the reliability
of each can be calculated in advance According to the methodology developed in this
thesis a comparison can later on be made for a real case situation between the results of a
full reliability analysis and the current design procedure In that sense a tangible
improvement can be suggested for the design concepts and even to introduce a new way
for the future design of dikes with retaining walls
It should be stressed that the conclusions are based on the results of this case study with its
geometry and set of material parameters and the variation coefficients Changes in the geometry
the material parameters or the statistical properties could lead to different results The
generalization of these conclusions should thus be treated carefully
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 120
MSc Thesis A Rippi 121
8 Conclusions and Recommendations
81 Conclusions The advances in structural reliability analysis have made a large variety of methods available The
applicability and efficiency of these methods depends on the problem that is analyzed and on the
models that are used One the other hand the necessity for optimizing the design procedure of
several structures is becoming more and more intense A solution to this demand can be the
reliability analysis of a system for which rare information about its application on real-world
problems is available This work can contribute to making probabilistic analysis more accessible
for dikes reinforced with a sheet pile wall and at the same time to analyzing such a complex
system in terms of the factors that mainly influence the system behavior In particular the main
research question of this research was
How can the probability of failure due to global instability of a dike with a sheet pile wall modeled
by a Finite Element Model be analyzed
For that purpose in this research a reliability analysis was conducted by making use of a soil-
structure model that was simulated in FEM The output of the FEM calculations was utilized as an
input in the probabilistic model that was used in this thesis in order to assess the reliability of a
dike with an anchored sheet pile wall In that case the load was basically the earth and water
pressure whose impact was investigated on the anchor (LSF 1) the sheet pile (LSF 2) and the soil
body (LSF 3) respectively as well as on the system as a whole
It was demonstrated that the calculation effort of this kind of analysis does not necessarily have to
be extremely high Especially when previous knowledge about the system behavior and the
reliability models used is available This can be achieved by carrying out a sensitivity analysis In
this thesis a global SA was conducted that was proved capable of giving a representative
estimation of the most influencing variables and give a general overview of the systemrsquos
performance However close to the design point local SA such as FORM becomes more suitable
As far as the reliability of the structural elements is concerned their limit state was evaluated in
terms of the exceedance of their yielding stress The reliability assessment was carried out with
FORM from which the probability of failure and the influence factors were concluded More
precisely the anchor failure mode seemed to be the most determinant one with the shear
stiffness of the soft soils to be the most important soil properties The sheet pile failure is quite
rare and it is unlikely to occur before the soil failure since the soil body has already entered its
plastic domain
The soil and the system failure were challenging to assess as there are multiple criteria that
someone can choose In this thesis regarding the soil failure the Plaxis definition of soil collapse
was used and it was evaluated with DS In that case the unit weight and the strength parameters
of the soft soil layers seemed to be influencing The detected failure directions were investigated
Conclusions amp Recommendations
MSc Thesis A Rippi 122
in terms of their validity More precisely the failure points were illustrated in Plaxis in order to
visualize the failure and see if it is real failure or not (real if dike crest settles) That procedure
revealed the possibilities and limitations of FEM modeling The advantage is that a potential
failure can be detected and its visualization is possible On the other hand one should be aware of
the limitations related with the constitutive model that is used In this study the Mohr Coulomb
method was applied that although it can efficiently identify possible failure mechanisms the
representation of the deformations is not so realistic
The failure of the system was considered as a serial system of LSF 1 2 and 3 and it was also
assessed with DS The importance factors stemmed from the reliability analysis indicated soil
features that were important in all the LSFs However it should be stressed that the frequency of
the Plaxis numerical errors was such that quite a few failure points seem to be unrealistic In this
thesis the identification of those errors and their partially exclusion from the analysis was
achieved The latter was attained by the construction of the so-called Response Surfaces (RS)
based on a preliminary sensitivity analysis Such a technique found to be valuable in order to
overcome such a problem however a study for fitting more suitable RS in each problem is
recommended Nevertheless the further elimination of those numerical errors is strongly
recommended to be investigated in future research given the results from the current study
Retrieving the importance factors from a DS analysis was also a challenging part In this study
additional methods were developed in order to evaluate the importance factors and define the
design point after a DS analysis It revealed that alternative possible approaches of the
importance factors (Average 10 and Average all methods) can presumably provide a better
understanding of the systemrsquos behaviour Such methods seemed to be valuable for excluding the
ldquonoiserdquo of the possible fake failure points that originated from FEM numerical errors However
due to the inherent randomness of DS in terms of the selection of the directions the design point
shall be inspected and compared with the other failure points that were found Then someone can
contemplate if it is necessary to proceed with an averaging of the importance factors in order to
get a clearer view the variables influence Of course such a procedure takes time that is not always
available but according to the requirements of each research the most preferable assumptions
should be made
Last but not least the reliability analysis came up with relatively low probabilities of failure which
implies that there is still room for optimizing the design concept of a dike with retaining wall
More precisely the original case study of Breedeveld (2011) has been altered by reducing the
strength properties of the subsoil and installing a less heavy sheet pile wall after following a less
strict deterministic design It was then observed that even then the reliability analysis came up
with a respectively high reliability index It should be stressed that the failure probability of the
system should be also coupled with the uncertainty of the water level in order to obtain an overall
picture of the total probability of failure Nevertheless this research proves the potentials to
optimize the design of such a system which presumably will lead to a less expensive construction
However additional failure mechanisms such as piping internal erosion etc should be also
investigated and the design properties of the system (such as the length and cross section area of
the sheet pile wall) shall be based on them
Conclusions amp Recommendations
MSc Thesis A Rippi 123
82 Recommendations The following topics are recommended for further research
Additional research is required on FEM modeling both for the structural elements and the
soil body More in particular in this research it has been proved that Mohr Coulomb
model cannot realistically estimate the deformations in a soft soil In that case there are
other more advanced models such Hardening Soil model and Soft Soil model that are
capable of simulating the soil displacements and the stress-strain relationship closer to
the actual ones However these methods require more parameters to be specified and
their coupling with reliability packages might not be so robust and functional
As far as the structural elements modeling is concerned there are two major topics that
future research should focus on Firstly the anchor rod can be simulated in Plaxis as a
node-to-node element and at the bottom an embedded beam can be installed in order to
represent the anchorage body This would provide a better reaction with the surrounding
soil body and it could probably simulate the anchor-soil interaction more realistically
Another alternative would be to simulate it as a plate element in order to be able to test
the moments that are also developed on the anchor separately Secondly it is
recommended that the elements be investigated close to their plastic area and a reliability
analysis to be carried out with plastic elements instead In this research an attempt was
made for such an analysis but it appeared not to be functional in cooperation with FORM
analysis
Another issue that is recommended to be investigated is the inspection and the validation
of the failure that Plaxis is giving During the soil failure analysis different types of failure
were observed from which some might not be realistic but due to the modeling limitations
Therefore an attempt should be made in order to further eliminate the effect of Plaxis
numerical errors on the reliability analysis that can lead to fake failure modes A check of
the failure points can be also conducted by using a more suitable constitutive model for
the deformations rather than Mohr Coulomb However this is not always feasible for each
individual case Therefore the illustration of the failure mechanisms in large scale
experiments can give a better insight into the systemrsquos behavior and valuable knowledge
can be acquired of the expected failure modes In this way also 3D effects can be
investigated see what is their effect to the dike instability and at what extend they shall be
considered
Apart from the soil body the stresses and the displacements of the structural elements
shall be also verified It is thus recommended that field tests and measurements take place
in embedded elements like sheet piles and anchor in order to validate Plaxis results and
learn what is the most suitable manner for the their simulation (fixed-end-anchor node-
to-node element plate etc)
Concerning the reliability analysis of the specific case study a research on the potential
relevant LSFs is recommended for the different elements Specifically in this thesis the
LSF were chosen in order to represent the ULS of the element There is the possibility that
the SLS is used and certain acceptable deformations can be settled as a threshold for the
Conclusions amp Recommendations
MSc Thesis A Rippi 124
limit states However the choice of this displacement limit is not straightforward and a
qualitative research shall also be conducted in order to identify the impacts of different
values Besides it is not only the value that should be selected but also the location on the
structure that is going to be checked
It is also advised that the potential correlations among the soil parameters to be included
In Wolters (2012) and Bach (2014) correlation matrices are introduced for specific soil
properties that can be included in a future research and reveal what is their effect on the
reliability analysis The influence of correlations depends on what kind of variables are
correlated and what is the influence of those variables on the system behavior However a
preliminary analysis with independent variables is advisable to be carried out at the
beginning so as for someone to be able to analyze the effect of the variables individually
and then evaluate the correlation effect more wisely
It would be also advisable a comparison to be made between the design of a real case
according to a full probabilistic analysis and partial safety factors In that way a direct
insight into the advantages and disadvantages of each method can be gained and
fundamental improvements of the design concept can be made for dikes with retaining
walls
At this moment the most reliable way of conducting a reliability analysis is with sampling
methods like MC or DS However such methods are usually unattractive due to their large
computational time Thus the tolerance of sampling methods accuracy is lowered in order
also to reduce the number of iterations or approximating methods are used like FORM
analysis However even if approximating methods are used they are usually preferred to
be validated with sampling methods in order for their results to be acceptable A real
validation of the failure probability would mean that a structure should be experimentally
tested in different positions and under the same conditions in order to see if the
probability of failure that is calculated in each case is the same Therefore a real
validation of the failure probability cannot yet be initiated but the computational effort of
the sampling methods can be eliminated with the development of technology and the
improvement of the computer science
MSc Thesis A Rippi 125
References
Abdo T amp Rackwitz R (1990) A new beta-point algorithm for large time-invariant and time
variant reliability problems Proceedings of the 3rd IFIP WG 75 Conference Berkeley pp1-12
California USA
Ang AHS and Tang W (1984) Probability Concepts in Engineering Planning and Design Volume
IBasic Principles John Wiley and Sons New York USA
Baecher GB and Christian JT (2003) Reliability and Statistics in Geotechnical Engineering
Chichester West Sussex John Wiley amp Sons Ltd
Behaviour of Soils and Rocks GEO-ENGINEERING - CIE4361 (Vol 2015) Retrieved from
httpsblackboardtudelftnlbbcswebdavpid-2386735-dt-content-rid-
7926659_2courses32281-14150221-possibilities26limitationspdf
Breedeveld J (2011) Technisch Rapport Analyse Macrostabiliteit Dijken met de Eindige Elementen
Methode Bijlage Voorbeeldcase Deltares Delft the Netherlands
Brinkgreve RBJ (2015) Personal discussion TU Delft the Netherlands
Brinkgreve RBJ and Bakker HL (1991) Non-linear finite element analysis of safety factors In
G Beer JR Booker and J P Carter (Eds) Proceedings of the Seventh International Conference on
computer methods and advances in Geomechanics Rotterdam A A Balkema Brookfield
Cukier RI Fortuin M Shuler KE Petschek AG and Schaibly JH (1973) A study of the
sensitivity of coupled reaction systems to uncertainties in rate coefficients I Theory California
University USA
Cundall PA (2002) A discontinuous future for numerical modeling in soil and rock Published in
Proceedings of the third international conference Discrete Element Methods Numerical Modeling
of Discontinua Edited by Cook and Jensen Santa Fe pp 3-4
CUR (2005) Handbook Quay Walls Gouda the Netherlands
CUR-publication 190 (1997) Probabilities in civil engineering Part 1 Probabilistic design in theory
Stichting CUR Gouda
CUR 166 ndash Damwandconstructies (2005) Technical Recommendation 4e herziene uitgave 2005
The Netherlands
Deltaprogramma (2014) Synthesedocument Veiligheid Achtergronddocument B1 Ministrie van
Infrastructuur en Milieu
EN 1990 2002 Basis of Structural Design European Committee for Standardization
References
MSc Thesis A Rippi 126
Farrokh N (2007) Tools and Strategies for Dealing with Uncertainty in Geotechnics In DV
Griffiths and GA Fenton (Ed) Probabilistic methods in Geotechnical Engineering (pp 71-95) New
York NY Springer Vienna
Fenton GA and Vanmarcke EH (1990) Simulation of random fields via local average
subdivision J Eng Mech ASCE 116(8)1733ndash1749
Flood Defences HYDRAULIC ENGINEERING ndash CIE5314 (Vol 2015) Retrieved from
httpsblackboardtudelftnlbbcswebdavpid-2125084-dt-content-rid-
7621002_2courses29759-
131404CIE5314_FloodDefences_05_failure_mechanisms_stability_Schweckendiekpdf
Gemeentewerken-Rotterdam (2003) Alle triaxiaal proeven rotterdam Technical report Gemeentewerken Rotterdam Rotterdam the Netherlands
Ghanem R and Spanos P (1991) Stochastic Finite Elements ndashA Spectral Approach
Springer New York
Gonzaacutelez A Cuadrado A Josa A and Olivella S (2013) Safety assessment of limit equilibrium
methods for the design of cantilever and anchored sheet pile walls Escola de Camins Barcelona
Griffiths DV and Fenton GA (Eds) (2007) Probabilistic Methods in Geotechnical Engineering Lecture notes monographs edited works and proceedings in the field of Mechanics Engineering Computer Science and Applied Mathematics Italy Hasofer AM and Lind NC (1974) An exact and invariant first-order reliability format Journal of
Engineering Mechanics Division ASCE 100(1) 111-121
Hoek E Read J Karzulovic A and Chen ZY (2000) Rock slopes in civil and mining engineering
Published in Proceedings of the International Conference on Geotechnical and Geological
Engineering GeoEng2000 19-24 November Melbourne
Homma T and Saltelli A (1996) Importance measures in in global sensitivity analysis of
nonlinear models Reliab Eng Syst Saf 52(1) pp1-17
Hydra-Ring 12 (2015) Probabilistic toolbox for the WTI2017 Technical Reference Manual Version
12 Deltares Delft
Jaky J (1944) ldquoThe coefficient of earth pressure at rest In Hungarian (A nyugalmi nyomas tenyezoje)rdquo J Soc Hung Eng Arch (Magyar Mernok es Epitesz-Egylet Kozlonye) 355ndash358
Johansson E amp Sandeman E (2014) Modelling of Deep Excavation in Soft Clay A comparison of different calculation methods to in-situ measurements Chalmers University of Technology Gothenburg Sweden
Joint Committee on Structural Safety (1981) General principles on reliability for structural design
International Association for Bridge and Structural Engineering
References
MSc Thesis A Rippi 127
Larsen H Lubking P and Breedeveld J (2013) Ontwerp stabiliteitsschermen (type II) in primaire
waterkeringen (groene versie) Deltares Delft
Melchers R E (2002) Safety and risk in structural engineering Prog Struct Engng Mater
4 193ndash202
Moumlllman A amp Vermeer P (2008) Reliability analysis of a dike failure case study Elbe river
University of Stuttgart Germany
Nelson RB (1999) Lecture Notes in Statistics An Introduction to Copulas Springer New York
OpenTURNS 15 (2015a) Reference Guide OpenTURNSrsquo methods for Step C uncertainty
propagation pp155-161
OpenTURNS 15 (2015b) Use Cases Guide for the Textual User Interface pp190-192
Pandya N Marsdea E Floquet P and Gabas N (2008) Sensitivity analysis of a model for
atmospheric dispersion of toxic gases In B Braunschweig and X Joulia (Ed) 18th European
Symposium on Computer Aided Process Engineering-ESCAPE18 (pp1143-1148) Elsevier BV
Phoon K (Ed) (2008) Reliability-based Design in Geotechnical Engineering Computations and Applications Taylor amp Francis Britain
Plaxis 2D (2015a) Tutorial manual Delft the Netherlands
Plaxis 2D (2015b) Reference Manual The Netherlands
Plaxis (2015) Material Models Manual The Netherlands
Powell MJD (1994) A direct search optimization method that models the objective and
constraint functions by linear interpolation In S Gomez amp JP Hennart (Ed) Advances in
Optimization and Numerical Analysis (pp51-67) Dordrecht Kluwer Academic
Rankine W (1857) On the stability of loose earth Philosophical Transactions of the Royal Society
of London Vol 147
Saltelli A (2004) What is Sensitivity Analysis In A Saltelli K Chan and EM Scott (Ed)
Sensitivity Analysis (pp 3-13) John Wiley amp Sons Publication
Saltelli A S Tarantola F Campolongo and M Ratto (2004) Sensitivity Analysis in Practice
A Guide to Assessing Scientific Models John Wiley amp Sons Ltd
Schittkowski K (1985) NLPQL A Fortran subroutine solving constrained nonlinear
programming problems Annals of Operations Research 5 485-500
Schweckendiek T (2006) Structural Reliability applied to Deep Excavations Coupling Reliability
Methods with Finite Elements TU Delft Delft
References
MSc Thesis A Rippi 128
Schweckendiek T Vrouwenvelder ACWM Calle EOF Jongejan RB and Kanning W (2013)
Partial Factors for Flood Defenses in the Netherlands Modern Geotechnical Codes of Practice ndash
Development and Calibration Fenton GA et al (eds) Special Geotechnical Publication Taylor amp
Francis
Seed RB Bea R G Abdelmalak R I Athanasopoulos A G Boutwell G P Bray J D Briaud J-L Cheung C Cobos-Roa D Cohen-Waeber J Collins B D Ehrensing L Farber D Hanemann M Harder L F Inkabi K S Kammerer A M Karadeniz D Kayen RE Moss R E S Nicks J Nimmala S Pestana J M Porter J Rhee K Riemer M F Roberts K Rogers J D Storesund R Govindasamy A V Vera-Grunauer X Wartman J E Watkins C M Wenk Jr E and Yim SC (2006) Investigation of the Performance of the New Orleans Flood Protection Systems in Hurricane Katrina on August 29 2005 Volume I Main Text and Executive Summary New Orleans USA
Steenbergen RDJM (2013) Lecture notes Probabilistic Design Delft the Netherlands TU Delft
TAW (Technische Adviescommissie voor de Waterkeringen) (2004)Technisch Rapport Waterspanningen bij dijken The Netherlands
Terzaghi K (1942) Theoretical Soil Mechanics New York Wiley ISBN 978-0-471-85305-3 Teunissen JAM (2009) SBW Analyse macrostabiliteit van dijken met Eindige Elementen
Modellen Achtergronden bij Activiteit EEM 04a Opstellen stappenplan Deltares the Netherlands
Terzaghi K (1943) Theoretical Soil Mechanics John Wiley and Sons New York
Turaacutengi T amp Rabitz H (2004) Local Methods In A Saltelli K Chan and EM Scott (Ed)
Sensitivity Analysis (pp 81-99) John Wiley amp Sons Publication
Vergouwe R Huting RJM and van der Scheer P (2014) Veiligheid Nederland in Kaart 2 Overstromingsrisico dijkring 43 Betuwe Tieler- en Culemborgerwaarden The Netherlands
Vrijling JK Schweckendiek T and Kanning W (2011) Safety Standards of Flood Defenses Keynote at ISGSR 2011 (International Symposium on Geotechnical Safety and Risk) Munich Vogt Schuppener Straub amp Braumlu (eds)
Vrijling JK and Gelder PHAJM van (2002) Probabilistic Design in Hydraulic Engineering Lecture notes TUDelft Faculty of Civil Engineering and Geosciences Delft Vrouwenvelder ACWM Steenbergen HMGM and Slijkhuis KAH (1999) Theoriehandleiding PC-Ring Delft the Netherlands TNO
Waarts PH (2000) Structural reliability using Finite Element Analysis An appraisal of DARS
(Directional Adaptive Response surface Sampling) Delft University of Technology Delft the
Netherlands
Wyllie DC amp May CW (2004) Rock Slope Engineering (4th edition) Spon Press London
MSc Thesis A Rippi 129
Appendix A OpenTURNS features
In this Appendix an overview of the most relevant of OT is given Firstly in section 112 an
introduction of the applicable methods in SA is made while in section 52 the results from such an
analysis are presented In Appendix A1 a more detailed elaboration is made of how the FAST
sensitivity method that was applied in this project works
In Appendix A2 an explanation of the different optimization algorithms available for FORM
analysis is given The theory behind these algorithms and their convergence criteria are discussed
while finally an evaluation of their performance after applying them in case of the anchor failure
is presented This evaluation has been conducted in order to decide about the most relatively
reliable algorithm that should be adopted for the reliability analysis
Finally in Appendix A3 the types of probability distributions that have been used in this study
for the random variables and which are available in OT are displayed
A1 Fourier Amplitude Sensitivity Test (FAST) In this section the extended FAST sensitivity method is discussed as it is also described in the
Reference Guide of OpenTURNS 15 (2015) FAST is based on the Fourier decomposition of the
variance of the model response ( ) the latter being represented by its Fourier expansion
is an input random vector of independent components Its key idea is to
recast this representation as a function of a scalar parameter by defining
exploring the support of the input random vector
For each input the same procedure is realized in three steps
1) Sampling
Deterministic space-filling paths with random starting points are defined ie each input Xi is
transformed as follows
( ( )) (A1)
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 130
Figure A1 Search curves in the input space as they are defined from the transformation (the blue
dot indicates the direction of the current search path)
where is the number of input variables N is the length of the discretization of the s-space with
s varying in (-π π) by step 2πN is a random phase-shift chosen uniformly in [0 2π]
is a set of integer frequencies assigned to each input Xi The frequency
associated with the input of interest is set to the maximum admissible frequency satisfying the
Nyquist criterion (which ensures to avoid aliasing effects)
(A2)
with M the interference factor usually equal to 4 or higher It corresponds to the truncation level
of the Fourier series ie the number of harmonics that are retained in the decomposition realised
in the third step of the procedure And the maximum frequency of the complementary set of
frequencies is
(A3)
with the index ldquo-irdquo which meaning ldquoall but irdquo
2) Simulations
Output is computed such as
Then is expanded onto a Fourier series
sum [ ] (A4)
where and are Fourier coefficients defined as follows
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 131
int
(A5)
int
(A6)
These coefficients are estimated thanks to the following discrete formulation
sum
(A7)
sum
(A8)
3) Estimations by frequency analysis
The quantitative measure of sensitivity is represented by Sensitivity Indices The first-order
sensitivity index of an input pi is the measure of the main (direct) effect of pi on the output
variance (where i ne j) the second-order sensitivity indices measures the interaction effect of
pi and pj on the output variance Other higher-order indices are defined in the same manner The
total sensitivity index is the sum of all sensitivity indices involving factor pi (Homma amp Saltelli
1996) For example the total sensitivity index of factor 1 for a model with 3 input factors is
given as
The first-order (Si) and total ( ) indices can be interpreted as following (Pandya et al 2008)
high pi is an influent parameter
and both small pi is not an influent parameter (neither alone nor in interaction with
other parameters)
and nearly the same no interaction of pi with the other parameters
and very different high interaction of pi with the other parameters
Total indices are especially suited to apportion the model output variation to the input factors in a
comprehensive manner The FAST method calculates the first-order and the total sensitivity
indices whereas Sobolrsquo method in addition to these also provides all higher-order sensitivity
indices to determine quantitatively the interaction between parameters However the
computational cost and calculation time of Sobolrsquos method tends to be higher than that of the
FAST method
Combining equations A1-A8 the first-order indices are estimated as follows
sum (
)
sum (
)
(A9)
where is the total variance the portion of D arising from the uncertainty of the ith input and N
the size of the sample using to compute the Fourier series
Subsequently the total-order indices are estimated as follows
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 132
sum (
)
sum (
)
(A10)
where is the part of the variance due to all inputs except the ith input
A2 Optimization Algorithms in FORM
Principles of optimization algorithms
The possible optimization algorithms in OT are
Abdo-Rackwitz (ARF)
Cobyla
Sequential Quadratic Programming (SQP)
Below a description of the main principles of the optimization methods mentioned above is
quoted ARF and SQP methods are elaborated according to Abdo and Rackwitz (1990) while
Cobylarsquos principles according to Powell (1994) Finally an evaluation of the performance of the
different methods follows after applying them in case of the anchor failure
The general optimization problem (objective function) can be written in the classical form
subject to inequality constraints (constraint function) (A11)
The Lagrangian function of the general problem is defined by
sum (A12)
where are the so-called Lagrangian multipliers The SQP method (developed by Schittkowski
(1985)) replaces the original problem by a sequence of quadratic programming problems which
are exactly solvable and which approximate the original one This is done by approximating the
Lagrangian function by its second order Taylor expansion in an initial point u0
(A13)
where
sum
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 133
sum
sum
in which is the gradient operator and represents the Hessian matrix27 of the function f in
the point u0 The optimality conditions for any iteration point k of the sequence of quadratic
expansions are
sum [ sum
] (A14)
(A15)
The exact calculation of the Hessian matrix is generally too expensive and cannot be efficiently
implemented for a general case Therefore the gradient information obtained in each point
during iteration is used to build up an approximation of this matrix using one of the known
update formulas
The new iteration point is defined by
(A16)
where is the step length and is a direction in which a line search is performed The process
stops when the optimality conditions of the original problem are satisfied
The most time consuming part in this algorithm is the updating of the Hessian matrix and the
solution of the system of equations A fair approximation of the Hessian of non-quadratic
functions is also obtained with about n updates of the matrix This means that the approximation
used in the few (say ten) iterations to reach convergence cannot be very good when the problem
has large number of variables The rounding errors during the updating process in large problems
can make the approximate Hessian to become singular Close to singularity the search direction
can be significantly distorted In this case the algorithm has to restart the iteration with a unit
Hessian matrix in the point where singularity occurred
The posterior algorithm ARF developed by Abdo and Rackwitz (1990) overcomes this problem
by obtaining a constant approximation of the true Hessian matrix Only the contribution of the
objective function to the Hessian is considered The numerical decomposition of the matrix
contains the scalar product of the gradients of the constraints as elements in each iteration
27 Hessian is a square matrix of second-order partial derivatives of a scalar-valued function or scalar field It describes the local curvature of a function of many variables
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 134
Cobyla algorithm has been developed by Powell (1994) and it constructs linear polynomial
approximation to the objective and constraint functions by interpolation at the vertices of
simplices (a simplex in n dimensions is the convex hull of n+1 points n being the number of
variables) It generates the next vector of variables from function values at the vertices
of a nondegenerate simplex in In this case there are unique linear
functions and say that interpolate f and at the vertices
and the optimization problem A15 by the linear programming problem
( ) (A18)
( )
The iterative use of expression A18 puts this method in the class of ldquosequential linear
programming algorithmsrdquo It is also essential to notice that in this method the gradients are
derived by interpolation at the vertices of simplices and not analytically as it happens in SQP and
ARF
Additionally Cobyla algorithm makes use of a trust region bound ρ so that the trust region
condition on the new vector of variables is
(A19)
Such a region gives the user some control over the steps that are taken automatically and which
respond satisfactorily to the fact that there may be no finite solution to the linear programming
problem The algorithm also employs a merit function of the form
( ) ( ) [ ( ) ] (A20)
in order to compare the goodness of two different vectors of variables Here is a parameter that
is adjusted automatically depending on the how close to optimizing the objective function the
analysis is and the subscript ldquo+rdquo means that the expression in square brackets is replaced by zero
if and only if its value is negative so we have ( ) ( ) whenever is feasible Parameters
and are changing ( only reduces) automatically according to the improvement of the
optimization problem that sequential vectors might cause Therefore if the change in the merit
function does not happen to improve the optimization then these values are changing Especially
regarding the parameter user gives the initial and the final values of it namely and (it
is recommended that be a reasonable change to make the variables for a coarse exploration
of the calculation while should be approximately the required distance from the final vector
of variables to the solution of the optimization problem)
Convergence criteria
The convergence of the aforementioned algorithms is controlled by the evaluation of the
following errors expressed in the standard space (for the relationship between the real space (x-
space) and the standard space (u-space) see section 422) (OpenTURNS 15 (b) 2015 pp 190-
192)
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 135
The absolute error which is the distance between two successive iterates
(A21)
The constraint error which is the absolute value of the limit state function minus the
threshold
(A22)
The relative error which is the relative distance between two successive iterates (with
regards to the second iterate)
(A23)
The residual error which is the orthogonality error (lack of orthogonality between the
vector linking the center and the iterate and the limit state function)
( ) (A24)
The algorithm converges if all the final error values are less than the maximum value specified by
the user The convergence can be also controlled by altering the maximum number of iterations
which should be higher in case of stricter required errors What is more there is the possibility to
manipulate the starting point that the algorithm uses for its first iteration Usually the default
starting point in a FORM analysis consists of the mean values of the random variables However if
there is already some knowledge about the potential design point it can be set as the starting
point of the algorithm in order to accelerate the convergence and save a lot of time
The aforementioned errors the maximum number of iterations and the starting point were
manipulated during the analysis in order to understand the functionality of the different
algorithms and to improve their efficiency
Evaluation of the algorithms performance
An evaluation of the above optimization algorithms has been made in order to make a decision of
the most suitable for the current project In principle a literature overview regarding the
performance of the algorithms is presented Then the 3 algorithms were tested considering the
anchor failure limit state function (see Eq 55) with a stress limit of 450000 Pa and their results
are discussed below
According to Abdo and Rackwitz (1990) the SQP method seems to have better convergence
behavior for highly curved constraint functions which is especially true when very expensive
structural state functions are involved such as finite element structural analyses However the
mentioned algorithm fails to reach convergence at a problem dimension (amount of random
variables) of around n=50 Therefore for increased dimensionality SQP is less efficient in terms
of storage and CPU time compared with other methods due to singularity of the updated Hessians
matrices In OT a warning is introduced since an analysis has started that a default
implementation of Hessian is being used and that the computation can be severely wrong
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 136
Abdo and Rackwitz (1990) also proved that ARF algorithm can handle problems of up to 2000
variables while the storage requirements and the CPU time are much smaller than with SQP
Therefore optimization problems of high dimensionality can be evaluated
In OT a warning appears at the beginning of a FORM calculation saying that a default
implementation of the gradient is being used and that the computation can be wrong A similar
message also appears for the SQP algorithm saying that a default implementation of the Hessian
matrix is used and that the computationrsquos results might be unreliable Moreover in both
algorithms it is advisable to check the values of the random variables that were used during the
analysis in order to ensure that they fluctuate within acceptable boundaries (Figure A4 is an
example of a variable that surpassed the minimum boundary at some stage)
Firstly the Cobyla algorithm was applied on the case study in combination with a FORM analysis
Cobyla does not require a gradient evaluation of the limit state function and does not use a default
implementation of the Hessians matrix as ARF and SQP does while the computational time is
quite small comparable to the aforementioned methods
The Cobyla algorithm was tested with (i) 3 (ii) 10 and (iii) 19 parameters respectively in order to
check its efficiency28 As it has been mentioned in the previous section the convergence criteria to
be steered are the errors (absolute relative residual and constraint) the maximum number of
iterations and the starting point At first the default errors were used with a value of 10-3 but it
was realised that such an order of errors lead to non-convergence of the algorithm Therefore the
algorithm was then checked after relaxing the maximum errors to values of 01 05 10 50 and
100 Finally the combination of the errors that led to convergence was 01 05 01 02 for the
maximum absolute relative residual and constraint error respectively After the analysis the
number of LSF calls in case (i) (ii) and (iii) was 25 even if the maximum one was set to 100 and
the probability of failure was around 035 The probability of failure was quite high and that is
why an inspection of the output file was carried out In Figures A2 and A3 the values of the
anchor limit state function are presented in the right graph during the iterations It was then
observed that the threshold of 450 kPa was never surpassed during the iterations
28 All of the algorithms were first checked with very simple LSF (linear and quadratic) and their results were verified by MC and DS
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 137
Figure A2 Friction angle and cohesion progress during a FORM analysis with Cobyla and with 3 random variables The figure at the right illustrates the limit state function evolution
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 138
Figure A3 Friction angle and cohesion progress during a FORM analysis with Cobyla and with 19 input parameters The last figure at the right illustrates the limit state function evolution
An additional action was then taken by relaxing the maximum number of iterations from 100 to
200 but no improvement was noticed (the LSF calls were again 25) It was also given extreme
values to the errors and especially to the absolute constraint and relative error in order to check
the response of the algorithm but that did not lead to better results Furthermore the
parameter (see theory part for explanation) was manipulated but even then no further
improvement was noticed in terms of the results In Figures A2 and A3 the performance of the
algorithm during the analysis can be seen through the values of the variables that were assigned
during the iterations
As someone can notice from the graphs above it is obvious that Cobyla does not perform very
well in high dimensionality problems as the variables do not seem to vary a lot More specifically
in Figure A3 with the 19 parameters it can be noticed that the random variable does not change
at all apart from two peaks at the beginning and at the end of the analysis respectively The
incapability of the algorithm to change the parameters in order to reach the limit of the anchor
stress ie 450 kPa (while here a maximum of approximately 267 kPa is reached) leads to a poor
approximation of both the design point and the probability of failure In Figure A2 with only 3
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 139
parameters the variability of the parameter is higher during the analysis which leads to a better
approach of the limit (around 370 kPa which is larger than with the 19 parmaters) but even then
a failure probability of 035 was not a realistic value On the other hand Cobyla is easy to use only
for small number of variables (ie less than 9) with linear LSF otherwise the linear
approximations can be highly inefficient Apparently the limit state function of the anchor stress
level as a function of the soil parameters is much more complex in order to be approximated by
this algorithm
Therefore the other available optimization algorithms ie ARF and SQP were examined as well
Because of the similarity of the two algorithms only the ARF performance is presented below and
a link with SQP is made when it is needed
The ARF algorithm was tested initially with 19 parameters The settings for the errors that were
used were 01 for the absolute relative residual and constraint error respectively and the
maximum number of iterations was set to 100
Figure A4 Friction angle and cohesion progress during a FORM analysis with ARF algorithm and with 19 input parameters The figure at the right illustrates the limit state function evolution while the red spots indicate the points where the 450 kPa threshold was surpassed
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 140
The algorithm worked efficiently which means that the algorithm managed to converge towards
the threshold of 450 kPa of the limit state function as it is depicted in Figure A4 (see red dots) It
was also noticed that the random variables fluctuated a lot during the iteration steps even if the
amount was significant (19 parameters) The capability of the algorithm to vary all the 19
variables evenly and in a wide range makes it possible to find the real design point and thus to
come up with a more reliable failure probability The resulted probability of failure was 3810-9
and the reliability index was ϐ=58
However the analysis took almost 4 days to finish which is a long duration and the number of LSF
calls counted at 4151 Moreover even if the results were promising after tracing all the 19
parameters in order to inspect the algorithmrsquos performance it was found out that some variables
took unacceptable values29 during the analysis As an example in Figure A5 Poissonrsquos ratio
received negative values (see red circle) at the beginning of the iterations There were also other
such examples in the same analysis such as the friction angle which took sky-high values and the
interface strength which took negative values as well as values above 1 which is unacceptable It
should be mentioned that the design point of the algorithm was completely reasonable but even
then such a performance can prevent the algorithm from converging to the right design point
quickly and increase the computational effort
Figure A5 Poisson ratio progress during the FORM analysis with ARF algorithm
Taking into account the above implications there are two main concerns at that stage (a) the
high computational time and (b) the unreasonable values that the variables took Firstly in order
to reduce the computational time the maximum errors were relaxed More precisely it was
noticed that after relaxing the maximum relative error a lot which is more related to the
29 In some cases a value of ldquoinfrdquo was given to the variables that led to termination of the analysis without results
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 141
relaxation factor (in this case an extreme value of 1000030 was tried) and increasing the absolute
error to 05 the number of LSF calls was reduced to 724 and the analysis was finished after
almost 10 hours In Figure A6 the values that the cohesion and the friction angle took during the
iterative procedure while the red spots in the right figure indicate the points that are included in
the failure space Figure A4 shows the progress of friction angle after relaxing the errors
Figure A6 Progress of the friction angle during the FORM analysis with ARF algorithm with errors
05 10000 01 01 for the absolute relative residual and constraint error respectively
As a general conclusion after the test of different combination of maximum errors by relaxing the
constraint absolute and relative error too much successive iterations of a certain variablersquos value
seem to have a wide difference which can detain the algorithm from converging Therefore the
wisest in this case study is to keep the maximum errors below or around 10 except for the
relative and the constraint error that can relax more than 1 (depending on the case) in order to
reduce the computational time to the minimum possible
To prevent having unacceptable values of the parameters and in order to increase the efficiency of
the algorithm the random variables were reduced to 11 after a SA and some of them were
truncated in a certain range according to observations and experience until now However the
algorithm then started to give an error due to zero gradient of the objective function at a specific
point (similarly SQP gave an error due to zero hessian matrix at a specific point) while the upper
limit of the LSFs could not be reached As a countermeasure the errors were relaxed (05 10 05
10 for the absolute relative residual and constraint error respectively) in order for the algorithm
not to be trapped in a specific point while also the step of the algorithm was manipulated It
was finally proved that this kind of error stemmed from the fact that the structural elements were
considered as elastoplastic and thus a limit of the stress level was considered that misled the
algorithm The elements were eventually considered as elastic and the numerical problem was
solved
30 Practically such a value has the meaning that there is no limitation of the relative error during the optimization In the standard space (where these errors are expressed) a value also greater than 10 can be also quite high However here extreme values were given due to the time limit to test multiple values
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 142
Before figuring out that the elastoplastic behaviour was giving the problem Cobyla as a gradient
free algorithm was tested again as a potential solution but with changing the starting point close
to the design one that was found with ARF It is worthwhile to take a look into the performance of
the algorithm with 19 random variables in Figure A7 In that way we boost Cobyla to converge
and it eventually performed quite differently than in Figure A3 and without giving convergence
problems The computational time was estimated to be 9 hours and the probability of failure was
very small in the order of 10-120 However the validation of the failure probability with DS
showed that the result of Cobyla was far from correct and thus ARF was preferred for continuing
with the rest of the analysis For educational reasons a FORM analysis with ARF was again
performed after changing the starting point and it was found that the computational time was
reduced to 4 hours Therefore it is still worthy to be aware of the fact that by changing the
starting point the performance of the algorithm can improve a lot However that requires a prior
knowledge of the potential design point that usually does not exist
Figure A 7 Cobyla performance after changing the starting point closer to the failure point
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 143
It must be also mentioned that by reducing the number of the variables to 11 as it was mentioned
before the algorithms could not converge to the errors that has been set and the computational
time increased compared with the case of 19 variables Taking into account the high interaction
among the variables that was proved in section 73 through the SA it can be possible that the
combination of the soil variables could lead to failure rather than the variables individually
Therefore an analysis was tried again with the complete set of the parameters and it was realised
that the algorithm did converge to the limit value
A3 Distribution Types
Uniform Distribution
If the probability of a result is between a lower limit a and an upper limit b and if every result
between these limits is equally probable a uniform distribution should be used This distribution
is defined by the two limit parameters and the probability density function is written
(A25)
and the CDF is
(A26)
Figure A8 Probability density function of a uniform distribution with a=-2 and b=2
The mean and the variance of the distribution are also defined by the limit parameters
(A27)
(A28)
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 144
Normal Distribution
One of the most commonly used distribution types is the normal distribution or Gauss-
distribution and it is defined by its first two central moments the mean micro and the standard
deviation σ The PDF is given by
radic
(A29)
and its CDF is
int
radic
(A30)
Figure A9 Probability density function of a normal distribution with different mean and standard
deviation values (notice that the curve with micro=0 and σ2=1 belongs to the standard normal distribution)
The normal distribution with mean and standard deviation is called the standard
normal distribution In the most reliability methods such as FORM SORM DS etc the analysis is
carried out in the standard normal space which actually means that the stochastic parameters are
transformed into standard normal variables
Lognormal Distribution
If a variable X has a lognormal distribution the variable Y=ln(X) has a normal distribution The
probability density function is defined as
radic (
) (A31)
where and the mean and the standard deviation of the normally distributed random
variable Y Sometimes it is useful to work with the Yrsquos parameters rather than Xrsquos These
parameters can be expressed as follows
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 145
radic
(A32)
(A33)
where and are the mean and the standard deviation of Y In Figure A10 an example is given
of a log-normally distributed variable with different values of mean and standard deviation
Figure A10 Probability density function of a log-normal distribution with different mean and standard deviation values
Truncated Normal Distribution
The truncated normal distribution is the probability distribution of a normally distributed
random variable whose value is either bounded below or above or both
Suppose has a normal distribution and lies within the interval
Then conditional on has a truncated normal distribution whose probability
density function is
(
) (
)
(A34)
and otherwise and is the probability density function and the CDF of the standard
normal distribution similar to Eq (1) and (2) If
and
then the CDF is
defined as
(A35)
Then the mean value and the variance of the conditional will be respectively
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 146
(A36)
[
(
) ] (A37)
In the next figure the PDF histograms of a variable with a mean of 0 and a standard deviation 1
were plotted for two cases (a) variable is normally distributed (b) variable is truncated-normally
distributed with =-2 and b=2
Figure A 11 Truncated normal distribution in comparison with normal distribution
MSc Thesis A Rippi 147
Appendix B Plaxis 2D (2015) features
In this Appendix the basic features of Plaxis 2D that were treated in this thesis are discussed First
of all an introduction over the Mohr Coulomb model is made as the soil failure criteria stem from
that model Then the initial stress generation in Plaxis is discussed and how that was applied in
the current master thesis while also the φ-c reduction technique is described which was used in
the SA of the soil failure Last but not least a more detailed explanation of the interface strength
Rinter is given which represents the modeling of the soil-structure interaction in Plaxis and it has
been considered as a random variable in the current project
B1 Mohr Coulomb failure criterion Soil tends to behave in a highly non-linear way under load This non-linear stress-strain behaviour
can be modelled at several levels of sophistication There are different constitutive models
available for simulating the behaviour of the soil whose parameters can change depending on the
level of sophistication that is required in each case
As someone can notice in Table B1 different models can be applied in different cases of soil
behaviour Hardening Soil model is becoming more and more interesting in soil structures as it
has been proved to approach the plastic behaviour of the soil and the non-linear stress-strain
relationship in a reliable and promising level Moreover UBCSAND31 and the Hypoplastic model
are more sophisticated models capable of representing several aspects of soil behaviour
however the amount of the model parameters is large and for that reason they are recommended
only in case that enough soil data are available for determining these parameters
Table B 1 Features of material models (source Behaviour of Soils and Rocks 2015)
ModelFeature Elasticity Failure Hardening Softening Small strain stiffness
Cyclic loading
Liquefaction Creep
Mohr-Coulomb x x
Duncan-Chang x x
Hardening Soil x x x
HSsmall x x x
x
Modified Cam-Clay
x x x x
Soft Soil x x x
Soft Soil Creep x x x
x
UBCSAND x x x
x x
Hypoplastic (x) x x x
x (x)
31 UBC are the initials for University of British Columbia in Vancouver in Canada
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 148
In this master thesis the Mohr-Coulomb model is going to be applied which is simple but an
efficient model to describe the soilrsquos failures condition Below a description of that model is made
The Mohr-Coulomb model is a linear elastic perfectly plastic model (see Figure B1) The linear
elastic part is based on Hookersquos law while the perfectly plastic part is based on the Mohr-Coulomb
failure criterion formulated in a non-associated plasticity framework
The main principle of elastoplasticity is that total strains and strain rates consist of elastic and
plastic strain components
(B1)
(B2)
in which the elastic strain rate component abides by the Hookersquos law and thus it can be expressed
as follows
(B3)
where is the elastic stress-strain matrix and the effective stress rate The plastic strain
component is given by
(B4)
where is a scalar defining the magnitude of the plastic strains called plastic multiplier and
is
a vector describing their direction with to be the so-called plastic potential function
Figure B1 Basic principle of elastoplasticity
The occurrence of (elasto)plasticity can be indicated by the yield function which is
illustrated in Figure B2 in principle stress space and it is derived according to the Mohr-Coulomb
failure criterion expressed in the next formula
(B5)
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 149
Considering the sign of this yield function the different soil states can be recognised as follows
if flt0 (inside yield contour) pure elastic behaviour
if f=0 and dflt0 unloading from aplastic state (=elastic behaviour)
if f=0 and df=0 elastoplastic behaviour
if fgt0 (outside yield contour) impossible stress state
Figure B2 Mohr-Coulomb yield surface in principle stress space
Finally in Mohr-Coulomb model the calculation of the plastic strains is achieved by the so-called
non-associated flow rule where with the dilatancy angle
This makes a difference between the friction and the dilatancy angle which prevents the model
from overestimating potential uplift
Concluding this model uses 5 input parameters
Stiffness parameters E (Youngrsquos modulus) and v (Poissonrsquos ratio) as they are included in
the elastic stress-strain matrix according to Hookersquos law
Strength parameters c (cohesion) and (friction angle)
Dilative behaviour ψ (dilatancy angle)
The author is aware of the possibilities and limitations of this model In particular the model is
capable of
a first order approach of the soil behaviour in general
describing quite well the (drained) failure behaviour
It should be mentioned at that point that the Youngrsquos modulus E cannot be steered via the
Python interface (as it appears to be a readonly parameter) and thus the shear stiffness G was
considered instead that is defined as follows
[kPa]
f=0 (yield contour)
flt0
fgt0
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 150
On the other hand the model has also limitations that are referring to
the linear elastic behaviour until failure (no strain- or stress path-dependent stiffness
behaviour)
overestimation of the shear strength in case of an undrained behaviour
However this model is considered reliable for detecting failure and giving realistic stresses
magnitudes in the soil and structural elements
Moreover in this project a non-associative behaviour was assumed by considering a zero
dilatancy angle However it has been proved by Teunissen (2009) that in case of deep failure
mechanisms in a dike structure with soil materials of a friction angle above 30deg problems arise in
the numerical analysis in finite element calculations Particularly the strength in the sliding
surface decreases due to rotation and the ground stresses are relieving around it by releasing
energy which might lead to calculation implications These phenomena are not treated in this
thesis but it is advisable to check and review the results after such an analysis
B2 φ-c Reduction Technique In this section the ldquoSafety Calculationrdquo is applied as it is described in Plaxis (Plaxis 2D (a) 2015)
In the Safety Calculation the φ-c reduction technique (Brinkgreve amp Bakker 1991) is adopted
where the shear strength parameters tanφ and c of the soil as well as the tensile strength are
successively reduced until failure of the structure occurs The dilatancy angle ψ is in principle not
affected by the phic reduction procedure However the dilatancy angle can never be larger than
the friction angle When the friction angle φ has reduced so much that it becomes equal to the
(given) dilatancy angle any further reduction of the friction angle will lead to the same reduction
of the dilatancy angle The strength of interfaces if used is reduced in the same way
The total multiplier is used to define the value of the soil strength parameters at a given
stage in the analysis
(B6)
where the strength parameters with the subscript ldquoinputrdquo refer to the properties entered in the
material sets and the parameters with the subscript ldquoreducedrdquo refer to the reduced values used in
the analysis is set to 10 at the beginning of a calculation to set all material strengths to
their input values
The incremental multiplier is used to specify the increment of the strength reduction of the
first calculation step (this increment is by default 01 but user can change it) The strength
parameters are successively reduced automatically until all required steps have been performed
(by default these steps are set to 100 but this can be subjected to changes depending on the
situation) It must always be checked whether the final step has resulted in a fully developed
failure mechanism In other words if has reached an almost constant value while the
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 151
deformation is continuing that indicates an equilibrium state In that case the factor of safety is
given by
(B7)
B3 Initial Stress Generation The first construction phase in Plaxis was the initial stress generation before the dike has been
constructed In that phase for horizontal soil surfaces with any soil layers parallel to this surface
and horizontal phreatic level32 a K0-procedure can be applied where K0 is the ratio between
horizontal and vertical stresses
(B8)
In practice the value of for a normally consolidated soil is often assumed to be related to the
friction angle by Jakyrsquos empirical expression (Jaky 1944)
Using very low or very high K0-values in the K0 procedure may lead to stresses that violate the
Mohr-Coulomb failure criterion In this case Plaxis automatically reduces the lateral stresses such
that the failure condition is obeyed Hence these stresses points are in a plastic state and are thus
indicated as plastic points Although the corrected stress state obeys the failure condition it may
result in a stress field which is not in equilibrium
B4 Interface Strength In FEM for a proper modelling of soil-structure interaction interface elements can be added to
plates in order to simulate the thin zone of intensely shearing material at the contact between the
plate and the soil
The strength properties of the interface are linked to the strength properties of the adjacent soil
layer and they are reduced by a factor of as follows
(B9)
(B10)
(B11)
where index ldquoirdquo refers to the interfacersquos parameters whereas index ldquosoilrdquo refers to the adjacent soil
parameters The behaviour of the interfaces is defined according to the Coulomb criterion which
distinguishes between elastic behaviour where small displacements can occur within the
interface and plastic interface behaviour when permanent slip may occur
32 For non-horizontal surfaces a Gravity loading is recommended
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 152
When the interface is elastic then both slipping (relative movement parallel to the interface) and
gapping (relative displacements perpendicular to the interface) can occur The magnitudes of the
interface displacements are
(B12)
(B13)
where is the shear modulus of the interface is the 1-D compression modulus of the
interface is the virtual thickness of the interface (generated automatically from Plaxis) is
the elastic interface normal stiffness and is the elastic interface shear stiffness
The shear and compression moduli are related as follows
(B14)
(B15)
Taking into account the above relations the displacements can be highly influenced by the elastic
parameters and thus the strength reduction factor can have a considerable impact on the
systemrsquos behaviour In the Reference manual of Plaxis 2D (2015) a value of 23 is generally
recommended while a value of 1 would not normally be used It should be stated that in the
framework of this thesis there are no available measurements in order to define the sensitivity to
this parameter and hence it is considered uncertain variable following a truncated normal
distribution
Appendix C ndash NEN 6740 ndash Table 1
MSc Thesis A Rippi 153
Appendix C NEN 6740 - Table 1
MSc Thesis A Rippi 154
MSc Thesis A Rippi 155
Appendix D Input Files for the Reliability Analysis
Below an example of the input files that were used for the anchor the sheet pile the soil body and
the system failure is presented in Figures D1 D2 D3 and D4 respectively In particular
regarding the input file of the anchor failure the meaning and the scope of each line is briefly
explained in Table D1
Figure D 1 Anchor analysis input file (with FORM method)
APPENIDX D ndash Input Files for the Reliability Analysis
MSc Thesis A Rippi 156
Table D1 Explanation of input components
A Reliability method and its parameters eg FORM with a maximum number of iterations 100 a maximum absolute error of 05 a relative error of 1000 a residual error of 05 and a constraint error of 05
B Number of random variables
C Soil material as it is specified in Plaxis eg AKleiVenigOnder is the clay layer under the dike
D Soil parameter eg Rinter is the interface strength of the clay layer
E Probability distribution and its parameters eg Truncated normal distribution with mean 066 standard deviation 013 and bounds from 00 to 10
F Copula type
G I Limit state function I is the threshold of the function
H Response surface function that is activated in case of Plaxis errors
Figure D 2 Sheet pile wall analysis input file (with FORM method)
APPENIDX D ndash Input Files for the Reliability Analysis
MSc Thesis A Rippi 157
Figure D 3 Soil analysis input file (with DS method)
APPENIDX D ndash Input Files for the Reliability Analysis
MSc Thesis A Rippi 158
Figure D 4 System analysis input file (with DS method)
MSc Thesis A Rippi 159
Appendix E Characteristic and mean values
The characteristic value of a parameter implies that there is 5 probability that the real value is
higher (load) or lower (resistance) than the characteristic In Figure 22 the probability density
function of a standard normal distributed parameter (mean value equals to zero and standard
deviation 1) is depicted In case of such a distribution the characteristic value is 164σ higher than
the mean value
Figure E1 Characteristic value illustration for a normal distribution (source Wolters 2012)
In this study full probabilistic calculations are taking place in order to estimate the corresponding
structural reliability According to the probabilistic design concept the mean values and the
standard deviations are used instead of characteristic values In that case characteristic values
and shall be converted to their mean values and as follows
where
(E1)
(E2)
and are the values for load or resistance respectively to which a probability of (non-
)exceedance of 95 corresponds (non-exceedance for the load and exceedance for the resistance)
(for a standard normal distributed parameter this is 1645) The definition of the rest parameters
is given in section 423
MSc Thesis A Rippi 160
MSc Thesis A Rippi 161
Appendix F Reliability methods
In this appendix an overview of different reliability methods is given More precisely in F1 an
explanation is given concerning the generation of random samples that is applied in MC as well as
in other sampling methods In Appendix F2 more advanced sampling methods than crude MC are
presented that can be used as accelerating reliability methods in the sense that the computational
time can be reduced Finally in Appendix F3 some basic principles of the FOSM are elaborated
that can help to better understand the concept of FORM as well
F1 Generation of random samples in Monte Carlo
The non-exceedance probability of an arbitrary random variable is uniformly distributed between
zero and one regardless of the distribution of the variable In formula (CUR ndashpublication 190
1997)
(F1)
is the uniformly distributed variable between zero and one and is the non-exceedance
probability Thus for the variable X
(F2)
in which is the inverse of the PDF of X Using this formula a random number X can be
generated from an arbitrary distribution by drawing a number of from the uniform
distribution between zero and one
More or less the same way base variables of a statistical vector can be drawn from a known joint
probability distribution function However the joint PDF must then be formulated as the product
of the conditional probability distributions of the base variables of the vector In formula this is
( )
By taking m realizations of the uniform probability distribution between zero and one a value can
be determined for every
(F3)
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 162
If the variables are statistically independent this can be simplified to
(F4)
By inserting the values for the LSF one can check whether the obtained vector ( ) is
located in the safe area
F2 Other Sampling Methods
Quasi-Monte Carlo Simulation
As an alternative to pseudo-random number generation (used in Crude MC) the quasi-random
number sequences are introduced from where the method Quasi-MC took its name Quasi-MC is a
technique which approximates the probability of failure using low discrepancy sequences33
x1hellipxN instead of randomly generated sequences
If the domain of failure is the goal is to estimate the following
probability
int ( )
(F5)
The main principle of this method is the integral of the function on [ ] can be
approximated by using some low discrepancy sequence x1hellipxN as follows
int ( )
sum
(F6)
The low discrepancy sequence is generated in Δ according to Lebesgue measure and then may be
transformed to any measure micro thanks to the inverse CDF technique in order to approximate the
integral
int ( )
sum
(F7)
In Figure 4 a comparison is made between quasi-random sample and pseudo-random sample in
order to illustrate the difference in sampling for the two methods
33 The discrepancy of a sequence is low if the number of points in the sequence falling into an arbitrary set B is close to proportional to the measure of B as would happen on average (but not for particular samples) in the case of a uniform distribution
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 163
Figure F1 Example of Quasi random Sample (left) and Pseudo Random Sample (right) (source
Waarts 2000)
This method a priori provides an asymptotically higher convergence rate than traditional MC but
no general rule can guarantee a better efficiency of the Quasi-MC sampling than the classical MC
sampling The advantages of Quasi-MC sampling tend to disappear with the increase of the
number of dimensions of the input variables It is recommended to use the Quasi-MC technique
with high sampling sizes or with very low dimensionality (in OT this method is valid only if the
input random variables are independent) (OpenTURNS 15 2015)
Importance Simulation
The main idea of the Importance Sampling method is to replace the initial probability distribution
of the input variables by a more efficient one Efficient means that more events will be counted
in the failure domain and thus reduce the variance of the estimator of the probability of
exceeding a threshold In other words importance sampling starts with the observation that if we
are going to sample randomly we should distribute the points to put the most of them in the
region that contains information (ie the failure region) and to waste as few as possible
In this method a sampling density is used instead of the actual probability density function
For a MC simulation the probability of failure is calculated by
sum ( ( )) ( )
( )
(F8)
where ( ( )) ( )
( ( )) ( )
The problem is that prior knowledge is needed of the failure area which in most cases is not
available In Figure 5 examples of the different sampling methods are shown
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 164
(a) (b)
(c)
Figure F2 Comparison of Importance Sampling (a) MC (b) and Numerical integration (c) (source
Latin Hypercube Simulation
Latin Hypercube Sampling is a sampling method enabling to better cover the domain of variations
of the input variables thanks to a stratified sampling strategy (in OT this method is applicable in
the case of independent input variables) Stratified sampling tries to place the sampling points so
that more of them will be found in regions where the variance of the limit state function g(x) is
largest In Figure F3 a comparison is made between the LHS and MC in terms of their sampling
technique The sampling procedure according to Baecher and Christian (2003) is based on the
next steps
1 Divide the range of each variable into several intervals of equal probability
2 If there are D variables to be sampled each in k regions then for each variable the k
sampling points are placed in independent random order
3 K samplings are made by choosing the first randomized value for each variable the second
randomized value for each variable and so on until k sets of randomized variables have
been chosen
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 165
Figure F 3 Comparison between Latin Hypercube sampling and MC (source OpenTurns 15 2015a)
F3 First Order Second Moment (FOSM) Method The First-Order Second-Moment (FOSM) approach (Ang amp tang 1984) provides analytical
approximation for the mean and the standard deviation of a parameter of interest as a function of
the mean and standard deviations of the various input factors and their correlations
Consider Z to be a function of random variables x1 x2 hellip xn that is
(F9)
In general case are correlated with covariance matrix [C] ie [C]=[σ][R][σ] where [σ]
is a diagonal matrix of standard deviations and [R] is the (positive-definite and symmetric)
correlation matrix with diagonal elements Rii=1 and non-diagonal elements Rij=ρij (ρij is the
correlation coefficient between variables i and j) In scalar notation Cij= σiσjRij (Farrokh 2007)
The first order approximation to the mean variance and standard deviation of the Z function is
based on the first terms of a Taylor series expansion of Z The following first-order estimates of
mean and variance are obtained
[ ]
(F10)
[ ] (F11)
Where the vector b denotes evaluated at the mean values of ie
(F12)
If there is no correlation among the variables Eq 223 can be written as
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 166
sum
(F13)
Then the reliability index can be calculated in 6 steps (Baecher amp Christian 2003)
1 Identify all variables that affect the mechanism that is researched
2 Determine the best estimate (usually the mean value) of each variable E[ ] and use these
to calculate the best estimate of the function E[Z]
3 Estimate the uncertainty in each variable and in particular its variance Var[ ]
4 Perform sensitivity analyses by calculating the partial derivatives of Z with respect to each
of the uncertain variables or by approximating each derivative by the divided difference
5 Use the equation of Var[Z] to obtain the variance of the function Z
6 Calculate the reliability index
[ ]
(F14)
MSc Thesis A Rippi 167
MSc Thesis A Rippi i
MSc Thesis A Rippi ii
Preface
This thesis is the final challenge in the master Hydraulic Engineering at Delft University of
Technology The report ldquoStructural reliability analysis of a dike with a sheet pile wall Coupling
Reliability methods with Finite Elementsrdquo was completed at Deltares as a part of a larger
research project namely TO2 in collaboration with Toegepast Natuurwetenschappelijk
Onderzoek (TNO) I chose that subject first of all because it combines two things that I enjoyed
a lot at TU Delft probabilities and flood defences systems Secondly it was an opportunity for
me to get acquainted with FEM and geotechnics that I was always interested in
Different people have contributed to the successful completion of this thesis First of all I would
like to express my appreciation to my graduation committee for their guidance and particularly
my daily supervisor Timo Schweckendiek Together we had many fruitful and interesting
discussions on the subject that triggered and motivated me for keep searching Especially I
want to express my gratitude and my thanks to Ana Teixeira and Jonathan Nuttall employees of
Deltares as they were also some of the main contributors to this research They stood by me not
only as colleagues and thesis mentors but also as friends I would like also to thank the rest of
Deltares employees whom ensured a friendly and easy going working environment Last but not
least I want to thank my friends and especially Panagiotis Apostolidis and my family for their
love support and advice throughout all my studies
Katerina Rippi
Delft November 2015
MSc Thesis A Rippi iii
MSc Thesis A Rippi iv
Abstract
Some dike sections in the Netherlands failed to comply with the safety standards and one of the
most promising countermeasures is the construction of retaining walls inside the dike The
Dutch design codes for dikes with retaining walls rely on Finite Element Analysis (FEM) in
combination with partial safety factors However this can lead to conservative designs For this
reason in this research a reliability analysis is carried out with FEM calculations aiming to
demonstrate the feasibility of reliability analysis for such a soil-structure interaction problem
The case study concerns a (river) dike with an anchored sheet pile wall modelled in PLAXIS The
sensitivity and reliability analyses were enabled by coupling the uncertainty software package
OpenTURNS and PLAXIS through a Python interface The most relevant (ultimate) limit states
concern the anchor the sheet pile wall and global instability (soil body failure) The case was
used to investigate the applicability of the First Order Reliability Method (FORM) and
Directional Sampling (DS) to analysing these limit states Finally also the system reliability was
evaluated using sampling-based methods (DS)
Due to the considerable number of random variables before starting the reliability analysis a
sensitivity analysis was conducted for each limit state This indicated the most important soil
layers to be accounted for and the variables to be considered as stochastic The sensitivity
analysis and later on the reliability analysis were based on analytical formulations of the limit
state functions The anchor and the sheet pile limit states were formulated in terms of their
yield stress for global instability loss of equilibrium in the FEM analysis was used to define
failure Moreover the systemrsquos reliability was evaluated by taking into account all the three limit
states that were mentioned previously
The goal is to implement the coupling between FEM and reliability methods in order to analyse
the components of such a system (ie anchor sheet pile wall and dikersquos soil body) estimate the
probability of failure and identify the most important soil properties that affect the behaviour of
each component and the system as a whole The results of this research can be used to assess
and optimize the current design procedure for dikes with retaining walls
MSc Thesis A Rippi v
MSc Thesis A Rippi vi
Table of Contents
Preface ii
Abstract iv
List of abbreviations x
1 Introduction 1
11 Project objective and main research questions 2
12 Research approach and outline 3
2 System description and current design concept 7
21 System description and forces configuration 7
22 Current design concept 12
23 Safety standards 17
3 Literature study 21
31 Background 21
311 Finite Element Modeling 21
312 Uncertainties and Sensitivity analysis 24
32 Previous Studies 29
33 Overview 34
4 Structural Reliability Analysis 35
41 Basics of Reliability Analysis 35
42 Overview of Reliability Analysis Methods 36
421 Level III Methods 36
422 Level II Methods 40
423 Level I Methods (semi-probabilistic) 42
424 Response Surface Techniques (RS) 44
43 Coupling Reliability Analysis with FEM 46
431 The functionality and possibilities of OT 46
432 Coupling OpenTURNS-Plaxis 47
44 Overview 49
5 Failure Mechanisms and Limit State Functions 53
51 Introduction to the system analysis and the limit states 53
52 Limit State Functions 55
521 Serviceability Limit State 55
522 ULS for Structural Members 57
MSc Thesis A Rippi vii
523 ULS for Soil Failure 60
53 Overview 66
6 Case Study-Dike with an anchored sheet pile wall 69
61 Case Description 69
62 Soil Parameters 70
63 Finite Element Model 73
64 Deterministic Analysis 74
641 Calculation Scheme and Design Values 75
642 Construction Stages 77
643 Determination of the structural elementsrsquo characteristics 78
65 Overview 81
7 Reliability analysis results with stochastic soil properties 85
71 Method description 85
72 Mean values calculations 88
73 Sensitivity Analysis Results 91
74 Soil Shear Failure 98
75 Anchor Failure 102
76 Sheet pile wall failure 105
8 Conclusions and Recommendations 121
81 Conclusions 121
82 Recommendations 123
References 125
Appendix A 129
OpenTURNS features 129
A1 Fourier Amplitude Sensitivity Test (FAST) 129
A2 Optimization Algorithms in FORM 132
Principles of optimization algorithms 132
Convergence criteria 134
Evaluation of the algorithms performance 135
A3 Distribution Types 143
Uniform Distribution 143
Normal Distribution 144
Lognormal Distribution 144
Truncated Normal Distribution 145
Appendix B 147
MSc Thesis A Rippi viii
Plaxis 2D (2015) features 147
B1 Mohr Coulomb failure criterion 147
B2 φ-c Reduction Technique 150
B3 Initial Stress Generation 151
B4 Interface Strength 151
Appendix C NEN 6740 - Table 1 153
Appendix D 155
Input Files for the Reliability Analysis 155
Appendix E 159
Characteristic and mean values 159
Appendix F 161
Reliability methods 161
F1 Generation of random samples in Monte Carlo 161
F2 Other Sampling Methods 162
F3 First Order Second Moment (FOSM) Method 165
MSc Thesis A Rippi ix
MSc Thesis A Rippi x
List of abbreviations
CDF Cumulative Distribution Function
CoV Coefficient of Variation
CUR Civieltechnisch Centrum Uitvoering Research en Regelgeving
DS Directional Sampling
FAST Fourier Amplitude Sensitivity Analysis
FEA Finite Element Analysis
FEM Finite Element Model
FERM Finite Element Reliability Method
FORM First Order Reliability Method
FOSM First Order Second Moment
LEM Limit Equilibrium Method
LRFD Load and Resistance Factor Design
LSF Limit State Function
MC Monte Carlo
OT OpenTURNS
PDF Probability Distribution Function
RFEM Random Finite Element Method
RS Response Surface
SA Sensitivity Analysis
SLS Serviceability Limit State
SORM Second Order Reliability Method
ULS Ultimate Limit State
VNK Veiligheid Nederland in Kaart
MSc Thesis A Rippi xi
The roots of education are bitter but the fruit is sweet
Aristotle
MSc Thesis A Rippi 1
1 Introduction
In the Netherlands according to the Flood Protection Program (Hoogwater
Beschermingsprogramma1) and the Room for the River program (Ruimte voor de Rivier2)
alternative structural techniques for the reinforcement of existing dikes or for future dike
constructions additional to the conventional ones have been introduced and tend to be
attractive The heightening of the dike crest and the construction of a stability berm are some of
the most common current actions for dike strengthening Additionally filter layers geotextiles
and drainage systems can be applied in advance in order to prevent several failure mechanisms
Recently cantilever or anchored sheet pile walls and diaphragm walls tend to be used as an
alternative procedure for embankment reinforcing (see Figure 11) Such an alternative can be
chosen for an improvement of the macro stability of the dikersquos inner slope and additionally for
saving space in the land area that needs to be habited
Figure 11 Design options for dike reinforcement (source Flood Defences 2015)
For that purpose engineers need a concrete and unequivocal design methodology for such
combined structures In the meantime Deltares in cooperation with the Water Board of
Rivierenland is working on a design guideline which will be applicable to design the so-called
Type II stability fences3 A draft of such a guideline is elaborated in the report of Larsen et al
(2013) In this report suggestions and recommendations are outlined for the design of such
structures with Finite Element Modeling (FEM) and partial factors Remarks are also made for
the advantages and the limitations of such models and how they should be handled so that
reasonable and trustworthy results can be retrieved from the analysis
1 httpwwwhoogwaterbeschermingsprogrammanldefaultaspx 2 httpalbertawatercomhow-is-water-governedwhat-is-room-for-the-river 3 Type II structures are structures which in combination with a soil construction fulfil water retaining functionalities
Introduction
MSc Thesis A Rippi 1
The concept behind the recommended design criteria (Larsen et al 2013) is the definition and
the evaluation of several partial factors with the view to determine overall safety factors These
factors are further discussed in section 22 However validation of these factors shall be carried
out before being used in any case otherwise either the safety of the structure is jeopardized or
the structure will be overdesigned and cost inefficient
For improving the design criteria researchers for a long time focused on enhancing structural
models (beams shells etc) and constitutive laws (elasticity plasticity damage theories etc)
With the development of computer science a great amount of work has been devoted to
numerically evaluate approximated solutions of the boundary value problems describing the
mechanical system FEM is probably nowadays the most applied approach for the solution of
these problems
However the increasing level of detail of the constitutive models and the constant enhancement
of the computational tools do not solve the problem of identification of the model parameters
and the inherent physical and modelling uncertainties Moreover in most civil engineering
applications the intrinsic randomness of materials (soil rock concrete etc) or loads (water
elevation wind earthquake motion etc) is such that deterministic models are using average or
later on characteristic values of the properties at best lead to rough representations of the
reality
As a counteraction a semi-probabilistic methodology has been developed that was based on the
application of characteristic and design values by using partial factors Current design codes
such as Eurocode provide target reliabilities for different types of structures and structural
elements according to the potential consequences of failure However these partial factors are
not always equally suitable and efficient for all types of structural applications since they have
been calibrated under specific conditions Besides the consequences of failure of flood defences
such as dikes can be comparable to the investments in increasing the reliability of such systems
are For that reason it would be advisable for these systems to define target reliabilities based
on a risk assessment (ie tailor-made solution) rather than using the standard partial factors
coming from general geotechnical design codes which may be either too low or too high for a
given flood defence system
One step of such a risk assessment is accounting for randomness and spatial variability of the
mechanical properties of materials is one of the tasks of stochastic or probabilistic mechanics
which has developed fast in the last decade In this master thesis project the uncertainty of soil
properties is going to be treated in terms of its contribution to failure For that purpose
probabilistic methods are going to be implemented for a dike with an anchored sheet pile wall
(see Figure 12) simulated in a FEM software that is specialized in soil mechanics The successful
implementation and in future research the verification of such methods can be considered as
the most preferable and cost efficient way to design structures with high safety requirements
and not only for the validation of the partial factors Of course such a procedure tends to be
time consuming However the gradual improvement of the current probabilistic methods in
combination with the state-of-the-art computer capabilities as well as the scientific knowledge
gained in terms of different systems behaviour and failure modes can introduce a more
optimized way of designing structures with considerable investments
Introduction
MSc Thesis A Rippi 2
Figure 12 Reinforced dike section with an anchored sheet pile wall
11 Project objective and main research questions At the beginning of this research it has been observed that the design procedure being followed
for dikes with retaining walls modeled in FEM can lead to an overestimation of the critical loads
and thus to a potential cost inefficient final structure An example of the magnitude of the design
values in such a case study is given in the next chapter where also a more detailed explanation
of the current design concept is described
The objective of this study is to implement a full probabilistic analysis for evaluating the
reliability of a dike reinforced with an anchored sheet pile wall Such a research analysis can
subsequently come up with valuable recommendations for the improvement of the present
design approach Soilrsquos and structural elementsrsquo behavior should be taken into account both
separately and as a system Some of the reliability methods can deal with system reliability
problems such as Monte Carlo (MC) and Directional Sampling (DS) whereas for others like
FORM and SORM additional methods should be applied that use reliability information for each
individual limit state function to obtain the systemrsquos reliability
In principle the probability of failure of different limit states is to be computed individually
while in the sequence the system reliability is going to be estimated As far as the soil failure
mechanisms are concerned this research will focus mainly on the global stability of the dike
slope while for the sheet pile wall and the anchorrsquos failure the exceedance of the steel yield
stress is going to be considered
The scope of this thesis is not only to test different reliability models robustnessrsquo in conjunction
with FEM simulations but also to get a better insight into the specific system behavior (ie of a
dike with a retaining wall) analyze each component separately and investigate its response
under certain load conditions Furthermore the minimization of the computational effort and
time could also be carried out meaningfully under the constraint of sufficient accuracy The
accuracy should be such that the probability of failure is acceptable for the ultimate limit state
(ULS) in a normal design process and according to the current safety standards Finally the
Sheet pile
wall
Anchor
Dike section
Introduction
MSc Thesis A Rippi 3
robustness of the coupling between the reliability model and the FEM will be tested via their
capability of adapting to new input parameters without encountering convergence errors
during execution
The main research question of this master thesis project is thus formed as follows
How can the probability of failure of a dike with a sheet pile wall due to global instability modeled
by a Finite Element Model be analyzed
Essential questions regarding the soil models and failure criteria as well as the reliability
methods and the systems behavior are generated which are listed below
Subquestion 1 Which reliability methods are computationally tractable in
combination with FEM
Subquestion 2 How robust (convergence) are the tractable methods
Subquestion 3 Which limit states are relevant for the envisaged application of
retaining walls in dikes and how can they be formulated using FEM
analysis outcomes
Subquestion 4 What is the contribution of different uncertainties in the failure
mechanisms of the system
Subquestion 5 Can response surface techniques help to increase the efficiency and
robustness of the reliability model
Subquestion 6 How can the current design approach for dikes with sheet piles be
improved
In the next chapters the above research question and the related subquestions are going to be
answered by following the methodology that is described in the next section
12 Research approach and outline In this section an overview is given of how the aforementioned objective and sub-questions are
approached The thesis is mainly divided into 8 chapters and 6 appendices In Figure 13 the
thesis outline is illustrated and an indication of which subquestions (Sq1-6) are treated to
which chapter is given In Chapter 1 an introduction into this research content is made and a
first illustration of the system under investigation is presented together with the main research
questions
Chapter 2 has been devoted to the description of the system (ie a dike with an anchored sheet
pile wall) and to the case study that has been adopted for being analyzed in terms of its
reliability The forces configuration is also illustrated and the results of the case study according
to the current design concept are presented and evaluated Last but not least the new
recommended safety standards are included and the current required reliability of the specific
dike section is presented
Introduction
MSc Thesis A Rippi 4
In Chapter 3 a literature study related to FEM uncertainties and sensitivity analysis is
presented Moreover previous studies that are associated with the objective of this research are
mentioned whereas some of their results were also taken into account for proceeding with this
research
The main scope of this thesis is the implementation of different reliability methods on a specific
case study with the view to investigate and analyze its behavior Some of these methods are
continuously mentioned through the test and thus in Chapter 4 an overview is given of the
most well-known reliability methods Eventually an evaluation of these methods is made based
on literature and preliminary testing with simple case studies and the procedure of their
coupling with FEM is discussed The evaluation helped to qualitatively answer subquestion 1
while also an introduction of how special reliability methods such Response Surfaces were used
in this thesis that partially covers subquestion 5 In this chapter an introduction into the Limit
State Functions (LSF) concept is made and how they are considered for the reliability analysis of
a system An LSF actually represents a failure mode that can be detected in a structural or soil
element and it is expressed as a function of several variables In Chapter 5 the failure
mechanisms and the corresponding LSF that are related to the specific case study are identified
and formulated which answers subquestion 3
In Chapter 6 the case study whose reliability is to be evaluated is presented as it was modeled
in FEM The boundary conditions are specified and the soil and structural properties are
indicated and illustrated Moreover a deterministic design is also taking place in order to
roughly estimate the structural elementsrsquo properties This will help to make a qualitative
comparison between the current design procedure and the design according to a fully
probabilistic approach by referring thus to subquestion 6
Eventually the results of the aforementioned analysis are presented in Chapter 7 The results
mainly include the estimated probability of failure of each of the system components as well as
of the system as a whole by considering the soil properties as stochastic Moreover an
interpretation of the failure points is made and an assessment of the level of impact of the
random variables on the systemrsquos behavior is carried out At that chapter subquestions 2 4 and
5 are mainly treated
Last but not least in Chapter 8 the general conclusions are presented together with some
valuable recommendations for future consideration and research Furthermore a reflection on
the methodology and how the different research questions were approached is made
Introduction
MSc Thesis A Rippi 5
Figure 13 Thesis outline
Introduction
MSc Thesis A Rippi 6
MSc Thesis A Rippi 7
2 System description and current design concept
In this section the system to be analyzed is described more in detail and the individual
components are identified Moreover the forcesrsquo configuration is explained and illustrated as
they would have been calculated with the conventional way in parallel with a qualitative
evaluation of the potential expected deformations Finally the current design concept and the
safety standards that are referred to such a system are introduced
21 System description and forces configuration The system to be investigated in this thesis consists of a dike section reinforced with a one-layer
anchored sheet pile wall The system has been simulated by FEM and later on it was coupled
with a reliability package for carrying out a reliability analysis In the figure below an
illustration of the system is depicted and the several elements of an anchored sheet pile wall are
showed
Figure 21 System layout and different components of the anchored sheet pile wall
In particular an anchored sheet pile wall consists of the sheet piles which are anchored in the
soil via a tie rod The rod is grouted most of the times in a sandy layer with a cemented grout
body and it both alleviates the sheet piles from the axial forces exerted by the upper structure
and keeps the wall stable in case of excessive developed moments due to the lateral earth
pressure Last but not least a waling system is applied mainly downstream of the sheet piles in
order to transfer the loads from the piles to the anchors in such a manner so as to avoid
excessive local stresses on the intersection between the sheet piles and the tie rod
Cross-Section
Top View
System description and current design concept
MSc Thesis A Rippi 8
The main load configuration acting on a dike section as well as a sheet pile wall is depicted in
Figures 22-24 together with the possible displacements As far as the soil body stability is
concerned most of the slope stability analysis computer programs are based on the limit
equilibrium concept according which a soil mass tends to slide down under the influence of
gravity The transitional or rotational movement is considered on an assumed or known slip
surface below the soil There an equilibrium should be achieved among the driving and the
resisting forces In that case the driving moments consist of the soil weight the water pressures
and the loads around the center of the slip surface such as a possible vertical load on the top of
the dike crest (ie traffic load) The magnitude of the water pressures is controlled by the water
elevation on the river side
In Figure 22(a) the external and internal forces acting on a slice of the slip surface are
illustrated The driving forces are the soil weight W the water pressure U and any additional
load that can contribute to the rotation of the slip surface which in turn activate the lateral
active earth pressure The resisting forces consist of the lateral passive earth pressure
(Rankine 1857) and the shear strength (Terzaghi 1943) Their expressions are written as
follows
[kN] (21)
[kN] (22)
[kPa] (23)
where is the active lateral earth pressure coefficient and the passive lateral earth
pressure [kNm3] is the unit weight of the corresponding soil layer and H [m] is the thickness
of the soil layer that the lateral earth pressure comes from (in case of multiple soil layers there
should be a distinction among the different forces) and [kPa] is the effective cohesion of the
soil layer The safety factor that is written in the Figure 22(a) as FS is defined by the ratio of the
shear strength (excluding the pore water pressure) divided by the weight of the earth body
(including the pore water pressure)
The loss of equilibrium between the driving and the resisting moments lead to the rotation and
instability of a slip surface as it is illustrated in Figure 22(b) A slide surface can take various
shapes according to the Limit Equilibrium Method (LEM) that is followed According to Bishop
method this surface tends to be circular in Uplift-Van method a horizontally compressed zone
can be also considered whereas in Spencer method the shape of the slide body can be arbitrary
The slip surface can be located either in the landside or in the river side of the dike depending
on the load the soil characteristics and the design of the dike (ie inner or outer berm
reinforcing revetment on the outer slope etc) Most of the times the landside area of the dike is
jeopardized due to different failure mechanisms that are described in section 523 The
instability of the outer slope is mainly relevant in case of sea dikes and a potential damage can
be usually reconstructed until the next flooding
System description and current design concept
MSc Thesis A Rippi 9
(a)
(b) Figure 22 (a) External and internal forces acting on a slip surface of a dike section and (b) possible deformation pattern
As a counteraction for the slope instability a sheet pile wall can be installed inside the dike The
special thing about these structures is that they make possible a greater freedom in form and
functionality than a traditional dike design This structure derives its strength from the
materials used such as steel which are able to withstand higher pressures than clay for instance
The general stability is due to friction and wedging in the bottom
After the reinforcement of the inner side of the dike with a retaining wall the strength capacity
of the wall is also important for the global stability of the system The forces to be taken into
account for the sheet pile wall stability are the active and the passive earth pressures (effective
pressures) and the water pressure as they are illustrated in Figure 23(a) In that figure the
System description and current design concept
MSc Thesis A Rippi 10
forces were simplified in a singular triangular shape as in a homogeneous soil body in order to
explain and depict the overall picture of the acting forces However the stresses distribution
over depth can be more complex depending on the variety of the soil layers that are present In
Figure 23(a) it can be noticed that a sheet pile wall which is installed close to the inner berm
might not have significant instability issues as there is the passive side that contributes to
resisting forces However in the passive side the soil can differ and be weaker than this of the
active side Therefore the passive force in that case might not be very supportive and thus the
wall shall be designed cautiously
Additionally the anchor resistance shall be reassured for the sake of the stability of both the
sheet pile wall and the overlying soil body In Figure 23(a) the forces acting on the anchorage
are displayed The distributed load qs comes from the earth pressure of the upper soil layer (it
might not be uniformly distributed as it is depicted in the figure but more in a trapezoidally type
of load) whereas the tension force Ft4 stems both from the earth pressure and the displacement
of the sheet pile wall The bending and the tension capacity of the anchor are determinant for its
stability In Figure 23(b) a possible displacement pattern of the retaining wall and the
anchorage is illustrated
Furthermore corrosion is an additional weakening impact on the wall that depends on the
water level and the pore water pressures near the structure The thickness of the wall and the
material properties are playing a key role to the resistance towards corrosion This will not be a
subject of this thesis however it should be taken into account in case of design purposes
(a)
4 where O is the average circumference of the pile P [kPa] is the shaft friction at depth z and L is the length of the pile
System description and current design concept
MSc Thesis A Rippi 11
(b) Figure 23 (a) Mobilizing forces acting on the sheet pile wall and the anchorage and (b) possible deformation pattern
It should be mentioned that as far as the pore pressures are concerned in Figure 23(a) only the
hydrostatic pore pressures are depicted Generally the pore pressures (or active pore pressures)
are the sum of the steady state pore pressures and the excess pore pressures
[kNm2] (24)
Steady state pore pressures can be hydrostatic (based on a horizontal phreatic line) or non-
hydrostatic (based on steady state groundwater flow) Excess pore pressures are based on
loading of undrained soils In FEM these various pore pressures are taken into account
automatically In Figure 24 an example of a potential distribution of the active pressures
nearby the sheet pile wall is depicted as it was deduced from Plaxis Therefore as it can be
noticed from the figure the distribution can indeed sometimes approximated as triangular
System description and current design concept
MSc Thesis A Rippi 12
Figure 24 Example distribution of the active pore pressures acting nearby the sheet pile wall
The knowledge of the type of forces that are exerted on both the structural components and the
soil body as well as of the most expected form of displacements is valuable in order for
someone to be able to evaluate the results deduced from FEM and detect potential modelling
errors Therefore this validation was necessary before starting with the reliability analysis
In the next sections the current design approach of the system described above is elaborated
Additionally the Dutch recommended safety standards are presented for this type of structures
which show the need of carrying out a reliability analysis For that purpose a case study was
adopted that it is presented first as it was designed according to the current regulations while
finally a comparison is made between the original and the new case study as it was found to be
according to the reliability analysis
22 Current design concept In the Netherlands dike reinforcement is becoming a major issue concerning the design of flood
defenses Special reinforcing structures such as cofferdams anchored sheet piles and diaphragm
walls are used for strengthening the dike Especially in the context of the Flood Protection
Program and the Room for the River program many dike reinforcement projects have been
suggested
In the report of Larsen et al (2013) the design procedure of a dike reinforced with sheet pile
wall is prescribed using FEM The safety philosophy that is followed in this technical report
actually composes the current procedure for designing dikes with sheet pile walls and it is
linked with the usage of partial safety factors Particularly the required overall safety factor
FEM that should be compared with the one from FEM calculations is determined as follows
SFEM
b d m n
(25)
where
System description and current design concept
MSc Thesis A Rippi 13
Partial safety factor indicating the uncertainties of the soil composition and
the water pressures (also called schematization factor)
Partial safety factor which is related to the calculation model and the way the
calculations have been carried out (also called model factor)
Partial safety factor which is related to the material parameters (also called
material factor)
Partial safety factor associated with damage caused during the soil tests (also
called loss factor)
Safety factor of load
Each of these partial factors shall be at first calibrated by conducting a full probabilistic analysis
of the system under consideration In most of the cases these partial factors have been already
calibrated on previous similar projects and then they are reused for any similar case In
Schweckendiek et al (2013) a new approach of using partial factors for flood defences is
proposed whose application is still under consideration At the moment the aforementioned
partial factors for dike reinforced with sheet pile walls modelled in FEM are estimated
according to mostly engineering judgement and political negotiations
In sequence the FEM design procedure that is applied for sheet pile walls inside soil structures
is based on a so-called ldquosafety calculationrdquo (for information for this type of calculation see
Appendix B2) In this calculation an artificial reduction to the soil strength parameters (friction
angle and cohesion) is applied in order to evaluate the overall safety factor and confirm if this is
lower or higher than the required one that was estimated according to Eq 25 Furthermore the
resulting moments and forces developed on the structural elements during the specific
calculation are used for their design
This is a quite helpful and indicative calculation type available in Plaxis that can give a sense of
the structurersquos safety factor and the possible failure mechanisms that can occur under the
predefined load configuration However such a method of calculating the developed stresses
can be misleading in terms of the moments and forces acting on the structural elements and the
total deformations of the system
In this thesis the reliability of a dike with an anchored sheet pile wall is going to be evaluated
and in particular the case study of Breedeveld (2011) is to be used for that purpose In Figure
26 the location of the existing dike section is presented while in Figure 25 the structure as it
was modelled in FEM is illustrated It is essential to mention that the retaining wall does not
exist in reality but it was placed so as to implement and demonstrate the current design
regulations in the report of Breedeveld (2011) The anchorrsquos angle has been taken equal to 30deg
from the vertical which is generally a steep anchor inclination That is expected to reduce the
part of the horizontal force that is undertaken by the anchor and thus the most of it exerts on
the sheet pile wall The angle is usually taken 30deg-45deg from the horizontal axis (CUR 2005) but
this is of course depended on the load conditions and the construction requirements for each
System description and current design concept
MSc Thesis A Rippi 14
case The influence of the anchor orientation and the construction methods that are generally
followed for an anchored sheet pile wall are not considered in this thesis The case study was
taken as it was modelled in the aforementioned project and the reader should be aware that this
serves the proceedings of a reliability analysis and not of a perfectly constructed retaining wall
Figure 25 Original dike section in Varik-Heesselt as it is presented in Breedeveld (2011)
In particular based on this case study and trying to understand the design criteria a
comparison was carried out between the ldquosafety calculationrdquo and a normal ldquoplastic calculationrdquo
(ie elastoplastic drained or undrained analysis) In both cases the boundary conditions the
external loads and the input parameters are the same The difference between them is the
calculation procedure followed in order to come up with final stresses which in the case of the
ldquoplastic calculationrdquo does not include a reduction of the soil strength properties In Figures 27
and 28 the results from both calculations are presented In Figure 27 the total deformations
are exhibited while in Figure 28 the moments and the axial forces on the sheet pile are shown
Figure 26 Location of the Varik-Heeselt dike section (section number TG-078) in the Dike ring 43
BetuweTieler-en Culemborgerwaarden west (source Breedeveld 2011)
System description and current design concept
MSc Thesis A Rippi 15
As it is realised from this figure the discrepancy between the two results is quite high while as
far as the deformations are concerned no realistic conclusions can be drawn from a ldquosafety
calculationrdquo Moreover taking as design moments the moments that are deduced from the
ldquosafety calculationrdquo will lead to a much heavier and more expensive sheet pile cross section than
the ldquoplastic calculationrdquo
It is also essential to mention that the required safety factor was estimated up to FEM =18
according to the partial-factor procedure that is defined in Eq 25 Such a safety factor as
prerequisite for coming up with the design values of the moments and forces of the sheet pile
and the procedure of the ldquosafety calculationrdquo in FEM can lead to a conservative and financially
inaccessible design
(a) ldquoPlastic calculationrdquo Deformed mesh and maximum deformations (umax= 008 m)
(b) ldquoSafety calculationrdquo Deformed mesh and maximum deformations (umax= 9041103 m)
Figure 27 Deformed mesh and maximum deformations in the end of (a) plastic calculation and
(b) safety calculation
30deg
System description and current design concept
MSc Thesis A Rippi 16
(a) ldquoSafety
calculationrdquo
Mmax = 9392 kNmm Nmax= -4402 kNm
(b) ldquoPlastic
calculationrdquo
Mmax = 6833 kNmm Nmax= -1372 kNm
Figure 28 Bending moments and axial forces exerted on the sheet pile during the (a) safety
calculation and (b) plastic calculation
On the other hand the aforementioned design procedure guaranties a strong structure capable
of probably undertaking more than the expected loads and thus ensuring the safety against
flooding of the landside However the knowledge on fully probabilistic methods that has been
obtained as well as the development of the technology can introduce a reliability analysis of the
system as a preferable way for the future design concept Such a procedure should be
accompanied also by field test that would be able to validate FEM results as well as in-situ
measurements and inspections of the soil properties for calibrating their statistical
characteristics
Due to the conservative results that the aforementioned procedure came up with and because of
the simplifications that this case study was later subjected to in order to be used in this thesis a
new design was carried out in Chapter 6 and the properties of the structural elements were
redefined
The inclination towards probabilistic methods has not only been created due to the incomplete
design regulations for the system under consideration but also the current proposed safety
standards which are discussed in the section below recommends a risk based safety assessment
of the primary flood defenses which in principle implies the evaluation of the failure probability
In the next section an overview of the new recommended safety standards related to primary
flood defenses is given and the required reliability for the stability of the aforementioned case
study is elaborated
System description and current design concept
MSc Thesis A Rippi 17
23 Safety standards An important factor for the evaluation of a structurersquos reliability is the safety standards that
have been set and according which the acceptance or not of the failure probability is made After
the safety assessment5 of 2011 it has been declared that 33 (1225 km) of the primary flood
defenses (3767 km) in the Netherlands does not comply with the safety standards For this
purpose a new session of investigations has commenced in order to assess the reliability of the
existing dikes and the potential amendment of the safety standards for future constructions
Taking into account this latest information and within the framework of the Flood Risk in the
Netherlands 2 project6 (VNK2) a new Delta program was founded whose one of the main
decisions was the ldquoDelta Safety Decisionrdquo This decision contains a proposal for new safety
standards of the primary dikes as they are depicted in Figure 29 which are stricter than the
previous ones and formulated in terms of failure probability
According to this project a new policy regarding the safety assessment of the flood defenses has
been settled and proposed for the future design criteria which is based on the evaluation of the
acceptable flood risk (probability of failure (x) consequences) of each dike ring rather than the
probability of exceedance in order to achieve a level of protection that is in balance with the
societal value (Cost Benefit Analysis and life loss)
VNK has been already using reliability evaluation techniques in order to calibrate partial safety
factors Especially in geotechnical engineering the high inherent uncertainty of the soil
properties renders the evaluation of the structural reliability essential for the safety assessment
of the structure Subsequently this implies the investigation of the response of the different
reliability methods applied in a real case study of a dike The techniques that are used to
evaluate the probability of failure are discussed thoroughly in Chapter 4 and in Appendix F In
case of a complicated structure such as a dike with a retaining wall both the reliability of the
dike body and the reliability of the wall shall be evaluated and eventually a common standard
for the reliability of the system shall be defined
5 A safety assessment is an evaluation of the function of the flood protection structure and it is being conducted every approximately 6 years after the reassessment of the existing hydraulic boundary conditions 6 in Dutch Veiligheid Nederland in Kaart 2 or VNK2 It has been initiated by the Ministry of Infrastructure and the Environment (IampM) the Association of Regional Water Authorities (UvW) and the Association of Provincial Authorities (IPO)
System description and current design concept
MSc Thesis A Rippi 18
Figure 29 Maximum admissible flooding probability for primary defences according to Delta
programme 2015 (Deltaprogramma 2014)
As far as the retaining walls are concerned CUR (2005) design procedure distinguishes the
following three safety classes for retaining walls with corresponding reliability indexes
Table 21 Safety classes and corresponding reliability indices (CUR 2005)
Class I Relatively simple constructions no personal safety risks and relatively minor
damage in the case of overall failure ϐ=25
Class II Considerable damage in the case of overall failure minor personal safety
risks ϐ=34
Class III Major damage in the case of overall failure andor considerable personal
safety risks ϐ=42
System description and current design concept
MSc Thesis A Rippi 19
Additionally as far as a dike section is concerned the required overall reliability index shall be
concluded according to the new norms that are shown in Figure 29 However the reliability
index of the different failure mechanisms of a dike shall be estimated separately
According to Vergouwe et al (2014) that is part of the VNK2 report the major failure
mechanisms in dike ring 43 where the case study under investigation is located (see Figure 26)
is uplift and piping and macro-instability of the inner slope Especially in Figure 210 the
estimated percentage of the contribution of each failure mechanism to the overall probability of
flooding7 is indicated In Vergouwe et al (2014) the overall flooding probability of the specific
dike section that the case study concerns was evaluated to 17800 Therefore the macro-
instability on which this thesis is concentrated with a contribution percentage of 226 has a
probability of about 310-5 (
) and thus a target reliability index β of
approximately 4 This reliability level can be considered as a benchmark for evaluating the one
that will be entailed in the end of this thesis
Figure 210 Percentage contribution of the different failure mechanisms to the probability of flooding (source Vergouwe et al 2014)
It should be mentioned though that the aforementioned safety standards concerning the dike
ring 43 were defined according to the probability of exceedance of a certain water level that the
dike has to retain Therefore this complies with the old norms and not with those appeared in
Figure 29 However for large engineered systems such as flood defence systems with large
potential consequences and substantial investments it is worthwhile to assign target reliability
levels based on a risk assessment of the area surrounded by the specific dike ring There are
three widely used types of criteria for evaluating the risks related to floods and major industrial
hazards (Vrijling et al 2011)
Economic criteria
Individual Risk Criteria
Societal Risk criteria
7 The dependencies among the failure mechanisms were not taken into account Therefore this probability stated in Vergouwe et al (2014) is not exactly equal to the probability of flooding
System description and current design concept
MSc Thesis A Rippi 20
In order to check then if a certain system abides by the target reliability that is required the
failure probability of each failure mechanism shall be calculated based on a LSF The
aggregation of the failure probabilities of all the modes gives the failure probability of the
system which is supposed to be compared with the required in order to reassure the reliability
of the structure Therefore this thesis is dealing with the calculation of this failure probability of
a certain system modelled in FEM analysis and understanding of the systemrsquos behaviour and
finally the investigation of the robustness of coupling FEM with reliability methods for
evaluating the reliability of a structure In the next chapter an overview over the research has
been done until now concerning coupling and reliability methods applications is presented and
a discussion over the most important is made
MSc Thesis A Rippi 21
3 Literature study
In this chapter an overview of some principles related to FEM and the uncertainties in
geotechnical engineering is presented More precisely an introduction into FEM concept and a
discussion over the different types of FEM is made Moreover the uncertainties in geotechnical
engineering are stressed and the general framework according to which they are handled is
described Finally previous studies that are associated with the application of FEM on soil
structures and the reliability analysis are discussed
31 Background
311 Finite Element Modeling
FEM is a numerical method whose essence is to convert a problem described by partial
differential equations over space and time into one by dividing the space-time continuum into a
set of discrete elements and assuming that the unknowns vary over each element FEM solution
process is as follows
1 Divide structure into pieces (elements with nodes) (discretizationmeshing)
2 Connect the elements at the nodes to form an approximate system of equations for the
whole structure (forming element matrices)
3 Solve the system of equations involving unknown quantities at the nodes (eg
displacements)
4 Calculate desired quantities (ie strains and stresses) at selected elements
The properties of each element are set the same as the material properties that have been
defined by the user Then the Deterministic Finite Element Method (the finite element method
that was formulated with deterministic variables) can be used in conjunction with means and
standard deviations of the input variables to obtain reliability estimates
In the simple coupling of FEM with reliability analysis each parameter that is considered as
stochastic is given a particular probability density function estimated either by field tests or by
engineering judgement This approach is also called ldquosingle random variablerdquo and it assumes
that the spatial correlation length is infinite In other words the soil is considered to be
homogeneous and the stochastic property assigned to the soil is taken at random from a certain
probability distribution
Nevertheless a more realistic model should take into account the spatial correlation within
smaller regions where the stochastic property is allowed to vary For that purpose the Random
Finite Element Method (RFEM) was introduced by Fenton and Vanmarcke (1990) in which the
random variables are correlated to one another using auto-correlation functions
The most sophisticated method of FEM is the Stochastic Finite Element Method introduced by
Ghanem and Spanos (1991) which accounts for the uncertainties in the input parameters
Literature study
MSc Thesis A Rippi 22
implicitly within the finite element calculation This aims at representing the complete response
probability distribution function (PDF) in an intrinsic way Two steps are basically applied for
that purpose
Discretization of the problem and definition of the random vector with an unknown
joint PDF and
Expansion of the response in a particular basis of random vectors with a finite variance
called the polynomial chaos
In this thesis the Deterministic Finite Element Method is to be used coupled with reliability
analysis According to Waarts (2000) coupling FEM with reliability methods (FERM) would lead
to the following advantages
In comparison to standard finite element analysis it gives direct insight into the
structural reliability and decisive parameters
Structures designed using FERM will either be safer orand more economically built in
comparison to structures designed using safety factors and classical constitutive models
FERM is expected to be useful for the determination of LSF that cannot be explicitly formulated
and that differ in each case such as soil limit state in different structural schematizations
Moreover it is likely to be valuable in areas where little knowledge exists on the systems
reliability of structures with multiple components (ie a dike with a sheet pile wall)
FEM in practice
The conventional method for stability analysis in a soil body is represented by LEM although
FEM is increasingly used by designersresearchers The latter has been proved to be quite
realistic for the progressive behaviour (ie stress-strain development in different construction
phases) of a soil system under the effect of stress redistribution in comparison with classical
models Especially in their master thesis Johansson amp Sandeman (2014) compared the
deformations and the forces measured at a deep excavation supported by anchored sheet pile
wall in a railway tunnel located in Gothenburg in Sweden with 1D finite element software the
2D finite element software (Plaxis) and hand calculations They proved that Plaxis produces
reliable results for horizontal deformations in the sheet pile wall and anchor forces when
compared to in-situ measurements
Moreover in Gonzaacutelez et al (2013) it was proved that the simplification of reality done by the
classical methods such as Blumrsquos Engelrsquos Kreyrsquos methods etc (for further information for these
methods a reference is made to Gonzaacutelez et al (2013)) allows us to generally understand the
behaviour of the system wall-soil Nonetheless the results that came out of this analysis were
found to be quite conservative whereas FEM managed to give a more realistic interpretation of
the wallrsquos movement
Seed et al (2006) investigated the performance of the New Orleans Flood Protection Systems
during hurricane Katrina in August 2005 by simulating the different levee sections with FEM As
an example in Figure 31 pictures of the failure of the Lower Ninth ward IHNC East Bank in
New Orleans after the hurricane Katrina are presented FEM was proved to be an efficient tool
to realistically interpret the shape and the triggers of the failure mechanism
Literature study
MSc Thesis A Rippi 23
Figure 31 Failure of the Lower Ninth ward IHNC East Bank in New Orleans modelled in FEM with Plaxis (source Seed et al 2006) At the bottom right the shear strain contours are depicted that are developed through the embankment and at the bottom left the deformed mesh is presented illustrating the failure mode of the levee
In recent implementation of design concepts and technical recommendations such as CUR 166
ndash Damwandconstructies (2005) which refers to the design of retaining walls it can be observed
that guidelines for Finite Element Analysis (FEA) have been set parallel to the existing analytical
or empirical calculation methods Therefore FEA tends to become more and more accepted as
an alternative for Serviceability Limit State (SLS) as well as for ULS design
According to Wyllie and Mah (2004) LEM give an estimation of the factor of safety with no
information on deformation of the slope However in numerical analysis the failure surface can
evolve during the calculation in a way that is representative of the natural evolution of the
physical failure plane in the slope In that way a better insight into the evolution of failure
mechanisms can be gained
Cundall (2002) compared the characteristics of numerical solutions and LEM in solving the
factor of safety of slopes and concluded that continuum mechanics-based numerical methods ie
FEM have the following advantages
No pre-defined slip surface is needed
The slip surface can be of any shape (in contrast with Bishop and Kranz stability)
Multiple failure surfaces are possible
No static assumptions are needed
Literature study
MSc Thesis A Rippi 24
Structures (such as footings embankments etc) and structural elements (such as
beams cables sheet piles etc) and interface can be included without concern about
compatibility
It is also important to recognize that LEM only identifies the onset of failure whereas FEM
includes the effect of stress redistribution and progressive failure after its initiation Numerical
models can also be used to determine the factor of safety of a slope in which a number of failure
mechanisms can exist simultaneously or where the mechanism of failure may change as
progressive failure occurs (Hoek et al 2000)
On the other hand the accuracy of FEM is dependent on usersrsquo settings such as the refinement
of the mesh generation the loading steps etc Moreover even if FEM can be functional and easy
to use it is essential that the user has a deep knowledge of the subject under investigation (soil
and structural mechanics) in order to be sceptical and critical with the FEM results able to
interpret the output behaviour of the structure and changecorrect everything that seems
peculiar Moreover the usage of FEM implies also some basic knowledge on numerical methods
and the general background of the FEM so as to solve possible numerical errors
An ideal way of modelling in FEM would be a prior calibration of the program according to data
related with the structurersquos properties and performance (ie stress generation) However this
means that enough field measurements should be carried out and under specific conditions in
order to be considered as a base for FEM calibration Unfortunately such field tests are not
always available and if so their reliability is on doubt For example measurements related with
the developed moments or displacements of a retaining wall inside a dike section are rare to be
found but even if there are some available they cannot represent the situation near the failure
domain where most of the engineers are worried about
312 Uncertainties and Sensitivity analysis
Uncertainty in geotechnical engineering can be categorized into aleatoric epistemic and
decision model uncertainty as they are depicted in Figure 32 (Baecher amp Christian (2003))
Aleatoric uncertainty is associated with the natural variability which is defined as the intrinsic
randomness of natural processes Such variability can be expressed by changes of the soil
properties over time at a certain location (temporal variability) or over space at a single time
(spatial variability) Epistemic uncertainty can be divided into the site characterization model
and parameters uncertainty and it is attributed to lack of information about events or lack of
understanding the physical laws that prohibits the ability of creating a realistic model Finally
the decision uncertainty describes the difficulty of being aware of social objectives defining
social values such as discount rates and predicting the planning horizon
Literature study
MSc Thesis A Rippi 25
Figure 32 Categories of uncertainty in geotechnical engineering (modified from Baecher amp
Christian (2003))
Below the general steps of an uncertainty study are described and an introduction into the
sensitivity analysis concept is made
Global methodology of an uncertainty study
A first step of an uncertainty study8 can be described as ldquothe definition of the problemrdquo Initially
the variables of interest (or else the output variables) of which the uncertainty is to be
quantified shall be specified In sequence given several input variables for which the user may
have data andor expertengineering judgment a model denoted usually by a mathematical
function should be introduced that enables the computation of the set variable of interest
After the general context has been staged we should choose the criteria with which the
uncertainty can be evaluated The most complete measure of uncertainty when dealing with a
random vector is the probability distribution In order to assess the value of this distribution
function the following criteria can be followed
Probability of exceeding a threshold the aim is to assess the probability that the variable
of interest exceeds a threshold important for the goals at stake
Quantiles the aim is to assess the threshold that a variable of interest may exceed with a
probability equal to a given value
Central dispersion the aim is to obtain an ldquoorder of magnituderdquo of the uncertainty by
specifying the average value and the variance of a variable of interest
8 In the engineering world this is also called reliability analysis However in this section the general uncertainty study is described In particular reliability describes the ability of a system or component to function under stated conditions for a specified time period
Literature study
MSc Thesis A Rippi 26
The next step is to define a model to represent and quantify the uncertainties of the input
variables One shall investigate each variable as a singularity and come up with the most
suitable probability density function (mostly depending on available data) Besides it is
essential to assess also the potential correlations among the variables that can be included in a
joint probability density function which is discussed later in this section
Eventually the uncertainties of the input variables shall be translated in terms of uncertainty on
the variables of interest This procedure is called uncertainty propagation and can be carried
out via several reliability methods (approximation methods or sampling methods) that are
extensively elaborated in Chapter 4 In Figure 33 a schematization of the different steps during
an uncertainty study is given
Figure 33 Methodology of uncertainty study
Last but not least a better understanding of uncertaintiesrsquo influence can be achieved by
analyzing the contribution of the different uncertainty sources to the uncertainty of the
variables of interest via a sensitivity analysis Such an analysis aims at identifying important
parameters for the system response besides it gives a better insight into the model used to
quantify the uncertainties Before conducting an uncertainty analysis it is advisable to filter out
parameters of less significance in order to reduce the modeling computational effort Below
different methods for sensitivity analysis are described
Sensitivity analysis
Sensitivity analysis (SA) is the study of how the variation in the output of a model can be
apportioned quantitatively and qualitatively to the variation in the model parameters (Saltelli
et al 2004) Saltelli (2004) proposes one possible way of grouping these methods into three
classes (i) screening methods (ii) global SA methods and (iii) local SA methods
i Screening methods
Screening is a particular instance of a sampling-based method The objective here is to identify
which input variables are contributing significantly to the output uncertainty in high-
dimensionality models rather than exactly quantifying sensitivity (ie in terms of variance)
Screening tends to have a relatively low computational cost when compared to other
Literature study
MSc Thesis A Rippi 27
approaches and can be used in a preliminary analysis to weed out less influential variables
before applying a more informative analysis to the remaining set One of the most commonly
used screening methods is the elementary effect method
ii Global SA methods
Global SA techniques incorporate the whole range of variation and the probability density
function of the input parameters to calculate their influence on the output Many global
sensitivity analysis techniques are now available such as
Variance-based methods
o Sobolrsquos method9
o Fourier Amplitude Sensitivity Test (FAST)10
o Analysis of Covariance (ANCOVA)
Regression-based methods11
Both FAST and Sobolrsquos method rely on the assumption of parameter independence while
ANCOVA can also handle correlated input parameters The main principles of FAST method are
presented in Appendix A1 For more information about the other methods the reader can
betake himself to the related references
iii Local SA methods
Local SA methods provide the slope of the calculated model output in the parameter space at a
given set of values (Turanyi and Rabitz 2004) This basically means that local methods involve
taking the partial derivative of the output Y with respect to an input factor Xi |
|
where the
subscript indicates that the derivative is taken at some point in the space of the input
Examples for these are (Schweckendiek 2006)
the α-values in a FORM calculation (further discussion in section 422)
Local Probabilistic Sensitivity Measure
In reliability analysis the local sensitivities will be more important than the global ones in the
sense that the influence of all variables in specific points such as the design point cannot be
identified by the global methods However local SA can only inspect one point at a time and the
sensitivity index of a specific parameter is dependent on the central values of the other
parameters
In this thesis the influence of the different parameters on the response of the system is
evaluated according to local sensitivity indices However before the main part of the reliability
analysis starts it was necessary to filter out the less influencing variables in order to reduce the
number of the major variables and make the analysis more efficient and computationally
affordable For this purpose the global sensitivity method FAST was used whose results are
presented in Chapter 7
9 created by Sobolrsquo (1993) 10 introduced by Cukier et al (1973) 11 introduced by Helton et al (2005)
Literature study
MSc Thesis A Rippi 28
Probabilities as a measure of uncertainties
Most engineers today use the concept of probabilities as the standard way to talk about
uncertainty One way to quantify a probability of a random variable is the calculation of the
cumulative probability function (CDF) of non-exceedance which can be obtained by
the probability distribution function For a random vector this reads
( ) (11)
From this the joint probability density function of this random vector can be determined as
( )
( )
(12)
This function is depicted in Figure 34 by means of contour levels The joint probability density
function has the shape of a hill as it is showed in the 3D figure and the inclination of the ellipses
reveals a correlation between the two variables and
Figure 34 Joint probability density function in 2D and 3D
The functions that join or couple multivariate distribution functions to their one-dimensional
marginal distribution functions are called copulas These are going to be used in order to define
the joint density probability distribution from which the random variables are taking their
values However no detailed explanation is made in this report regarding copulas and its
characteristics as it is automatically implemented in OT whereas the user has to define only the
type of copula that heshe considers the most appropriate (a reference is made to Nelson (1999)
for further information) According to the type of copula that is selected the order of the values
of the variables can be chosen within an iterative reliability method In this thesis an
independent copula was used considering that the input variables are independent the ones
from the others due to the time limit
Literature study
MSc Thesis A Rippi 29
32 Previous Studies In this section a summary of the most relevant researches on the topic of FEM applied in soil
structures and coupling with reliability methods is given
To begin with an attempt to estimate the reliability of a structure modeled in FEM has been
made by Waarts (2000) by introducing an optimized reliability method in terms of
computational effort and efficiency
In particular in Waarts (2000) two adopted reliability methods are introduced both making
use of a response surface (a detailed explanation of the reliability methods and the response
surface techniques is made in Chapter 4 and Appendix F) These adaptive response surfaces are
used in combination with FORM and DS respectively The accuracy and the effectiveness of
these methods are investigated on the basis of artificial LSFs and a comparison is made with the
existing standard reliability methods The most efficient combinations of response surface
techniques and reliability methods were with FORM (FORM-ARS) and DS (DARS) Comparing
these two methods DARS predominated over FORM-ARS as it can cope with a much wider
range of limit state functions
In a later stage the above best performing reliability method (ie DARS) is further investigated
in terms of its efficiency on the basis of complex structures reliability In Figure 35 a couple of
case studies used to verify the performance of DARS are given
Figure 35 Left Two-story two-bay frame structure Right overview of a sheet pile installation
(source Waarts 2000)
The conclusions of this research showed that DARS serves its purpose and that the benefit from
using it increases with increasing number of random variables In Figure 36 the Limit State
Function Evaluations (LSFE) that are carried out as a function of the number of variables are
presented
Literature study
MSc Thesis A Rippi 30
Figure 36 Number of LSFE as a function of the number of variables n The upper line depicts the
required number of LSFE N=160 (it is supposed that on average 3 LSFE per sample are required
and consequently the number of LSFE equals 3x160n=480n) The lower line shows the
performance of DARS (source Waarts 2000)
More recently a methodology for assessing the reliability of geotechnical structures and in
particular deep excavations has been developed by Schweckendiek (2006) More precisely this
study deals with the application of different reliability methods in combination with FEM which
carries out the LSFE The reliability methods are applied initially on simple examples in order to
be calibrated and eventually the most efficient methods are used for the reliability assessment
of a deep excavation with a sheet pile wall The several failure mechanisms are treated in detail
concerning the structural elements of the excavation as well as the soil medium
Finally the failure probability of the system is estimated according to a fault tree
schematization of a deep excavation with one layer anchored sheet pile wall Uncertainties in
the soil properties the phreatic levels and also the corrosion of the sheet pile wall were taken
into account In Figure 37 a picture of the case study is shown together with a FEA illustration
and the importance factors of different soil properties in terms of the probability of failure after
a FORM analysis
Literature study
MSc Thesis A Rippi 31
Figure 37 Finite Element simulation and contribution of several structural and soil properties in
the failure mechanism of an anchored retaining wall in layered soil (source Schweckendiek 2006)
For a similar case study as in Schweckendiek (2006) a cantilever and an anchored sheet pile
wall the safety of the structure based on LEM is evaluated in a research conducted by Gonzaacutelez
et al (2013) Moreover in order to compare the efficiency of the classical methods with FEM
the finite element software Plaxis has been used This research shows the limitation of the
classical models and the efficiency of FEM as far as an anchored sheet pile wall in concerned
More precisely LEM results found to be conservative as they do not take into account the
confinement around the free length of the anchor rod that increases the passive pressure After
comparing numerical results with the classical methods in anchored walls the failure
mechanism given by the FEM suggests an intermediate failure state where the sheet pile wall
describes a translation movement on the deep zone and at the same time a rotation movement
around the anchor point that is closer to the reality
Furthermore Moumlllmann and Vermeer (2008) conducted a reliability analysis of a dike failure A
case study at river Elbe in Saxony Germany has been used and the failure probabilities of
different dike cross sections involving different failure modes were compared More precisely
overflowwave overtopping upliftpiping slope stability and damage of the revetment on the
waterside were taken into account For each failure mode the particular failure probability was
determined using the software package PC-Ring (Vrouwenvelder et al 1999) while the
reliability method used for the assessment of the failure probability was FORM That was
coupled with a slope stability software where Bishoprsquos approach was performed
However this method is limited to circular slip surfaces and prescribed pore pressure
distributions within the dike In order to overcome this limitation FEM was used for further
analysis In combination with FEM an adaptive response surface technique was used in order to
represent the limit state function of each failure mode In Figure 38 the simulated structure
Literature study
MSc Thesis A Rippi 32
together with the response surface of the slope instability failure mode is showed Note that this
structure is approached with a single soil layer (same soil within and under the dike) and that
no structural element is implemented inside the dike
Figure 38 Dike model for the probabilistic FEA of dike stability and the corresponding response
surface of the limit state function for 3 stochastic input parameters (source Moumlllmann and
Vermeer 2008)
The failure probability of each of the aforementioned failure mechanisms were finally calculated
and compared with those recommended in each case while the performance of the coupling
between FEM and response surface method was assessed
Additionally a more recent investigation in terms of dikes with sheet piles employing FEA was
carried out by Breedeveld (2011) The main objective of this study was to display the
implementation of partial factors for design purposes using FEM and especially the software
Plaxis in simulating dikes with a sheet pile wall The dike was firstly simulated as a single
structure without reinforcement while in sequence the occurring stresses and pore pressures
were calculated with an existing sheet pile wall The results of the distribution of the effective
stresses within and below the dike are illustrated in Figure 39 His case study has been
introduced in Chapter 2 while in Chapter 6 a more detailed explanation of how it was modelled
in FEM is given This case study will be used as an example case in this thesis
(a)
Literature study
MSc Thesis A Rippi 33
(b)
Figure 39 Effective stress distribution before (a) and after (b) the installation of the sheet pile wall (source Breedeveld 2011)
Last but not least Bach (2014) has been occupied with the reliability analysis of bored piles and
the case study that he examined is presented in Figure 310 The major objectives of this
research were (1) Propose models to calibrate resistance factors for the Load and Resistance
Factor Design (LRFD) (see section 423 for further explanation of this method) approach and
find a suitable model aiming to directly determine reliability of a bored pile considering some
types of defect that may occur in the bored pile (2) Select a quality control method and evaluate
its reliability when applied to bored piles
For that purpose he made a coupling calculation between the finite element software Plaxis
(version 90) and the numerical probabilistic toolbox Prob2B in order to design bored pile
foundations in light of the reliability-based design Two failure modes the geotechnical failure
mode and the structural failure mode were proposed in order to comprehensively assess the
reliability of an axially loaded pile The parameter uncertainty was considered through the use
of statistical parameters and probability distributions for material parameters in soil models
The soil parameters were treated as random variables The geometry parameters of pile were
used as deterministic quantities because a change in pile shape in the calculation process
requires establishing a new mesh which is now impossible with regard to the Plaxis software
Figure 310 Nine calculation cases considered with different pile toe levels (source Bach 2014)
Literature study
MSc Thesis A Rippi 34
For the reliability of the pile the ULS of the pile was investigated and finally a displacement
criterion of a percentage of the pile diameter used depending on the soil type and
characteristics that the structure is founded on
33 Overview In this chapter concepts that are related to FEM probabilities uncertainty and sensitivity
analysis were discussed The introduction of FEM applications and the principles of how to
handle uncertainties will help the reader to better understand and follow the process of this
thesis
Furthermore preliminary researches related to several aspects of this thesis are presented that
mainly concern
Reliability methods and their applications [Waarts 2000 Schweckendiek 2006 Bach
2014]
Coupling FEM with reliability methods [Waarts 2000 Schweckendiek 2006 Bach
2014]
(Reliability) Analysis of geotechnical structures and especially dikes reinforced with
retaining walls [Waarts 2000 Schweckendiek 2006 Moumlllmann and Vermeer 2008
Breedeveld 2011 Gonzaacutelez et al 2013 Bach 2014]
As someone might have noticed from the overview of the existing research on the topic both the
performance of different reliability methods and the simulation of a dike with a sheet pile wall
in FEM have been carried out These references were quite helpful in order to get an idea of
coupling FEM with reliability methods as well as FEM and reliability methods individually
Moreover a first dive into the system under investigation and the variables to be handled was
made However noone has analysed the reliability of such a complicated system that is more
and more often seemed to be applied in the Netherlands Thus it would be essential to evaluate
the implementation and the results of the reliability analysis of a dike with sheet piles simulated
in FEM The conclusions of such a research are going to contribute to design and optimization
concepts and hopefully to a better understanding of the system behaviour
MSc Thesis A Rippi 35
4 Structural Reliability Analysis
The aim of this chapter is to introduce the basic theory of a reliability analysis and to summarize
the principles of different reliability methods that are widely used for the uncertainty
evaluation of a system Particularly three of them (MC DS and FORM)12 were used in this thesis
and are described in this chapter Finally a brief evaluation of the selected reliability methods
and the way that they are applied on this thesis by coupling them with FEM are discussed
41 Basics of Reliability Analysis In the consideration of the reliability of an element the determination of the probability of
failure is the central issue The limit between failure and non-failure is defined as a limit state
and the reliability is the probability that this limit state is not exceeded The limit states are
interpreted through the so-called limit state functions (LSF) whose general form is
(41)
in which R is the strength or more general the resistance to failure and S is the load or that
which is conducive to failure (ldquosolicitationrdquo) The basic principle of structural design is that the
resistance needs to be higher than the load or in other words that the LSF is larger than zero
(Zgt0) The main objective of the design is to ensure that this performance criterion is valid
throughout the lifetime of a structure Nevertheless the majority of the quantities that both
resistance and load consists of are uncertain Therefore a probability of satisfying the preceding
criterion is estimated and expresses the reliability of the structure P(Zgt0) The probability of
failure is
(42)
Hence the probability of failure is complement to the reliability to the sense that
(43)
The LSF is plotted in Figure 41 indicating the failure space and the limit state function by Z=0
12 Besides RS techniques were used for proceeding with the analysis but most for solving numerical problems rather than evaluating the failure probability The RS techniques are still under investigation and need to be tested before being applied in real cases However it is highly possible that this would be a potential way of evaluating the failure probability in the future
Structural Reliability Analysis
MSc Thesis A Rippi 36
Figure 41 Limit Sate function
In this case the design concept is based on the evaluation of the so-called design point which is
the point in the failure space with the greatest probability density Generally this point is
located on the border between the safe and the unsafe areas
In the structural domain the Joint committee on structural safety (1981) proposed a level-
classification of the calculation methods This classification includes the following three levels
Level III these methods calculate the probability of failure by considering the probability
density functions of all strength and load variables The reliability of an element is linked
directly to the probability of failure
Level II this level comprises a number of methods for determining the probability of
failure and thus the reliability It entails a linearization of the LSF in a carefully selected
point These methods approximate the probability distribution of each variable by a
standard normal distribution
Level I at this level no failure probabilities are calculated The level I calculation is a
design method according to the standards which consider an element sufficiently reliable
if a certain margin is present between the representative values of the strength and the
loads This margin is created by taking the so-called partial safety factors into account in
the design
In the next section the different reliability methods to be used in this thesis are further
elaborated
42 Overview of Reliability Analysis Methods
421 Level III Methods
The foundation of the Level III failure probability calculation is the mathematical formulation of
the subset of the probability space which involves failure (Zlt0) Level III reliability methods
Structural Reliability Analysis
MSc Thesis A Rippi 37
(also known as fully probabilistic methods) compute the probability of failure based on the
exact probability density function and the exact limit state functions Therefore such methods
are considered to be the most accurate Well-known sampling methods are (Quasi-) MC
simulation Importance Simulation DS and the Latin Hybercube Simulation Below the former is
discussed as it represents the fundamental theory on which the sampling methods are based on
while DS is also further explained as it is continuously used in this thesis For the rest a brief
explanation is provided in Appendix F There are also other sampling methods such as the
Numerical and the Directional integration which will not be described in this study However
for further information about them a reference though is made to Waarts (2000) In Level III
methods errors can only occur by a too low number of simulations
Crude Monte Carlo Simulation
The MC method uses the possibility of drawing random numbers from a uniform probability
density function between zero and one If and are the
vectors for the resistance and the load respectively that consists of n variables then by taking
multiple realizations of the uniform probability distribution a value can be determined for
every and
By repeating this procedure a large number of times the failure probability can be estimated
(44)
where is the total number of simulations (nm draws from the uniform distribution in which
m is the number of base variables) and is the number of simulations for which Zlt0
The simulationrsquos relative error is
(45)
If the expected value of the relative error is zero the standard deviation is
radic
(46)
Based on the central limit theorem the error is normally distributed provided is sufficiently
large The probability that the relative error is smaller than the given value E is then
(47)
Thus for the reliability of the relative error is smaller than For the wanted k and
E the required number of simulations can be determined with
Structural Reliability Analysis
MSc Thesis A Rippi 38
(48)
If for instance a reliability of 95 is required for a maximum relative error E=01 the required
number of simulations amounts to
(49)
The number of simulations is therefore still dependent on the probability of failure
MC simulation is applicable whatever the complexity of the deterministic model is However the
number of samples required to get an accurate estimation of may be dissuasive especially
when the value of is small (10-k) Thus crude MC is not applicable when small values of are
sought andor when the CPU cost of each run of the model is non-negligible
In Appendix F and especially in sections F1 and F2 a further elaboration of the MC simulation
is presented Moreover the methods of Quasi-MC Importance sampling and Latin Hybercube
are described
Directional Sampling
The directional simulation method is an accelerated sampling method It implies a preliminary
iso-probabilistic transformation as for FORM and SORM methods however it remains based on
sampling and is thus not an approximation method Below the method as it is described in Open
TURNS 15 Reference Guide (2015 pp190-193) is presented
Each simulation of the directional simulation algorithm is made of three steps Using the
probability distribution of a random vector we seek to evaluate the following probability
( ) (410)
Here is a deterministic vector and ( ) the limit state function which enables the
definition of the event For the iteration these steps are the
following
Let A point is drawn randomly on S according to a uniform
distribution
In the direction starting from the origin and the passing through solutions of the
equation ( ) (ie limits of ) The set of values of that belong to is
deduced for these solutions it is a subset of
Then one calculates the probability By property of independent
standard variable
is a random variable distributed according to chi-square
distribution which makes the computation effortless
Structural Reliability Analysis
MSc Thesis A Rippi 39
Finally the estimate of the probability after N simulations is the following
sum
(411)
The main idea is that for each direction we go along the direction by step of a specified length
from the origin to the maximum distant point and we check if there is a sign change on each
segment so formed (see Figure 42) We go until the maximum distant point Then for all the
segments where a sign change is detected we research the root λ in the segment (there are
also other strategies available in OT however all of them follow the same principle of
searching) The following figure illustrates the principle of an iteration in dimension 2
Figure 42 DS of a 2-dimensional problem
Similar to MC method the outcome of the estimated probability of failure is a random variable
and the error in the estimate can be made as small as possible by taking a sufficient number of
samples For DS the standard deviation σ of the estimated failure probability can be quantified
as follows (Melchers 2002)
radic
sum
(412)
From this equation relative errors and the confidence intervals can be estimated As can be
seen the error in the estimated probability of failure will decrease with increasing number of
sampled directions Equation (412) can be used to determine the number of sampled directions
that is required for a reliable estimate of the failure probability The convergence criterion in
this method is usually the coefficient of variation (CoV) which is defined as
Structural Reliability Analysis
MSc Thesis A Rippi 40
422 Level II Methods
Level II methods (known also as fully probabilistic methods with approximations) can take all
the probabilistic properties of the random variables into account but they include
approximations of the limit state function and therefore their use and outcomes should be
inspected and evaluated in order to be considered as reliable However experience in that
methods has shown that the computational effort is profoundly decreasing in comparison with
Level III methods and that the application of them can provide important parameters such as
the reliability index (β) and the influence factors (α) Some of the most known Level II methods
are First Order Second Moment (FOSM) Method First and Second Order Reliability Methods
(FORM and SORM) and the Point Estimate method (PEM) In the next paragraphs FORM is
elaborated as it will be mainly used in this thesis while in Appendix F3 the principles of FOSM
are discussed
First Order Reliability Method (FORM)
The principles of FORM are based on the First-Order Second-Moment (FOSM) approach which
first introduced the reliability index concept This method is presented in Appendix F3 in detail
However the reliability index estimated using the FOSM approach is not ldquoinvariantrdquo which
means that the index value depends on the limit state function format (Farrokh 2007) Hasofer
and Lind (1974) proposed an invariant definition of the reliability index that leads to a
geometric interpretation known as first-order reliability method (FORM)13 The first step of this
method is the transformation of the random variables to equivalent standard normally
distributed variables and the whole procedure is carried out in u-space (or else standard space)
For variables that are normally distributed this step is as follows
(413)
For other types of distributions there are procedures available for carrying out this
transformation such as Generalised Nataf and Rosenblatt transformations In this study these
transformations are not discussed however for further explanation a reference is made to the
Reference Guide of OpenTURNS 15 (2015)
In the sequence the limit state function Z is expressed in terms of
The second step is the approximation of the function with the first two terms of the Taylor-
polynomial The approximation reads (CUR 1997)
( ) ( ) sum
(414)
where is the point where the linearization takes place This approximation of Z is linear and
according to the central limit theorem it is normally distributed The expected value of the LSF
can be approximated by the expected value of the linearized function
13 VNK2 uses FORM in view of the calculation time required In case of inaccuracies VNK2 switches to more time-consuming techniques such as DS whereby a large number of draws are made from the probability density functions of the various stochastic variables (uncertain quantities)
Structural Reliability Analysis
MSc Thesis A Rippi 41
( ) sum
(415)
While the standard deviation (sum of contributions of each variable to the variance of Z) is
defined as
radicsum
( )
(416)
Using the definition of FOSM for the reliability index [ ]
where [ ] is the mean value of
the limit state function and substituting Eq 415 and Eq 416 in it the reliability index can be
approximated However linearization in different points can lead to different values of the
reliability index Hasofer and Lind definition of the reliability index overcomes this limitation
and renders it equal to the minimum distance from the origin to the design point
(417)
where radic
Looking for the design point is basically an optimization problem Many analytical and
numerical approaches can be used for that purpose A relatively straightforward method to do
this is by firstly assuming that the design point is the mean value (the starting point can also be
another point according to an engineering judgement in order to accelerate the optimization
procedure) The obtained β-value is used to determine a new point in which the LSF is
linearized In this case the importance factors αi are calculated as
( )
radicsum (
( ) )
(418)
where is the design point and is the number of variables expresses the
contribution of the variance of each variable to the total variance of Z in the design point Figure
43 illustrates the linear approximation of the limit state function and the aforementioned
parameters for a two dimensional problem The probability that Zlt0 can be determined using
the standard normal CDF
(
) (419)
Structural Reliability Analysis
MSc Thesis A Rippi 42
The new calculation point is determined by
(420)
After some iterations the chosen optimization algorithm finds the final design point and the
new reliability index In Appendix A2 an overview of the different optimization algorithms
available in OT is presented together with an evaluation of their performance in the specific
case study
Figure 43 Two-dimensional illustration of u-space LSF and design point
423 Level I Methods (semi-probabilistic)
At the beginning of the probabilistic concepts incorporation in the field of structural
engineering the most notable development was the implementation of LFRD At the past single
factors of safety on the ratio of total resistance to total load were used in order to ensure that
the stresses developed from an applied load on a structure were lower than the allowable ones
LRFD replaces those factors by introducing a set of partial safety factors on the individual
components of load and resistance These partial factors can be selected such as they account
for uncertainties associated with the resistance and the load respectively
The current design philosophy in CUR 166 and Eurocode is characterized by the use of
characteristic values of the parameters (in Appendix E a further explanation of the
characteristic values is given together with the way that they are converted into the mean
values for being used in the reliability analysis later on) The values of the partial safety factors
are additional factors to the characteristic values Therefore the characteristic values are
multiplied with the aforementioned safety factors ( ) and the following criterion has
to be satisfied for a reliable structural performance (see Figure 44)
(421)
Structural Reliability Analysis
MSc Thesis A Rippi 43
The main disadvantage of LRFD is the fact that there is a set of load and resistance factors that
need to be calibrated in order to cover the specific cases that are likely to occur This is because
of the lack of sufficient statistical data that are necessary for this calibration procedure
Consequently it cannot reassure that all the designs to which the set of factors is applied result
in the reliability level that was aimed for As a result in majority of the cases a conservative
calibration that gives the required reliability is chosen which can lead to ldquoover-designrdquo
structures
Figure 44 Design with partial factors for the load and resistance parameters (source
Schweckendiek 2006)
The evaluation of an elementrsquos reliability starts in principle with calculating the probability of
failure and subsequently the reliability for the given strength and load In practise the problem
is often that the strength is unknown but it has to be determined for a given reliability The
determination of the required reliability can be estimated with the help of Level II and III
methods by iteratively adjusting the strength in the calculation until a sufficiently small
probability of failure is found
In the design domain regulations and guidelines follow the standard that the characteristic
value14 of the strength is divided by a factor and that the characteristic value of the load is
multiplied by a factor as it is described in Eq 421
The link between Level I and the estimation of the failure probability has been achieved through
Level II methods The design point that results from a Level II method calculation is the point
with the greatest joint probability density of the strength and the load and it is therefore
possible that when failure occurs the strength and load values will be close to that point The
design values can be specified as
(422)
(423)
14 The characteristic value of a parameter is its main representative value and its value can be specified as a mean value an upper or lower value or a nominal value according to EN 1990 definition (2002)
Structural Reliability Analysis
MSc Thesis A Rippi 44
where are the coefficient of variation for the resistance and the solicitation respectively
15 and are the importance factor and β is the reliability factor These can be also expressed
in terms of characteristic values as follows
(424)
By substituting function 423 with 421 and 422 the partial safety factors are defined as
(425)
(426)
where 16 and are the values for load or resistance respectively to which a probability of
(non-)exceedance of 95 corresponds (non-exceedance for the load and exceedance for the
resistance) (for a standard normal distributed parameter this is 1645)
424 Response Surface Techniques (RS)
In case that the models presented above tend to be time consuming for the limit state function
evaluation or convergence problems of the optimization algorithms occur a better methodology
is recommended known as response surface The RS is mainly used when (unknown) response is
only available from experiments or complex FEM computations (for example large highly-non-
linear FEM models) Indeed once a RS has been built up the various standard methods may be
applied at a negligible cost An analytical limit state function replaces the real response function
The main idea is that the response consisting of a complex function of input variables is
approximated by a simple function of the input variables
A list of possible response surfaces techniques is given below
Linear and quadratic Taylor expansions
Polynomial RS based on least squares
Kriging method
Polynomial chaos expansion
A detailed explanation of the aforementioned methods can be found in OpenTURNS 15 (2015a)
The first method is associated with the approximation of the model response around a specific
set of input parameters while the rest seek a global approximation of the model over its whole
domain of definition The most sophisticated one is the polynomial chaos expansion but one the
15 α values are negative for solicitation (ie RV for which increasing their values lead to adversely effects on the reliability) and positive for resistance (ie RV for which increasing their values implies a beneficial effect on the reliability) 16 is negative and can be negative or positive
Structural Reliability Analysis
MSc Thesis A Rippi 45
other hand its complexity does not make it attractive In this research a polynomial RS is
applied based on least squares
In combination with FEA the standard procedure is as follows
1 Select the most important random variables on the basis of engineering basis
2 A SA is carried out in combination with FEM
3 Reduce the stochastic variables if needed according to the SA
4 A RS is constructed through the response data
5 A reliability calculation is carried out using the RS instead of the real response
A polynomial RS is generally constructed by fitting a quadratic function to the sampling points
whose general expression is
sum
sum
sum sum
(427)
The type of the reliability method to be used is of little importance since the time consuming
LSFE (using FEM) are replaced by analytical expressions However of main importance is now
the accuracy of the RS compared to the real response There are several parameters in order to
evaluate the goodness of fit of the RS to the real model among which is the well-known R-
squared
In Figure 45 an example of a quadratic response surface is showed that is fitted on the data
response indicated with the black circles This was a 6-dimensional problem that means 6
different variables were included in the system (non-visualized dimensions are kept constant at
their mean value) In this thesis the response of the system is to be acquired after the coupling
of FEM with a reliability method and that would give the ldquoblackrdquo circles that are shown in Figure
45 The response surface technique applies a curve fitting on the real response of the system
(ldquoblackrdquo circles) whose analytical formula can later be used for the reliability analysis of similar
type of systems with negligible computation time
Structural Reliability Analysis
MSc Thesis A Rippi 46
Figure 45 Quadratic response surface with 28 terms (ie ) for 6 dimensions ( ie x
=[ x1 hellip x6])
43 Coupling Reliability Analysis with FEM The limit state function evaluations (LSFE) will be carried out with the software Plaxis 2D 2015
which is a special two-dimensional finite element software used to perform deformations and
stability analysis for various types of geotechnical applications Moreover considering the case
study under investigation in this thesis Plaxis in contrast with other FEM such as Abacus
Comsol DIANA etc offers several techniques to realistically simulate structural elements such
as sheet pile walls and anchors and their interaction with soil while the variety of the
constitutive models for the soil body that are available and the ability to include the history of
the construction phases can lead to a better analysis of the systemrsquos behavior in terms of the
stress level and the deformations It is essential at that point to mention that using FEM for this
purpose means that the limit state formulation is implicit and can only be solved numerically
The reliability analysis is carried out through an uncertainty package In this section firstly a
description of the reliability package and its possibilities are given Finally an explanation of the
coupling procedure between the reliability tool and FEM is given together with the calculation
process that was followed for the parameters manipulation
431 The functionality and possibilities of OT
The reliability analysis tool to be utilized in this research is OpenTURNS (OT) which is a
scientific library usable as a Python module dedicated to the treatment of uncertainties and it is
still under development during the work of this thesis Several reliability packages are available
such as OpenEarth (applied in Matlab) Prob2b (applied in Probox software of TNO) and
Probabilistic Toolkit (software of Deltares) Plaxis 2015 features Python connection possibility
and thus coupling was decided to be carried out with an uncertainty package in Python and OT
is the most developed one Moreover until now only OT is an open source package and can be
Structural Reliability Analysis
MSc Thesis A Rippi 47
used easily from anyone as there is a wide community of experienced people supporting with
their knowledge while many related manual reports are already available for starting learning
A list of the current available reliability methods in OT is given in Figure 46 In this figure apart
from the standard methods the alternative method of the Adaptive Response surfaces is
introduced that is mainly used when (unknown) response is only available from complex FE
computations
Figure 46 Reliability methods available in OpenTURNS
In this thesis the methods to be basically applied are the FORM and the DS analysis while MC is
also used mainly for confirming the application of FORM Additionally the RS technique has
been mainly implemented for enhancing the performance of FORM and DS analysis
OT handles 47 types of distributions amongst which the ones used in this research are Normal
Lognormal Uniform and Truncated Normal distribution Regarding the joint distributions that
are available to be selected for random vectors 11 types of copulas are existing in OT amongst
which the most known ones are the Independent the Gumbel and the Normal copula In this
thesis the independent copula has been used during the reliability analysis which means that
the variables were assumed to be independent among each other However it is strongly
advised a further elaboration and research considering correlation matrices for specific soil
parameters
432 Coupling OpenTURNS-Plaxis
The finite element simulation will be carried out in Plaxis 2015 Plaxis is a finite element
software for plane strain and axi-symmetric modelling of soil and rock behaviour Moreover it
supports a fully automatic mesh generation allowing for a virtually infinite number of 6-node
and 15-node elements
Structural Reliability Analysis
MSc Thesis A Rippi 48
The coupling of reliability analysis and FEA requires an interface for the communication
between each other When a reliability tool is coupled with another software program the
reliability program carries out the whole reliability analysis and it uses the other program only
for the evaluation of the limit state function More precisely OT should be able to read and
amend Plaxis output values for important variables such as material parameters pore pressures
generation and stresses development and corresponding deformations inside the dike
Respectively Plaxis has to be also capable of obtaining the new values that has been set by OT
for the variables that are treated as stochastic during an iterative process according to the
reliability assessment In Figure 47 an illustration of the coupling methodology and its function
is shown
Figure 47 Coupling scheme OpenTURNS-Plaxis
In principle an input file is firstly required where the user set the preferable reliability method
to be used the stochastic input parameters and their probability distributions the joint
probability distribution and the corresponding correlation matrix and finally the limit state
function is formed depending on the situation In Appendix D an example of the input files that
were used for soil sheet pile wall and anchor analysis is shown However such input files
should be interpreted so as to be readable by both Plaxis and OT Therefore an input interpreter
was created which is actually a python script that helps OT to start up the reliability analysis
according to the assigned method variables distributions and LSF As it was mentioned before
the evaluation of the limit state function is conducted by Plaxis For that purpose the input
interpreter should be also able to send the next set of input parameters to Plaxis However an
additional means of connecting the interpreter with Plaxis is also needed This is can be
Structural Reliability Analysis
MSc Thesis A Rippi 49
achieved via a Plaxis interface which actually call Plaxis to make the calculations while it also
transfers the required value of the limit state function to input interpreter and this in turn to OT
Eventually the probability of failure is obtained as a model result However it is also essential
that Plaxis simulation procedure converges to the desirable criteria and under the physical
boundary conditions that have been determined Likewise the convergence criteria of the
reliability methods shall be manipulated so as the optimization algorithms to able to converge
efficiently (see Appendix A2 for further explanation)
44 Overview In section 42 a summary of the principles of the main reliability methods was made From
Level III methods MC and DS were presented whereas from Level II the basics of a FORM
analysis were introduced Besides the concept of the Level I method was explained Even
though this method was not used in this research for the evaluation of the probability of failure
it was applied in section 643 for the deterministic analysis of the system under investigation
and the preliminary calculations of the dimensions In Appendix F more reliability methods are
introduced and they are further described In the sequence their implementation in conjunction
with Plaxis was discussed through the coupling of FEM with OT For more information
concerning OT special features (ie SA optimization algorithms and probability distributions) a
reference is made to Appendices A and E
Before any of these reliability methods was applied on the case study of this thesis their
performance and their compatibility with Plaxis were tested with simple examples At first the
probability of failure was evaluated with FORM SORM Importance Sampling (IS) crude MC and
DS for simple artificial linear and non-linear limit state functions
and
respectively where R B and F are the random variables distributed normally
or lognormally The crude MC technique has been applied in order to validate the failure
probability in case the various methods give different results Secondly a flexible circular
footing over a one-layer soil material was used from Plaxis tutorial in order to test the coupling
between Plaxis and the reliability method and reassure the functionality of the interface In this
pilot example the objective was to determine the probability that the settlements of the footing
exceed a certain threshold of settlements and thus The respective
components that were taken into account for the failure mechanism was the cohesion (c) the
friction angle (φ) and the specific weight (γ) of the soil
From a qualitative evaluation of the reliability methods that has been done and according to the
experience gained from the aforementioned applications of some methods valuable conclusions
were drawn in terms of their efficiency More precisely regarding the sampling methods it does
not required previous knowledge of the failure domain apart from the IS method in which a
starting point inside the failure domain shall be provided Moreover the calculation effort
depends on the magnitude of the failure probability and the required accuracy whereas for the
IS it depends also on the selection of the sampling region In Waarts (2000) it has been proved
that DS tends to be more efficient than MC for low dimensional problems as it is indicated in
Figure 48 (number of random variables nlt100)
Structural Reliability Analysis
MSc Thesis A Rippi 50
Figure 48 Required number of samples for MC and DS as a function of the random variables
(source Waarts 2000)
As far as the approximation reliability methods (FORM and SORM) are concerned it is not
necessary to be aware of the failure region in advance Furthermore it has been observed that
the required iterations and the calculation time is quite lower in comparison with the sampling
methods However the accuracy of the method is highly dependent on the shape of the LSF
In Table 41 a summary of the most applied reliability methods evaluation is presented
According to this evaluation it was decided that the methods of FORM and DS as well as the
combination of them with RS techniques are going to be tested and evaluated in terms of their
efficiency and robustness These methods were chosen from both Level II and Level III methods
due to their expected reduced computational time the non-requirement of previous knowledge
about the LSF and their satisfactory performance in similar case studies In Chapter 5 the
different failure mechanisms and the relevant LSF that are going to be utilized in this thesis are
presented whereas in Chapter 7 the results deduced from the reliability analysis are elaborated
In Chapter 7 there is also an indication of which reliability method was used for each systemrsquos
component the selection of which was based on the aforementioned evaluation
Table 41Evaluation of reliability methods
Method Previous knowledge Accuracy Calculation Effort
FORM not required full accuracy for
Gaussian variables and linear LSF
depends on LSF linearity and the number of random
variables
SORM not required exact up to 2nd order
LSF error dependent on the shape of LSF
depends on LSF linearity and the number of random
variables
Structural Reliability Analysis
MSc Thesis A Rippi 51
Crude MC not required can be controlled by convergence criteria
depends on the magnitude of failure probability and the
required accuracy
IS required can be controlled by convergence criteria
depends on the choices made for the sampling region
DS not required can be controlled by convergence criteria
for low dimensional problems (nlt100) DS is more efficient
than MC
Especially RS were created based on a SA sampling in order basically to increase the efficiency
of FROM and DS analysis in case of FEM numerical errors Precisely an analytical formula or a
response surface was needed that can approximately represent the performance of the structure
under investigation and that is called in case of Plaxis errors This formula was essential in
order to keep the reliability analysis running by providing a response value for the LSF when
Plaxis calculation is unsuccessful and thus incapable of returning a result It should be
mentioned that in case a RS is implemented for the reliability analysis Figure 47 can be
reduced to a graph consisting of only the input file OT and the input interpreter as no Plaxis
calculations would be necessary any more The LSFE will then be performed through the
response surface that is provided by the user Due to this limited use of the RS method it will
not be further treated in this thesis
Structural Reliability Analysis
MSc Thesis A Rippi 52
MSc Thesis A Rippi 53
5 Failure Mechanisms and Limit State Functions
The reliability analysis of a structure requires the definition of the different failure modes that
are relevant to the corresponding structural elements In this chapter the possible failure
mechanisms of the systemrsquos elements expressed in LSF are identified and their role in the
system reliability is explained Based on these LSF the reliability of the system components was
evaluated These LSFs focus on the ULS whose exceedance implies failure of the corresponding
component
51 Introduction to the system analysis and the limit states For system reliability analysis all the possible failure mechanisms have to be identified and
summarized in a fault tree In Figures 51 and 52 the fault trees of a dike section without
structural elements and a retaining structure with sheet piles are depicted respectively In this
project the combination of these two separate structures ie a reinforced dike section with an
anchored sheet pile wall is to be simulated and studied in terms of its failure modes
As it is described later in this chapter failure can be expressed in different ways depending on
the structure and stakeholders demands and safety standards This research is mainly focused
on the ULS of failure and Mohr Coulomb was chosen as the soil failure criterion due to its
efficiency on detecting failure Such a failure criterion is actually defining the stress strain
relationship and the gradual weakening of the soil (ie the gradual reduction of the strength
parameters such as the friction angle and the cohesion due to the stresses development until
soil collapses) under a stress condition In Plaxis new version there is the ability of gradually
reducing structural properties such the strength of steel elements However the results
retrieved of such a safety analysis have not been yet investigated and for that reason such
calculations are not included in this project
As far as the dike section is concerned the failure mechanism to be investigated in this project is
the macro instability of the slope rather than the other mechanisms depicted in Figure 51 On
the other hand because of Plaxis limitations to distinguish among the different structural
elements failure mechanisms each element was considered as a singular case for investigating
its failure mode In the next section a distinction is also made between the ULS and the SLS
For a reinforced dike with retaining walls there are basically four classes of structural elements
(see Figures 12 and 21)
∙ Retaining wall (ie sheet piles)
∙ Anchors
∙ Walings
∙ Soil structure (ie dike section)
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 54
For the reliability analysis of each of the elements the probability of failure is to be determined
In the next sections of this chapter the respective LSF that are considered for each class are
further elaborated according to the ULS criterion while in the end the general combined fault
tree is to be schematized
Figure 51 Fault tree of a dike section
Figure 52 Fault tree for retaining structures using anchored sheet piles (source Schweckendiek 2006)
The system failure in this research is considered as a serial system of the anchor sheet pile
wall and soil body LSFs The fault tree of the system failure is presented in Figure 53 together
with the LSF of each component as they are formulated according to sections 522 and 523
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 55
Figure 53 Fault tree of the system failure
It should be mentioned at that point that the probability of failure of the system differs from the
overall probability of flooding which takes into account any potential failure mechanism of the
system combined with the uncertainties of the water elevation More precisely as far as the dike
safety is concerned a probability of failure (where ) under a certain
water level is estimated from the fault tree of Figure 51 which in sequence is multiplied
with the occurrence probability of the corresponding water level in order to estimate the
overall probability of failure for the specific water elevation Then that product is accumulated
over a required range of water levels that can jeopardise the overall stability of the structure in
order to calculate the overall probability of failure or else the probability of flooding This
probability is then compared to the one established from the safety standards so as to reassure
the safety or not of the structure can be estimated as follows
int
sum (51)
In this thesis the probability is to be estimated where is a specific water level
and especially the design water level as it is considered to be the most challenging part of the
procedure described above gives the probability of occurrence of the water level (ie it is
the PDF of ) After setting up the steps and implementing them successfully for the estimation
of the calculation of the overall probability of failure is just a repetition of the same
procedure for more water levels It should be mentioned that in this thesis failure consists only
of the macro-instability failure mode whereas the rest are excluded for the time being A
simplified alternative for the overcoming the integral of Eq51 is to repeat the calculations for
certain water levels and then sum their products instead of integrating the full range of the
them
52 Limit State Functions
521 Serviceability Limit State
The SLS is evaluated in design calculations in order to avoid excessive deformations of a
structure that could lead to the loss of its service requirements and its functionality In some
cases the SLS criteria are used to prevent damage to adjacent structures in the surroundings
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 56
For example concerning the design of a dike section attention should be paid to deformations
caused to buildings that are located in the inland part of the dike in case of an extreme event
In Figure 54 an example of a deformed dike is depicted In this figure some of the potential
locations that excessive deformations might be experienced are illustrated such as
∙ the vertical settlement of top of dike (arrow 1)
∙ the horizontal displacement of the top of the sheet pile (it might also be the middle part
of the sheet pile that ends up with the largest deformation according to the loading
conditions and the specific soil structure different deformations might occur) (arrow 2)
∙ Uplift upstream of the dike (arrow 3) special attention shall be paid to this type of
deformation not only due to the direct effect on inland structures but also as a sign of
developing piping mechanism
(a)
(b)
Figure 54 Figure (a) shows the initial undeformed structure while (b) illustrates the deformed mesh (not to scale) (the plotted length of the anchor does not have any physical meaning) In (b) the possible locations and directions (black arrows) of potentially excessive deformations
According to the current Dutch regulations (Larsen et al 2013) the SLS criteria for designing
reinforced dikes with sheet pile walls require that
∙ the settlement on the top of dike does not exceed 10 cm and
∙ the horizontal displacements of the sheet pile wall does not overcome 10 cm
In this research the SLS criteria and especially the dike settlements were used roughly for the
preliminary determination of the structures characteristics that are discussed in section 643
Dike Anchor
Sheet pile wall
1
2 3
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 57
More precisely displacements at the top of the dike were also taken into account as
complementary to the safety factor in order to determine the required length of the sheet pile
wall As far as the reliability analysis in concerned the LSF were formulated based on the ULS of
the different systemrsquos components that are presented in the next section
522 ULS for Structural Members
In the design process one is most interested in the ULS of a failure mechanism This state
describes the situation wherein the acting extreme loads are just balanced by the strength of the
construction If that limit state is exceeded the construction will lose its functionality and thus
collapse or fail In general most attention is paid to the behaviour of the structure after
completion However during construction there are also periods in which the construction may
fail The different phases of construction are listed in section 642 In the present section the
analytical LSF of the structural elements are given as they are going to be used in the reliability
analysis
Sheet pile wall
The most relevant failure mode for the sheet pile wall is the exceedance of the yield strength
which corresponds to the ultimate steel strength The response of the structure is mainly due to
bending moments and the axial forces (shear forces are considered to be negligible) Where an
axial force is present allowance should be made for its effect on the moment resistance
Figure 55 The water and earth pressures as well as the anchor forces lead to bending moments and axial forces on the sheet pile wall
In Figure 55 an example of the axial forces and the bending moments that can be developed in
the sheet pile wall with one anchor layer are illustrated after an extreme water level loading
Accordingly the maximum stresses on the sheet pile wall are composed of a bending moment
and a normal force component17
[
] (52)
where [kNm] and [kN] are the bending moment and the axial normal force
respectively that depend on the depth level where they are calculated over the sheet pile length
[m3] is the elastic section modulus and [m2] the cross-sectional area of the sheet pile
wall
17 the vertical anchor force component is reducing by its interaction with the soil over depth
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 58
Bending moment and axial force can be variable over the depth and that is why they are
expressed as a function of z-depth FEM has the advantage to take into account second order
effects ie a stiffer structure will experience higher bending moments than a more flexible one
Taking into account the above the limit state function can be formed as the difference between
the maximum developed stress and the yield stress
[
] (53)
where and can be characterized as the load variables while and can be considered
as the resistance variables and are assumed to be constant over depth
Concerning the permissible displacements for the ULS an upper limit for the maximum
horizontal displacements of the top of the sheet pile has been set to 150L where L is the
vertical length of the sheet pile wall (Larsen et al 2013) This requirement was mainly used in
section 643 where the structural properties were determined after the deterministic analysis
Anchors
Anchors are loaded by their reaction to the horizontal loads on the retaining walls The failure of
the anchor element is actually represented by the failure of the steel members of the anchor
(tubes bars cables etc) that are loaded by traction forces
As it is illustrated in Figure 56 the anchor experiences an axial tensile N that is nearly constant
over its length
Figure 56 Axial loading of anchor inside a dike
The elastic behaviour of an anchor involves only a relationship between axial force N and
displacement (elongation) u of the form
[ ] (54)
where EA [kN] is the anchor stiffness consisting of the steel Youngrsquos modulus E [kNm2] and the
anchor cross section A [m2] and L [m] in the length of the anchor
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 59
Similarly to the sheet pile wall the limit state function of the anchor involves the certain yield or
ultimate strength of the steel members and the maximum stress that the anchor experiences
during its loading Consequently the LSF is as following
(55)
where [kN] is the calculated anchor force and [m2] is the cross sectional area of the
anchor (both of them considered to be constant over the depth) It is essential to mention that
the anchor is also subjected to bending moments due to soil settlements (that are implicitly
illustrated via the uniformly distributed load q over the tie rod) that should be taken into
account in order to investigate the displacements of the tie rod itself However in this thesis
only the axial forces on the anchor are considered without taking into account the individual
deformations and its reaction with the surrounding soil
Walings
The waling is the element that transfers the loads from the retaining wall to the anchors (see
Figure 21) The loading of the walings can be schematized as a continuous beam on several
supports as it is depicted in Figure 57
Figure 57 Loading of walings
If Nα [kNm] is the calculated anchor force per m then the developed bending in the support Ms
and in the opening Mo can be approximated as follows
[ ] (56)
(57)
where La is the mutual anchor distance Considering as the design moment of the waling to be
the limit state function can be formulated as follows
Waling Sheet pile Anchor
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 60
(58)
For a conservative design of the waling the limit state function will give the same or lower
failure probability than the anchor itself Therefore it will actually not be necessary to carry out
this analysis The anchor failure will be the determinant mechanism (Schweckendiek 2006)
523 ULS for Soil Failure
In this thesis emphasis is given on the dike global instability which actually consists of several
failure modes Figure 58 indicates the possible failure mechanisms of flood defences The
combination if these can lead to the overall instability of the dike and thus to soil body failure
However Plaxis is not capable of accounting for all of them More precisely Plaxis assumes the
soil to be a continuous body and thus it can model movements in the scale of soil bodies (using a
relative fine mesh in relation to the moving soil body) but not on a soil particle level Plaxis can
simulate the groundwater flow in a soil body and from this someone can deduce input for a
piping analysis (see failure mechanism G in Figure 58) (using other softwaremethods) but as
it has been mentioned above it cannot determine movement of soil particles due to
groundwater flow Moreover Plaxis cannot deal with the flow and waves occurring in ldquoopen
waterrdquo ie water outside the soil in a canal a lake or sea for instance
Figure 58 Possible failure mechanisms of flood defences (source Flood Defences 2015)
Therefore Plaxis can handle the failure modes (C) (D) (E) (J) (K) and (L) of those illustrated in
Figure 57 In this thesis loads due to extreme temperature conditions or ships are not
considered while the settlements are taken into account implicitly though the mechanisms (C)
(D) and (E) These mechanisms are thus investigated in this thesis and they mainly refer to the
macro-instability of the inner and outer slope of the dike or the overall dike body (ie horizontal
sliding (mechanism D))
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 61
For these failure modes soil failure is defined as shear failure In Figure 59 the most relevant
patterns of the macro-instability regarding the shear strength of the soil are illustrated
Applying FEA the most critical failure mode is determined automatically However it is not
always straightforward what the trigger mechanism of the failure was and that is why more
investigation and FEA tests are needed in order to obtain a clear view
(a) Outer slope failure (b) ldquoActive siderdquo failure
(c) ldquoPassive siderdquo failure (d) Overall failure
(e) ldquoKranzrdquo stability
Figure 59 Macro-instability failure modes in a dike with an anchored sheet pile wall
In Figure 59 failure modes (a)-(d) follows Mohr-Coulomb failure criterion In principle the
total stress state inside a dike section consists of the effective stresses and the pore pressures
according to Terzaghi principle
[
] (59)
Mohr-Coulomb failure criterion basically considers as limit for the failure to occur the
maximum shear stress that the soil can withstand This is determined from the friction
angle the cohesion of the soil and the current stress state The drained shear strength
(similarly in for the undrained shear strength) in the soil according to Mohr-Coulomb is defined
as follows
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 62
[
] (510)
where is the effective friction angle In Appendix B1 a detailed explanation of this failure
criterion is attached
Taking into account the above mentioned failure mechanism (a) is mainly triggered by the
increase of the pore pressures in the outer slope which subsequently causes a decrease in the
effective stresses ( ) which at the same time leads to a reduction of the shear strength
(considering Mohr-Coulomb failure criterion) When the shear stresses in the soil exceed the
shear strength a slip plane forms and a soil wedge collapses
Failure modes (b) and (c) are primarily determined from the active and passive effective
stresses respectively As far as the ldquoactive siderdquo failure is concerned displacements of the sheet
pile wall towards the downstream side of the dike lead to development of active stresses
where is the active pressure coefficient During the ldquoactiverdquo failure the
retaining soil is allowed to relax which leads to a decrease of horizontal stresses and
simultaneously to a decrease in shear strength Similarly to failure mode (a) a soil wedge
collapses The result is an increased earth pressure on the wall for which it is not designed
However this type of failure in dikes is not such determinant for the soil and the sheet pile wall
failure because there is always the resistance of the passive side of the dike that keeps the active
side stable enough Such a mechanism is thus more relevant for deep excavations with retaining
walls
In mechanism (c) the passive soil resistance is exceeded by the horizontal loads In this case the
wall moves inland and a wedge of soil compresses The shear strength is larger due to the
deformation of the wall and the horizontal stresses increase stresses where is
the passive pressure coefficient ( ) This failure mechanism usually occurs due to an
underestimation of the sheet pile length or due to the presence of a weak soil layer in the
passive side
Failure mode (d) is actually a combination of (b) and (c) possibly accompanied also with the
development of excess pore pressures as it was discussed in failure mechanism (a)
Moreover too short anchors could cause a soil block to collapse as it is showed in Figure 59(e)
(ldquoKranzrdquo stability)
As it has been already mentioned for mechanism (d) failure modes can be correlated and
combined with each other in order to lead to a final failure state Therefore the common failure
probability would be smaller than the sum of the singular probabilities In this thesis the total
probability of failure of the soil body due to global instability is to be determined This is due to
the fact that the application of FEM is capable of simulating the combination of the failure
modes depicted in Figure 58 and thus the overall failure (related to (C) (D) and (E) modes of
Figure 58) can be evaluated However with the classical engineering approach for the stresses
calculation each one of the failure modes depicted in Figure 59 should be separately evaluated
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 63
Below a description of the available methods to formulate the LSF of the soil failure is given
After the evaluation of these alternatives a selection was made regarding the most suitable LSF
for the current case study
Limit State functions for soil failure
Soil mechanical failure can be analysed in several ways (Schweckendiek 2006)
1 Excessive Deformations
2 φ-c Reduction
3 Relative Shear Resistance
4 Plaxis definition of soil collapse
The possibilities and limitations of the aforementioned methods are briefly discussed below
1 Excessive Deformations
Similarly to the SLS a limit state function can be formulated by deformations that are
unacceptable As it is illustrated in Figure 54(b) there are some points (the top of the dike the
sheet pile top and the inland soil level behind the dike) the displacements of which shall be
limited to the minimum possible during the structurersquos lifetime in order for the structure to
meet the required service standards Therefore the calculated deformations at locations 1 2
and 3 could give a clear indication of failure given the maximum acceptable deformation
In case that there is a limit value for each location then the limit state function can be formed as
follows
[ ] (511)
with this criterion designer can control the developed deformation on the system and improve
the structure so as to limit the displacements to the required level and subsequently increase
the reliability
On the other hand such an approach might exhibit several problems during the reliability
analysis that are listed below
Many failure mechanisms suddenly happen and as a result no significant displacements
are observed before moving very close to failure In sequence this may cause problems
for the iterative procedures of some reliability methods such as FORM and DS
The determination of suitable maximum admissible deformations is not
straightforward They have to be large enough to serve as failure criterion and
simultaneously they shall not be larger than the values that can be calculated within
the limits of equilibrium in FEM-calculations This requires previous knowledge on the
analysed system as well as on the feasibility of FEM-calculations on the specific subject
Last but not least the location of the undesirable displacements in a dike section is also
under investigation as there can be multiple vulnerable spots on the dike that should
be kept under a certain range of deformations in order not to cause a collapse and that
they are not always known in advance
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 64
2 φ-c Reduction
An available type of calculation in Plaxis is the so-called ldquoSafety Calculationrdquo with which global
safety factors can be computed In Appendix B2 a more detailed elaboration of this method is
exhibited This feature can also be used in reliability analysis by formulating the LSF as follows
(512)
where is the Multiplier Safety Factor obtained by the φ-c reduction The main concept is
that if the safety factor is smaller than 1 then it is considered as failure This method can provide
us with the probability of failure for a general soil body collapse However even if this method
seems to be quite simple there are some issues during its implementation in a reliability
analysis
For complex limit state functions there can be convergence problems for some
reliability methods with iterative procedures (ie FORM) This can be explained from the
fact that in a φ-c reduction soil strength properties follows a certain path as it is
described in Appendix B2 independently of the values for the random variables that
have been set from the reliability method
The safety factor is a general safety factor regarding the failure of the system
Therefore sometimes it is not straightforward what the ldquotriggerrdquo factors are that led to a
certain type of failure
A safety calculation in Plaxis is time-consuming and the outcome can be unstable (Plaxis
calculationsrsquo convergence depends on the number of calculation steps)
Plaxis cannot handle safety factors below 1 Therefore a new limit of the safety factor
shall be introduced in this limit state function
3 Relative Shear Resistance
In this method the basic idea is to define soil failure according to the failure criterion that Plaxis
is set to use for the analysis In this research the Mohr Coulomb failure criterion is applied in
combination with a linear ndashelastic perfectly-plastic stress-strain relationship (that is the so-
called Mohr-Coulomb model) For this model the initiation of plasticity is considered as failure
Therefore the maximum shear resistance is defined just before plastic yielding occur for any
given stress state Of course the occurrence of plasticity does not directly indicate the failure of
the soil structure however this model can give a first estimate for the stress state and the
deformations
According to this method the relative shear resistance is defined as the ratio between the
mobilized shear stress and the maximum shear resistance In Figure 510 the Mohr-Coulomb
model is illustrated in a Cartesian stress-plane In the graph the principle stresses σ1 and σ3 are
indicated More precisely the principle stresses are defined as follows
radic
(513)
(514)
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 65
radic
(515)
Figure 510 Mohrs stress circle and Mohr-Coulomb failure criterion
Hence the mobilized shear stress is
radic
(516)
And the maximum shear resistance which is the distance of the mean stress point to the yield
surface is defined as
(517)
Then the relative shear resistance is a measure for load-resistance ratio that can be determined
in any integration point in the soil continuum
(518)
However this criterion requires prior knowledge of the possible relevant failure mechanisms
and their potential locations on the soil structure This is necessary in this method in order to
choose a suitable cluster of integration points where the average value of the relative shear
strength is to be determined Therefore the limit state function to be considered is not
straightforward and it needs problem investigation in advance
4 Plaxis definition of soil collapse
In Plaxis the construction stages are analysed by performing a Load advancement ultimate level
procedure which is controlled by a total multiplier
(519)
where is the load that is applied by Plaxis is the load at the beginning of the
calculation phase (ie the load that has been reached at the end of the previous calculation
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 66
phase) and is the defined load configuration (ie water level soil weight vertical load
on the top of the dike etc) Thus the main idea is that Plaxis applies an incremental load until it
reaches the defined one When has reached an ultimate level (which by default is 1)
the current phase is finished However if a staged construction calculation has not properly
finished the multiplier fails to converge to the desired ultimate level at the end of the
phase and a warning of ldquoSoil collapsesrdquo appears in the Log info box
In other words a collapse load has been reached In this case the total specified load has not
been applied In physical terms this means that the current value of the stiffness parameter CSP
is less than 0015 (Plaxis 2D 2015b) It is a measure for the amount of plasticity that occurs
during the calculation When the solution is fully elastic CSP is equal to unity whereas at failure
it approaches zero Therefore soil reaches its upper limit of plasticity and it collapses which can
be visualized as a settlement of the dike body Such a warning in the analysis is thus assumed to
be a possible failure situation
53 Overview In this chapter an overview of the relevant possible LSF of the structural elements and the dike
body respectively has been given The different LSFs are formulated and discussed whereas the
failure of each component is going to be evaluated in the ULS of each component
The system failure in this research is considered as the combination of the anchor sheet pile
wall and soil body LSF Particularly for the anchor and the sheet pile failure their yield stress
was chosen as a limitation for their failure by adopting the next LSFs
[(
)] sheet pile wall LSF (520)
anchor LSF (521)
As far as the soil body is concerned in this research method 4 is to be used in order to identify
soil failure The limit state function is then formed as a single value in case of a successful and an
unsuccessful computation Therefore the limit state function was chosen to be equal to 1
(actually the value of is retrieved from Plaxis that in case of a successful calculation is
1) in case of a successful Plaxis analysis and -1 in case that soil collapses in any phase (error in
Plaxis Log file ldquoSoil Collapsesrdquo) This formula can be written as follows
(522)
This accounts for the global instability of the dike body in contrast with the other methods that
are related to a certain dike area (methods 1 and 3) or they are cannot handle overall failure
due to numerical issues (method 2) More precisely methods 1 and 3 were excluded due to the
prior knowledge that is required in method 3 while in method 1 a definition of the maximum
admissible deformations shall be first introduced However attention shall be paid to the type
of the warning that Plaxis gives because it might also be that the maximum number of load steps
was not sufficient In that case the phase must be recalculated using a larger number of steps
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 67
Someone could argue that such a failure can be considered as a system failure rather than a soil
failure alone Partially this is true as the incapability of the sheet pile wall or the anchor to keep
the upstream part of the dike stable can lead to non-convergence and thus according to Plaxis
definition to failure This incapability can be translated into various scenarios of failures that
stem from either the structural elements or the soil body weakness and that are illustrated in
Figure 511 This picture shows the different scenarios that can take place due to macro
instability that was discussed in 523 and the failure of the retaining wallrsquos elements
Figure 511 Combined failure scenarios (initial and developed mechanisms) (source Larsen et al 2013)
However this does not mean that the sheet pile wall or the anchor fail because of exceeding
their yield stress but rather due to possible unacceptable deformations This is also an
advantage of the inherent residual strength of the structural elements that allows a large
deformation before they fail However the residual strength of the elements was not taken into
account in this thesis as they were considered as elastic Moreover the message for the ldquosoil
collapsesrdquo can appear in a phase prior to the installation of the wall which means that this type
of failure is exclusively attributed to the deficiency of the soil strength Therefore it is not
always straightforward what is failing first and lead to the system failure Consequently by
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 68
considering such an analysis representative of a system failure it is likely to underestimate the
overall probability of failure whereas if it is considered as a soil failure only it might
overestimate the total probability of failure since it is later combined with the rest LSF of the
anchor and the sheet pile wall In this research it will be considered only for the soil failure that
brings the results to the safe side even if it does not lead to the best optimization In this thesis a
serial system of the anchor the sheet pile and the soil body LSF was considered the fault tree of
which is depicted in Figure 53
Another way to evaluate the reliability of the different elements could be the reach of a
threshold displacement according to the SLS criteria Such a threshold could be also considered
for the system as a whole rather than the individual elements However in that case someone
should be cautious regarding the choice of the value for the maximum deformations as well as
the location of their occurrence
Concluding in this section the possible failure mechanisms of a flood defence (Figure 58) were
shown Then it was clarified which of these mechanisms can be modelled in Plaxis and they
were further explained and schematized in Figure 58 with the presence of the retaining wall as
they would have to be handled in case of hand calculations Plaxis can incorporate these
mechanisms automatically and thus no distinguish among the different patterns of these
specific modes is necessary Last but not least in Figure 511 the possible failure modes for the
system of the dike and the retaining wall that can be simulated in Plaxis are depicted In this
thesis the plasticity of the structural elements is not considered and thus possible plastic hinges
in the anchor and the sheet pile are not taken into account Therefore scenario 3 is excluded
In Chapter 7 the results after the reliability analysis of the aforementioned LSFs are given It
was essential though before proceeding with the reliability analysis to get a better
understanding into the system under investigation For this purpose in the next chapter a
description of the case study as it has been simulated in Plaxis is given while also the different
structural properties are defined for the following reliability analysis The soil variables to be
considered as stochastic are presented and a first insight into of the systemrsquos behaviour is
obtained through the mean values calculation
MSc Thesis A Rippi 69
6 Case Study-Dike with an anchored sheet pile wall
In this chapter a description of the case study is presented together with the relevant soil and
structural parameters In sequence a deterministic analysis follows in order to specify the basic
structural characteristics and obtain a first sense of the stresses magnitude developed on the
structure This was carried out based on the recommendations given in the CUR 166 (2005)
The aim of this deterministic analysis is to redefine the structural properties in order to avoid
having an overdesigned structure understand the current design procedure and see if there are
any possibilities of improvement by applying the proposed reliability analysis
61 Case Description The case study concerns a dike section with an anchored sheet pile wall which was initially used
to showcase the current methodology of designing dikes with sheet piles using partial factors
For more information a reference is made to Breedeveld (2011)
In this research this case study has been modified and simplified for the needs of the thesis and
it is illustrated in Figure 61 while in Table 61 the soil materials are identified Details for the
properties of the soil layers are given in section 62 We assume the structure to be
homogeneous to the third dimension and therefore a plane-strain model is applied The dike
soil materials (layers 3 and 4 in Figure 61) were adopted from the case study used in
Breedeveld (2011) that was presented in Chapter 2 Especially in Figure 25 the original case
study as it was modeled in Plaxis is shown
Table 61 Soil materials that corresponds to Figure 61
1 Aquifer Sand layer 2 Aquitard Clay layer 3 Dike new material Clay layer 4 Dike old material Clay layer
The geometrical properties the groundwater level and the design water level were taken
identical to the original case study as well as the soil properties of the dike materials The soil
layers under the dike were modified for the need of the thesis and were modeled using random
average quantities for which the statistics were chosen arbitrarily but in a realistic range
according to the NEN6740- Table 1 (see Appendix C) included in the Dutch code for
geotechnical structures Moreover the vertical load was removed in order to examine the
impact of the water elevation on the dike stability
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 70
Figure 61 Case study Dike with anchored sheet pile wall Dimensions and Soil materials
62 Soil Parameters The soil parameters for the dike material were obtained from the report of Breedeveld (2011)
where for some parameters the characteristic values were available whereas for some others
the mean values were given For the soil layers below the dike Table 1 of NEN6740 was used as
it was mentioned before which gives characteristic values For those parameters whose value
was a characteristic 18 one a transformation was made according to their probability
distributions in order to acquire the mean values that are required for the reliability analysis
Only the properties required for the use of the Mohr-Coulomb model are presented
The parameters of each soil layer are presented in Tables 63(a) and (b) and the front number of
each layer indicates the corresponding soil layer as showed in Figure 61
The distribution types and the coefficients of variation for each parameter were chosen
according to the knowledge that has been obtained until now about the physically possible
ranges of such parameters and the recommendations that have been given in several researches
up to now (Baecheramp Christian 2003 Schweckendiek 2006 GriffithsampFenton 2007 Phoon
2008 Wolters 2012 Bach 2014) The distributions and the CoV of all parameters are displayed
in Table 62 They were considered to be the same for all soil layers Particularly for the
saturated and unsaturated volumetric weight γsat and γunsat respectively a relationship was
established in order to derive the one from the other In general γunsat varies between the real
dry weight and γsat depending each time on the degree of saturation An estimated maximum
difference between γunsat and γsat is approximately 4 kNm3 (Brinkgreve 2015) In this thesis γsat
was expressed as the summation of γunsat and a variable with uniform distribution in the range
of [0 2]19 Initially a deterministic relationship was defined between the two variables but later
on it was proved the variable was important for the soil failure and thus it was decided to
consider it as random
18 for the definition of the characteristic value see Appendix E 19 The range of [0 4] was not chosen in order to avoid extreme conditions
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 71
Table 62 Soil parameter distributions
Soil parameter Symbol DistributionRelation COV Unit
Unsaturated
Volumetric weight γunsat Normal (microσ) 5 [kNm3]
Saturated Volumetric
weight γsat γsat= γunsat+U(02) 5 [kNm3]
Cohesion c Lognormal (microσ0) 20 [kPa]
Friction angle φ Truncated normal (microσ045) 10 [ ˚]
Youngrsquos Modulus E Lognormal (microσ0) 25 [kPa]
Poissonrsquos ratio ν Truncated normal (microσ005) 10 [-]
Interface strength Rinter Truncated normal (microσ0099) 20 [-]
Table 63 (a) Soil statistical properties under the dike and (b) soil statistical properties in the dike
(a) 1 SAND (very silty)
Soil property Symbol Characteristic
value
Mean
value STD Unit
Saturated
Volumetric
weight
γsat 20 22 11 [kNm3]
Unsaturated
Volumetric
weight
γunsat 19 21 103 [kNm3]
Cohesion c 0 0 0 [kPa]
Friction angle φ 30 36 36 [ ˚]
Dilatancy angle ψ 0 0 0 [ ˚]
Youngrsquos Modulus E 20000 30769 7692 [kPa]
Poissonrsquos ratio ν 025 03 003 [-]
Interface strength Rinter 044 066 013 [-]
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 72
2 CLAY (clean medium)
Soil property Symbol Characteristic
value
Mean
value STD Unit
Saturated
Volumetric
weight
γsat 17 185 093 [kNm3]
Unsaturated
Volumetric
weight
γunsat 17 185 093 [kNm3]
Cohesion c 10 14 282 [kPa]
Friction angle φ 175 21 21 [ ˚]
Dilatancy angle ψ 0 0 0 [ ˚]
Youngrsquos Modulus E 2000 3077 769 [kPa]
Poissonrsquos ratio ν 029 035 0035 [-]
Interface strength Rinter 034 05 01 [-]
(b) 3 DIKE NEW (very sandy clay)
Soil property Symbol Characteristic
value
Mean
value STD Unit
Saturated
Volumetric
weight
γsat 17 185 093 [kNm3]
Unsaturated
Volumetric
weight
γunsat 17 185 093 [kNm3]
Cohesion c 4 564 113 [kPa]
Friction angle φ 29 347 347 [ ˚]
Dilatancy angle ψ 0 0 0 [ ˚]
Youngrsquos Modulus E 1625 2500 625 [kPa]
Poissonrsquos ratio ν 029 035 0035 [-]
Interface strength Rinter 034 05 01 [-]
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 73
4 DIKE OLD (little sandy medium clay)
Soil property Symbol Characteristic
value
Mean
value STD Unit
Saturated
Volumetric
weight
γsat 195 212 106 [kNm3]
Unsaturated
Volumetric
weight
γunsat 19 207 103 [kNm3]
Cohesion c 13 183 367 [kPa]
Friction angle φ 28 335 335 [ ˚]
Dilatancy angle ψ 0 0 0 [ ˚]
Youngrsquos Modulus E 2925 4500 1125 [kPa]
Poissonrsquos ratio ν 029 035 0035 [-]
Interface strength Rinter 034 05 01 [-]
63 Finite Element Model The structure has been modelled with the finite element software Plaxis 2D 2015 In Figure 62
the model as it was created and the automatically generated mesh are presented The ldquoboxrdquo
around the dike section was used to refine the mesh in this area The mesh in the interfaces
between the sheet pile and the soil was also refined for a better representation of the potential
high stresses that can develop in that area
Figure 62 FEM model and generated mesh for case study
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 74
The sheet pile has been modelled with elastic20 plate elements and the free anchor length with
an elastic fixed-end-anchor element The grout body of the anchor was not explicitly modelled
but it is implied in the end of the anchor as its final point is considered to be fixed in the soil A
fixed-end-anchor was chosen because even if it is a simplification of an anchorage system it can
still give reliable results in terms of the anchor axial force which is actually included in the limit
state function However the interaction of the soil and the grout body in that case cannot be
simulated The interaction and the modelling of a grouted body in the soil are still under
investigation due to several difficulties in modelling the real conditions
For all the steel members a Youngrsquos modulus of E=210 GPa was used while the specific
structural parameters are to be specified after the deterministic analysis that is described in the
next section
64 Deterministic Analysis The design concept of dikes with sheet piles modelled in FEM is based on the combination of
several partial factors in order to evaluate the overall required safety factor Additionally the
design recommendations of CUR 166 (2005) are considered for using the design or the
characteristic values of the structural parameters and the loads The design procedure (see
section 22) that has been initially followed in this structure is reported in Breedeveld (2011)
and the following structural elements have been defined accordingly
Cross sectional area of the sheet pile (per meter)
Length of the sheet pile wall
Steel quality of the piles
Cross sectional area of the anchor (per meter)
Free length of the anchor (without the grounded body)
Steel quality of the anchorrsquos reinforcement
Profile of waling
The required overall safety factor was calculated up to 18 according to the current design
procedure described in section 22 according which the above mentioned structural parameters
were defined Moreover apart from the overall safety factor in Breedeveld (2011) additional
partial factors are applied on the deduced forces and bending moments of the anchor and the
sheet pile wall as they are defined in CUR 166 (2005) In particular in Table 64 the different
partial factors that are applied on the several structural developed actions are given
Table 64 Partial factors for the different structural actions according to CUR 166 (2005)
Structural action Partial factor γ
Sheet pile normal force Fn 115
Sheet pile bending moment M 115
Anchor normal force FA 125
Waling bending moment Mw 110
20 The elements were considered as elastic in order to avoid numerical problems in FORM analysis For a better explanation a reference is made to the evaluation of the optimization algorithms in Appendix A2
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 75
However as it is mentioned in section 62 alterations have been made in the soil layers below
the dike while also the vertical load has been removed On the top of that in section 22 it was
shown that the large applied partial factors led to an overestimation of the design values for the
moments of the sheet pile Therefore a redesign of the sheet pile and the anchor is carried out
in this section in order to determine the new required characteristics of the structural elements
according to the new loading situation with less strict partial factors In particular in this thesis
the required overall safety factor was considered 12 according to CUR 166 (2005) calculation
scheme that is described in the next section The reason for using a lower safety factor was to
avoid an overestimation of the design forces and moments of the sheet pile wall that would
render a reliability analysis meaningless as the structure would be quite safe Such an analysis is
also important in order to show the discrepancy between the two different design procedures
and indicate the optimal one
However the reliability analysis was carried out for the mean values of the soil parameters and
for a safety factor of 10 in order to account only for soil properties uncertainties A safety factor
of 10 was not directly used as the design procedure followed in section 643 accounted only for
the macro-instability of the system and thus other failure mechanisms such as piping
overtopping infiltration etc were not taken into account In that way we prevent coming up
with a retaining wall that would be volatile under other crucial failure mechanisms
In section 641 the calculation scheme to be used in this thesis for the deterministic analysis is
further described In the sequence in section 642 the construction phases that were followed
in Plaxis simulation are listed and illustrated and finally in section 643 the results of the
preliminary design are displayed
641 Calculation Scheme and Design Values
In this research only the design philosophy suggested in CUR 166 is applied and the desired
safety factor of the structure is considered to be 12 for structures classified in ldquosafety class IIIrdquo
in order to avoid an overestimation of the structurersquos dimensions as it happened in the original
case study At this point it should be also clarified that the redesign was conducted considering
only the overall stability of the structure and no other failure mechanisms such as piping
internal erosion or overtopping
The prescribed calculation schemes in CUR 166 (2005) for designing structures with retaining
walls are basically the following
Calculation scheme A Calculations with design values The calculations are executed
using the design values for the soil parameters retaining height water levels and
stiffness of the structure It uses two different soil stiffnesses When using a high soil
stiffness the anchor force is generally relatively high The advantage of this scheme is
that it requires relative little effort The disadvantage is that the deformations may be
overestimated due to the use of design values in every stage However a designer is not
interested in deformations when performing a ULS calculation
Calculation scheme B Calculations with characteristic values The calculations are
executed using characteristic values of the soil strength (crsquo and φrsquo) Design values are
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 76
used for the retaining height water levels external loads and stiffness of the soil Here
again two different soil stiffnesses are used In the end of the governing stage a phi-c
reduction should be carried out As the characteristic values for the soil parameters are
used the deformations are most probably smaller than in case of calculation scheme A
The safety margin lies in the fact that phi-c reduction is imposed to obtain a safety factor
of 115 (class II) or 12 (class III)
For the soil stiffness parameters two options are given in CUR 166 a low and a high value
(61)
where is the (low and the high) design value the (low and the high) characteristic value
and is the mean value of the stiffness parameter In Woltersrsquos master thesis (2012) both
have been applied and the appeared to be governing for the bending moment in the wall
and the anchor force
In this research the Calculation scheme B was used and the as the soil stiffness
parameter It should be mentioned that regarding the soil parameters the characteristic values
were available with an exemption for the stiffness for which the mean values were given The
groundwater level was manually schematized according to TAW (2004) because of lack of data
while the maximum water level in the river side was taken equal to the design water level
according the Dutch regulations for the specific dike region
In Tables 65 and 66 the partial factors according to CUR 166 and the corresponding design
values for the different soil properties are presented As someone can notice in these tables
only the strength parameters (phi and c) and the stiffness parameter (E) are exhibited For the
rest parameters the design value is identical to the characteristic value (see Tables 63(a) and
(b)) as a partial factor of 1 is used
Table 65 Partial factors and design values for the soil layers under the dike
1 SAND (very silty)
Soil property Symbol Partial factor γk Design value Unit
Cohesion c 11 0 [kPa]
Friction angle φ 12 2569 [ ˚]
Youngrsquos Modulus E 13 1538462 [kPa]
2 CLAY (medium)
Soil property Symbol Partial factor γk Design value Unit
Cohesion c 11 91 [kPa]
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 77
Friction angle φ 12 1472 [ ˚]
Youngrsquos Modulus E 13 153846 [kPa]
Table 66 Partial factors and design values for the soil materials in the dike
3 DIKE NEW (medium clay)
Soil property Symbol Partial factor γk Design value Unit
Cohesion c 11 364 [kPa]
Friction angle φ 12 2479 [ ˚]
Youngrsquos Modulus E 13 1250 [kPa]
4 DIKE OLD (stiff clay)
Soil property Symbol Partial factor γk Design value Unit
Cohesion c 11 1182 [kPa]
Friction angle φ 12 239 [ ˚]
Youngrsquos Modulus E 13 2250 [kPa]
642 Construction Stages
The construction and the gradual loading of the dike ware modelled as follows (see Figure 63)
1 K0-procedure for the generation of the initial stresses under horizontal groundwater
level
2 Dike self-weight under horizontal groundwater level
3 Rise groundwater level to the phreatic water elevation
4 Interpolation21 of the clay layer under the dike and activation of the sheet pile wall and
the anchor
5 Apply extreme water level conditions on the dike structure
6 (φ-c reduction for the determination of the safety factor)
21 Due to different permeability between the clay and the sand layer an approximation of the developed pore pressures is carried out through interpolation between the two layers
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 78
1
2
3
4
5
Figure 63 Construction and loading stages
For the soil modelling the Mohr-Coulomb model was used with a non-associative flow rule
(φneψ) For phases 1-4 the calculations were conducted under drained conditions whereas in
phases 5 and 6 the undrained behaviour of clay layers was considered
643 Determination of the structural elementsrsquo characteristics
According to the calculation scheme that was described above the design characteristics of the
structural elements were obtained from phase 6 In particular the target overall safety factor is
12 with which initially the length of the sheet pile wall was calculated
+370 NAP
+700 NAP
+1050 NAP
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 79
Firstly a length of 10 m and a profile of AZ 12 with a steel quality of 240 kNm2 yield stress
were used while for the anchor a cross-section of 628 mm2 and S355 were chosen for the first
calculation However even if the safety factor reached 12 it seemed to be quite unstable during
the calculation steps while the developed anchor stress was above the ultimate yield stress
What is more the deformations resulted in that case were more than 30 cm Therefore in order
to stabilize the safety factor and to decrease the displacements the length of the sheet pile was
increased 1 m (so 11 m in total) while also a better steel quality for the sheet pile was chosen
S355 Furthermore the anchor cross section was increased to 933 mm2 with a steel quality of
MW450 Last but not least the waling was designed based on the anchor force
In the following paragraphs the design procedure of each element is described providing also
the results of moments and forces that were retrieved from Plaxis
Sheet Pile wall
As it was partially mentioned in the previous paragraph the length of the sheet piles is
considered to be sufficient when the safety factor of the system22 reaches at least 12 For the
reasons discussed above the sheet pile length was decided to be 11 m and with a profile of AZ
12 S355
With this configuration the safety factor reached is 13 as it is depicted at the right graph of
Figure 64 Someone can notice that the safety factor from 117 in the situation without a sheet
pile structure increased to 13 with the structure
Figure 64 Safety factor deduced from the phi-c reduction calculation without the anchored sheet pile wall at the left (safety factor=117) and with the wall at the right (safety factor=13) The steps on the x-axis represent the calculation steps have been carried out in order to reach an equilibrium after incrementally reducing the strength parameters ie phi and c
The design moment is to be taken for a safety factor of 12 that is reached in step 34 where it
was found to be 1519 kNmm and the corresponding axial force 9133 kNm Therefore
considering Eq 52 for checking the exceedance of the ultimate yield stress and applying the
prescribed partial factor of 115 for both the moment and the axial force we get
22 Plaxis always gives the safety factor of the system which is under investigation
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 80
In Figure 65 the maximum bending moments with a safety factor 10 and 12 are respectively
showed From the values being shown in that figure someone can easily distinguish the
difference between the two moments It must be also mentioned in that point that the required
safety factor defined in Chapter 2 was estimated at 18 (and not 12 that is considered here)
which gives excessive design moments (almost 7 times larger) For that reason in this research
a lower safety factor was assumed avoiding thus the design of a quite conservative structure
Figure 65 Bending moments on the sheet pile for a safety factor safety factor=10 at the left and
safety factor=12 at the right
The explanation behind this discrepancy is that in case of the phi-c reduction procedure the
artificially decreasing strength of the soil and on the other hand the non-reduced strength and
stiffness23 parameters of the structural elements tends to lead in an increased arching
phenomenon that makes the structure attract the most of the load
In this research the reliability of the structure will be evaluated for a safety factor of 1 or else
during phase 5 and by considering the mean values of the soil parameters As someone can
notice from Figure 65 the magnitude of the maximum bending moment for a safety factor =10
is quite low which implies a possible minor contribution to the overall failure
At this point it is essential to mention that Mohr-Coulomb model is capable of simulating the
elastic behaviour and the state at failure Therefore in case that plasticity of the soil or the
structural elements plays an important role a different constitutive model is recommended to
be applied
Anchor
For a safety factor 12 the anchor force was found to be 10387 kNm Considering a mutual
anchor distance of 3 m the total axial force that is undertaken by the rod multiplied also with
the prescribed design partial factor is
23 ldquoStrengthrdquo and ldquostiffnessrdquo are usually confused terms Strength is a measure of the maximum load that can be placed on a material before it permanently deforms or breaks and it is represented by the yield stress σy in metals and by φ and c in soil Stiffness is a measure of the amount of deflection that a load causes in a material and it is represented by the Youngrsquos modulus E and the Poisson ratio ν both in soil and metals
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 81
Therefore choosing an anchor with a cross sectional area of A = 933 mm2 and a steel quality of
MW450 ( ) the design stress level is
In comparison with the sheet piles the anchor does develop large stresses as it was concluded
from the results in the deterministic calculations This is probably attributed to the high
stiffness of the anchor that apparently attracts the load and thus the anchor develops large axial
force
Waling
The type of the waling system is defined according to the anchor force as it was described in
section 522 Therefore by choosing a waling type of 2UPE 200 the developed bending
moment can be estimated as follows
where is the design bending moment of the walling system according to the chosen profile
In Table 67 the characteristics of each structural element that was described above are
displayed in detail
65 Overview In this chapter an introduction of the case study as it has been used in this research is made The
soil materials and their properties are summarized and their mean and design values are
defined The probability distributions of the soil parameters to be considered as stochastic were
proposed according to the literature study and the engineering judgement Moreover the mesh
generation and the modelling features are discussed
Last but not least due the observed overestimation of the retaining wallrsquos design moment and
axial force according to the current design regulation a new deterministic analysis was carried
out in order to redefine the structural properties and come up with a less conservative structure
This makes the reliability analysis of the system meaningful in a sense that an extremely safe
structure would give a failure probability of zero Besides the scope of a reliability analysis is
also the optimization of the construction that corresponds to a financially attractive and
simultaneously safe structure This can be made by conducting a risk analysis whose first step
would be the analysis that is carried out in this research Therefore the aim is to start with the
marginal required structural properties according to the safety standards In a risk analysis that
should conducted for several dimensions of the structure until the most optimized one is
identified However this is out of the scope of this thesis
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 82
In Table 67 the structural properties of the sheet pile wall the anchor and the waling are
summarized as they were found in the deterministic analysis
Table 67 Design parameters of structural elements
SHEET PILES
Property Symbol Value Unit
Profile - AZ 12 [-]
Steel quality - S355 [-]
Length L 11 [m]
Moment of inertia I 21430 [cm4m]
Elastic section modulus Wel 1245 [cm3m]
Mass w 943 [kgm]
Sheet pile thickness d 85 [mm]
Cross sectional area A 1201 [cm2m]
ANCHOR
Property Symbol Value Unit
Steel quality - MW450 [-]
Free length Lafree 104 [m]
angle φ 30 [deg]
Cross sectional area A 933 [mm2]
Mutual anchor distance s 3 [m]
WALING
Property Symbol Value Unit
Profile - 2UPE200 [-]
Steel quality - S355 [-]
Elastic section modulus Wel 191 [cm3m]
In the original case study the sheet pile wall is 162 m long and it has a profile of AZ 28 (S240)
while the anchor is an MW450 with a cross-section of 1134 mm2 As it is noticed both sheet
pilesrsquo and anchorrsquos properties have been reduced after the new deterministic analysis However
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 83
it should be mentioned that in the original case there was a vertical load on the top of the dike
that was displaced while the clay layers under the dike was merged into one in the new case
study These changes might lead to less heavy structure However the difference among the
design properties between the new and the original case study is mainly due to the required
safety factor that was assumed In the new case study the safety factor was set to 12 whereas in
the original cases study a safety factor of 18 was considered as it was calculated according to
the partial safety factors
In the next chapter the results after the reliability analysis of the just designed case study are
presented
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 84
MSc Thesis A Rippi 85
7 Reliability analysis results with stochastic soil properties
In this chapter the probability of failure of the different structural elements that were discussed
in Chapter 4 will be estimated by varying the soil properties of the several soil layers under and
inside the dike At the beginning a better insight into the system behavior is gained by
inspecting the deterministic Plaxis calculations for the mean values of the soil variables After
that a global SA is carried out in order to evaluate the most important soil parameters and
inspect their impact on the different structural elements in combination with each other and
individually Finally the results regarding the probability of failure of the anchor the sheet pile
wall and the soil are presented as they deduced by the reliability analysis (concerning the ULS of
the elements) Finally the systemrsquos reliability is evaluated The methodology that was followed
during the probabilistic analysis is described below and it is mainly dedicated to the DS
rationale that was used in order to obtain the influence factors and to evaluate the results of
Plaxis calculations
71 Method description For the calculation of the failure probability of the systemrsquos different components FORM and DS
were used for the retaining wall and the soil body respectively From these methods the
probability of failure and the important factors are going to be deduced for each component (ie
anchor sheet pile wall and soil body) However before starting with the reliability analysis it is
essential to carry out a sensitivity analysis in order to obtain a first impression of what are the
most influencing variables and thus determine the input random parameters for the reliability
analysis In the figure below the steps for the reliability analysis as they were followed in this
thesis are presented
Figure 71 Steps of the reliability analysis
Below the methodology of both the sensitivity and the reliability analysis are discussed and
explained in detail
Sensitivity analysis
In this thesis the FAST method was utilized for conducting the sensitivity analysis which is
further elaborated in Appendix A1 Here an explanation is given for the quantitative measure of
sensitivity which is represented by Sensitivity Indices The first-order sensitivity index of an
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 86
input variable pi is the measure of the main (direct) effect of pi on the output variance
(where i ne j) the second-order sensitivity indices measures the interaction effect of pi and pj on
the output variance Other higher-order indices are defined in the same manner The total
sensitivity index is the sum of all sensitivity indices involving factor pi (Homma amp Saltelli
1996) For example the total sensitivity index of factor 1 for a model with 3 input factors is
given as
The first-order (Si) and total ( ) indices can be interpreted as following (Pandya et al 2008)
high pi is an influent parameter
and both small pi is not an influent parameter (neither alone nor in interaction
with other parameters)
and nearly the same no interaction of pi with the other parameters
and very different high interaction of pi with the other parameters
Reliability analysis
As far as the FORM analysis is concerned section 422 gives sufficient information about the
methodology that is followed in order to obtain the failure probability and the influence factors
(ie α-values) The main concept and principles of DS are briefly explained in section 421
whereas here a more precise description of the implementation of DS and the how the α2-
values can be deduced is given
In particular DS is a method in which ldquodirectionsrdquo are sampled The concept of this method is
schematically depicted in Figure 72(a) For each sampled direction it is evaluated whereas
along this line the LSF Z equals to zero or not This procedure is repeated for a number of
directions and this method is applied in the standard normal space (u-space see also section
422) using equations to transform the standard normal variables u1 hellip un to their
corresponding ldquorealrdquo variables (in the x-space) X1 hellipXn and vice versa
(a) (b)
Figure 72 (a) Schematic view of the DS method in the u-space The numbers 0 1 2 3 4 show successive LSF evaluations along one direction (b) Schematic view of the line search algorithm β is the distance from the origin in the u-space along one direction The numbers 0 1 2 3 4 correspond to the ones shown in figure (a) (source Hydra-Ring 12 2015)
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 87
The search procedure for the location on the line where Z=0 is often referred to as the ldquoline
search methodrdquo and it is depicted in Figure 72 (a) and (b) In both figures the numbers 0-4 refer
to successive Z-function evaluations along one direction Figure 72(a) shows 3 evaluated
directions in the u-space whereas Figure 72(b) shows the (fictional) value of the Z-function
along one direction (where ||u|| is the distance from the origin along the line) The maximum
length of the direction line as well as the step size along the direction (for example the distance
between 0 and 1 along one line) can be steered in order to increase the efficiency of the method
and presumably decrease the computational time
In the sequence a method was developed in this thesis in order to estimate the α2-values of the
random variables out of a DS analysis as OT does not have an available method so far For that
purpose a transformation of the output samples to u-space was firstly carried out Then the
distance to the origin of all samples (in u-space) that are located on a direction where failure
(Z=0) was detected is calculated as follows
radicsum
i=1hellip (71)
where n is the number of random variables and uij is the ith ldquofailurerdquo sample of the jth random
variable is equivalent with the reliability index of this direction βi The α-values for each
random variable can then be calculated as (for more information about these values see
sections 422- Eq 417 and 423)
(72)
After that three different methods were created in order to evaluate the influence of the
random variables on the response of the LSF ie the α2-values (square of Eq 72)
Shortest distance ldquoβminrdquo
Average 10
Average all
In the Shortest distance method the sample with the smallest distance to the origin is
considered to be the design point and thus the corresponding influence factors are estimated
according to that point The smallest distance is also equivalent to the reliability index βmin This
method is well-known for estimating the design point out of MC techniques (Vrijling amp Van
Gelder 2002) However for a soil failure modeled in FEM this might not be always indicative of
the real influence that the soil properties can have on the system failure Therefore the last two
techniques were used as additional in order to get a better insight into the dominant soil
properties In Figure 73 a general scheme of the methodology that was followed for obtaining
the α2-values is depicted where also the different techniques are presented
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 88
Figure 73 DS methodology for calculating α2-values
The second method Average 10 takes into account the α2-values of the samples in the failure
domain whose distance to the origin lies within the 10 higher than βmin and averages them
Therefore it is firstly required the calculation of βmin as it was described for the Shortest distance
method Such a technique helps to identify the other directions that are close to the one which
gives the shortest distance to the origin and thus inspect if the important variables that were
indicated according to the first method are influential for other failure mechanisms as well
Last but not least the third method Average all averages the α2-values of all the samples
located in the failure domain This shows the overall contribution of the random variables to the
response of the LSF under investigation Concerning the soil failure LSF that is described in
section 73 such a method can be more suitable for distinguishing the soil properties that are
involved in the majority of the failure mechanisms rather than only in the one that the first
method indicates
72 Mean values calculations The results of the deterministic analysis that are presented in the previous chapter included a
degree of conservatism due to the load and material factors (partial safety factors) In order to
overcome this limitation structural reliability calculations were carried out and are discussed in
the next sections by considering relevant soil parameters as random However before starting
the probabilistic calculations by coupling Plaxis with OT that cannot be fully interpreted and
verified to a certain extent it is essential to first carry out some deterministic calculation Such
calculations can help to get a better insight into the behavior of the system and show
qualitatively what kind of response to wait later in the analyses These calculations were taken
place based on the mean values of the soil properties
Deformations
The analysis of this case study focuses on the ULS of the structure For the stresses and
deformation calculations the Mohr-Coulomb model is used that is generally accepted for the
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 89
simulation of the soil behavior until failure and for detecting the failure modes (a more detailed
explanation of this model is provided in Appendix B1) However this model is not so suitable
for analyzing the deformations of soft soils or the settlements around the wall
The deformation pattern that would more likely occur due to primarily the soil body failure can
be illustrated by conducting a ϕ-c reduction calculation In Figure 74 (a)-(c) the displacement
pattern after the ϕ-c reduction of the mean values of the friction angle and the cohesion
respectively As it can be seen a slip surface in the inner side of the dike has been created that
rotates towards the inland
(a) deformed mesh
(b) shadings
(c) arrows
Figure 74 Deformations after the ϕ-c reduction (mean values)
Stresses on the structural elements and the soil body
The anchor and the sheet pile wall are installed in phase 4 In phase 5 the anchor normal force
increases and especially for the mean values it amounts to Na = 7983kNm
As far as the sheet pile wall is concerned the bending moments and the normal forces that are
developed on the wall are indicated in Figure 75 as they were deduced from phase 5 Similarly
to the anchor the maximum stresses occur in phase 5 and are mainly located in the area of the
wall which is surrounded by the clay layer (brownish soil layer in Figure 74(a)) As someone
can observe the bending moments are not significant taking into account that the maximum one
observed reaches 2132 kNm
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 90
Figure 75 Bending moments and normal forces for the mean values calculation
For the interpretation of the stress field of the soil body the effective principle stresses and the
relative shear stress τrel (see Eq 517) were illustrated in Figure 76 (a) and (b) respectively In
Figure 76 (a) the crosses indicate the direction and the magnitude of the principle stresses σ1rsquo
and σ3rsquo and from which an active soil behavior (
) can be noticed on the river side of
the retaining wall whereas a passive behavior (
) is dominant in the inland side
(a) Effective principle stresses
(b) Relative
shear stress τrel
Figure 76 Effective principle stresses and relative shear stress in mean values calculation
In Figure 76 (b) the distribution of the relative shear strength shows the potential developing
plastic area in case of failure that expands from the river side of the wall and around it It can
also betoken the possible shape of formation of a slip surface that starts from the outer side of
the dike and goes around the wall
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 91
73 Sensitivity Analysis Results In section 312 different types of SA are presented and that can be applied through OT In this
research due to the time limit and the amount of parameters FAST method was used for
carrying out the SA as it allows the calculation of the first order sensitivity indices as well as the
total order indices and the computational time required is less than the other available methods
In this section the results from the conducted SA are presented separately for the anchor the
sheet pile and the soil performance Regarding the sheet pile wall and the anchor the sensitivity
of the total developed stresses towards the soil properties was evaluated according to Eq 53
and 55 respectively The influence of the soil parameters on the soil behaviour was evaluated
based on the calculation of the overall safety factor after a safety calculation (ϕ-c reduction
method) The total order indices are considered for distinguishing the most important
parameters while their difference with the first order indices Si indicates the level of interaction
effect amongst the parameters on the output variance
As far as the anchor and the sheet pile wall are concerned the SA was carried out in two steps
(1) SA with all the soil parameters (listed in Table 62) and (2) SA with the most important
parameters found in step 1 In Figure 77(a) and (b) the sensitivity indices from step (1) are
presented as they were found for the anchor and the sheet pile respectively Step 1 mainly
provides an overview of the contribution of the variablesrsquo variance to the total variance of the
limit state response and it helps to identify the most influencing parameters Step 2 was
basically carried out in order to build the response surface based on the SA output that was
later utilized in the FORM analysis as a complementary function in case of Plaxis errors At that
point it should be mentioned that the Dikenew material was not considered neither for the SA
nor for the reliability analysis of the structural elements as its influence was assumed to be
insignificant relative to the other soil layers However Dikenew was considered in case of the
soil body investigation
From Figure 77(a) it can be concluded that as far as the anchor is concerned the shear stiffness
G of the clay layer seems to be the most crucial soil property In particular the anchorrsquos stress
level seems also to be sensitive towards the Poisson ratio of the clay layer ν the friction angle φ
of the sand layer and the strength parameters φ c of the Dikeold material Moreover the
interface strength Rinter between the Dikeold material and the sheet pile wall seems to display
an additional notable influence as well The unit soil weight γ does not appear to affect the
anchor performance significantly
As far as the sheet pile wall is concerned the shear stiffness of the clay layer and the Dikeold
material seem to be dominant according to Figure 77(b) From these two soil layers it appears
that also the strength properties (φ c and Rinter) as well as the unit soil weight γ have a relative
impact on the stress level of the wall Furthermore the sand layer contributes via especially its
friction angle and shear stiffness
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 92
(a)
(b)
Figure 77 Total and First order sensitivity indices for the (a) anchor and the (b) sheet pile in step 1
As it was mentioned before step 2 was used for building the response surfaces for the anchor
and the sheet pile limit states The input random variables in this step are the most influencing
ones that were found in step 1 In Table 71 the soil properties that were considered as random
during step 2 are presented for both anchor and sheet pile It is essential to mention that the RS
was constructed by fitting a quadratic polynomial function to the sampling points of the SA
whose general expression is given in Eq 425 In Figure 78(a) and (b) an illustration of these
response surfaces is made However the figures depict the response of the anchor and the sheet
pile stress level over the fluctuation of only two variables amongst the 14 that are presented in
Table 71
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 93
Table 71 Soil properties considered for the RS construction of the anchor and the sheet pile respectively
Soil layer Anchor Sheet pile
Clay Sand
Dikeold
The linear regression of all the parameters was based on least squares and the R2 parameter
was found 0989 for the sheet pile and 0996 for the anchor Therefore the regression was quite
representative for the domain of the samples over which it was adjusted However there were
many residuals (ie difference between the observed and the predicted value) in the order of
103 and 104 which means that a possible expansion of this RS to an extended domain of the
input variables may lead to an underestimation of the limit state response
(a)
(b)
Figure 78 Polynomial response surface adjusted for the (a) anchor and (b) sheet pile stress level For the anchor the friction angle of the sand layer and the shear stiffness of the clay layer were accounted for this illustration whereas for the sheet pile wall the shear stiffness of the sand and the cohesion of the clay layer were used
The SA for the soil performance was also conducted in two steps However in that case there
was no need of a response surface construction since the LSF for the reliability analysis (Eq
522) takes the value of -1 in case of Plaxis error due to soil collapse Eq 522 though takes only
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 94
two different values -1 and 1 for soil collapse or not respectively which cannot be used for
carrying a SA For that reason the φ-c reduction method was used and the sensitivity of the
value of the safety factor (Msf) was investigated instead The second step of the SA was carried
out in order to divide the soil properties into two different groups of variables and thus
investigate the sensitivity of the soil in a more efficient way rather than including all the
parameters at one SA It should be mentioned that in the soil analysis soil properties of the
Dikenew material were added In Figure 79(a) and (b) the total and the first order indices are
depicted for step 1 and 2 respectively Step 1 includes all the stiffness and strength parameters
of all the soil layers whereas in step 2 the most important variables that were found in step 1
together with the unit soil weight γ are included Subsequently step 1 detects the most
influencing stiffness and strength properties and in step 2 the impact of γ is investigated
As it is observed in Figure 79(a) and (b) the safety factor appeared to be sensitive towards the
sandrsquos friction angle while also the strength properties (c and φ) and the shear stiffness (G) of
the clay layer play an important role Moreover the cohesion of the Dikeold material and the
unit soil weight of the clay sand and Dikenew layer seem to be influencing as well As it has
been mentioned above for this SA the LSF was different than the one considered in the
reliability analysis Therefore even if the SA at that point helps to evaluate in general the most
dominant soil properties the engineering judgement should also be included for the selection of
random variables that are going to be utilized in the reliability analysis
(a)
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 95
(b)
Figure 79 Sensitivity indices for the soil in step 1 and 2
It is also advisable that the difference between the total and the first order sensitivity indices be
investigated in order to have an insight into the most interactive variables The magnitude of the
interaction effect of a variable is crucial for the system behaviour because a variable might not
be important as a singularity but it is possible its combination with another variable to have a
considerable effect on the limit state under investigation Therefore this can be considered as
an additional criterion for choosing the set of the random variables to be used in the reliability
analysis In Figure 710(a) and (b) the difference between the two indices is presented for the
anchor and the sheet pile in (a) and the soil in (b) as it was deduced from the first step of the SA
Regarding the sheet pile wall the most interactive variables are the shear stiffness and the unit
weight of the clay layer the friction angle of the sand and the cohesion shear stiffness interface
strength and unit weight of the Dikeold material Concerning the anchorrsquos limit state the
cohesion of the Dikeold material seems to have the highest interaction with the rest of the
variables Last but not least regarding the soil performance the stiffness of all the soil layers
and the friction angle of the clay and the sand layer have a higher interaction effect amongst the
other parameters The importance of the aforementioned variables can be enhanced
considering both total indices and the difference that they display between their first and total
indices
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 96
(a)
(b)
Figure 710 Interaction effect amongst the parameters in (a) anchor and sheet pile and in (b) soil This effect is expressed as the difference between the total order indices T and the first order indices S as they were deduced from the first step of the SA for both anchorsheet pile and soil respectively
Considering the discussion above some preliminary conclusions can be drawn regarding the
influence of the several variables to the different components of the system
1 both anchor and sheet pile wallrsquos limit states are sensitive mostly towards the shear
stiffness of the soil
2 sheet pile wall stress level is also affected by the unit soil weight more than the anchor
does
3 soil body is mostly influenced by the strength properties and the unit soil weight
4 the interaction effect among the soil properties is more profound for the soil and the
sheet pile limit states rather than the anchorrsquos
It must be mentioned that this SA is a global one that gives an overview of the impact of the soil
parameters in terms of the system behaviour However near the design point a local sensitivity
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 97
can possibly give a better estimation of the variablesrsquo impact Near the design point the
structural elements as well as the soil body usually reaches plasticity that implies a non-linear
behaviour Such a behaviour is unlikely to be initiated during a SA that usually samples around
the mean values This can be counteracted by conducting more iterations so as to have a larger
amount of available samplings but this can make the SA to be quite time consuming and thus to
lose its benefit to give a fast preliminary estimation of the most influencing soil layers and
properties
Besides the purpose of a preliminary global SA is to reduce the number of the stochastic
parameters to a reasonable amount so as the initial reliability analysis to be more efficient and
computationally accessible The parameters that were eventually considered as stochastic in
each LSF are summarized in the table below and they were chosen according to both the SA and
the engineering judgement In the last column the stochastic variables used also for the
systemrsquos analysis are presented
Table 72 Stochastic variables for each LSF
Soil
Failure Anchor Failure
Sheet Pile failure
System failure
Sa
nd
γunsat radic
c
φ radic radic radic radic
G radic radic radic radic
v
radic
Rinter radic
Cla
y
γunsat radic
radic radic
c radic radic radic radic
φ radic radic radic radic
G radic radic radic radic
v
radic
Rinter radic radic radic
Dik
e_n
ew
γunsat radic
c radic
φ radic
G
v
Rinter
Dik
e_o
ld
γunsat
radic radic
c radic radic radic radic
φ radic radic radic radic
G
radic radic radic
v
radic
Rinter radic radic radic
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 98
74 Soil Shear Failure As far as the soil failure is concerned the limit state function that was implemented is given
below (see also sections 52 and 53) It was formulated as a single value depending on the
success of computation in Plaxis
(71)
The reliability method that was chosen in that case is DS due to the formulation of the limit state
function and the convergence issues of FORM There were used 12 input stochastic soil
parameters as it is indicated in Table 72 (see also input file of Soil Failure in Appendix D) The
variables were chosen based on the SA that was presented in the previous section For the soil
failure Poisson ratio was initially included in the analysis but it was noticed that values close to
04 during Plaxis calculations were leading to unrealistic behaviour of the dike More precisely
it has been observed that Poisson ratio equal to 04 can lead to swelling effects on the dike
which consequently lead to a raise of the dike crest instead of collapsing as it would be expected
in a soil failure analysis Therefore Poisson ratio was excluded from the reliability analysis of
the soil failure
The results of the reliability analysis according to DS are presented below In particular 300
iterations were carried out and 3 different failure directions were found Specifically in Table
73 the probability of failure the reliability index the number of iterations and the duration of
the analysis are firstly shown In the sequence the points from each failing direction with the
lowest reliability index are shown
Table 73 Reliability results of DS for the soil failure
DS Pf 1310-8 β 55 Number of LSF calls 1840 Number of iterations (or directions)
300
Elapsed time (hr) 34
Failure points (in real space) Parameter 1st direction 2nd direction 3rd direction Dikenew_phi [deg] 2277 2543 3021 Dikenew_c [kPa] 529 557 444 Dikenew_gammaunsat [kNm3]
1748 1848 1973
Dikenew_gammasat [kNm3]
1948 1984 1997
Sand_phi [deg] 2782 2039 3899 Sand_G [kPa] 808437 715473 715785 Sand_gammaunsat [kNm3]
2183 2122 2097
Sand_gammasat [kNm3] 2352 2145 2283 Clay_G [kPa] 56934 90063 80134 Clay_c [kPa] 704 1593 794
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 99
Clay_phi [deg] 1039 1629 1077 Clay_gammaunsat [kNm3]
1772 1413 1612
Clay_gammasat [kNm3] 1833 1433 1719 Dikeold_c [kPa] 1621 2978 915 Dikeold_phi [deg] 2474 3149 3281
According to the Shortest distance method the failure point from the 2nd direction turned out to
be the design point the importance factors of which are depicted in Figure 714 (first pie chart
at the left) However before concluding for the final design point an assessment of the above
failing directions was made
In particular the combination of the variables from each direction was later implemented in
Plaxis in order to illustrate the failure This is also necessary for verifying and assessing the
failure that Plaxis is giving and inspecting the error in the logfile in order to check if it is a ldquorealrdquo
(ie error=Soil collapses) or a ldquofakerdquo failure (ie error=Load advancement failed or Not enough
load steps) In Figure 711 the ldquorealrdquo failure is defined
Moreover a demonstration of the displacements is also essential in order to certify the failure
shape that normally has the dike crest settled down A different output of the displacements
formation (ie dike crest goes up) would probably imply a ldquofakerdquo failure However it should be
mentioned that a ldquofakerdquo failure in Mohr Coulomb model can be a ldquorealrdquo failure in another model
that can more realistically simulate the deformation patterns of the structure In general even
though Mohr Coulomb can be successfully applied for detecting failure it is not very suitable for
the determination of the displacements and especially for soft soils such that the dike under
investigation is constructed by It should be also mentioned at that point that in this case study
the system is subjected to unloading conditions24 (ie decrease of effective stresses) in soft soil
layers after phase 4 which cannot be modelled successfully by the Mohr-Coulomb model In that
case other constitutive models such as the Soft Soil Model and the Hardening Soil model are
presumably more reliable for simulating such situations For further information for this model
a reference is made to Plaxis (2015) Therefore someone should be always critical and verify
the failure points based on hisher engineering judgement
Figure 711 Prerequisites for real failure in Plaxis
24 From phase 4 to 5 there is an increase in water pressures and thus due to constant total stresses there is a decrease of the effective stresses This can be characterised as an unloading situation
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 100
In order to make it more understandable the different failure points were illustrated in Plaxis
and their deformed mesh and total displacements are presented in Figure 712 (a)-(c) for each
failure direction As one can observe in the first two directions the dike crest settles down
whereas in the 3rd one the dike body seems to swell and move to the right (the blue lines that
surrounds the dike soil layers indicate the initial dike geometry) However even if the 2nd
direction gave a normal pattern of displacements the error message was ldquoNot enough load
stepsrdquo The combination of the variables of the 3rd direction was then tested again with the
Hardening Soil model (HS) in order to illustrate the deformations with a more suitable model
and it was proved that the deformations seem to be realistic as it is depicted in Figure 713 With
Mohr Coulomb only the 1st direction seems to comply with the diagram in Figure 711 and that
would properly be considered as the real design point
(a) Deformed mesh
(b) Total displacements |u| in m (shadings)
(c) Total displacements direction (arrows)
1st direction 2nd direction 3rd direction
Figure 712 Illustration of the design point in soil body failure
(a) Deformed mesh (b) Total displacements |u| in m (shadings)
(c) Total displacements direction (arrows)
Figure 713 Displacements pattern of the 3rd failure direction according to HS
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 101
However taking into account that the other directions showed characteristics of failing behavior
they should also be accounted for the determination of the importance factors Excluding them
from the analysis would not be wise as the error message or the deformation schematization can
be a limitation of FEM or Mohr Coulomb model respectively In any case the points that lead to
failure shall be treated critically in order to decide if they should be included or not In this case
the points from the three directions seem to be realistic as both strength and stiffness
parameters are low enough in order to lead to a potential failure Therefore they were finally
considered for the analysis and the importance factors were averaged over all the points as it is
depicted in Figure 714 (last graph at the right)
Figure 714 Importance factors α2 for soil failure
Another option is also to average over the 10 range of close to the lowest reliability index that
was calculated with the shortest distance method (ie Average 10 method) in order to come
up with the α2 values However such methodology can lead to averaging over points of the same
direction that are not very different with each other Therefore the rest of the failure points are
automatically omitted from the consideration This can be also verified from Figure 714 where
the pie chart of the Shortest Distance and the Average 10 are quite similar whereas the Average
all chart introduces the importance of other soil variables as well
From a general perspective the soil properties of the clay layer seems to be determinant
whereas also sand and Dikeold material play an important role to the soil failure More precisely
according to the Average all method the unit weight and the friction angle of the clay layer
turned out to be the most influencing while also the cohesion of the Dikeold material and
friction angle of the sand layer contribute to the failure domain From a qualitative point of view
the weakening and consequently the settlement of the Dikeold material pushes the subsoil
creating an additional surcharge for the underlying clay and sand layer which act like the
foundation soil the incapability then of the clay layer to withstand the overlying load due to the
low unit weight in combination with the low friction angle of the clay as well as the low stiffness
and strength of the sand layer can lead to the creation of an inner slip surface and thus failure
Since the clay layer can be considered as a foundation soil the importance of the friction angle
and the unit weight can be also explained from the Terzaghirsquos bearing capacity theory (Terzaghi
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 102
1943) where these two factors are crucial for the determination of the ultimate bearing capacity
of the soil However these can be better verified by conducting large scale experiments in dikes
and inspecting the sensitivity of the dikersquos stability towards the soil properties
75 Anchor Failure The reliability of the support system is estimated based on the anchor and the waling probability
of failure However the load on the waling is proportional to the anchor force as it appeared in Eq
57 Therefore after the estimation of the anchor reliability by calculating the probability of
exceedance of a certain admissible anchor force the waling can be designed deterministically
Besides the failure probability of the waling must be lower than the anchorrsquos (given that no
uncertainties of the strength and the structural properties are considered) For the anchor failure
the LSF was formulated as follows
(72)
where [kN] is the anchor force is the yield stress and the cross
section area The reliability method to be utilized in this analysis is FORM and the selection of the
random variables to be used was made according to the sensitivity analysis results and the
engineering judgment and they are listed in Table 72 In particular the variables from the second
step of the sensitivity analysis were considered which will also help at a later stage to compare
FORM sensitivity method and FAST In Table 74 the results of the reliability analysis with FORM
are listed Additionally the design point together with the importance factors of each variable is
presented It should be mentioned at that point that the sensitivity analysis helped to define a
starting point closer to the design point and thus reduce the computational time required for the
analysis More precisely the analysis lasted approximately 16 hours whereas by starting from
mean values from which by default FORM starts the iterations it would take almost a day
Table 74 Reliability results for the anchor failure with FORM
FORM Pf 28410-6 β 45 Number of LSF calls 119 Maximum number of iterations
100
Elapsed time (hr) 16
Design point Parameter value α-value Clay_phi [deg] 1814 -0059 Clay_nu [-] 033 0086 Clay_c [kPa] 1487 -0089 Clay_G [kPa] 41739 0911 Clay_Rinter [-] 051 -0021 Sand_Rinter [-] 069 -0051 Sand_nu [-] 030 0007 Sand_phi [deg] 3259 -0174 Sand_G [kPa] 1130770 0048 Dikeold_phi [deg] 2842 -0020 Dikeold_nu [-] 036 -0045
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 103
Dikeold_c [kPa] 1873 -0045 Dikeold_G [kPa] 243183 -0326 Dikeold_Rinter [-] 053 -0072
Figure 715 Importance factors α2 in anchor failure from FORM analysis
The influence coefficients in Figure 715 indicate that this limit state is governed by the shear
stiffness of the clay layer G Furthermore the shear stiffness of the Dikeold material seems to
contribute significantly to the anchor yield stress exceedance It can be concluded that the
problem is still in the elastic domain as far as the soil behaviour is concerned In case of
predominately plastic behaviour the strength properties of the soil (φ and c) become more
important In Figure 716 a demonstration of the design point as it was acquired by the FORM
analysis is presented The figures indicate that presumably due to the clay layer withdrawal in the
passive side as it is depicted in Figure 716(b) with the red shadings sheet pile moves to the right
and subsequently anchor is tensioned enough to reach its yield stress25
(a) Deformed mesh (b) Total displacements |u| in m (shadings)
(c) Total displacements direction (arrows)
Figure 716 Design point illustration from the anchor failure
25 A better representation of the soil-anchor interaction could be possibly obtained if the anchor was modelled as a node-to-node element and the anchorage body as an embedded element
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 104
The calculations were repeated with the 6 most influencing parameters both with FORM and DS
in order to investigate the effect of reducing the number of random variables and in order to
validate the results of FORM analysis In Table 75 the reliability results of respectively FORM and
DS are presented while in Figure 717 the importance factors of both methods are juxtaposed The
design point and the importance factors of DS were estimated based on the Average 10 method
(see section 71)
Table 75 Reliability results of FORM and DS with reduced variables in anchor failure
FORM Pf 8510-6 DS Pf 1510-5 β 43 β 42 Number of LSF calls
28 Number of LSF calls
776
Maximum number of iterations
100 Number of iterations
100
Elapsed time (hr) 034 (starting point close to the design point)
Elapsed time (hr)
115
Design point Design point Parameter value α-value value α-value Clay_nu [-] 033 0119 033 0242 Clay_c [kPa] 1404 -0026 1535 -0212 Clay_G [kPa] 43236 0928 44548 0807 Sand_phi [deg] 3163 -0110 2954 0203 Dikeold_G [kPa] 239465 -0330 251723 -0366 Dikeold_Rinter [-] 053 -0064 060 -0263
Figure 717 Importance factors α2 with FORM and DS analysis respectively
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 105
As it can be noticed comparing FORM results from Table 74 and 75 the reliability index changes
from β=45 with 14 parameters to β=43 with the 6 most influencing parameters This verifies the
importance of the most influencing parameters that it was found in the analysis with the 14
parameters The small discrepancy though between the two reliability indexes might be
attributed to the interaction effect of the soil variables on the output performance
The estimation of failure probability with DS is almost the same with FORM As far as the
importance factors are concerned DS also came up with the conclusion that the shear stiffness of
the clay and the Dikeold material are the most crucial soil properties for the anchor stress level
However according to DS the stiffness of clay ν and the interface strength Rinter between the
Dikeold material and the sheet pile wall seem to also contribute to the limit state of the anchor
The reason why the strength of the clay layer under the dike and the Dikeold material (the part of
the dike that comes in contact with the anchor and the sheet pile wall) are the most important
variables can be explained by the principle of soil arching According to this principle the soil
columns on both sides of the rigid sheet pile wall are more compressive than the soil columns on
the top of the wall (ie overlying dike body) because of the higher stiffness of the wall when
compared with soils As such soil columns on both sides tend to settle more than the soils on top
of the rigid wall and this differential settlement causes a downward shear force acting along the
sides of soil columns on top of the wall As such the vertical load on the wall becomes larger than
the sole weight of soil columns on its top and the anchor that actually receives the most of this
vertical load reaches eventually its yielding stress
Last but not least it should be mentioned that the response surface which was constructed based
on the SA worked efficiently in most of the Plaxis errors cases in a sense that it gave reasonable
results of the anchor stress level
76 Sheet pile wall failure For the reliability analysis of the sheet pile wall the strength property ie the yield stress was
considered constant and equal to 355 kPa Therefore the LSF that was applied for this case was
formulated as follows
[
] (73)
where [kNm] and [kN] are the maximum bending moment and the axial force
that are exerted on the wall respectively is the elastic section modulus ( =1245 cm3m)
and is the cross sectional area ( =1201 cm2m for a AZ12 profile)
In sections 643 and 72 the maximum bending moments of the sheet pile wall were calculated
with the design values of the soil properties and the mean values respectively It has been
observed that the bending moments were not so significant and determinant in comparison with
the anchor force and the safety factor of the system that were also considered (in section 643) in
order to determine the structural properties
In a later stage a SA was carried out for the sheet pile wall that has been further discussed in
section 73 Figure 718 shows the stress level of the wall that was developed after 3150 iterations
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 106
by considering as random variables the ones that are defined in Table 72 As someone can
observe the stress level is generally quite low and even an order lower than the yield stress (ie
355105 Pa)
That small magnitude of bending moments can be probably attributed to the counteraction of the
active earth pressures from the passive ones that due to the homogeneity of the soil in both sides
of the wall are quite similar
Figure 718 Stress level [in Pa] of the sheet pile wall after 3150 iterations of the SA
The reliability analysis has been carried out with FORM method and considering as random
variables those listed in Table 72 However due to the low level of stresses that are developed
and the simultaneously high yield stress the method converged after almost 35 days of
calculations As it can be noticed in Table 76 the analysis came up with an extremely low
probability of failure It should be mentioned at that point that the convergence of the analysis
was successful thanks to the RS that was fitted This means that a Plaxis error of soil collapse
preceded and it activated the RS which implies that the soil failure is more likely to occur before
the sheet pile failure
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 107
Table 76 Reliability results for the sheet pile wall failure with FORM
FORM Pf 7210-38 β 13 Number of LSF calls 2619 Maximum number of iterations
100
Elapsed time (hr) 84 (35 days)
Design point Parameter value α-value Clay_phi [deg] 1164 0261 Clay_c [kPa] 586 0336 Clay_G [kPa] 14275 0663 Clay_Rinter [-] 052 -0016 Clay_gammaunsat [kNm3]
1848 0000
Clay_gammasat [kNm3] 1948 0000 Sand_phi [deg] 3020 0000 Sand_G [kPa] 1193650 0000 Dikeold_phi [deg] 1360 0406 Dikeold_c [kPa] 592 0438 Dikeold_G [kPa] 168798 0000 Dikeold_Rinter [-] 053 -0024 Dikeold_gammaunsat [kNm3]
2275 -0150
Dikeold_gammasat [kNm3]
2375 0000
This can be also verified from the importance factors that are depicted in Figure 719 where apart
from the shear stiffness of the clay layer which seems to be the predominant soil property for the
sheet pile LSF the strength parameters (φ and c) of the clay and the Dikeold materials also play
an important role This implies that the soil behaviour has already passed into its plastic domain
where the governing properties are the strength parameters before the sheet pile exceeds its
yield stress
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 108
Figure 719 Importance factors α2 of the sheet pile stress level in the plastic domain of soil
Nevertheless in order to investigate the potential most influencing soil properties for the sheet
pile LSF in the elastic domain as well the available output data of the FORM analysis have been
assessed The highest observed value of the sheet pile stress before its failure was around
235000 kPa After tracing the results it seems that when FORM tries to reduce further the
stiffness and strength properties of the soil layers soil collapses before the sheet pile failure The
vectors of the random variable that led to a stress level on the sheet pile wall in the range of
230000-235000 kPa were chosen in order to evaluate the importance factors The method that
was followed in that case is the Shortest Distance and the results are shown in Figure 720
Figure 720 Importance factors α2 of the sheet pile stress level in the elastic domain of soil
As it can be noticed from Figure 720 the results in the elastic domain regarding the influence of
the stochastic parameters are comparable with anchorrsquos analysis The stiffness parameter of the
clay layer (69) is the most dominant variable for the system behaviour whereas the stiffness of
the sand (15) and the Dikeold (4) layers seems also to have a significant influence This is not
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 109
surprising as the nature of the loads that generate high stresses on the anchor as well as the sheet
pile is the same namely the horizontal earth and water forces
Concluding in this case study the sheet pile wall is not expected to show a decisive behaviour to
the systemrsquos reliability according to the reliability analysis and the deterministic analysis (see
section 643) It should be stressed the fact that the fitted RS in a lot of calculations seemed not to
perform suitably for the sheet pile and this can be attributed to the fact that the it was created
based on the SA which was far from the plastic domain In that case a better fit of a RS is
recommended However such an action is still under investigation that for future considerations
in the design concept will be contemplated as a valuable technique for complex structures
modelled in FEM whose reliability analysis with conventional methods is time consuming and the
coupling may lead to numerical problems In the next section the analysis of the systemrsquos
reliability is presented
77 System Failure The estimation of the systemrsquos reliability was based on the systemrsquos fault tree of Figure 53
according to which the general LSF was formulated as the minimum of the three LSF that were
presented above for the soil the sheet pile and the anchor and it is given below
[ ]
[(
)]
(74)
According to this LSF the failure that occurs first is counted as a system failure The analysis was
carried out with DS reliability method and the variables that were taken as stochastic are listed in
Table 72 As it was discussed in section 51 the systemrsquos probability of failure differs from the
probability of flooding for which a combination between the systems reliability and the water
level uncertainty shall be made
In Table 77 the reliability analysis results are presented as they were deduced from DS Similarly
to the soil failure in Figure 721 the importance factors according to the Shortest Distance the 10
Average and the Average all methods are illustrated
Table 77 Reliability analysis results from the system failure
DS Pf 2410-3 β 30 Number of LSF calls 1807 Number of iterations (or directions)
300
Elapsed time (hr) 39
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 110
Figure 721 Importance factors α2 of the system failure
The analysis took almost one and a half days and it came with indicative valuable though results
about the system behaviour According to the averaged importance factors in Figure 721
(deduced by the Average all method) we can notice features from both soil and anchor failure
For example the unit weight (14) the friction angle (7) and the cohesion (8) of the clay
layer as well as the friction angle (8) and the stiffness (7) of the sand layer that played an
essential role in the soil failure analysis Moreover the stiffness of the clay (11) and the Dikeold
(9) layers that appeared to be important for the anchor and sheet pile LSF They seem to be also
essential in the system analysis
However the other two methods (Shortest Distance and Average 10) came up with the cohesion
and the interface strength of the Dikeold material to be the most influencing soil properties for
the system For that reason an investigation was conducted regarding the output results of the
system analysis in order to obtain an insight into the different reasons of the system failure More
precisely the failure points were divided into those that come from Plaxis Error and those that
come from the yield stress exceedance of the structural elements It was then observed that many
failure points originate from Plaxis errors that are related to ldquoNot enough load stepsrdquo (Error codes
102 and 112) rather than ldquoSoil seems to collapserdquo (Error code 101)
Therefore in order to estimate the actual important soil properties the vectors of the random
variables that led to errors of ldquoSoil seems to collapserdquo and to anchorsheet pile (SP) failure were
studied separately The importance factors are depicted in Figure 722 which actually verifies the
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 111
importance of the variables that were also deduced from the Averaged all method in Figure 721
(which are basically the unit weight and the stiffness of the clay layer and the friction angle of the
sand layer) It should be mentioned that the failure of the structural elements contributes 65
whereas the soil failure contributes 35 on the total probability of failure The latter would be
possibly lower than 2410-3 that was initially found as several failure points are not attributed to
real failure but to Plaxis numerical errors
Figure 722 Influence factors α2 for the structural elements (left) and the soil LSF (right) respectively
The system analysis can lead to valuable conclusions for the system behaviour as a whole and
individually for the different elements given though that an inspection of the results is followed
Moreover such an analysis is taking into account correlations among the three sub-failure
mechanisms implicitly For example the different LSFs may be affected by the same soil variable
such as the shear stiffness of the clay layer for the anchor and the sheet pile LSF In that case the
most dominant failure mechanism will occur first and the rest are excluded26 However during
the system analysis someone is not aware of what failed or not and therefore heshe is not
capable of distinguishing the weaknesses of the system that need further improvement Besides
with this technique is not also easy to inspect the failure mechanisms that occur in order to get
some knowledge about the systemrsquos behaviour Therefore the reliability analysis of the
components is recommended to be carried out separately if time is available
78 Comparison between Global Sensitivity Analysis and Local Sensitivity
Analysis (FORM)
The aim of this section is a comparison between the importance factors that were deduced in
section 75 according to a FORM analysis and the results of the global SA that were presented in
section 73 for the anchor failure
The output of the global SA is actually the total indices that show the influence of the random
variables The importance factors of FORM and the total indices of global SA cannot be directly
26 This method is not capable though of accounting for the simultaneous occurrence of different failure mechanisms as the system was considered a serial system
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 112
compared as they are different values More precisely the total index of a random variable is the
ratio between the total variance and the portion of the total variance that stems from the
uncertainty of the specific variable and its interactions with the rest (see Eq A10 in Appendix A)
On the other hand importance factors of FORM express the ratio between the normalized value of
the variable in the design point and the reliability index (see Eq 72) Therefore the comparison
between these values is carried out qualitatively in terms of the variables that are distinguished
as the most influencing
In particular according to global SA the shear stiffness G of the clay layer and the cohesion c of
the Dikeold material seem to be the most determinant soil properties Moreover the strength
reduction parameters Rinter of the Dikeold and the clay layers and the friction angle φ of the
Dikeold and the sand layer appear to have a relatively notable contribution to the anchors limit
state
On the other hand FORM analysis has concluded to the same results as the global SA with the
difference that Dikeold material affects the anchorrsquos stress level more with its shear stiffness
rather than its cohesion In Table 78 the results of both FORM and global SA are presented
separately for each soil layer and property (in blue for the global SA and in black for the FORM
analysis)
Table 78 Important soil properties according to global SA and FORM analysis for the anchor failure (radicradic=very important radic=less important but still considerable)
FORM global SA Sand Clay Dikeold
ν
radicradic
ϕ radicradic
radic
c radic radic radicradic
G radicradicradicradic radicradic
Rinter
radic radicradic
The validation of the most important parameters was carried out with the repetition of the
analysis with DS as it was showed in section 75 The small difference of the probability of failure
as well as the same tendency of the important factors of both DS and FORM verified the reliability
of FORM analysis Therefore the similarity of the results deduced according to global SA with the
ones given by FORM indicates the fact that in this case study the global SA should be considered
as a trustable way of SA for obtaining a first overview of the dominant variables that affect the
limit state under investigation Furthermore it is capable of giving a general insight into the
outputrsquos (in this case anchor stress level) behavior and the soil layers that can potentially stand as
predominant during the reliability analysis For example in order to obtain a first sense of the
reliability and the magnitude of the expected probability of failure someone can inspect the
output samples of the LSF deduced from the global SA In Figure 723 the samples of the anchor
stress level are presented as they were deduced after 3150 iterations during the global SA
(computation time =75 hr = 3 days) The fact that after 3150 iterations and variables
combinations the anchor did not reach its ultimate stress level (ie 450 kPa) implies that the
probability of failure might be small This can also help for qualitative decision making concerning
a systemrsquos behavior without conducting a reliability analysis Moreover global SA proved the high
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 113
interaction (large difference between first and total order indices) among the parameters that can
affect the reliability analysis and can help to decide the amount of the random variables
Figure 723 Anchor stress level during global SA (step 1)
However as someone can observe from Table 78 there are also some discrepancies concerning
the dominance of some variables This can be attributed to the fact that the influence of all
parameters in specific points such as the design point cannot be identified by the global SA
Moreover it is possible that the samples used in the SA could not reach the field where the
system has a plastic behavior and where presumably other variables may also affect the limit
state
As general conclusions for global SA some possibilities and limitations are listed below
Possibilities
bull Indicates the level of interaction effect on the output variance
bull Indicate the amount of the random variables that should be considered in the reliability
analysis
bull Gives a general insight into the output behavior
bull Valuable for future use in the construction of response surfaces
Limitations
bull Cannot rely on global SA for the evaluation of the most important parameters near the
design point
bull Difficult to reach plastic zone (large amount of samples are probably needed)
79 Discussion
In this chapter the reliability analysis results were presented for the soil the anchor the sheet
pile and eventually the system failure taking soil properties of the several soil layers as stochastic
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 114
Initially a SA was carried out for each system component in order to identify the major soil
properties and reduce the number of the stochastic parameters The reliability analysis was
performed either with FORM or DS and in both cases the design point and the importance factors
were estimated Eventually the list of subquestions that is proposed in section 11 was answered
The way they were treated and the corresponding conclusions are elaborated below
Which reliability methods are computationally tractable in combination with FEM
Before starting analyzing the reliability of the system it was essential to evaluate and
assess the performance of the available reliability methods in order to be aware of their
performance in terms of computational time and flexibility in different kind of LSFs After
a literature review and the implementation of some of them in simple LSF problems a
general overview has been obtained for the most applied reliability methods concerning
the previous knowledge required for the system behavior in order to provide the right
input parameters its accuracy and the computational effort Eventually FORM and DS
were chosen to be employed for this case study considering the no preliminary knowledge
of the LSF is required their reliable accuracy and the limited computational effort that
they have in comparison with other methods
The anchorrsquos and sheet pilersquos reliability has been evaluated with FORM method whereas
soil and systemrsquos reliability have been assessed by DS Carrying out a FORM analysis and
steering the algorithmrsquos parameters in order to converge proved to be more difficult than
it was expected However an efficient performance was finally achieved and the failure
probability of the anchor failure was successfully estimated together with the importance
factors (or else the α values) The analysis was repeated with the most important variables
and they were validated with DS The FORM analysis took approximately one and half
hour to converge (with 14 random variables) after providing a starting point that it was
estimated to be close to the design point whereas for the validation it took almost half an
hour On the other hand in case of the sheet pile wall FORM analysis took almost 4 days
The computational time also depends on the LSF to be approximated the number of
variable and the complexity of the problem that each time is analyzed As far as the soil
analysis is concerned DS took approximately one and a half day to converge for 300
iterations and 14 random variables
Which limit states are relevant for the envisaged application of retaining walls in dikes and
how can they be formulated using FEM analysis outcomes
The relevant LSFs of a dike with an anchored sheet pile wall were investigated For each
element ie the sheet pile the anchor and the soil body a separate LSF has been set up
The focus of this research was on the ULS that in case of the structural elements it was
represented by the ultimate yield stress of the anchor and the sheet pile respectively as
they were expressed in Eq 72 and 73 The determination of the soil failure has been also
elaborated and the available alternatives that can be considered in that case were
elaborated After a short evaluation the Plaxis definition of soil failure was selected that it
was formulated as it is shown in Eq 71 Finally the systemrsquos LSF was formed as a
combination of the aforementioned limit states More precisely the system was
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 115
considered as a serial and thus system failure occurs when one of the elements fails first
(see Eq 74)
How robust (convergence) are the tractable methods
The robustness of the different reliability methods can be controlled by the convergence
criteria that are defined by the user and depend on the required accuracy In case of FORM
the convergence criteria consist of a set of errors that has mostly to do with the
approximation of the reliability index and the limit state threshold whereas in DS the
convergence is controlled through the coefficient of variation (CoV)
In order to reassure the robustness of FORM the method has been validated with DS The
validation showed that the FORM method complies with the result of DS with a small
difference though in the design point and the importance factors FORM analysis seemed
to converge sufficiently and faster than DS However the choice of the optimization
algorithm plays a key role to the methodrsquos efficiency In this research the convergence
errors were manipulated for increasing the efficiency of the method More precisely it has
been observed that by relaxing the relative error and the constraint error (the explanation
of these errors and an evaluation of the different optimization algorithms are available in
Appendix A2) FORM convergence can be accelerated
The reliability analysis of the soil body and the system was performed with DS Similarly
to FORM analysis in order to improve the performance of the searching algorithm some
of the methodrsquos parameters should be stipulated for improving its efficiency and
reliability Therefore the maximum step size of the algorithms was tripled and it was
proved that for the same time period double iterations were feasible to be carried out
This does not mean that the larger the step size the better it is The choice of the step size
must be made wisely depending on the reliability index that it is expected and
respectively it is allowed to change for achieving a better performance In this case study
it was proved that after 300 iterations the CoV of DS tends to be high and it can therefore
still be improved by carrying out more iterations In Figure 724 an example of a DS
convergence graph concerning the system analysis is shown The middle red line shows
the probability estimation whereas the green lines show the boundaries At that example
the CoV was 040
DS seems to perform successfully both in soil failure with one LSF and in system failure
where 3 different LSF were included In such cases performing an approximation method
such as FORM would not be wise as the output from each iteration does not provide
information for going to the next that a FORM analysis normally needs In contrast it just
gives an indication of failure or not This means that only a random sampling method can
handle it and converge successfully as it was proved
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 116
Figure 724 Directional Sampling convergence graph (at confidence level 095) of the system LSF
What is the contribution of different uncertainties in the failure mechanisms of the system
According to the global SA and the reliability analysis of FORM and DS the friction angle
the cohesion and the shear stiffness properties appeared to have an influence on the
different LSFs This is not surprising as they all have large CoV of 10 20 and 25
respectively Thus their uncertainty generally showed an impact on the systemrsquos
behaviour It is also essential to mention that the global SA has shown that the uncertainty
of the interface strength Rinter (CoV=20) has also a notable contribution to the structural
elements LSFs However the FORM analysis which is more representative close to the
design point came up with a less important impact of that variable
The reliability analysis showed that as far as the structural elements (anchor and sheet
pile) are concerned the shear stiffness of the clay and the Dikeold material is determinant
and especially for the sheet pile the strength parameters of the soft soils seem also to play
a crucial role In contrast with the anchor where the soil body seems still to behave in an
elastic manner during the sheet pile failure soil appears to reach plasticity in more
locations where the strength properties (ϕ and c) are also influential The most dominant
failure mode appeared to be the anchor failure whereas the soil failure is always preceded
the sheet pile failure The dominance of the anchor failure can be explained from the fact
that the most influent parameter in this LSF is the stiffness that is also the variable with
the highest uncertainty (CoV=25) Therefore the plastic limit of the anchor is more
likely to be reached
Regarding the soil body reliability the analysis detected 3 different failure directions in
which the clay and the sand layer below the dike seem to be the most influent More
specifically the friction angle φ the cohesion c the unit weight γ of the clay layer the
friction angle of the sand layer and the cohesion of the Dikeold material turned out to play
an important role The decrease of the clay layerrsquos unit weight lead to the incapability of
the layer to withstand the overlying load of the dike body and at the same time the
weakening of the sand layer contribute to the creation of an inner sliding surface Even if
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 117
the uncertainty of the unit weight is small (CoV=5) it is still important for the stability
of the system as the maximum deformations appear in that layer (see Figure 712) Last
but not least the illustration of the several failure points in Plaxis allows someone to
inspect the failure mode and realize if it is a real failure of the structure or if it is a
weakness of the modelling
Finally during the system analysis multiple failure directions were detected whereas the
pie charts of the importance factors include soil properties that affect both soil and
structural elements as it was expected However due to Plaxis numerical errors several
failure points were found to be not realistic and they should be excluded in order to come
up with a more reliable failure probability which will be presumably lower The most
suitable method for estimating the importance factors though seems to be the Average all
method which is not distorted from the Plaxis numerical errors However it is
recommended that a further research be done on how realistic are the failure points
coming from Plaxis warnings and to verify if they comply with Figure 711 failure
definition
In Table 79 a synopsis of the most important soil properties for the different LSFs is
presented It is also apparent the common contribution that some of them have in
different LSFs In the last two columns of the table the results of the influencing soil
variables for the system LSF are shown with the Average 10 and Average all method
respectively For the reasons that were explained in the previous paragraph the Average
all method is considered to be the most suitable in that case and as it can be noticed from
the table it gives the most representative picture of the governing soil variables of each
LSF
Table 79 Important soil properties for the different LSFs (radicradic=very important radic=less important but still considerable the most important parameters for each LSF is painted in orange)
Soil material
Soil property
Anchor Sheet pile
Soil body
System Average
10
System Average
all
Clay
G radicradic radicradic
radic radicradic
φ
radic radicradic
c
radic radic radic
Rinter γ
radicradic
radicradic
Sand
G
radic radic
φ
radicradic radicradic
c
Rinter γ
Dikeold
G radicradic
radic
φ
radic
c
radicradic radic radicradic radic
Rinter radic
radicradic radic
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 118
γ
radic radic
Rinter was taken from the validation of FORM with DS Figure 717 (right graph)
taken from the sensitivity of the sheet pile in the soil elastic domain Figure 720
Can response surface techniques help to increase the efficiency and robustness of the
reliability model
In this research RS techniques were used only as a source for providing a reasonable
value to the FORM analysis in order to continue running in case of Plaxis errors These
response surfaces were constructed based on the SA results for both the anchor and the
sheet pile wall Apparently the domain of the SA is limited and most of the times it is
incapable of approaching the design point However RS worked efficiently in many cases
where Plaxis failed to give an output result such as in case of the sheet pile LSF Moreover
the RS could be used individually for the estimation of the failure probability (instead of
Plaxis calculations) but a better fit is advised to be firstly achieved that would be
representative in a wider plastic domain of the structural elements Therefore the RS
technique can be used to enhance the efficiency of the reliability analysis of a complex
system in the sense that it can be considered as solution for FEM errors What is more RS
techniques are a quite promising method for the reliability analysis of a system as it can
also be used alone without coupling with FEM and eliminate the computational effort
However this requires more investigation of such a technique
How can the current design approach be improved
In section 22 the current design concept of the case study that is under investigation in
this research is described At that section the design values of the moments and the forces
on the structural elements are presented as well as the required total safety factor FEM
=18 is given as it was initially calculated by Breedeveld (2011) It has been observed that
a significant overestimation is made in terms of the design values that might later lead to
heavy and cost inefficient structures For that purpose alterations was made on the
original case study in terms of the structural elements and some soil properties and a new
deterministic design was made in order to come up with a less heavy structure This
would also render it possible to investigate the reliability of the system rather than taking
zero probabilities or non-convergence Moreover it would finally help to see if there are
any potentials for optimizing the design by applying a full probabilistic analysis
According to the soil reliability analysis the probability of failure was found 1310-8 and
the reliability index β=55 As far as the anchor failure is concerned the probability of
failure found to be 28410-6 with a reliability index β=45 according to the FORM analysis
The sheet pile failure seem not be significant as it will more likely occur after the soil
failure Eventually the system analysis was performed the probability of failure was
found to be 2410-3 and the reliability index β=30 which is high enough to consider the
system safe
The systemrsquos reliability index that came out of this analysis cannot be directly compared
with the required one (ie β =40) because they stem from different probabilities The
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 119
former refers to the probability of failure due to the global instability given a design water
level whereas the latter refers to the probability of flooding that incorporates also the
water level uncertainty It must be mentioned that in reality there is no sheet pile wall
inside the dike section in the dike ring 43 but this section is under investigation for
reinforcement purposes as it does not comply with the safety standards after a recent
inspection that has been carried out Moreover the original case study taken from
Breedeveld (2011) was subjected to multiple alterations for the needs of the current
research Therefore a comparison of the reliability index with the required one from the
regulations would not be meaningful to make However what it can be compared in that
case is the order of magnitude of the reliability index After the analysis a β of 30 was
estimated for the probability of failure whereas for the overall probability of flooding a β
of 40 is required whose magnitude is the comparable with the former This can firstly be
considered a sort of the probability ldquovalidationrdquo in the sense that the analysis gave
realistic results and secondly it can imply that the structure that has been investigated in
this research is likely to be safe towards macro-instability as its coupling with the water
level uncertainty would certainly give a reliability index above 40 Furthermore a
reliability index of 30 indicates that there is room for optimizing the design concept of
such a system if someone considers also the fact that the soil strength of the subsoil was
further reduced for the needs of the current study
Due to the limitations of the modeling and the alterations of the original case study these
reliability indices should not be taken into account as an indicative number for the safety
of this dike section but rather as an order of magnitude in order to realize what the
potentials for improving the design procedure are This research can also contribute to
the way of designing not only the system but also each structural element as the reliability
of each can be calculated in advance According to the methodology developed in this
thesis a comparison can later on be made for a real case situation between the results of a
full reliability analysis and the current design procedure In that sense a tangible
improvement can be suggested for the design concepts and even to introduce a new way
for the future design of dikes with retaining walls
It should be stressed that the conclusions are based on the results of this case study with its
geometry and set of material parameters and the variation coefficients Changes in the geometry
the material parameters or the statistical properties could lead to different results The
generalization of these conclusions should thus be treated carefully
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 120
MSc Thesis A Rippi 121
8 Conclusions and Recommendations
81 Conclusions The advances in structural reliability analysis have made a large variety of methods available The
applicability and efficiency of these methods depends on the problem that is analyzed and on the
models that are used One the other hand the necessity for optimizing the design procedure of
several structures is becoming more and more intense A solution to this demand can be the
reliability analysis of a system for which rare information about its application on real-world
problems is available This work can contribute to making probabilistic analysis more accessible
for dikes reinforced with a sheet pile wall and at the same time to analyzing such a complex
system in terms of the factors that mainly influence the system behavior In particular the main
research question of this research was
How can the probability of failure due to global instability of a dike with a sheet pile wall modeled
by a Finite Element Model be analyzed
For that purpose in this research a reliability analysis was conducted by making use of a soil-
structure model that was simulated in FEM The output of the FEM calculations was utilized as an
input in the probabilistic model that was used in this thesis in order to assess the reliability of a
dike with an anchored sheet pile wall In that case the load was basically the earth and water
pressure whose impact was investigated on the anchor (LSF 1) the sheet pile (LSF 2) and the soil
body (LSF 3) respectively as well as on the system as a whole
It was demonstrated that the calculation effort of this kind of analysis does not necessarily have to
be extremely high Especially when previous knowledge about the system behavior and the
reliability models used is available This can be achieved by carrying out a sensitivity analysis In
this thesis a global SA was conducted that was proved capable of giving a representative
estimation of the most influencing variables and give a general overview of the systemrsquos
performance However close to the design point local SA such as FORM becomes more suitable
As far as the reliability of the structural elements is concerned their limit state was evaluated in
terms of the exceedance of their yielding stress The reliability assessment was carried out with
FORM from which the probability of failure and the influence factors were concluded More
precisely the anchor failure mode seemed to be the most determinant one with the shear
stiffness of the soft soils to be the most important soil properties The sheet pile failure is quite
rare and it is unlikely to occur before the soil failure since the soil body has already entered its
plastic domain
The soil and the system failure were challenging to assess as there are multiple criteria that
someone can choose In this thesis regarding the soil failure the Plaxis definition of soil collapse
was used and it was evaluated with DS In that case the unit weight and the strength parameters
of the soft soil layers seemed to be influencing The detected failure directions were investigated
Conclusions amp Recommendations
MSc Thesis A Rippi 122
in terms of their validity More precisely the failure points were illustrated in Plaxis in order to
visualize the failure and see if it is real failure or not (real if dike crest settles) That procedure
revealed the possibilities and limitations of FEM modeling The advantage is that a potential
failure can be detected and its visualization is possible On the other hand one should be aware of
the limitations related with the constitutive model that is used In this study the Mohr Coulomb
method was applied that although it can efficiently identify possible failure mechanisms the
representation of the deformations is not so realistic
The failure of the system was considered as a serial system of LSF 1 2 and 3 and it was also
assessed with DS The importance factors stemmed from the reliability analysis indicated soil
features that were important in all the LSFs However it should be stressed that the frequency of
the Plaxis numerical errors was such that quite a few failure points seem to be unrealistic In this
thesis the identification of those errors and their partially exclusion from the analysis was
achieved The latter was attained by the construction of the so-called Response Surfaces (RS)
based on a preliminary sensitivity analysis Such a technique found to be valuable in order to
overcome such a problem however a study for fitting more suitable RS in each problem is
recommended Nevertheless the further elimination of those numerical errors is strongly
recommended to be investigated in future research given the results from the current study
Retrieving the importance factors from a DS analysis was also a challenging part In this study
additional methods were developed in order to evaluate the importance factors and define the
design point after a DS analysis It revealed that alternative possible approaches of the
importance factors (Average 10 and Average all methods) can presumably provide a better
understanding of the systemrsquos behaviour Such methods seemed to be valuable for excluding the
ldquonoiserdquo of the possible fake failure points that originated from FEM numerical errors However
due to the inherent randomness of DS in terms of the selection of the directions the design point
shall be inspected and compared with the other failure points that were found Then someone can
contemplate if it is necessary to proceed with an averaging of the importance factors in order to
get a clearer view the variables influence Of course such a procedure takes time that is not always
available but according to the requirements of each research the most preferable assumptions
should be made
Last but not least the reliability analysis came up with relatively low probabilities of failure which
implies that there is still room for optimizing the design concept of a dike with retaining wall
More precisely the original case study of Breedeveld (2011) has been altered by reducing the
strength properties of the subsoil and installing a less heavy sheet pile wall after following a less
strict deterministic design It was then observed that even then the reliability analysis came up
with a respectively high reliability index It should be stressed that the failure probability of the
system should be also coupled with the uncertainty of the water level in order to obtain an overall
picture of the total probability of failure Nevertheless this research proves the potentials to
optimize the design of such a system which presumably will lead to a less expensive construction
However additional failure mechanisms such as piping internal erosion etc should be also
investigated and the design properties of the system (such as the length and cross section area of
the sheet pile wall) shall be based on them
Conclusions amp Recommendations
MSc Thesis A Rippi 123
82 Recommendations The following topics are recommended for further research
Additional research is required on FEM modeling both for the structural elements and the
soil body More in particular in this research it has been proved that Mohr Coulomb
model cannot realistically estimate the deformations in a soft soil In that case there are
other more advanced models such Hardening Soil model and Soft Soil model that are
capable of simulating the soil displacements and the stress-strain relationship closer to
the actual ones However these methods require more parameters to be specified and
their coupling with reliability packages might not be so robust and functional
As far as the structural elements modeling is concerned there are two major topics that
future research should focus on Firstly the anchor rod can be simulated in Plaxis as a
node-to-node element and at the bottom an embedded beam can be installed in order to
represent the anchorage body This would provide a better reaction with the surrounding
soil body and it could probably simulate the anchor-soil interaction more realistically
Another alternative would be to simulate it as a plate element in order to be able to test
the moments that are also developed on the anchor separately Secondly it is
recommended that the elements be investigated close to their plastic area and a reliability
analysis to be carried out with plastic elements instead In this research an attempt was
made for such an analysis but it appeared not to be functional in cooperation with FORM
analysis
Another issue that is recommended to be investigated is the inspection and the validation
of the failure that Plaxis is giving During the soil failure analysis different types of failure
were observed from which some might not be realistic but due to the modeling limitations
Therefore an attempt should be made in order to further eliminate the effect of Plaxis
numerical errors on the reliability analysis that can lead to fake failure modes A check of
the failure points can be also conducted by using a more suitable constitutive model for
the deformations rather than Mohr Coulomb However this is not always feasible for each
individual case Therefore the illustration of the failure mechanisms in large scale
experiments can give a better insight into the systemrsquos behavior and valuable knowledge
can be acquired of the expected failure modes In this way also 3D effects can be
investigated see what is their effect to the dike instability and at what extend they shall be
considered
Apart from the soil body the stresses and the displacements of the structural elements
shall be also verified It is thus recommended that field tests and measurements take place
in embedded elements like sheet piles and anchor in order to validate Plaxis results and
learn what is the most suitable manner for the their simulation (fixed-end-anchor node-
to-node element plate etc)
Concerning the reliability analysis of the specific case study a research on the potential
relevant LSFs is recommended for the different elements Specifically in this thesis the
LSF were chosen in order to represent the ULS of the element There is the possibility that
the SLS is used and certain acceptable deformations can be settled as a threshold for the
Conclusions amp Recommendations
MSc Thesis A Rippi 124
limit states However the choice of this displacement limit is not straightforward and a
qualitative research shall also be conducted in order to identify the impacts of different
values Besides it is not only the value that should be selected but also the location on the
structure that is going to be checked
It is also advised that the potential correlations among the soil parameters to be included
In Wolters (2012) and Bach (2014) correlation matrices are introduced for specific soil
properties that can be included in a future research and reveal what is their effect on the
reliability analysis The influence of correlations depends on what kind of variables are
correlated and what is the influence of those variables on the system behavior However a
preliminary analysis with independent variables is advisable to be carried out at the
beginning so as for someone to be able to analyze the effect of the variables individually
and then evaluate the correlation effect more wisely
It would be also advisable a comparison to be made between the design of a real case
according to a full probabilistic analysis and partial safety factors In that way a direct
insight into the advantages and disadvantages of each method can be gained and
fundamental improvements of the design concept can be made for dikes with retaining
walls
At this moment the most reliable way of conducting a reliability analysis is with sampling
methods like MC or DS However such methods are usually unattractive due to their large
computational time Thus the tolerance of sampling methods accuracy is lowered in order
also to reduce the number of iterations or approximating methods are used like FORM
analysis However even if approximating methods are used they are usually preferred to
be validated with sampling methods in order for their results to be acceptable A real
validation of the failure probability would mean that a structure should be experimentally
tested in different positions and under the same conditions in order to see if the
probability of failure that is calculated in each case is the same Therefore a real
validation of the failure probability cannot yet be initiated but the computational effort of
the sampling methods can be eliminated with the development of technology and the
improvement of the computer science
MSc Thesis A Rippi 125
References
Abdo T amp Rackwitz R (1990) A new beta-point algorithm for large time-invariant and time
variant reliability problems Proceedings of the 3rd IFIP WG 75 Conference Berkeley pp1-12
California USA
Ang AHS and Tang W (1984) Probability Concepts in Engineering Planning and Design Volume
IBasic Principles John Wiley and Sons New York USA
Baecher GB and Christian JT (2003) Reliability and Statistics in Geotechnical Engineering
Chichester West Sussex John Wiley amp Sons Ltd
Behaviour of Soils and Rocks GEO-ENGINEERING - CIE4361 (Vol 2015) Retrieved from
httpsblackboardtudelftnlbbcswebdavpid-2386735-dt-content-rid-
7926659_2courses32281-14150221-possibilities26limitationspdf
Breedeveld J (2011) Technisch Rapport Analyse Macrostabiliteit Dijken met de Eindige Elementen
Methode Bijlage Voorbeeldcase Deltares Delft the Netherlands
Brinkgreve RBJ (2015) Personal discussion TU Delft the Netherlands
Brinkgreve RBJ and Bakker HL (1991) Non-linear finite element analysis of safety factors In
G Beer JR Booker and J P Carter (Eds) Proceedings of the Seventh International Conference on
computer methods and advances in Geomechanics Rotterdam A A Balkema Brookfield
Cukier RI Fortuin M Shuler KE Petschek AG and Schaibly JH (1973) A study of the
sensitivity of coupled reaction systems to uncertainties in rate coefficients I Theory California
University USA
Cundall PA (2002) A discontinuous future for numerical modeling in soil and rock Published in
Proceedings of the third international conference Discrete Element Methods Numerical Modeling
of Discontinua Edited by Cook and Jensen Santa Fe pp 3-4
CUR (2005) Handbook Quay Walls Gouda the Netherlands
CUR-publication 190 (1997) Probabilities in civil engineering Part 1 Probabilistic design in theory
Stichting CUR Gouda
CUR 166 ndash Damwandconstructies (2005) Technical Recommendation 4e herziene uitgave 2005
The Netherlands
Deltaprogramma (2014) Synthesedocument Veiligheid Achtergronddocument B1 Ministrie van
Infrastructuur en Milieu
EN 1990 2002 Basis of Structural Design European Committee for Standardization
References
MSc Thesis A Rippi 126
Farrokh N (2007) Tools and Strategies for Dealing with Uncertainty in Geotechnics In DV
Griffiths and GA Fenton (Ed) Probabilistic methods in Geotechnical Engineering (pp 71-95) New
York NY Springer Vienna
Fenton GA and Vanmarcke EH (1990) Simulation of random fields via local average
subdivision J Eng Mech ASCE 116(8)1733ndash1749
Flood Defences HYDRAULIC ENGINEERING ndash CIE5314 (Vol 2015) Retrieved from
httpsblackboardtudelftnlbbcswebdavpid-2125084-dt-content-rid-
7621002_2courses29759-
131404CIE5314_FloodDefences_05_failure_mechanisms_stability_Schweckendiekpdf
Gemeentewerken-Rotterdam (2003) Alle triaxiaal proeven rotterdam Technical report Gemeentewerken Rotterdam Rotterdam the Netherlands
Ghanem R and Spanos P (1991) Stochastic Finite Elements ndashA Spectral Approach
Springer New York
Gonzaacutelez A Cuadrado A Josa A and Olivella S (2013) Safety assessment of limit equilibrium
methods for the design of cantilever and anchored sheet pile walls Escola de Camins Barcelona
Griffiths DV and Fenton GA (Eds) (2007) Probabilistic Methods in Geotechnical Engineering Lecture notes monographs edited works and proceedings in the field of Mechanics Engineering Computer Science and Applied Mathematics Italy Hasofer AM and Lind NC (1974) An exact and invariant first-order reliability format Journal of
Engineering Mechanics Division ASCE 100(1) 111-121
Hoek E Read J Karzulovic A and Chen ZY (2000) Rock slopes in civil and mining engineering
Published in Proceedings of the International Conference on Geotechnical and Geological
Engineering GeoEng2000 19-24 November Melbourne
Homma T and Saltelli A (1996) Importance measures in in global sensitivity analysis of
nonlinear models Reliab Eng Syst Saf 52(1) pp1-17
Hydra-Ring 12 (2015) Probabilistic toolbox for the WTI2017 Technical Reference Manual Version
12 Deltares Delft
Jaky J (1944) ldquoThe coefficient of earth pressure at rest In Hungarian (A nyugalmi nyomas tenyezoje)rdquo J Soc Hung Eng Arch (Magyar Mernok es Epitesz-Egylet Kozlonye) 355ndash358
Johansson E amp Sandeman E (2014) Modelling of Deep Excavation in Soft Clay A comparison of different calculation methods to in-situ measurements Chalmers University of Technology Gothenburg Sweden
Joint Committee on Structural Safety (1981) General principles on reliability for structural design
International Association for Bridge and Structural Engineering
References
MSc Thesis A Rippi 127
Larsen H Lubking P and Breedeveld J (2013) Ontwerp stabiliteitsschermen (type II) in primaire
waterkeringen (groene versie) Deltares Delft
Melchers R E (2002) Safety and risk in structural engineering Prog Struct Engng Mater
4 193ndash202
Moumlllman A amp Vermeer P (2008) Reliability analysis of a dike failure case study Elbe river
University of Stuttgart Germany
Nelson RB (1999) Lecture Notes in Statistics An Introduction to Copulas Springer New York
OpenTURNS 15 (2015a) Reference Guide OpenTURNSrsquo methods for Step C uncertainty
propagation pp155-161
OpenTURNS 15 (2015b) Use Cases Guide for the Textual User Interface pp190-192
Pandya N Marsdea E Floquet P and Gabas N (2008) Sensitivity analysis of a model for
atmospheric dispersion of toxic gases In B Braunschweig and X Joulia (Ed) 18th European
Symposium on Computer Aided Process Engineering-ESCAPE18 (pp1143-1148) Elsevier BV
Phoon K (Ed) (2008) Reliability-based Design in Geotechnical Engineering Computations and Applications Taylor amp Francis Britain
Plaxis 2D (2015a) Tutorial manual Delft the Netherlands
Plaxis 2D (2015b) Reference Manual The Netherlands
Plaxis (2015) Material Models Manual The Netherlands
Powell MJD (1994) A direct search optimization method that models the objective and
constraint functions by linear interpolation In S Gomez amp JP Hennart (Ed) Advances in
Optimization and Numerical Analysis (pp51-67) Dordrecht Kluwer Academic
Rankine W (1857) On the stability of loose earth Philosophical Transactions of the Royal Society
of London Vol 147
Saltelli A (2004) What is Sensitivity Analysis In A Saltelli K Chan and EM Scott (Ed)
Sensitivity Analysis (pp 3-13) John Wiley amp Sons Publication
Saltelli A S Tarantola F Campolongo and M Ratto (2004) Sensitivity Analysis in Practice
A Guide to Assessing Scientific Models John Wiley amp Sons Ltd
Schittkowski K (1985) NLPQL A Fortran subroutine solving constrained nonlinear
programming problems Annals of Operations Research 5 485-500
Schweckendiek T (2006) Structural Reliability applied to Deep Excavations Coupling Reliability
Methods with Finite Elements TU Delft Delft
References
MSc Thesis A Rippi 128
Schweckendiek T Vrouwenvelder ACWM Calle EOF Jongejan RB and Kanning W (2013)
Partial Factors for Flood Defenses in the Netherlands Modern Geotechnical Codes of Practice ndash
Development and Calibration Fenton GA et al (eds) Special Geotechnical Publication Taylor amp
Francis
Seed RB Bea R G Abdelmalak R I Athanasopoulos A G Boutwell G P Bray J D Briaud J-L Cheung C Cobos-Roa D Cohen-Waeber J Collins B D Ehrensing L Farber D Hanemann M Harder L F Inkabi K S Kammerer A M Karadeniz D Kayen RE Moss R E S Nicks J Nimmala S Pestana J M Porter J Rhee K Riemer M F Roberts K Rogers J D Storesund R Govindasamy A V Vera-Grunauer X Wartman J E Watkins C M Wenk Jr E and Yim SC (2006) Investigation of the Performance of the New Orleans Flood Protection Systems in Hurricane Katrina on August 29 2005 Volume I Main Text and Executive Summary New Orleans USA
Steenbergen RDJM (2013) Lecture notes Probabilistic Design Delft the Netherlands TU Delft
TAW (Technische Adviescommissie voor de Waterkeringen) (2004)Technisch Rapport Waterspanningen bij dijken The Netherlands
Terzaghi K (1942) Theoretical Soil Mechanics New York Wiley ISBN 978-0-471-85305-3 Teunissen JAM (2009) SBW Analyse macrostabiliteit van dijken met Eindige Elementen
Modellen Achtergronden bij Activiteit EEM 04a Opstellen stappenplan Deltares the Netherlands
Terzaghi K (1943) Theoretical Soil Mechanics John Wiley and Sons New York
Turaacutengi T amp Rabitz H (2004) Local Methods In A Saltelli K Chan and EM Scott (Ed)
Sensitivity Analysis (pp 81-99) John Wiley amp Sons Publication
Vergouwe R Huting RJM and van der Scheer P (2014) Veiligheid Nederland in Kaart 2 Overstromingsrisico dijkring 43 Betuwe Tieler- en Culemborgerwaarden The Netherlands
Vrijling JK Schweckendiek T and Kanning W (2011) Safety Standards of Flood Defenses Keynote at ISGSR 2011 (International Symposium on Geotechnical Safety and Risk) Munich Vogt Schuppener Straub amp Braumlu (eds)
Vrijling JK and Gelder PHAJM van (2002) Probabilistic Design in Hydraulic Engineering Lecture notes TUDelft Faculty of Civil Engineering and Geosciences Delft Vrouwenvelder ACWM Steenbergen HMGM and Slijkhuis KAH (1999) Theoriehandleiding PC-Ring Delft the Netherlands TNO
Waarts PH (2000) Structural reliability using Finite Element Analysis An appraisal of DARS
(Directional Adaptive Response surface Sampling) Delft University of Technology Delft the
Netherlands
Wyllie DC amp May CW (2004) Rock Slope Engineering (4th edition) Spon Press London
MSc Thesis A Rippi 129
Appendix A OpenTURNS features
In this Appendix an overview of the most relevant of OT is given Firstly in section 112 an
introduction of the applicable methods in SA is made while in section 52 the results from such an
analysis are presented In Appendix A1 a more detailed elaboration is made of how the FAST
sensitivity method that was applied in this project works
In Appendix A2 an explanation of the different optimization algorithms available for FORM
analysis is given The theory behind these algorithms and their convergence criteria are discussed
while finally an evaluation of their performance after applying them in case of the anchor failure
is presented This evaluation has been conducted in order to decide about the most relatively
reliable algorithm that should be adopted for the reliability analysis
Finally in Appendix A3 the types of probability distributions that have been used in this study
for the random variables and which are available in OT are displayed
A1 Fourier Amplitude Sensitivity Test (FAST) In this section the extended FAST sensitivity method is discussed as it is also described in the
Reference Guide of OpenTURNS 15 (2015) FAST is based on the Fourier decomposition of the
variance of the model response ( ) the latter being represented by its Fourier expansion
is an input random vector of independent components Its key idea is to
recast this representation as a function of a scalar parameter by defining
exploring the support of the input random vector
For each input the same procedure is realized in three steps
1) Sampling
Deterministic space-filling paths with random starting points are defined ie each input Xi is
transformed as follows
( ( )) (A1)
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 130
Figure A1 Search curves in the input space as they are defined from the transformation (the blue
dot indicates the direction of the current search path)
where is the number of input variables N is the length of the discretization of the s-space with
s varying in (-π π) by step 2πN is a random phase-shift chosen uniformly in [0 2π]
is a set of integer frequencies assigned to each input Xi The frequency
associated with the input of interest is set to the maximum admissible frequency satisfying the
Nyquist criterion (which ensures to avoid aliasing effects)
(A2)
with M the interference factor usually equal to 4 or higher It corresponds to the truncation level
of the Fourier series ie the number of harmonics that are retained in the decomposition realised
in the third step of the procedure And the maximum frequency of the complementary set of
frequencies is
(A3)
with the index ldquo-irdquo which meaning ldquoall but irdquo
2) Simulations
Output is computed such as
Then is expanded onto a Fourier series
sum [ ] (A4)
where and are Fourier coefficients defined as follows
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 131
int
(A5)
int
(A6)
These coefficients are estimated thanks to the following discrete formulation
sum
(A7)
sum
(A8)
3) Estimations by frequency analysis
The quantitative measure of sensitivity is represented by Sensitivity Indices The first-order
sensitivity index of an input pi is the measure of the main (direct) effect of pi on the output
variance (where i ne j) the second-order sensitivity indices measures the interaction effect of
pi and pj on the output variance Other higher-order indices are defined in the same manner The
total sensitivity index is the sum of all sensitivity indices involving factor pi (Homma amp Saltelli
1996) For example the total sensitivity index of factor 1 for a model with 3 input factors is
given as
The first-order (Si) and total ( ) indices can be interpreted as following (Pandya et al 2008)
high pi is an influent parameter
and both small pi is not an influent parameter (neither alone nor in interaction with
other parameters)
and nearly the same no interaction of pi with the other parameters
and very different high interaction of pi with the other parameters
Total indices are especially suited to apportion the model output variation to the input factors in a
comprehensive manner The FAST method calculates the first-order and the total sensitivity
indices whereas Sobolrsquo method in addition to these also provides all higher-order sensitivity
indices to determine quantitatively the interaction between parameters However the
computational cost and calculation time of Sobolrsquos method tends to be higher than that of the
FAST method
Combining equations A1-A8 the first-order indices are estimated as follows
sum (
)
sum (
)
(A9)
where is the total variance the portion of D arising from the uncertainty of the ith input and N
the size of the sample using to compute the Fourier series
Subsequently the total-order indices are estimated as follows
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 132
sum (
)
sum (
)
(A10)
where is the part of the variance due to all inputs except the ith input
A2 Optimization Algorithms in FORM
Principles of optimization algorithms
The possible optimization algorithms in OT are
Abdo-Rackwitz (ARF)
Cobyla
Sequential Quadratic Programming (SQP)
Below a description of the main principles of the optimization methods mentioned above is
quoted ARF and SQP methods are elaborated according to Abdo and Rackwitz (1990) while
Cobylarsquos principles according to Powell (1994) Finally an evaluation of the performance of the
different methods follows after applying them in case of the anchor failure
The general optimization problem (objective function) can be written in the classical form
subject to inequality constraints (constraint function) (A11)
The Lagrangian function of the general problem is defined by
sum (A12)
where are the so-called Lagrangian multipliers The SQP method (developed by Schittkowski
(1985)) replaces the original problem by a sequence of quadratic programming problems which
are exactly solvable and which approximate the original one This is done by approximating the
Lagrangian function by its second order Taylor expansion in an initial point u0
(A13)
where
sum
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 133
sum
sum
in which is the gradient operator and represents the Hessian matrix27 of the function f in
the point u0 The optimality conditions for any iteration point k of the sequence of quadratic
expansions are
sum [ sum
] (A14)
(A15)
The exact calculation of the Hessian matrix is generally too expensive and cannot be efficiently
implemented for a general case Therefore the gradient information obtained in each point
during iteration is used to build up an approximation of this matrix using one of the known
update formulas
The new iteration point is defined by
(A16)
where is the step length and is a direction in which a line search is performed The process
stops when the optimality conditions of the original problem are satisfied
The most time consuming part in this algorithm is the updating of the Hessian matrix and the
solution of the system of equations A fair approximation of the Hessian of non-quadratic
functions is also obtained with about n updates of the matrix This means that the approximation
used in the few (say ten) iterations to reach convergence cannot be very good when the problem
has large number of variables The rounding errors during the updating process in large problems
can make the approximate Hessian to become singular Close to singularity the search direction
can be significantly distorted In this case the algorithm has to restart the iteration with a unit
Hessian matrix in the point where singularity occurred
The posterior algorithm ARF developed by Abdo and Rackwitz (1990) overcomes this problem
by obtaining a constant approximation of the true Hessian matrix Only the contribution of the
objective function to the Hessian is considered The numerical decomposition of the matrix
contains the scalar product of the gradients of the constraints as elements in each iteration
27 Hessian is a square matrix of second-order partial derivatives of a scalar-valued function or scalar field It describes the local curvature of a function of many variables
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 134
Cobyla algorithm has been developed by Powell (1994) and it constructs linear polynomial
approximation to the objective and constraint functions by interpolation at the vertices of
simplices (a simplex in n dimensions is the convex hull of n+1 points n being the number of
variables) It generates the next vector of variables from function values at the vertices
of a nondegenerate simplex in In this case there are unique linear
functions and say that interpolate f and at the vertices
and the optimization problem A15 by the linear programming problem
( ) (A18)
( )
The iterative use of expression A18 puts this method in the class of ldquosequential linear
programming algorithmsrdquo It is also essential to notice that in this method the gradients are
derived by interpolation at the vertices of simplices and not analytically as it happens in SQP and
ARF
Additionally Cobyla algorithm makes use of a trust region bound ρ so that the trust region
condition on the new vector of variables is
(A19)
Such a region gives the user some control over the steps that are taken automatically and which
respond satisfactorily to the fact that there may be no finite solution to the linear programming
problem The algorithm also employs a merit function of the form
( ) ( ) [ ( ) ] (A20)
in order to compare the goodness of two different vectors of variables Here is a parameter that
is adjusted automatically depending on the how close to optimizing the objective function the
analysis is and the subscript ldquo+rdquo means that the expression in square brackets is replaced by zero
if and only if its value is negative so we have ( ) ( ) whenever is feasible Parameters
and are changing ( only reduces) automatically according to the improvement of the
optimization problem that sequential vectors might cause Therefore if the change in the merit
function does not happen to improve the optimization then these values are changing Especially
regarding the parameter user gives the initial and the final values of it namely and (it
is recommended that be a reasonable change to make the variables for a coarse exploration
of the calculation while should be approximately the required distance from the final vector
of variables to the solution of the optimization problem)
Convergence criteria
The convergence of the aforementioned algorithms is controlled by the evaluation of the
following errors expressed in the standard space (for the relationship between the real space (x-
space) and the standard space (u-space) see section 422) (OpenTURNS 15 (b) 2015 pp 190-
192)
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 135
The absolute error which is the distance between two successive iterates
(A21)
The constraint error which is the absolute value of the limit state function minus the
threshold
(A22)
The relative error which is the relative distance between two successive iterates (with
regards to the second iterate)
(A23)
The residual error which is the orthogonality error (lack of orthogonality between the
vector linking the center and the iterate and the limit state function)
( ) (A24)
The algorithm converges if all the final error values are less than the maximum value specified by
the user The convergence can be also controlled by altering the maximum number of iterations
which should be higher in case of stricter required errors What is more there is the possibility to
manipulate the starting point that the algorithm uses for its first iteration Usually the default
starting point in a FORM analysis consists of the mean values of the random variables However if
there is already some knowledge about the potential design point it can be set as the starting
point of the algorithm in order to accelerate the convergence and save a lot of time
The aforementioned errors the maximum number of iterations and the starting point were
manipulated during the analysis in order to understand the functionality of the different
algorithms and to improve their efficiency
Evaluation of the algorithms performance
An evaluation of the above optimization algorithms has been made in order to make a decision of
the most suitable for the current project In principle a literature overview regarding the
performance of the algorithms is presented Then the 3 algorithms were tested considering the
anchor failure limit state function (see Eq 55) with a stress limit of 450000 Pa and their results
are discussed below
According to Abdo and Rackwitz (1990) the SQP method seems to have better convergence
behavior for highly curved constraint functions which is especially true when very expensive
structural state functions are involved such as finite element structural analyses However the
mentioned algorithm fails to reach convergence at a problem dimension (amount of random
variables) of around n=50 Therefore for increased dimensionality SQP is less efficient in terms
of storage and CPU time compared with other methods due to singularity of the updated Hessians
matrices In OT a warning is introduced since an analysis has started that a default
implementation of Hessian is being used and that the computation can be severely wrong
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 136
Abdo and Rackwitz (1990) also proved that ARF algorithm can handle problems of up to 2000
variables while the storage requirements and the CPU time are much smaller than with SQP
Therefore optimization problems of high dimensionality can be evaluated
In OT a warning appears at the beginning of a FORM calculation saying that a default
implementation of the gradient is being used and that the computation can be wrong A similar
message also appears for the SQP algorithm saying that a default implementation of the Hessian
matrix is used and that the computationrsquos results might be unreliable Moreover in both
algorithms it is advisable to check the values of the random variables that were used during the
analysis in order to ensure that they fluctuate within acceptable boundaries (Figure A4 is an
example of a variable that surpassed the minimum boundary at some stage)
Firstly the Cobyla algorithm was applied on the case study in combination with a FORM analysis
Cobyla does not require a gradient evaluation of the limit state function and does not use a default
implementation of the Hessians matrix as ARF and SQP does while the computational time is
quite small comparable to the aforementioned methods
The Cobyla algorithm was tested with (i) 3 (ii) 10 and (iii) 19 parameters respectively in order to
check its efficiency28 As it has been mentioned in the previous section the convergence criteria to
be steered are the errors (absolute relative residual and constraint) the maximum number of
iterations and the starting point At first the default errors were used with a value of 10-3 but it
was realised that such an order of errors lead to non-convergence of the algorithm Therefore the
algorithm was then checked after relaxing the maximum errors to values of 01 05 10 50 and
100 Finally the combination of the errors that led to convergence was 01 05 01 02 for the
maximum absolute relative residual and constraint error respectively After the analysis the
number of LSF calls in case (i) (ii) and (iii) was 25 even if the maximum one was set to 100 and
the probability of failure was around 035 The probability of failure was quite high and that is
why an inspection of the output file was carried out In Figures A2 and A3 the values of the
anchor limit state function are presented in the right graph during the iterations It was then
observed that the threshold of 450 kPa was never surpassed during the iterations
28 All of the algorithms were first checked with very simple LSF (linear and quadratic) and their results were verified by MC and DS
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 137
Figure A2 Friction angle and cohesion progress during a FORM analysis with Cobyla and with 3 random variables The figure at the right illustrates the limit state function evolution
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 138
Figure A3 Friction angle and cohesion progress during a FORM analysis with Cobyla and with 19 input parameters The last figure at the right illustrates the limit state function evolution
An additional action was then taken by relaxing the maximum number of iterations from 100 to
200 but no improvement was noticed (the LSF calls were again 25) It was also given extreme
values to the errors and especially to the absolute constraint and relative error in order to check
the response of the algorithm but that did not lead to better results Furthermore the
parameter (see theory part for explanation) was manipulated but even then no further
improvement was noticed in terms of the results In Figures A2 and A3 the performance of the
algorithm during the analysis can be seen through the values of the variables that were assigned
during the iterations
As someone can notice from the graphs above it is obvious that Cobyla does not perform very
well in high dimensionality problems as the variables do not seem to vary a lot More specifically
in Figure A3 with the 19 parameters it can be noticed that the random variable does not change
at all apart from two peaks at the beginning and at the end of the analysis respectively The
incapability of the algorithm to change the parameters in order to reach the limit of the anchor
stress ie 450 kPa (while here a maximum of approximately 267 kPa is reached) leads to a poor
approximation of both the design point and the probability of failure In Figure A2 with only 3
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 139
parameters the variability of the parameter is higher during the analysis which leads to a better
approach of the limit (around 370 kPa which is larger than with the 19 parmaters) but even then
a failure probability of 035 was not a realistic value On the other hand Cobyla is easy to use only
for small number of variables (ie less than 9) with linear LSF otherwise the linear
approximations can be highly inefficient Apparently the limit state function of the anchor stress
level as a function of the soil parameters is much more complex in order to be approximated by
this algorithm
Therefore the other available optimization algorithms ie ARF and SQP were examined as well
Because of the similarity of the two algorithms only the ARF performance is presented below and
a link with SQP is made when it is needed
The ARF algorithm was tested initially with 19 parameters The settings for the errors that were
used were 01 for the absolute relative residual and constraint error respectively and the
maximum number of iterations was set to 100
Figure A4 Friction angle and cohesion progress during a FORM analysis with ARF algorithm and with 19 input parameters The figure at the right illustrates the limit state function evolution while the red spots indicate the points where the 450 kPa threshold was surpassed
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 140
The algorithm worked efficiently which means that the algorithm managed to converge towards
the threshold of 450 kPa of the limit state function as it is depicted in Figure A4 (see red dots) It
was also noticed that the random variables fluctuated a lot during the iteration steps even if the
amount was significant (19 parameters) The capability of the algorithm to vary all the 19
variables evenly and in a wide range makes it possible to find the real design point and thus to
come up with a more reliable failure probability The resulted probability of failure was 3810-9
and the reliability index was ϐ=58
However the analysis took almost 4 days to finish which is a long duration and the number of LSF
calls counted at 4151 Moreover even if the results were promising after tracing all the 19
parameters in order to inspect the algorithmrsquos performance it was found out that some variables
took unacceptable values29 during the analysis As an example in Figure A5 Poissonrsquos ratio
received negative values (see red circle) at the beginning of the iterations There were also other
such examples in the same analysis such as the friction angle which took sky-high values and the
interface strength which took negative values as well as values above 1 which is unacceptable It
should be mentioned that the design point of the algorithm was completely reasonable but even
then such a performance can prevent the algorithm from converging to the right design point
quickly and increase the computational effort
Figure A5 Poisson ratio progress during the FORM analysis with ARF algorithm
Taking into account the above implications there are two main concerns at that stage (a) the
high computational time and (b) the unreasonable values that the variables took Firstly in order
to reduce the computational time the maximum errors were relaxed More precisely it was
noticed that after relaxing the maximum relative error a lot which is more related to the
29 In some cases a value of ldquoinfrdquo was given to the variables that led to termination of the analysis without results
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 141
relaxation factor (in this case an extreme value of 1000030 was tried) and increasing the absolute
error to 05 the number of LSF calls was reduced to 724 and the analysis was finished after
almost 10 hours In Figure A6 the values that the cohesion and the friction angle took during the
iterative procedure while the red spots in the right figure indicate the points that are included in
the failure space Figure A4 shows the progress of friction angle after relaxing the errors
Figure A6 Progress of the friction angle during the FORM analysis with ARF algorithm with errors
05 10000 01 01 for the absolute relative residual and constraint error respectively
As a general conclusion after the test of different combination of maximum errors by relaxing the
constraint absolute and relative error too much successive iterations of a certain variablersquos value
seem to have a wide difference which can detain the algorithm from converging Therefore the
wisest in this case study is to keep the maximum errors below or around 10 except for the
relative and the constraint error that can relax more than 1 (depending on the case) in order to
reduce the computational time to the minimum possible
To prevent having unacceptable values of the parameters and in order to increase the efficiency of
the algorithm the random variables were reduced to 11 after a SA and some of them were
truncated in a certain range according to observations and experience until now However the
algorithm then started to give an error due to zero gradient of the objective function at a specific
point (similarly SQP gave an error due to zero hessian matrix at a specific point) while the upper
limit of the LSFs could not be reached As a countermeasure the errors were relaxed (05 10 05
10 for the absolute relative residual and constraint error respectively) in order for the algorithm
not to be trapped in a specific point while also the step of the algorithm was manipulated It
was finally proved that this kind of error stemmed from the fact that the structural elements were
considered as elastoplastic and thus a limit of the stress level was considered that misled the
algorithm The elements were eventually considered as elastic and the numerical problem was
solved
30 Practically such a value has the meaning that there is no limitation of the relative error during the optimization In the standard space (where these errors are expressed) a value also greater than 10 can be also quite high However here extreme values were given due to the time limit to test multiple values
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 142
Before figuring out that the elastoplastic behaviour was giving the problem Cobyla as a gradient
free algorithm was tested again as a potential solution but with changing the starting point close
to the design one that was found with ARF It is worthwhile to take a look into the performance of
the algorithm with 19 random variables in Figure A7 In that way we boost Cobyla to converge
and it eventually performed quite differently than in Figure A3 and without giving convergence
problems The computational time was estimated to be 9 hours and the probability of failure was
very small in the order of 10-120 However the validation of the failure probability with DS
showed that the result of Cobyla was far from correct and thus ARF was preferred for continuing
with the rest of the analysis For educational reasons a FORM analysis with ARF was again
performed after changing the starting point and it was found that the computational time was
reduced to 4 hours Therefore it is still worthy to be aware of the fact that by changing the
starting point the performance of the algorithm can improve a lot However that requires a prior
knowledge of the potential design point that usually does not exist
Figure A 7 Cobyla performance after changing the starting point closer to the failure point
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 143
It must be also mentioned that by reducing the number of the variables to 11 as it was mentioned
before the algorithms could not converge to the errors that has been set and the computational
time increased compared with the case of 19 variables Taking into account the high interaction
among the variables that was proved in section 73 through the SA it can be possible that the
combination of the soil variables could lead to failure rather than the variables individually
Therefore an analysis was tried again with the complete set of the parameters and it was realised
that the algorithm did converge to the limit value
A3 Distribution Types
Uniform Distribution
If the probability of a result is between a lower limit a and an upper limit b and if every result
between these limits is equally probable a uniform distribution should be used This distribution
is defined by the two limit parameters and the probability density function is written
(A25)
and the CDF is
(A26)
Figure A8 Probability density function of a uniform distribution with a=-2 and b=2
The mean and the variance of the distribution are also defined by the limit parameters
(A27)
(A28)
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 144
Normal Distribution
One of the most commonly used distribution types is the normal distribution or Gauss-
distribution and it is defined by its first two central moments the mean micro and the standard
deviation σ The PDF is given by
radic
(A29)
and its CDF is
int
radic
(A30)
Figure A9 Probability density function of a normal distribution with different mean and standard
deviation values (notice that the curve with micro=0 and σ2=1 belongs to the standard normal distribution)
The normal distribution with mean and standard deviation is called the standard
normal distribution In the most reliability methods such as FORM SORM DS etc the analysis is
carried out in the standard normal space which actually means that the stochastic parameters are
transformed into standard normal variables
Lognormal Distribution
If a variable X has a lognormal distribution the variable Y=ln(X) has a normal distribution The
probability density function is defined as
radic (
) (A31)
where and the mean and the standard deviation of the normally distributed random
variable Y Sometimes it is useful to work with the Yrsquos parameters rather than Xrsquos These
parameters can be expressed as follows
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 145
radic
(A32)
(A33)
where and are the mean and the standard deviation of Y In Figure A10 an example is given
of a log-normally distributed variable with different values of mean and standard deviation
Figure A10 Probability density function of a log-normal distribution with different mean and standard deviation values
Truncated Normal Distribution
The truncated normal distribution is the probability distribution of a normally distributed
random variable whose value is either bounded below or above or both
Suppose has a normal distribution and lies within the interval
Then conditional on has a truncated normal distribution whose probability
density function is
(
) (
)
(A34)
and otherwise and is the probability density function and the CDF of the standard
normal distribution similar to Eq (1) and (2) If
and
then the CDF is
defined as
(A35)
Then the mean value and the variance of the conditional will be respectively
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 146
(A36)
[
(
) ] (A37)
In the next figure the PDF histograms of a variable with a mean of 0 and a standard deviation 1
were plotted for two cases (a) variable is normally distributed (b) variable is truncated-normally
distributed with =-2 and b=2
Figure A 11 Truncated normal distribution in comparison with normal distribution
MSc Thesis A Rippi 147
Appendix B Plaxis 2D (2015) features
In this Appendix the basic features of Plaxis 2D that were treated in this thesis are discussed First
of all an introduction over the Mohr Coulomb model is made as the soil failure criteria stem from
that model Then the initial stress generation in Plaxis is discussed and how that was applied in
the current master thesis while also the φ-c reduction technique is described which was used in
the SA of the soil failure Last but not least a more detailed explanation of the interface strength
Rinter is given which represents the modeling of the soil-structure interaction in Plaxis and it has
been considered as a random variable in the current project
B1 Mohr Coulomb failure criterion Soil tends to behave in a highly non-linear way under load This non-linear stress-strain behaviour
can be modelled at several levels of sophistication There are different constitutive models
available for simulating the behaviour of the soil whose parameters can change depending on the
level of sophistication that is required in each case
As someone can notice in Table B1 different models can be applied in different cases of soil
behaviour Hardening Soil model is becoming more and more interesting in soil structures as it
has been proved to approach the plastic behaviour of the soil and the non-linear stress-strain
relationship in a reliable and promising level Moreover UBCSAND31 and the Hypoplastic model
are more sophisticated models capable of representing several aspects of soil behaviour
however the amount of the model parameters is large and for that reason they are recommended
only in case that enough soil data are available for determining these parameters
Table B 1 Features of material models (source Behaviour of Soils and Rocks 2015)
ModelFeature Elasticity Failure Hardening Softening Small strain stiffness
Cyclic loading
Liquefaction Creep
Mohr-Coulomb x x
Duncan-Chang x x
Hardening Soil x x x
HSsmall x x x
x
Modified Cam-Clay
x x x x
Soft Soil x x x
Soft Soil Creep x x x
x
UBCSAND x x x
x x
Hypoplastic (x) x x x
x (x)
31 UBC are the initials for University of British Columbia in Vancouver in Canada
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 148
In this master thesis the Mohr-Coulomb model is going to be applied which is simple but an
efficient model to describe the soilrsquos failures condition Below a description of that model is made
The Mohr-Coulomb model is a linear elastic perfectly plastic model (see Figure B1) The linear
elastic part is based on Hookersquos law while the perfectly plastic part is based on the Mohr-Coulomb
failure criterion formulated in a non-associated plasticity framework
The main principle of elastoplasticity is that total strains and strain rates consist of elastic and
plastic strain components
(B1)
(B2)
in which the elastic strain rate component abides by the Hookersquos law and thus it can be expressed
as follows
(B3)
where is the elastic stress-strain matrix and the effective stress rate The plastic strain
component is given by
(B4)
where is a scalar defining the magnitude of the plastic strains called plastic multiplier and
is
a vector describing their direction with to be the so-called plastic potential function
Figure B1 Basic principle of elastoplasticity
The occurrence of (elasto)plasticity can be indicated by the yield function which is
illustrated in Figure B2 in principle stress space and it is derived according to the Mohr-Coulomb
failure criterion expressed in the next formula
(B5)
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 149
Considering the sign of this yield function the different soil states can be recognised as follows
if flt0 (inside yield contour) pure elastic behaviour
if f=0 and dflt0 unloading from aplastic state (=elastic behaviour)
if f=0 and df=0 elastoplastic behaviour
if fgt0 (outside yield contour) impossible stress state
Figure B2 Mohr-Coulomb yield surface in principle stress space
Finally in Mohr-Coulomb model the calculation of the plastic strains is achieved by the so-called
non-associated flow rule where with the dilatancy angle
This makes a difference between the friction and the dilatancy angle which prevents the model
from overestimating potential uplift
Concluding this model uses 5 input parameters
Stiffness parameters E (Youngrsquos modulus) and v (Poissonrsquos ratio) as they are included in
the elastic stress-strain matrix according to Hookersquos law
Strength parameters c (cohesion) and (friction angle)
Dilative behaviour ψ (dilatancy angle)
The author is aware of the possibilities and limitations of this model In particular the model is
capable of
a first order approach of the soil behaviour in general
describing quite well the (drained) failure behaviour
It should be mentioned at that point that the Youngrsquos modulus E cannot be steered via the
Python interface (as it appears to be a readonly parameter) and thus the shear stiffness G was
considered instead that is defined as follows
[kPa]
f=0 (yield contour)
flt0
fgt0
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 150
On the other hand the model has also limitations that are referring to
the linear elastic behaviour until failure (no strain- or stress path-dependent stiffness
behaviour)
overestimation of the shear strength in case of an undrained behaviour
However this model is considered reliable for detecting failure and giving realistic stresses
magnitudes in the soil and structural elements
Moreover in this project a non-associative behaviour was assumed by considering a zero
dilatancy angle However it has been proved by Teunissen (2009) that in case of deep failure
mechanisms in a dike structure with soil materials of a friction angle above 30deg problems arise in
the numerical analysis in finite element calculations Particularly the strength in the sliding
surface decreases due to rotation and the ground stresses are relieving around it by releasing
energy which might lead to calculation implications These phenomena are not treated in this
thesis but it is advisable to check and review the results after such an analysis
B2 φ-c Reduction Technique In this section the ldquoSafety Calculationrdquo is applied as it is described in Plaxis (Plaxis 2D (a) 2015)
In the Safety Calculation the φ-c reduction technique (Brinkgreve amp Bakker 1991) is adopted
where the shear strength parameters tanφ and c of the soil as well as the tensile strength are
successively reduced until failure of the structure occurs The dilatancy angle ψ is in principle not
affected by the phic reduction procedure However the dilatancy angle can never be larger than
the friction angle When the friction angle φ has reduced so much that it becomes equal to the
(given) dilatancy angle any further reduction of the friction angle will lead to the same reduction
of the dilatancy angle The strength of interfaces if used is reduced in the same way
The total multiplier is used to define the value of the soil strength parameters at a given
stage in the analysis
(B6)
where the strength parameters with the subscript ldquoinputrdquo refer to the properties entered in the
material sets and the parameters with the subscript ldquoreducedrdquo refer to the reduced values used in
the analysis is set to 10 at the beginning of a calculation to set all material strengths to
their input values
The incremental multiplier is used to specify the increment of the strength reduction of the
first calculation step (this increment is by default 01 but user can change it) The strength
parameters are successively reduced automatically until all required steps have been performed
(by default these steps are set to 100 but this can be subjected to changes depending on the
situation) It must always be checked whether the final step has resulted in a fully developed
failure mechanism In other words if has reached an almost constant value while the
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 151
deformation is continuing that indicates an equilibrium state In that case the factor of safety is
given by
(B7)
B3 Initial Stress Generation The first construction phase in Plaxis was the initial stress generation before the dike has been
constructed In that phase for horizontal soil surfaces with any soil layers parallel to this surface
and horizontal phreatic level32 a K0-procedure can be applied where K0 is the ratio between
horizontal and vertical stresses
(B8)
In practice the value of for a normally consolidated soil is often assumed to be related to the
friction angle by Jakyrsquos empirical expression (Jaky 1944)
Using very low or very high K0-values in the K0 procedure may lead to stresses that violate the
Mohr-Coulomb failure criterion In this case Plaxis automatically reduces the lateral stresses such
that the failure condition is obeyed Hence these stresses points are in a plastic state and are thus
indicated as plastic points Although the corrected stress state obeys the failure condition it may
result in a stress field which is not in equilibrium
B4 Interface Strength In FEM for a proper modelling of soil-structure interaction interface elements can be added to
plates in order to simulate the thin zone of intensely shearing material at the contact between the
plate and the soil
The strength properties of the interface are linked to the strength properties of the adjacent soil
layer and they are reduced by a factor of as follows
(B9)
(B10)
(B11)
where index ldquoirdquo refers to the interfacersquos parameters whereas index ldquosoilrdquo refers to the adjacent soil
parameters The behaviour of the interfaces is defined according to the Coulomb criterion which
distinguishes between elastic behaviour where small displacements can occur within the
interface and plastic interface behaviour when permanent slip may occur
32 For non-horizontal surfaces a Gravity loading is recommended
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 152
When the interface is elastic then both slipping (relative movement parallel to the interface) and
gapping (relative displacements perpendicular to the interface) can occur The magnitudes of the
interface displacements are
(B12)
(B13)
where is the shear modulus of the interface is the 1-D compression modulus of the
interface is the virtual thickness of the interface (generated automatically from Plaxis) is
the elastic interface normal stiffness and is the elastic interface shear stiffness
The shear and compression moduli are related as follows
(B14)
(B15)
Taking into account the above relations the displacements can be highly influenced by the elastic
parameters and thus the strength reduction factor can have a considerable impact on the
systemrsquos behaviour In the Reference manual of Plaxis 2D (2015) a value of 23 is generally
recommended while a value of 1 would not normally be used It should be stated that in the
framework of this thesis there are no available measurements in order to define the sensitivity to
this parameter and hence it is considered uncertain variable following a truncated normal
distribution
Appendix C ndash NEN 6740 ndash Table 1
MSc Thesis A Rippi 153
Appendix C NEN 6740 - Table 1
MSc Thesis A Rippi 154
MSc Thesis A Rippi 155
Appendix D Input Files for the Reliability Analysis
Below an example of the input files that were used for the anchor the sheet pile the soil body and
the system failure is presented in Figures D1 D2 D3 and D4 respectively In particular
regarding the input file of the anchor failure the meaning and the scope of each line is briefly
explained in Table D1
Figure D 1 Anchor analysis input file (with FORM method)
APPENIDX D ndash Input Files for the Reliability Analysis
MSc Thesis A Rippi 156
Table D1 Explanation of input components
A Reliability method and its parameters eg FORM with a maximum number of iterations 100 a maximum absolute error of 05 a relative error of 1000 a residual error of 05 and a constraint error of 05
B Number of random variables
C Soil material as it is specified in Plaxis eg AKleiVenigOnder is the clay layer under the dike
D Soil parameter eg Rinter is the interface strength of the clay layer
E Probability distribution and its parameters eg Truncated normal distribution with mean 066 standard deviation 013 and bounds from 00 to 10
F Copula type
G I Limit state function I is the threshold of the function
H Response surface function that is activated in case of Plaxis errors
Figure D 2 Sheet pile wall analysis input file (with FORM method)
APPENIDX D ndash Input Files for the Reliability Analysis
MSc Thesis A Rippi 157
Figure D 3 Soil analysis input file (with DS method)
APPENIDX D ndash Input Files for the Reliability Analysis
MSc Thesis A Rippi 158
Figure D 4 System analysis input file (with DS method)
MSc Thesis A Rippi 159
Appendix E Characteristic and mean values
The characteristic value of a parameter implies that there is 5 probability that the real value is
higher (load) or lower (resistance) than the characteristic In Figure 22 the probability density
function of a standard normal distributed parameter (mean value equals to zero and standard
deviation 1) is depicted In case of such a distribution the characteristic value is 164σ higher than
the mean value
Figure E1 Characteristic value illustration for a normal distribution (source Wolters 2012)
In this study full probabilistic calculations are taking place in order to estimate the corresponding
structural reliability According to the probabilistic design concept the mean values and the
standard deviations are used instead of characteristic values In that case characteristic values
and shall be converted to their mean values and as follows
where
(E1)
(E2)
and are the values for load or resistance respectively to which a probability of (non-
)exceedance of 95 corresponds (non-exceedance for the load and exceedance for the resistance)
(for a standard normal distributed parameter this is 1645) The definition of the rest parameters
is given in section 423
MSc Thesis A Rippi 160
MSc Thesis A Rippi 161
Appendix F Reliability methods
In this appendix an overview of different reliability methods is given More precisely in F1 an
explanation is given concerning the generation of random samples that is applied in MC as well as
in other sampling methods In Appendix F2 more advanced sampling methods than crude MC are
presented that can be used as accelerating reliability methods in the sense that the computational
time can be reduced Finally in Appendix F3 some basic principles of the FOSM are elaborated
that can help to better understand the concept of FORM as well
F1 Generation of random samples in Monte Carlo
The non-exceedance probability of an arbitrary random variable is uniformly distributed between
zero and one regardless of the distribution of the variable In formula (CUR ndashpublication 190
1997)
(F1)
is the uniformly distributed variable between zero and one and is the non-exceedance
probability Thus for the variable X
(F2)
in which is the inverse of the PDF of X Using this formula a random number X can be
generated from an arbitrary distribution by drawing a number of from the uniform
distribution between zero and one
More or less the same way base variables of a statistical vector can be drawn from a known joint
probability distribution function However the joint PDF must then be formulated as the product
of the conditional probability distributions of the base variables of the vector In formula this is
( )
By taking m realizations of the uniform probability distribution between zero and one a value can
be determined for every
(F3)
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 162
If the variables are statistically independent this can be simplified to
(F4)
By inserting the values for the LSF one can check whether the obtained vector ( ) is
located in the safe area
F2 Other Sampling Methods
Quasi-Monte Carlo Simulation
As an alternative to pseudo-random number generation (used in Crude MC) the quasi-random
number sequences are introduced from where the method Quasi-MC took its name Quasi-MC is a
technique which approximates the probability of failure using low discrepancy sequences33
x1hellipxN instead of randomly generated sequences
If the domain of failure is the goal is to estimate the following
probability
int ( )
(F5)
The main principle of this method is the integral of the function on [ ] can be
approximated by using some low discrepancy sequence x1hellipxN as follows
int ( )
sum
(F6)
The low discrepancy sequence is generated in Δ according to Lebesgue measure and then may be
transformed to any measure micro thanks to the inverse CDF technique in order to approximate the
integral
int ( )
sum
(F7)
In Figure 4 a comparison is made between quasi-random sample and pseudo-random sample in
order to illustrate the difference in sampling for the two methods
33 The discrepancy of a sequence is low if the number of points in the sequence falling into an arbitrary set B is close to proportional to the measure of B as would happen on average (but not for particular samples) in the case of a uniform distribution
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 163
Figure F1 Example of Quasi random Sample (left) and Pseudo Random Sample (right) (source
Waarts 2000)
This method a priori provides an asymptotically higher convergence rate than traditional MC but
no general rule can guarantee a better efficiency of the Quasi-MC sampling than the classical MC
sampling The advantages of Quasi-MC sampling tend to disappear with the increase of the
number of dimensions of the input variables It is recommended to use the Quasi-MC technique
with high sampling sizes or with very low dimensionality (in OT this method is valid only if the
input random variables are independent) (OpenTURNS 15 2015)
Importance Simulation
The main idea of the Importance Sampling method is to replace the initial probability distribution
of the input variables by a more efficient one Efficient means that more events will be counted
in the failure domain and thus reduce the variance of the estimator of the probability of
exceeding a threshold In other words importance sampling starts with the observation that if we
are going to sample randomly we should distribute the points to put the most of them in the
region that contains information (ie the failure region) and to waste as few as possible
In this method a sampling density is used instead of the actual probability density function
For a MC simulation the probability of failure is calculated by
sum ( ( )) ( )
( )
(F8)
where ( ( )) ( )
( ( )) ( )
The problem is that prior knowledge is needed of the failure area which in most cases is not
available In Figure 5 examples of the different sampling methods are shown
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 164
(a) (b)
(c)
Figure F2 Comparison of Importance Sampling (a) MC (b) and Numerical integration (c) (source
Latin Hypercube Simulation
Latin Hypercube Sampling is a sampling method enabling to better cover the domain of variations
of the input variables thanks to a stratified sampling strategy (in OT this method is applicable in
the case of independent input variables) Stratified sampling tries to place the sampling points so
that more of them will be found in regions where the variance of the limit state function g(x) is
largest In Figure F3 a comparison is made between the LHS and MC in terms of their sampling
technique The sampling procedure according to Baecher and Christian (2003) is based on the
next steps
1 Divide the range of each variable into several intervals of equal probability
2 If there are D variables to be sampled each in k regions then for each variable the k
sampling points are placed in independent random order
3 K samplings are made by choosing the first randomized value for each variable the second
randomized value for each variable and so on until k sets of randomized variables have
been chosen
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 165
Figure F 3 Comparison between Latin Hypercube sampling and MC (source OpenTurns 15 2015a)
F3 First Order Second Moment (FOSM) Method The First-Order Second-Moment (FOSM) approach (Ang amp tang 1984) provides analytical
approximation for the mean and the standard deviation of a parameter of interest as a function of
the mean and standard deviations of the various input factors and their correlations
Consider Z to be a function of random variables x1 x2 hellip xn that is
(F9)
In general case are correlated with covariance matrix [C] ie [C]=[σ][R][σ] where [σ]
is a diagonal matrix of standard deviations and [R] is the (positive-definite and symmetric)
correlation matrix with diagonal elements Rii=1 and non-diagonal elements Rij=ρij (ρij is the
correlation coefficient between variables i and j) In scalar notation Cij= σiσjRij (Farrokh 2007)
The first order approximation to the mean variance and standard deviation of the Z function is
based on the first terms of a Taylor series expansion of Z The following first-order estimates of
mean and variance are obtained
[ ]
(F10)
[ ] (F11)
Where the vector b denotes evaluated at the mean values of ie
(F12)
If there is no correlation among the variables Eq 223 can be written as
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 166
sum
(F13)
Then the reliability index can be calculated in 6 steps (Baecher amp Christian 2003)
1 Identify all variables that affect the mechanism that is researched
2 Determine the best estimate (usually the mean value) of each variable E[ ] and use these
to calculate the best estimate of the function E[Z]
3 Estimate the uncertainty in each variable and in particular its variance Var[ ]
4 Perform sensitivity analyses by calculating the partial derivatives of Z with respect to each
of the uncertain variables or by approximating each derivative by the divided difference
5 Use the equation of Var[Z] to obtain the variance of the function Z
6 Calculate the reliability index
[ ]
(F14)
MSc Thesis A Rippi 167
MSc Thesis A Rippi ii
Preface
This thesis is the final challenge in the master Hydraulic Engineering at Delft University of
Technology The report ldquoStructural reliability analysis of a dike with a sheet pile wall Coupling
Reliability methods with Finite Elementsrdquo was completed at Deltares as a part of a larger
research project namely TO2 in collaboration with Toegepast Natuurwetenschappelijk
Onderzoek (TNO) I chose that subject first of all because it combines two things that I enjoyed
a lot at TU Delft probabilities and flood defences systems Secondly it was an opportunity for
me to get acquainted with FEM and geotechnics that I was always interested in
Different people have contributed to the successful completion of this thesis First of all I would
like to express my appreciation to my graduation committee for their guidance and particularly
my daily supervisor Timo Schweckendiek Together we had many fruitful and interesting
discussions on the subject that triggered and motivated me for keep searching Especially I
want to express my gratitude and my thanks to Ana Teixeira and Jonathan Nuttall employees of
Deltares as they were also some of the main contributors to this research They stood by me not
only as colleagues and thesis mentors but also as friends I would like also to thank the rest of
Deltares employees whom ensured a friendly and easy going working environment Last but not
least I want to thank my friends and especially Panagiotis Apostolidis and my family for their
love support and advice throughout all my studies
Katerina Rippi
Delft November 2015
MSc Thesis A Rippi iii
MSc Thesis A Rippi iv
Abstract
Some dike sections in the Netherlands failed to comply with the safety standards and one of the
most promising countermeasures is the construction of retaining walls inside the dike The
Dutch design codes for dikes with retaining walls rely on Finite Element Analysis (FEM) in
combination with partial safety factors However this can lead to conservative designs For this
reason in this research a reliability analysis is carried out with FEM calculations aiming to
demonstrate the feasibility of reliability analysis for such a soil-structure interaction problem
The case study concerns a (river) dike with an anchored sheet pile wall modelled in PLAXIS The
sensitivity and reliability analyses were enabled by coupling the uncertainty software package
OpenTURNS and PLAXIS through a Python interface The most relevant (ultimate) limit states
concern the anchor the sheet pile wall and global instability (soil body failure) The case was
used to investigate the applicability of the First Order Reliability Method (FORM) and
Directional Sampling (DS) to analysing these limit states Finally also the system reliability was
evaluated using sampling-based methods (DS)
Due to the considerable number of random variables before starting the reliability analysis a
sensitivity analysis was conducted for each limit state This indicated the most important soil
layers to be accounted for and the variables to be considered as stochastic The sensitivity
analysis and later on the reliability analysis were based on analytical formulations of the limit
state functions The anchor and the sheet pile limit states were formulated in terms of their
yield stress for global instability loss of equilibrium in the FEM analysis was used to define
failure Moreover the systemrsquos reliability was evaluated by taking into account all the three limit
states that were mentioned previously
The goal is to implement the coupling between FEM and reliability methods in order to analyse
the components of such a system (ie anchor sheet pile wall and dikersquos soil body) estimate the
probability of failure and identify the most important soil properties that affect the behaviour of
each component and the system as a whole The results of this research can be used to assess
and optimize the current design procedure for dikes with retaining walls
MSc Thesis A Rippi v
MSc Thesis A Rippi vi
Table of Contents
Preface ii
Abstract iv
List of abbreviations x
1 Introduction 1
11 Project objective and main research questions 2
12 Research approach and outline 3
2 System description and current design concept 7
21 System description and forces configuration 7
22 Current design concept 12
23 Safety standards 17
3 Literature study 21
31 Background 21
311 Finite Element Modeling 21
312 Uncertainties and Sensitivity analysis 24
32 Previous Studies 29
33 Overview 34
4 Structural Reliability Analysis 35
41 Basics of Reliability Analysis 35
42 Overview of Reliability Analysis Methods 36
421 Level III Methods 36
422 Level II Methods 40
423 Level I Methods (semi-probabilistic) 42
424 Response Surface Techniques (RS) 44
43 Coupling Reliability Analysis with FEM 46
431 The functionality and possibilities of OT 46
432 Coupling OpenTURNS-Plaxis 47
44 Overview 49
5 Failure Mechanisms and Limit State Functions 53
51 Introduction to the system analysis and the limit states 53
52 Limit State Functions 55
521 Serviceability Limit State 55
522 ULS for Structural Members 57
MSc Thesis A Rippi vii
523 ULS for Soil Failure 60
53 Overview 66
6 Case Study-Dike with an anchored sheet pile wall 69
61 Case Description 69
62 Soil Parameters 70
63 Finite Element Model 73
64 Deterministic Analysis 74
641 Calculation Scheme and Design Values 75
642 Construction Stages 77
643 Determination of the structural elementsrsquo characteristics 78
65 Overview 81
7 Reliability analysis results with stochastic soil properties 85
71 Method description 85
72 Mean values calculations 88
73 Sensitivity Analysis Results 91
74 Soil Shear Failure 98
75 Anchor Failure 102
76 Sheet pile wall failure 105
8 Conclusions and Recommendations 121
81 Conclusions 121
82 Recommendations 123
References 125
Appendix A 129
OpenTURNS features 129
A1 Fourier Amplitude Sensitivity Test (FAST) 129
A2 Optimization Algorithms in FORM 132
Principles of optimization algorithms 132
Convergence criteria 134
Evaluation of the algorithms performance 135
A3 Distribution Types 143
Uniform Distribution 143
Normal Distribution 144
Lognormal Distribution 144
Truncated Normal Distribution 145
Appendix B 147
MSc Thesis A Rippi viii
Plaxis 2D (2015) features 147
B1 Mohr Coulomb failure criterion 147
B2 φ-c Reduction Technique 150
B3 Initial Stress Generation 151
B4 Interface Strength 151
Appendix C NEN 6740 - Table 1 153
Appendix D 155
Input Files for the Reliability Analysis 155
Appendix E 159
Characteristic and mean values 159
Appendix F 161
Reliability methods 161
F1 Generation of random samples in Monte Carlo 161
F2 Other Sampling Methods 162
F3 First Order Second Moment (FOSM) Method 165
MSc Thesis A Rippi ix
MSc Thesis A Rippi x
List of abbreviations
CDF Cumulative Distribution Function
CoV Coefficient of Variation
CUR Civieltechnisch Centrum Uitvoering Research en Regelgeving
DS Directional Sampling
FAST Fourier Amplitude Sensitivity Analysis
FEA Finite Element Analysis
FEM Finite Element Model
FERM Finite Element Reliability Method
FORM First Order Reliability Method
FOSM First Order Second Moment
LEM Limit Equilibrium Method
LRFD Load and Resistance Factor Design
LSF Limit State Function
MC Monte Carlo
OT OpenTURNS
PDF Probability Distribution Function
RFEM Random Finite Element Method
RS Response Surface
SA Sensitivity Analysis
SLS Serviceability Limit State
SORM Second Order Reliability Method
ULS Ultimate Limit State
VNK Veiligheid Nederland in Kaart
MSc Thesis A Rippi xi
The roots of education are bitter but the fruit is sweet
Aristotle
MSc Thesis A Rippi 1
1 Introduction
In the Netherlands according to the Flood Protection Program (Hoogwater
Beschermingsprogramma1) and the Room for the River program (Ruimte voor de Rivier2)
alternative structural techniques for the reinforcement of existing dikes or for future dike
constructions additional to the conventional ones have been introduced and tend to be
attractive The heightening of the dike crest and the construction of a stability berm are some of
the most common current actions for dike strengthening Additionally filter layers geotextiles
and drainage systems can be applied in advance in order to prevent several failure mechanisms
Recently cantilever or anchored sheet pile walls and diaphragm walls tend to be used as an
alternative procedure for embankment reinforcing (see Figure 11) Such an alternative can be
chosen for an improvement of the macro stability of the dikersquos inner slope and additionally for
saving space in the land area that needs to be habited
Figure 11 Design options for dike reinforcement (source Flood Defences 2015)
For that purpose engineers need a concrete and unequivocal design methodology for such
combined structures In the meantime Deltares in cooperation with the Water Board of
Rivierenland is working on a design guideline which will be applicable to design the so-called
Type II stability fences3 A draft of such a guideline is elaborated in the report of Larsen et al
(2013) In this report suggestions and recommendations are outlined for the design of such
structures with Finite Element Modeling (FEM) and partial factors Remarks are also made for
the advantages and the limitations of such models and how they should be handled so that
reasonable and trustworthy results can be retrieved from the analysis
1 httpwwwhoogwaterbeschermingsprogrammanldefaultaspx 2 httpalbertawatercomhow-is-water-governedwhat-is-room-for-the-river 3 Type II structures are structures which in combination with a soil construction fulfil water retaining functionalities
Introduction
MSc Thesis A Rippi 1
The concept behind the recommended design criteria (Larsen et al 2013) is the definition and
the evaluation of several partial factors with the view to determine overall safety factors These
factors are further discussed in section 22 However validation of these factors shall be carried
out before being used in any case otherwise either the safety of the structure is jeopardized or
the structure will be overdesigned and cost inefficient
For improving the design criteria researchers for a long time focused on enhancing structural
models (beams shells etc) and constitutive laws (elasticity plasticity damage theories etc)
With the development of computer science a great amount of work has been devoted to
numerically evaluate approximated solutions of the boundary value problems describing the
mechanical system FEM is probably nowadays the most applied approach for the solution of
these problems
However the increasing level of detail of the constitutive models and the constant enhancement
of the computational tools do not solve the problem of identification of the model parameters
and the inherent physical and modelling uncertainties Moreover in most civil engineering
applications the intrinsic randomness of materials (soil rock concrete etc) or loads (water
elevation wind earthquake motion etc) is such that deterministic models are using average or
later on characteristic values of the properties at best lead to rough representations of the
reality
As a counteraction a semi-probabilistic methodology has been developed that was based on the
application of characteristic and design values by using partial factors Current design codes
such as Eurocode provide target reliabilities for different types of structures and structural
elements according to the potential consequences of failure However these partial factors are
not always equally suitable and efficient for all types of structural applications since they have
been calibrated under specific conditions Besides the consequences of failure of flood defences
such as dikes can be comparable to the investments in increasing the reliability of such systems
are For that reason it would be advisable for these systems to define target reliabilities based
on a risk assessment (ie tailor-made solution) rather than using the standard partial factors
coming from general geotechnical design codes which may be either too low or too high for a
given flood defence system
One step of such a risk assessment is accounting for randomness and spatial variability of the
mechanical properties of materials is one of the tasks of stochastic or probabilistic mechanics
which has developed fast in the last decade In this master thesis project the uncertainty of soil
properties is going to be treated in terms of its contribution to failure For that purpose
probabilistic methods are going to be implemented for a dike with an anchored sheet pile wall
(see Figure 12) simulated in a FEM software that is specialized in soil mechanics The successful
implementation and in future research the verification of such methods can be considered as
the most preferable and cost efficient way to design structures with high safety requirements
and not only for the validation of the partial factors Of course such a procedure tends to be
time consuming However the gradual improvement of the current probabilistic methods in
combination with the state-of-the-art computer capabilities as well as the scientific knowledge
gained in terms of different systems behaviour and failure modes can introduce a more
optimized way of designing structures with considerable investments
Introduction
MSc Thesis A Rippi 2
Figure 12 Reinforced dike section with an anchored sheet pile wall
11 Project objective and main research questions At the beginning of this research it has been observed that the design procedure being followed
for dikes with retaining walls modeled in FEM can lead to an overestimation of the critical loads
and thus to a potential cost inefficient final structure An example of the magnitude of the design
values in such a case study is given in the next chapter where also a more detailed explanation
of the current design concept is described
The objective of this study is to implement a full probabilistic analysis for evaluating the
reliability of a dike reinforced with an anchored sheet pile wall Such a research analysis can
subsequently come up with valuable recommendations for the improvement of the present
design approach Soilrsquos and structural elementsrsquo behavior should be taken into account both
separately and as a system Some of the reliability methods can deal with system reliability
problems such as Monte Carlo (MC) and Directional Sampling (DS) whereas for others like
FORM and SORM additional methods should be applied that use reliability information for each
individual limit state function to obtain the systemrsquos reliability
In principle the probability of failure of different limit states is to be computed individually
while in the sequence the system reliability is going to be estimated As far as the soil failure
mechanisms are concerned this research will focus mainly on the global stability of the dike
slope while for the sheet pile wall and the anchorrsquos failure the exceedance of the steel yield
stress is going to be considered
The scope of this thesis is not only to test different reliability models robustnessrsquo in conjunction
with FEM simulations but also to get a better insight into the specific system behavior (ie of a
dike with a retaining wall) analyze each component separately and investigate its response
under certain load conditions Furthermore the minimization of the computational effort and
time could also be carried out meaningfully under the constraint of sufficient accuracy The
accuracy should be such that the probability of failure is acceptable for the ultimate limit state
(ULS) in a normal design process and according to the current safety standards Finally the
Sheet pile
wall
Anchor
Dike section
Introduction
MSc Thesis A Rippi 3
robustness of the coupling between the reliability model and the FEM will be tested via their
capability of adapting to new input parameters without encountering convergence errors
during execution
The main research question of this master thesis project is thus formed as follows
How can the probability of failure of a dike with a sheet pile wall due to global instability modeled
by a Finite Element Model be analyzed
Essential questions regarding the soil models and failure criteria as well as the reliability
methods and the systems behavior are generated which are listed below
Subquestion 1 Which reliability methods are computationally tractable in
combination with FEM
Subquestion 2 How robust (convergence) are the tractable methods
Subquestion 3 Which limit states are relevant for the envisaged application of
retaining walls in dikes and how can they be formulated using FEM
analysis outcomes
Subquestion 4 What is the contribution of different uncertainties in the failure
mechanisms of the system
Subquestion 5 Can response surface techniques help to increase the efficiency and
robustness of the reliability model
Subquestion 6 How can the current design approach for dikes with sheet piles be
improved
In the next chapters the above research question and the related subquestions are going to be
answered by following the methodology that is described in the next section
12 Research approach and outline In this section an overview is given of how the aforementioned objective and sub-questions are
approached The thesis is mainly divided into 8 chapters and 6 appendices In Figure 13 the
thesis outline is illustrated and an indication of which subquestions (Sq1-6) are treated to
which chapter is given In Chapter 1 an introduction into this research content is made and a
first illustration of the system under investigation is presented together with the main research
questions
Chapter 2 has been devoted to the description of the system (ie a dike with an anchored sheet
pile wall) and to the case study that has been adopted for being analyzed in terms of its
reliability The forces configuration is also illustrated and the results of the case study according
to the current design concept are presented and evaluated Last but not least the new
recommended safety standards are included and the current required reliability of the specific
dike section is presented
Introduction
MSc Thesis A Rippi 4
In Chapter 3 a literature study related to FEM uncertainties and sensitivity analysis is
presented Moreover previous studies that are associated with the objective of this research are
mentioned whereas some of their results were also taken into account for proceeding with this
research
The main scope of this thesis is the implementation of different reliability methods on a specific
case study with the view to investigate and analyze its behavior Some of these methods are
continuously mentioned through the test and thus in Chapter 4 an overview is given of the
most well-known reliability methods Eventually an evaluation of these methods is made based
on literature and preliminary testing with simple case studies and the procedure of their
coupling with FEM is discussed The evaluation helped to qualitatively answer subquestion 1
while also an introduction of how special reliability methods such Response Surfaces were used
in this thesis that partially covers subquestion 5 In this chapter an introduction into the Limit
State Functions (LSF) concept is made and how they are considered for the reliability analysis of
a system An LSF actually represents a failure mode that can be detected in a structural or soil
element and it is expressed as a function of several variables In Chapter 5 the failure
mechanisms and the corresponding LSF that are related to the specific case study are identified
and formulated which answers subquestion 3
In Chapter 6 the case study whose reliability is to be evaluated is presented as it was modeled
in FEM The boundary conditions are specified and the soil and structural properties are
indicated and illustrated Moreover a deterministic design is also taking place in order to
roughly estimate the structural elementsrsquo properties This will help to make a qualitative
comparison between the current design procedure and the design according to a fully
probabilistic approach by referring thus to subquestion 6
Eventually the results of the aforementioned analysis are presented in Chapter 7 The results
mainly include the estimated probability of failure of each of the system components as well as
of the system as a whole by considering the soil properties as stochastic Moreover an
interpretation of the failure points is made and an assessment of the level of impact of the
random variables on the systemrsquos behavior is carried out At that chapter subquestions 2 4 and
5 are mainly treated
Last but not least in Chapter 8 the general conclusions are presented together with some
valuable recommendations for future consideration and research Furthermore a reflection on
the methodology and how the different research questions were approached is made
Introduction
MSc Thesis A Rippi 5
Figure 13 Thesis outline
Introduction
MSc Thesis A Rippi 6
MSc Thesis A Rippi 7
2 System description and current design concept
In this section the system to be analyzed is described more in detail and the individual
components are identified Moreover the forcesrsquo configuration is explained and illustrated as
they would have been calculated with the conventional way in parallel with a qualitative
evaluation of the potential expected deformations Finally the current design concept and the
safety standards that are referred to such a system are introduced
21 System description and forces configuration The system to be investigated in this thesis consists of a dike section reinforced with a one-layer
anchored sheet pile wall The system has been simulated by FEM and later on it was coupled
with a reliability package for carrying out a reliability analysis In the figure below an
illustration of the system is depicted and the several elements of an anchored sheet pile wall are
showed
Figure 21 System layout and different components of the anchored sheet pile wall
In particular an anchored sheet pile wall consists of the sheet piles which are anchored in the
soil via a tie rod The rod is grouted most of the times in a sandy layer with a cemented grout
body and it both alleviates the sheet piles from the axial forces exerted by the upper structure
and keeps the wall stable in case of excessive developed moments due to the lateral earth
pressure Last but not least a waling system is applied mainly downstream of the sheet piles in
order to transfer the loads from the piles to the anchors in such a manner so as to avoid
excessive local stresses on the intersection between the sheet piles and the tie rod
Cross-Section
Top View
System description and current design concept
MSc Thesis A Rippi 8
The main load configuration acting on a dike section as well as a sheet pile wall is depicted in
Figures 22-24 together with the possible displacements As far as the soil body stability is
concerned most of the slope stability analysis computer programs are based on the limit
equilibrium concept according which a soil mass tends to slide down under the influence of
gravity The transitional or rotational movement is considered on an assumed or known slip
surface below the soil There an equilibrium should be achieved among the driving and the
resisting forces In that case the driving moments consist of the soil weight the water pressures
and the loads around the center of the slip surface such as a possible vertical load on the top of
the dike crest (ie traffic load) The magnitude of the water pressures is controlled by the water
elevation on the river side
In Figure 22(a) the external and internal forces acting on a slice of the slip surface are
illustrated The driving forces are the soil weight W the water pressure U and any additional
load that can contribute to the rotation of the slip surface which in turn activate the lateral
active earth pressure The resisting forces consist of the lateral passive earth pressure
(Rankine 1857) and the shear strength (Terzaghi 1943) Their expressions are written as
follows
[kN] (21)
[kN] (22)
[kPa] (23)
where is the active lateral earth pressure coefficient and the passive lateral earth
pressure [kNm3] is the unit weight of the corresponding soil layer and H [m] is the thickness
of the soil layer that the lateral earth pressure comes from (in case of multiple soil layers there
should be a distinction among the different forces) and [kPa] is the effective cohesion of the
soil layer The safety factor that is written in the Figure 22(a) as FS is defined by the ratio of the
shear strength (excluding the pore water pressure) divided by the weight of the earth body
(including the pore water pressure)
The loss of equilibrium between the driving and the resisting moments lead to the rotation and
instability of a slip surface as it is illustrated in Figure 22(b) A slide surface can take various
shapes according to the Limit Equilibrium Method (LEM) that is followed According to Bishop
method this surface tends to be circular in Uplift-Van method a horizontally compressed zone
can be also considered whereas in Spencer method the shape of the slide body can be arbitrary
The slip surface can be located either in the landside or in the river side of the dike depending
on the load the soil characteristics and the design of the dike (ie inner or outer berm
reinforcing revetment on the outer slope etc) Most of the times the landside area of the dike is
jeopardized due to different failure mechanisms that are described in section 523 The
instability of the outer slope is mainly relevant in case of sea dikes and a potential damage can
be usually reconstructed until the next flooding
System description and current design concept
MSc Thesis A Rippi 9
(a)
(b) Figure 22 (a) External and internal forces acting on a slip surface of a dike section and (b) possible deformation pattern
As a counteraction for the slope instability a sheet pile wall can be installed inside the dike The
special thing about these structures is that they make possible a greater freedom in form and
functionality than a traditional dike design This structure derives its strength from the
materials used such as steel which are able to withstand higher pressures than clay for instance
The general stability is due to friction and wedging in the bottom
After the reinforcement of the inner side of the dike with a retaining wall the strength capacity
of the wall is also important for the global stability of the system The forces to be taken into
account for the sheet pile wall stability are the active and the passive earth pressures (effective
pressures) and the water pressure as they are illustrated in Figure 23(a) In that figure the
System description and current design concept
MSc Thesis A Rippi 10
forces were simplified in a singular triangular shape as in a homogeneous soil body in order to
explain and depict the overall picture of the acting forces However the stresses distribution
over depth can be more complex depending on the variety of the soil layers that are present In
Figure 23(a) it can be noticed that a sheet pile wall which is installed close to the inner berm
might not have significant instability issues as there is the passive side that contributes to
resisting forces However in the passive side the soil can differ and be weaker than this of the
active side Therefore the passive force in that case might not be very supportive and thus the
wall shall be designed cautiously
Additionally the anchor resistance shall be reassured for the sake of the stability of both the
sheet pile wall and the overlying soil body In Figure 23(a) the forces acting on the anchorage
are displayed The distributed load qs comes from the earth pressure of the upper soil layer (it
might not be uniformly distributed as it is depicted in the figure but more in a trapezoidally type
of load) whereas the tension force Ft4 stems both from the earth pressure and the displacement
of the sheet pile wall The bending and the tension capacity of the anchor are determinant for its
stability In Figure 23(b) a possible displacement pattern of the retaining wall and the
anchorage is illustrated
Furthermore corrosion is an additional weakening impact on the wall that depends on the
water level and the pore water pressures near the structure The thickness of the wall and the
material properties are playing a key role to the resistance towards corrosion This will not be a
subject of this thesis however it should be taken into account in case of design purposes
(a)
4 where O is the average circumference of the pile P [kPa] is the shaft friction at depth z and L is the length of the pile
System description and current design concept
MSc Thesis A Rippi 11
(b) Figure 23 (a) Mobilizing forces acting on the sheet pile wall and the anchorage and (b) possible deformation pattern
It should be mentioned that as far as the pore pressures are concerned in Figure 23(a) only the
hydrostatic pore pressures are depicted Generally the pore pressures (or active pore pressures)
are the sum of the steady state pore pressures and the excess pore pressures
[kNm2] (24)
Steady state pore pressures can be hydrostatic (based on a horizontal phreatic line) or non-
hydrostatic (based on steady state groundwater flow) Excess pore pressures are based on
loading of undrained soils In FEM these various pore pressures are taken into account
automatically In Figure 24 an example of a potential distribution of the active pressures
nearby the sheet pile wall is depicted as it was deduced from Plaxis Therefore as it can be
noticed from the figure the distribution can indeed sometimes approximated as triangular
System description and current design concept
MSc Thesis A Rippi 12
Figure 24 Example distribution of the active pore pressures acting nearby the sheet pile wall
The knowledge of the type of forces that are exerted on both the structural components and the
soil body as well as of the most expected form of displacements is valuable in order for
someone to be able to evaluate the results deduced from FEM and detect potential modelling
errors Therefore this validation was necessary before starting with the reliability analysis
In the next sections the current design approach of the system described above is elaborated
Additionally the Dutch recommended safety standards are presented for this type of structures
which show the need of carrying out a reliability analysis For that purpose a case study was
adopted that it is presented first as it was designed according to the current regulations while
finally a comparison is made between the original and the new case study as it was found to be
according to the reliability analysis
22 Current design concept In the Netherlands dike reinforcement is becoming a major issue concerning the design of flood
defenses Special reinforcing structures such as cofferdams anchored sheet piles and diaphragm
walls are used for strengthening the dike Especially in the context of the Flood Protection
Program and the Room for the River program many dike reinforcement projects have been
suggested
In the report of Larsen et al (2013) the design procedure of a dike reinforced with sheet pile
wall is prescribed using FEM The safety philosophy that is followed in this technical report
actually composes the current procedure for designing dikes with sheet pile walls and it is
linked with the usage of partial safety factors Particularly the required overall safety factor
FEM that should be compared with the one from FEM calculations is determined as follows
SFEM
b d m n
(25)
where
System description and current design concept
MSc Thesis A Rippi 13
Partial safety factor indicating the uncertainties of the soil composition and
the water pressures (also called schematization factor)
Partial safety factor which is related to the calculation model and the way the
calculations have been carried out (also called model factor)
Partial safety factor which is related to the material parameters (also called
material factor)
Partial safety factor associated with damage caused during the soil tests (also
called loss factor)
Safety factor of load
Each of these partial factors shall be at first calibrated by conducting a full probabilistic analysis
of the system under consideration In most of the cases these partial factors have been already
calibrated on previous similar projects and then they are reused for any similar case In
Schweckendiek et al (2013) a new approach of using partial factors for flood defences is
proposed whose application is still under consideration At the moment the aforementioned
partial factors for dike reinforced with sheet pile walls modelled in FEM are estimated
according to mostly engineering judgement and political negotiations
In sequence the FEM design procedure that is applied for sheet pile walls inside soil structures
is based on a so-called ldquosafety calculationrdquo (for information for this type of calculation see
Appendix B2) In this calculation an artificial reduction to the soil strength parameters (friction
angle and cohesion) is applied in order to evaluate the overall safety factor and confirm if this is
lower or higher than the required one that was estimated according to Eq 25 Furthermore the
resulting moments and forces developed on the structural elements during the specific
calculation are used for their design
This is a quite helpful and indicative calculation type available in Plaxis that can give a sense of
the structurersquos safety factor and the possible failure mechanisms that can occur under the
predefined load configuration However such a method of calculating the developed stresses
can be misleading in terms of the moments and forces acting on the structural elements and the
total deformations of the system
In this thesis the reliability of a dike with an anchored sheet pile wall is going to be evaluated
and in particular the case study of Breedeveld (2011) is to be used for that purpose In Figure
26 the location of the existing dike section is presented while in Figure 25 the structure as it
was modelled in FEM is illustrated It is essential to mention that the retaining wall does not
exist in reality but it was placed so as to implement and demonstrate the current design
regulations in the report of Breedeveld (2011) The anchorrsquos angle has been taken equal to 30deg
from the vertical which is generally a steep anchor inclination That is expected to reduce the
part of the horizontal force that is undertaken by the anchor and thus the most of it exerts on
the sheet pile wall The angle is usually taken 30deg-45deg from the horizontal axis (CUR 2005) but
this is of course depended on the load conditions and the construction requirements for each
System description and current design concept
MSc Thesis A Rippi 14
case The influence of the anchor orientation and the construction methods that are generally
followed for an anchored sheet pile wall are not considered in this thesis The case study was
taken as it was modelled in the aforementioned project and the reader should be aware that this
serves the proceedings of a reliability analysis and not of a perfectly constructed retaining wall
Figure 25 Original dike section in Varik-Heesselt as it is presented in Breedeveld (2011)
In particular based on this case study and trying to understand the design criteria a
comparison was carried out between the ldquosafety calculationrdquo and a normal ldquoplastic calculationrdquo
(ie elastoplastic drained or undrained analysis) In both cases the boundary conditions the
external loads and the input parameters are the same The difference between them is the
calculation procedure followed in order to come up with final stresses which in the case of the
ldquoplastic calculationrdquo does not include a reduction of the soil strength properties In Figures 27
and 28 the results from both calculations are presented In Figure 27 the total deformations
are exhibited while in Figure 28 the moments and the axial forces on the sheet pile are shown
Figure 26 Location of the Varik-Heeselt dike section (section number TG-078) in the Dike ring 43
BetuweTieler-en Culemborgerwaarden west (source Breedeveld 2011)
System description and current design concept
MSc Thesis A Rippi 15
As it is realised from this figure the discrepancy between the two results is quite high while as
far as the deformations are concerned no realistic conclusions can be drawn from a ldquosafety
calculationrdquo Moreover taking as design moments the moments that are deduced from the
ldquosafety calculationrdquo will lead to a much heavier and more expensive sheet pile cross section than
the ldquoplastic calculationrdquo
It is also essential to mention that the required safety factor was estimated up to FEM =18
according to the partial-factor procedure that is defined in Eq 25 Such a safety factor as
prerequisite for coming up with the design values of the moments and forces of the sheet pile
and the procedure of the ldquosafety calculationrdquo in FEM can lead to a conservative and financially
inaccessible design
(a) ldquoPlastic calculationrdquo Deformed mesh and maximum deformations (umax= 008 m)
(b) ldquoSafety calculationrdquo Deformed mesh and maximum deformations (umax= 9041103 m)
Figure 27 Deformed mesh and maximum deformations in the end of (a) plastic calculation and
(b) safety calculation
30deg
System description and current design concept
MSc Thesis A Rippi 16
(a) ldquoSafety
calculationrdquo
Mmax = 9392 kNmm Nmax= -4402 kNm
(b) ldquoPlastic
calculationrdquo
Mmax = 6833 kNmm Nmax= -1372 kNm
Figure 28 Bending moments and axial forces exerted on the sheet pile during the (a) safety
calculation and (b) plastic calculation
On the other hand the aforementioned design procedure guaranties a strong structure capable
of probably undertaking more than the expected loads and thus ensuring the safety against
flooding of the landside However the knowledge on fully probabilistic methods that has been
obtained as well as the development of the technology can introduce a reliability analysis of the
system as a preferable way for the future design concept Such a procedure should be
accompanied also by field test that would be able to validate FEM results as well as in-situ
measurements and inspections of the soil properties for calibrating their statistical
characteristics
Due to the conservative results that the aforementioned procedure came up with and because of
the simplifications that this case study was later subjected to in order to be used in this thesis a
new design was carried out in Chapter 6 and the properties of the structural elements were
redefined
The inclination towards probabilistic methods has not only been created due to the incomplete
design regulations for the system under consideration but also the current proposed safety
standards which are discussed in the section below recommends a risk based safety assessment
of the primary flood defenses which in principle implies the evaluation of the failure probability
In the next section an overview of the new recommended safety standards related to primary
flood defenses is given and the required reliability for the stability of the aforementioned case
study is elaborated
System description and current design concept
MSc Thesis A Rippi 17
23 Safety standards An important factor for the evaluation of a structurersquos reliability is the safety standards that
have been set and according which the acceptance or not of the failure probability is made After
the safety assessment5 of 2011 it has been declared that 33 (1225 km) of the primary flood
defenses (3767 km) in the Netherlands does not comply with the safety standards For this
purpose a new session of investigations has commenced in order to assess the reliability of the
existing dikes and the potential amendment of the safety standards for future constructions
Taking into account this latest information and within the framework of the Flood Risk in the
Netherlands 2 project6 (VNK2) a new Delta program was founded whose one of the main
decisions was the ldquoDelta Safety Decisionrdquo This decision contains a proposal for new safety
standards of the primary dikes as they are depicted in Figure 29 which are stricter than the
previous ones and formulated in terms of failure probability
According to this project a new policy regarding the safety assessment of the flood defenses has
been settled and proposed for the future design criteria which is based on the evaluation of the
acceptable flood risk (probability of failure (x) consequences) of each dike ring rather than the
probability of exceedance in order to achieve a level of protection that is in balance with the
societal value (Cost Benefit Analysis and life loss)
VNK has been already using reliability evaluation techniques in order to calibrate partial safety
factors Especially in geotechnical engineering the high inherent uncertainty of the soil
properties renders the evaluation of the structural reliability essential for the safety assessment
of the structure Subsequently this implies the investigation of the response of the different
reliability methods applied in a real case study of a dike The techniques that are used to
evaluate the probability of failure are discussed thoroughly in Chapter 4 and in Appendix F In
case of a complicated structure such as a dike with a retaining wall both the reliability of the
dike body and the reliability of the wall shall be evaluated and eventually a common standard
for the reliability of the system shall be defined
5 A safety assessment is an evaluation of the function of the flood protection structure and it is being conducted every approximately 6 years after the reassessment of the existing hydraulic boundary conditions 6 in Dutch Veiligheid Nederland in Kaart 2 or VNK2 It has been initiated by the Ministry of Infrastructure and the Environment (IampM) the Association of Regional Water Authorities (UvW) and the Association of Provincial Authorities (IPO)
System description and current design concept
MSc Thesis A Rippi 18
Figure 29 Maximum admissible flooding probability for primary defences according to Delta
programme 2015 (Deltaprogramma 2014)
As far as the retaining walls are concerned CUR (2005) design procedure distinguishes the
following three safety classes for retaining walls with corresponding reliability indexes
Table 21 Safety classes and corresponding reliability indices (CUR 2005)
Class I Relatively simple constructions no personal safety risks and relatively minor
damage in the case of overall failure ϐ=25
Class II Considerable damage in the case of overall failure minor personal safety
risks ϐ=34
Class III Major damage in the case of overall failure andor considerable personal
safety risks ϐ=42
System description and current design concept
MSc Thesis A Rippi 19
Additionally as far as a dike section is concerned the required overall reliability index shall be
concluded according to the new norms that are shown in Figure 29 However the reliability
index of the different failure mechanisms of a dike shall be estimated separately
According to Vergouwe et al (2014) that is part of the VNK2 report the major failure
mechanisms in dike ring 43 where the case study under investigation is located (see Figure 26)
is uplift and piping and macro-instability of the inner slope Especially in Figure 210 the
estimated percentage of the contribution of each failure mechanism to the overall probability of
flooding7 is indicated In Vergouwe et al (2014) the overall flooding probability of the specific
dike section that the case study concerns was evaluated to 17800 Therefore the macro-
instability on which this thesis is concentrated with a contribution percentage of 226 has a
probability of about 310-5 (
) and thus a target reliability index β of
approximately 4 This reliability level can be considered as a benchmark for evaluating the one
that will be entailed in the end of this thesis
Figure 210 Percentage contribution of the different failure mechanisms to the probability of flooding (source Vergouwe et al 2014)
It should be mentioned though that the aforementioned safety standards concerning the dike
ring 43 were defined according to the probability of exceedance of a certain water level that the
dike has to retain Therefore this complies with the old norms and not with those appeared in
Figure 29 However for large engineered systems such as flood defence systems with large
potential consequences and substantial investments it is worthwhile to assign target reliability
levels based on a risk assessment of the area surrounded by the specific dike ring There are
three widely used types of criteria for evaluating the risks related to floods and major industrial
hazards (Vrijling et al 2011)
Economic criteria
Individual Risk Criteria
Societal Risk criteria
7 The dependencies among the failure mechanisms were not taken into account Therefore this probability stated in Vergouwe et al (2014) is not exactly equal to the probability of flooding
System description and current design concept
MSc Thesis A Rippi 20
In order to check then if a certain system abides by the target reliability that is required the
failure probability of each failure mechanism shall be calculated based on a LSF The
aggregation of the failure probabilities of all the modes gives the failure probability of the
system which is supposed to be compared with the required in order to reassure the reliability
of the structure Therefore this thesis is dealing with the calculation of this failure probability of
a certain system modelled in FEM analysis and understanding of the systemrsquos behaviour and
finally the investigation of the robustness of coupling FEM with reliability methods for
evaluating the reliability of a structure In the next chapter an overview over the research has
been done until now concerning coupling and reliability methods applications is presented and
a discussion over the most important is made
MSc Thesis A Rippi 21
3 Literature study
In this chapter an overview of some principles related to FEM and the uncertainties in
geotechnical engineering is presented More precisely an introduction into FEM concept and a
discussion over the different types of FEM is made Moreover the uncertainties in geotechnical
engineering are stressed and the general framework according to which they are handled is
described Finally previous studies that are associated with the application of FEM on soil
structures and the reliability analysis are discussed
31 Background
311 Finite Element Modeling
FEM is a numerical method whose essence is to convert a problem described by partial
differential equations over space and time into one by dividing the space-time continuum into a
set of discrete elements and assuming that the unknowns vary over each element FEM solution
process is as follows
1 Divide structure into pieces (elements with nodes) (discretizationmeshing)
2 Connect the elements at the nodes to form an approximate system of equations for the
whole structure (forming element matrices)
3 Solve the system of equations involving unknown quantities at the nodes (eg
displacements)
4 Calculate desired quantities (ie strains and stresses) at selected elements
The properties of each element are set the same as the material properties that have been
defined by the user Then the Deterministic Finite Element Method (the finite element method
that was formulated with deterministic variables) can be used in conjunction with means and
standard deviations of the input variables to obtain reliability estimates
In the simple coupling of FEM with reliability analysis each parameter that is considered as
stochastic is given a particular probability density function estimated either by field tests or by
engineering judgement This approach is also called ldquosingle random variablerdquo and it assumes
that the spatial correlation length is infinite In other words the soil is considered to be
homogeneous and the stochastic property assigned to the soil is taken at random from a certain
probability distribution
Nevertheless a more realistic model should take into account the spatial correlation within
smaller regions where the stochastic property is allowed to vary For that purpose the Random
Finite Element Method (RFEM) was introduced by Fenton and Vanmarcke (1990) in which the
random variables are correlated to one another using auto-correlation functions
The most sophisticated method of FEM is the Stochastic Finite Element Method introduced by
Ghanem and Spanos (1991) which accounts for the uncertainties in the input parameters
Literature study
MSc Thesis A Rippi 22
implicitly within the finite element calculation This aims at representing the complete response
probability distribution function (PDF) in an intrinsic way Two steps are basically applied for
that purpose
Discretization of the problem and definition of the random vector with an unknown
joint PDF and
Expansion of the response in a particular basis of random vectors with a finite variance
called the polynomial chaos
In this thesis the Deterministic Finite Element Method is to be used coupled with reliability
analysis According to Waarts (2000) coupling FEM with reliability methods (FERM) would lead
to the following advantages
In comparison to standard finite element analysis it gives direct insight into the
structural reliability and decisive parameters
Structures designed using FERM will either be safer orand more economically built in
comparison to structures designed using safety factors and classical constitutive models
FERM is expected to be useful for the determination of LSF that cannot be explicitly formulated
and that differ in each case such as soil limit state in different structural schematizations
Moreover it is likely to be valuable in areas where little knowledge exists on the systems
reliability of structures with multiple components (ie a dike with a sheet pile wall)
FEM in practice
The conventional method for stability analysis in a soil body is represented by LEM although
FEM is increasingly used by designersresearchers The latter has been proved to be quite
realistic for the progressive behaviour (ie stress-strain development in different construction
phases) of a soil system under the effect of stress redistribution in comparison with classical
models Especially in their master thesis Johansson amp Sandeman (2014) compared the
deformations and the forces measured at a deep excavation supported by anchored sheet pile
wall in a railway tunnel located in Gothenburg in Sweden with 1D finite element software the
2D finite element software (Plaxis) and hand calculations They proved that Plaxis produces
reliable results for horizontal deformations in the sheet pile wall and anchor forces when
compared to in-situ measurements
Moreover in Gonzaacutelez et al (2013) it was proved that the simplification of reality done by the
classical methods such as Blumrsquos Engelrsquos Kreyrsquos methods etc (for further information for these
methods a reference is made to Gonzaacutelez et al (2013)) allows us to generally understand the
behaviour of the system wall-soil Nonetheless the results that came out of this analysis were
found to be quite conservative whereas FEM managed to give a more realistic interpretation of
the wallrsquos movement
Seed et al (2006) investigated the performance of the New Orleans Flood Protection Systems
during hurricane Katrina in August 2005 by simulating the different levee sections with FEM As
an example in Figure 31 pictures of the failure of the Lower Ninth ward IHNC East Bank in
New Orleans after the hurricane Katrina are presented FEM was proved to be an efficient tool
to realistically interpret the shape and the triggers of the failure mechanism
Literature study
MSc Thesis A Rippi 23
Figure 31 Failure of the Lower Ninth ward IHNC East Bank in New Orleans modelled in FEM with Plaxis (source Seed et al 2006) At the bottom right the shear strain contours are depicted that are developed through the embankment and at the bottom left the deformed mesh is presented illustrating the failure mode of the levee
In recent implementation of design concepts and technical recommendations such as CUR 166
ndash Damwandconstructies (2005) which refers to the design of retaining walls it can be observed
that guidelines for Finite Element Analysis (FEA) have been set parallel to the existing analytical
or empirical calculation methods Therefore FEA tends to become more and more accepted as
an alternative for Serviceability Limit State (SLS) as well as for ULS design
According to Wyllie and Mah (2004) LEM give an estimation of the factor of safety with no
information on deformation of the slope However in numerical analysis the failure surface can
evolve during the calculation in a way that is representative of the natural evolution of the
physical failure plane in the slope In that way a better insight into the evolution of failure
mechanisms can be gained
Cundall (2002) compared the characteristics of numerical solutions and LEM in solving the
factor of safety of slopes and concluded that continuum mechanics-based numerical methods ie
FEM have the following advantages
No pre-defined slip surface is needed
The slip surface can be of any shape (in contrast with Bishop and Kranz stability)
Multiple failure surfaces are possible
No static assumptions are needed
Literature study
MSc Thesis A Rippi 24
Structures (such as footings embankments etc) and structural elements (such as
beams cables sheet piles etc) and interface can be included without concern about
compatibility
It is also important to recognize that LEM only identifies the onset of failure whereas FEM
includes the effect of stress redistribution and progressive failure after its initiation Numerical
models can also be used to determine the factor of safety of a slope in which a number of failure
mechanisms can exist simultaneously or where the mechanism of failure may change as
progressive failure occurs (Hoek et al 2000)
On the other hand the accuracy of FEM is dependent on usersrsquo settings such as the refinement
of the mesh generation the loading steps etc Moreover even if FEM can be functional and easy
to use it is essential that the user has a deep knowledge of the subject under investigation (soil
and structural mechanics) in order to be sceptical and critical with the FEM results able to
interpret the output behaviour of the structure and changecorrect everything that seems
peculiar Moreover the usage of FEM implies also some basic knowledge on numerical methods
and the general background of the FEM so as to solve possible numerical errors
An ideal way of modelling in FEM would be a prior calibration of the program according to data
related with the structurersquos properties and performance (ie stress generation) However this
means that enough field measurements should be carried out and under specific conditions in
order to be considered as a base for FEM calibration Unfortunately such field tests are not
always available and if so their reliability is on doubt For example measurements related with
the developed moments or displacements of a retaining wall inside a dike section are rare to be
found but even if there are some available they cannot represent the situation near the failure
domain where most of the engineers are worried about
312 Uncertainties and Sensitivity analysis
Uncertainty in geotechnical engineering can be categorized into aleatoric epistemic and
decision model uncertainty as they are depicted in Figure 32 (Baecher amp Christian (2003))
Aleatoric uncertainty is associated with the natural variability which is defined as the intrinsic
randomness of natural processes Such variability can be expressed by changes of the soil
properties over time at a certain location (temporal variability) or over space at a single time
(spatial variability) Epistemic uncertainty can be divided into the site characterization model
and parameters uncertainty and it is attributed to lack of information about events or lack of
understanding the physical laws that prohibits the ability of creating a realistic model Finally
the decision uncertainty describes the difficulty of being aware of social objectives defining
social values such as discount rates and predicting the planning horizon
Literature study
MSc Thesis A Rippi 25
Figure 32 Categories of uncertainty in geotechnical engineering (modified from Baecher amp
Christian (2003))
Below the general steps of an uncertainty study are described and an introduction into the
sensitivity analysis concept is made
Global methodology of an uncertainty study
A first step of an uncertainty study8 can be described as ldquothe definition of the problemrdquo Initially
the variables of interest (or else the output variables) of which the uncertainty is to be
quantified shall be specified In sequence given several input variables for which the user may
have data andor expertengineering judgment a model denoted usually by a mathematical
function should be introduced that enables the computation of the set variable of interest
After the general context has been staged we should choose the criteria with which the
uncertainty can be evaluated The most complete measure of uncertainty when dealing with a
random vector is the probability distribution In order to assess the value of this distribution
function the following criteria can be followed
Probability of exceeding a threshold the aim is to assess the probability that the variable
of interest exceeds a threshold important for the goals at stake
Quantiles the aim is to assess the threshold that a variable of interest may exceed with a
probability equal to a given value
Central dispersion the aim is to obtain an ldquoorder of magnituderdquo of the uncertainty by
specifying the average value and the variance of a variable of interest
8 In the engineering world this is also called reliability analysis However in this section the general uncertainty study is described In particular reliability describes the ability of a system or component to function under stated conditions for a specified time period
Literature study
MSc Thesis A Rippi 26
The next step is to define a model to represent and quantify the uncertainties of the input
variables One shall investigate each variable as a singularity and come up with the most
suitable probability density function (mostly depending on available data) Besides it is
essential to assess also the potential correlations among the variables that can be included in a
joint probability density function which is discussed later in this section
Eventually the uncertainties of the input variables shall be translated in terms of uncertainty on
the variables of interest This procedure is called uncertainty propagation and can be carried
out via several reliability methods (approximation methods or sampling methods) that are
extensively elaborated in Chapter 4 In Figure 33 a schematization of the different steps during
an uncertainty study is given
Figure 33 Methodology of uncertainty study
Last but not least a better understanding of uncertaintiesrsquo influence can be achieved by
analyzing the contribution of the different uncertainty sources to the uncertainty of the
variables of interest via a sensitivity analysis Such an analysis aims at identifying important
parameters for the system response besides it gives a better insight into the model used to
quantify the uncertainties Before conducting an uncertainty analysis it is advisable to filter out
parameters of less significance in order to reduce the modeling computational effort Below
different methods for sensitivity analysis are described
Sensitivity analysis
Sensitivity analysis (SA) is the study of how the variation in the output of a model can be
apportioned quantitatively and qualitatively to the variation in the model parameters (Saltelli
et al 2004) Saltelli (2004) proposes one possible way of grouping these methods into three
classes (i) screening methods (ii) global SA methods and (iii) local SA methods
i Screening methods
Screening is a particular instance of a sampling-based method The objective here is to identify
which input variables are contributing significantly to the output uncertainty in high-
dimensionality models rather than exactly quantifying sensitivity (ie in terms of variance)
Screening tends to have a relatively low computational cost when compared to other
Literature study
MSc Thesis A Rippi 27
approaches and can be used in a preliminary analysis to weed out less influential variables
before applying a more informative analysis to the remaining set One of the most commonly
used screening methods is the elementary effect method
ii Global SA methods
Global SA techniques incorporate the whole range of variation and the probability density
function of the input parameters to calculate their influence on the output Many global
sensitivity analysis techniques are now available such as
Variance-based methods
o Sobolrsquos method9
o Fourier Amplitude Sensitivity Test (FAST)10
o Analysis of Covariance (ANCOVA)
Regression-based methods11
Both FAST and Sobolrsquos method rely on the assumption of parameter independence while
ANCOVA can also handle correlated input parameters The main principles of FAST method are
presented in Appendix A1 For more information about the other methods the reader can
betake himself to the related references
iii Local SA methods
Local SA methods provide the slope of the calculated model output in the parameter space at a
given set of values (Turanyi and Rabitz 2004) This basically means that local methods involve
taking the partial derivative of the output Y with respect to an input factor Xi |
|
where the
subscript indicates that the derivative is taken at some point in the space of the input
Examples for these are (Schweckendiek 2006)
the α-values in a FORM calculation (further discussion in section 422)
Local Probabilistic Sensitivity Measure
In reliability analysis the local sensitivities will be more important than the global ones in the
sense that the influence of all variables in specific points such as the design point cannot be
identified by the global methods However local SA can only inspect one point at a time and the
sensitivity index of a specific parameter is dependent on the central values of the other
parameters
In this thesis the influence of the different parameters on the response of the system is
evaluated according to local sensitivity indices However before the main part of the reliability
analysis starts it was necessary to filter out the less influencing variables in order to reduce the
number of the major variables and make the analysis more efficient and computationally
affordable For this purpose the global sensitivity method FAST was used whose results are
presented in Chapter 7
9 created by Sobolrsquo (1993) 10 introduced by Cukier et al (1973) 11 introduced by Helton et al (2005)
Literature study
MSc Thesis A Rippi 28
Probabilities as a measure of uncertainties
Most engineers today use the concept of probabilities as the standard way to talk about
uncertainty One way to quantify a probability of a random variable is the calculation of the
cumulative probability function (CDF) of non-exceedance which can be obtained by
the probability distribution function For a random vector this reads
( ) (11)
From this the joint probability density function of this random vector can be determined as
( )
( )
(12)
This function is depicted in Figure 34 by means of contour levels The joint probability density
function has the shape of a hill as it is showed in the 3D figure and the inclination of the ellipses
reveals a correlation between the two variables and
Figure 34 Joint probability density function in 2D and 3D
The functions that join or couple multivariate distribution functions to their one-dimensional
marginal distribution functions are called copulas These are going to be used in order to define
the joint density probability distribution from which the random variables are taking their
values However no detailed explanation is made in this report regarding copulas and its
characteristics as it is automatically implemented in OT whereas the user has to define only the
type of copula that heshe considers the most appropriate (a reference is made to Nelson (1999)
for further information) According to the type of copula that is selected the order of the values
of the variables can be chosen within an iterative reliability method In this thesis an
independent copula was used considering that the input variables are independent the ones
from the others due to the time limit
Literature study
MSc Thesis A Rippi 29
32 Previous Studies In this section a summary of the most relevant researches on the topic of FEM applied in soil
structures and coupling with reliability methods is given
To begin with an attempt to estimate the reliability of a structure modeled in FEM has been
made by Waarts (2000) by introducing an optimized reliability method in terms of
computational effort and efficiency
In particular in Waarts (2000) two adopted reliability methods are introduced both making
use of a response surface (a detailed explanation of the reliability methods and the response
surface techniques is made in Chapter 4 and Appendix F) These adaptive response surfaces are
used in combination with FORM and DS respectively The accuracy and the effectiveness of
these methods are investigated on the basis of artificial LSFs and a comparison is made with the
existing standard reliability methods The most efficient combinations of response surface
techniques and reliability methods were with FORM (FORM-ARS) and DS (DARS) Comparing
these two methods DARS predominated over FORM-ARS as it can cope with a much wider
range of limit state functions
In a later stage the above best performing reliability method (ie DARS) is further investigated
in terms of its efficiency on the basis of complex structures reliability In Figure 35 a couple of
case studies used to verify the performance of DARS are given
Figure 35 Left Two-story two-bay frame structure Right overview of a sheet pile installation
(source Waarts 2000)
The conclusions of this research showed that DARS serves its purpose and that the benefit from
using it increases with increasing number of random variables In Figure 36 the Limit State
Function Evaluations (LSFE) that are carried out as a function of the number of variables are
presented
Literature study
MSc Thesis A Rippi 30
Figure 36 Number of LSFE as a function of the number of variables n The upper line depicts the
required number of LSFE N=160 (it is supposed that on average 3 LSFE per sample are required
and consequently the number of LSFE equals 3x160n=480n) The lower line shows the
performance of DARS (source Waarts 2000)
More recently a methodology for assessing the reliability of geotechnical structures and in
particular deep excavations has been developed by Schweckendiek (2006) More precisely this
study deals with the application of different reliability methods in combination with FEM which
carries out the LSFE The reliability methods are applied initially on simple examples in order to
be calibrated and eventually the most efficient methods are used for the reliability assessment
of a deep excavation with a sheet pile wall The several failure mechanisms are treated in detail
concerning the structural elements of the excavation as well as the soil medium
Finally the failure probability of the system is estimated according to a fault tree
schematization of a deep excavation with one layer anchored sheet pile wall Uncertainties in
the soil properties the phreatic levels and also the corrosion of the sheet pile wall were taken
into account In Figure 37 a picture of the case study is shown together with a FEA illustration
and the importance factors of different soil properties in terms of the probability of failure after
a FORM analysis
Literature study
MSc Thesis A Rippi 31
Figure 37 Finite Element simulation and contribution of several structural and soil properties in
the failure mechanism of an anchored retaining wall in layered soil (source Schweckendiek 2006)
For a similar case study as in Schweckendiek (2006) a cantilever and an anchored sheet pile
wall the safety of the structure based on LEM is evaluated in a research conducted by Gonzaacutelez
et al (2013) Moreover in order to compare the efficiency of the classical methods with FEM
the finite element software Plaxis has been used This research shows the limitation of the
classical models and the efficiency of FEM as far as an anchored sheet pile wall in concerned
More precisely LEM results found to be conservative as they do not take into account the
confinement around the free length of the anchor rod that increases the passive pressure After
comparing numerical results with the classical methods in anchored walls the failure
mechanism given by the FEM suggests an intermediate failure state where the sheet pile wall
describes a translation movement on the deep zone and at the same time a rotation movement
around the anchor point that is closer to the reality
Furthermore Moumlllmann and Vermeer (2008) conducted a reliability analysis of a dike failure A
case study at river Elbe in Saxony Germany has been used and the failure probabilities of
different dike cross sections involving different failure modes were compared More precisely
overflowwave overtopping upliftpiping slope stability and damage of the revetment on the
waterside were taken into account For each failure mode the particular failure probability was
determined using the software package PC-Ring (Vrouwenvelder et al 1999) while the
reliability method used for the assessment of the failure probability was FORM That was
coupled with a slope stability software where Bishoprsquos approach was performed
However this method is limited to circular slip surfaces and prescribed pore pressure
distributions within the dike In order to overcome this limitation FEM was used for further
analysis In combination with FEM an adaptive response surface technique was used in order to
represent the limit state function of each failure mode In Figure 38 the simulated structure
Literature study
MSc Thesis A Rippi 32
together with the response surface of the slope instability failure mode is showed Note that this
structure is approached with a single soil layer (same soil within and under the dike) and that
no structural element is implemented inside the dike
Figure 38 Dike model for the probabilistic FEA of dike stability and the corresponding response
surface of the limit state function for 3 stochastic input parameters (source Moumlllmann and
Vermeer 2008)
The failure probability of each of the aforementioned failure mechanisms were finally calculated
and compared with those recommended in each case while the performance of the coupling
between FEM and response surface method was assessed
Additionally a more recent investigation in terms of dikes with sheet piles employing FEA was
carried out by Breedeveld (2011) The main objective of this study was to display the
implementation of partial factors for design purposes using FEM and especially the software
Plaxis in simulating dikes with a sheet pile wall The dike was firstly simulated as a single
structure without reinforcement while in sequence the occurring stresses and pore pressures
were calculated with an existing sheet pile wall The results of the distribution of the effective
stresses within and below the dike are illustrated in Figure 39 His case study has been
introduced in Chapter 2 while in Chapter 6 a more detailed explanation of how it was modelled
in FEM is given This case study will be used as an example case in this thesis
(a)
Literature study
MSc Thesis A Rippi 33
(b)
Figure 39 Effective stress distribution before (a) and after (b) the installation of the sheet pile wall (source Breedeveld 2011)
Last but not least Bach (2014) has been occupied with the reliability analysis of bored piles and
the case study that he examined is presented in Figure 310 The major objectives of this
research were (1) Propose models to calibrate resistance factors for the Load and Resistance
Factor Design (LRFD) (see section 423 for further explanation of this method) approach and
find a suitable model aiming to directly determine reliability of a bored pile considering some
types of defect that may occur in the bored pile (2) Select a quality control method and evaluate
its reliability when applied to bored piles
For that purpose he made a coupling calculation between the finite element software Plaxis
(version 90) and the numerical probabilistic toolbox Prob2B in order to design bored pile
foundations in light of the reliability-based design Two failure modes the geotechnical failure
mode and the structural failure mode were proposed in order to comprehensively assess the
reliability of an axially loaded pile The parameter uncertainty was considered through the use
of statistical parameters and probability distributions for material parameters in soil models
The soil parameters were treated as random variables The geometry parameters of pile were
used as deterministic quantities because a change in pile shape in the calculation process
requires establishing a new mesh which is now impossible with regard to the Plaxis software
Figure 310 Nine calculation cases considered with different pile toe levels (source Bach 2014)
Literature study
MSc Thesis A Rippi 34
For the reliability of the pile the ULS of the pile was investigated and finally a displacement
criterion of a percentage of the pile diameter used depending on the soil type and
characteristics that the structure is founded on
33 Overview In this chapter concepts that are related to FEM probabilities uncertainty and sensitivity
analysis were discussed The introduction of FEM applications and the principles of how to
handle uncertainties will help the reader to better understand and follow the process of this
thesis
Furthermore preliminary researches related to several aspects of this thesis are presented that
mainly concern
Reliability methods and their applications [Waarts 2000 Schweckendiek 2006 Bach
2014]
Coupling FEM with reliability methods [Waarts 2000 Schweckendiek 2006 Bach
2014]
(Reliability) Analysis of geotechnical structures and especially dikes reinforced with
retaining walls [Waarts 2000 Schweckendiek 2006 Moumlllmann and Vermeer 2008
Breedeveld 2011 Gonzaacutelez et al 2013 Bach 2014]
As someone might have noticed from the overview of the existing research on the topic both the
performance of different reliability methods and the simulation of a dike with a sheet pile wall
in FEM have been carried out These references were quite helpful in order to get an idea of
coupling FEM with reliability methods as well as FEM and reliability methods individually
Moreover a first dive into the system under investigation and the variables to be handled was
made However noone has analysed the reliability of such a complicated system that is more
and more often seemed to be applied in the Netherlands Thus it would be essential to evaluate
the implementation and the results of the reliability analysis of a dike with sheet piles simulated
in FEM The conclusions of such a research are going to contribute to design and optimization
concepts and hopefully to a better understanding of the system behaviour
MSc Thesis A Rippi 35
4 Structural Reliability Analysis
The aim of this chapter is to introduce the basic theory of a reliability analysis and to summarize
the principles of different reliability methods that are widely used for the uncertainty
evaluation of a system Particularly three of them (MC DS and FORM)12 were used in this thesis
and are described in this chapter Finally a brief evaluation of the selected reliability methods
and the way that they are applied on this thesis by coupling them with FEM are discussed
41 Basics of Reliability Analysis In the consideration of the reliability of an element the determination of the probability of
failure is the central issue The limit between failure and non-failure is defined as a limit state
and the reliability is the probability that this limit state is not exceeded The limit states are
interpreted through the so-called limit state functions (LSF) whose general form is
(41)
in which R is the strength or more general the resistance to failure and S is the load or that
which is conducive to failure (ldquosolicitationrdquo) The basic principle of structural design is that the
resistance needs to be higher than the load or in other words that the LSF is larger than zero
(Zgt0) The main objective of the design is to ensure that this performance criterion is valid
throughout the lifetime of a structure Nevertheless the majority of the quantities that both
resistance and load consists of are uncertain Therefore a probability of satisfying the preceding
criterion is estimated and expresses the reliability of the structure P(Zgt0) The probability of
failure is
(42)
Hence the probability of failure is complement to the reliability to the sense that
(43)
The LSF is plotted in Figure 41 indicating the failure space and the limit state function by Z=0
12 Besides RS techniques were used for proceeding with the analysis but most for solving numerical problems rather than evaluating the failure probability The RS techniques are still under investigation and need to be tested before being applied in real cases However it is highly possible that this would be a potential way of evaluating the failure probability in the future
Structural Reliability Analysis
MSc Thesis A Rippi 36
Figure 41 Limit Sate function
In this case the design concept is based on the evaluation of the so-called design point which is
the point in the failure space with the greatest probability density Generally this point is
located on the border between the safe and the unsafe areas
In the structural domain the Joint committee on structural safety (1981) proposed a level-
classification of the calculation methods This classification includes the following three levels
Level III these methods calculate the probability of failure by considering the probability
density functions of all strength and load variables The reliability of an element is linked
directly to the probability of failure
Level II this level comprises a number of methods for determining the probability of
failure and thus the reliability It entails a linearization of the LSF in a carefully selected
point These methods approximate the probability distribution of each variable by a
standard normal distribution
Level I at this level no failure probabilities are calculated The level I calculation is a
design method according to the standards which consider an element sufficiently reliable
if a certain margin is present between the representative values of the strength and the
loads This margin is created by taking the so-called partial safety factors into account in
the design
In the next section the different reliability methods to be used in this thesis are further
elaborated
42 Overview of Reliability Analysis Methods
421 Level III Methods
The foundation of the Level III failure probability calculation is the mathematical formulation of
the subset of the probability space which involves failure (Zlt0) Level III reliability methods
Structural Reliability Analysis
MSc Thesis A Rippi 37
(also known as fully probabilistic methods) compute the probability of failure based on the
exact probability density function and the exact limit state functions Therefore such methods
are considered to be the most accurate Well-known sampling methods are (Quasi-) MC
simulation Importance Simulation DS and the Latin Hybercube Simulation Below the former is
discussed as it represents the fundamental theory on which the sampling methods are based on
while DS is also further explained as it is continuously used in this thesis For the rest a brief
explanation is provided in Appendix F There are also other sampling methods such as the
Numerical and the Directional integration which will not be described in this study However
for further information about them a reference though is made to Waarts (2000) In Level III
methods errors can only occur by a too low number of simulations
Crude Monte Carlo Simulation
The MC method uses the possibility of drawing random numbers from a uniform probability
density function between zero and one If and are the
vectors for the resistance and the load respectively that consists of n variables then by taking
multiple realizations of the uniform probability distribution a value can be determined for
every and
By repeating this procedure a large number of times the failure probability can be estimated
(44)
where is the total number of simulations (nm draws from the uniform distribution in which
m is the number of base variables) and is the number of simulations for which Zlt0
The simulationrsquos relative error is
(45)
If the expected value of the relative error is zero the standard deviation is
radic
(46)
Based on the central limit theorem the error is normally distributed provided is sufficiently
large The probability that the relative error is smaller than the given value E is then
(47)
Thus for the reliability of the relative error is smaller than For the wanted k and
E the required number of simulations can be determined with
Structural Reliability Analysis
MSc Thesis A Rippi 38
(48)
If for instance a reliability of 95 is required for a maximum relative error E=01 the required
number of simulations amounts to
(49)
The number of simulations is therefore still dependent on the probability of failure
MC simulation is applicable whatever the complexity of the deterministic model is However the
number of samples required to get an accurate estimation of may be dissuasive especially
when the value of is small (10-k) Thus crude MC is not applicable when small values of are
sought andor when the CPU cost of each run of the model is non-negligible
In Appendix F and especially in sections F1 and F2 a further elaboration of the MC simulation
is presented Moreover the methods of Quasi-MC Importance sampling and Latin Hybercube
are described
Directional Sampling
The directional simulation method is an accelerated sampling method It implies a preliminary
iso-probabilistic transformation as for FORM and SORM methods however it remains based on
sampling and is thus not an approximation method Below the method as it is described in Open
TURNS 15 Reference Guide (2015 pp190-193) is presented
Each simulation of the directional simulation algorithm is made of three steps Using the
probability distribution of a random vector we seek to evaluate the following probability
( ) (410)
Here is a deterministic vector and ( ) the limit state function which enables the
definition of the event For the iteration these steps are the
following
Let A point is drawn randomly on S according to a uniform
distribution
In the direction starting from the origin and the passing through solutions of the
equation ( ) (ie limits of ) The set of values of that belong to is
deduced for these solutions it is a subset of
Then one calculates the probability By property of independent
standard variable
is a random variable distributed according to chi-square
distribution which makes the computation effortless
Structural Reliability Analysis
MSc Thesis A Rippi 39
Finally the estimate of the probability after N simulations is the following
sum
(411)
The main idea is that for each direction we go along the direction by step of a specified length
from the origin to the maximum distant point and we check if there is a sign change on each
segment so formed (see Figure 42) We go until the maximum distant point Then for all the
segments where a sign change is detected we research the root λ in the segment (there are
also other strategies available in OT however all of them follow the same principle of
searching) The following figure illustrates the principle of an iteration in dimension 2
Figure 42 DS of a 2-dimensional problem
Similar to MC method the outcome of the estimated probability of failure is a random variable
and the error in the estimate can be made as small as possible by taking a sufficient number of
samples For DS the standard deviation σ of the estimated failure probability can be quantified
as follows (Melchers 2002)
radic
sum
(412)
From this equation relative errors and the confidence intervals can be estimated As can be
seen the error in the estimated probability of failure will decrease with increasing number of
sampled directions Equation (412) can be used to determine the number of sampled directions
that is required for a reliable estimate of the failure probability The convergence criterion in
this method is usually the coefficient of variation (CoV) which is defined as
Structural Reliability Analysis
MSc Thesis A Rippi 40
422 Level II Methods
Level II methods (known also as fully probabilistic methods with approximations) can take all
the probabilistic properties of the random variables into account but they include
approximations of the limit state function and therefore their use and outcomes should be
inspected and evaluated in order to be considered as reliable However experience in that
methods has shown that the computational effort is profoundly decreasing in comparison with
Level III methods and that the application of them can provide important parameters such as
the reliability index (β) and the influence factors (α) Some of the most known Level II methods
are First Order Second Moment (FOSM) Method First and Second Order Reliability Methods
(FORM and SORM) and the Point Estimate method (PEM) In the next paragraphs FORM is
elaborated as it will be mainly used in this thesis while in Appendix F3 the principles of FOSM
are discussed
First Order Reliability Method (FORM)
The principles of FORM are based on the First-Order Second-Moment (FOSM) approach which
first introduced the reliability index concept This method is presented in Appendix F3 in detail
However the reliability index estimated using the FOSM approach is not ldquoinvariantrdquo which
means that the index value depends on the limit state function format (Farrokh 2007) Hasofer
and Lind (1974) proposed an invariant definition of the reliability index that leads to a
geometric interpretation known as first-order reliability method (FORM)13 The first step of this
method is the transformation of the random variables to equivalent standard normally
distributed variables and the whole procedure is carried out in u-space (or else standard space)
For variables that are normally distributed this step is as follows
(413)
For other types of distributions there are procedures available for carrying out this
transformation such as Generalised Nataf and Rosenblatt transformations In this study these
transformations are not discussed however for further explanation a reference is made to the
Reference Guide of OpenTURNS 15 (2015)
In the sequence the limit state function Z is expressed in terms of
The second step is the approximation of the function with the first two terms of the Taylor-
polynomial The approximation reads (CUR 1997)
( ) ( ) sum
(414)
where is the point where the linearization takes place This approximation of Z is linear and
according to the central limit theorem it is normally distributed The expected value of the LSF
can be approximated by the expected value of the linearized function
13 VNK2 uses FORM in view of the calculation time required In case of inaccuracies VNK2 switches to more time-consuming techniques such as DS whereby a large number of draws are made from the probability density functions of the various stochastic variables (uncertain quantities)
Structural Reliability Analysis
MSc Thesis A Rippi 41
( ) sum
(415)
While the standard deviation (sum of contributions of each variable to the variance of Z) is
defined as
radicsum
( )
(416)
Using the definition of FOSM for the reliability index [ ]
where [ ] is the mean value of
the limit state function and substituting Eq 415 and Eq 416 in it the reliability index can be
approximated However linearization in different points can lead to different values of the
reliability index Hasofer and Lind definition of the reliability index overcomes this limitation
and renders it equal to the minimum distance from the origin to the design point
(417)
where radic
Looking for the design point is basically an optimization problem Many analytical and
numerical approaches can be used for that purpose A relatively straightforward method to do
this is by firstly assuming that the design point is the mean value (the starting point can also be
another point according to an engineering judgement in order to accelerate the optimization
procedure) The obtained β-value is used to determine a new point in which the LSF is
linearized In this case the importance factors αi are calculated as
( )
radicsum (
( ) )
(418)
where is the design point and is the number of variables expresses the
contribution of the variance of each variable to the total variance of Z in the design point Figure
43 illustrates the linear approximation of the limit state function and the aforementioned
parameters for a two dimensional problem The probability that Zlt0 can be determined using
the standard normal CDF
(
) (419)
Structural Reliability Analysis
MSc Thesis A Rippi 42
The new calculation point is determined by
(420)
After some iterations the chosen optimization algorithm finds the final design point and the
new reliability index In Appendix A2 an overview of the different optimization algorithms
available in OT is presented together with an evaluation of their performance in the specific
case study
Figure 43 Two-dimensional illustration of u-space LSF and design point
423 Level I Methods (semi-probabilistic)
At the beginning of the probabilistic concepts incorporation in the field of structural
engineering the most notable development was the implementation of LFRD At the past single
factors of safety on the ratio of total resistance to total load were used in order to ensure that
the stresses developed from an applied load on a structure were lower than the allowable ones
LRFD replaces those factors by introducing a set of partial safety factors on the individual
components of load and resistance These partial factors can be selected such as they account
for uncertainties associated with the resistance and the load respectively
The current design philosophy in CUR 166 and Eurocode is characterized by the use of
characteristic values of the parameters (in Appendix E a further explanation of the
characteristic values is given together with the way that they are converted into the mean
values for being used in the reliability analysis later on) The values of the partial safety factors
are additional factors to the characteristic values Therefore the characteristic values are
multiplied with the aforementioned safety factors ( ) and the following criterion has
to be satisfied for a reliable structural performance (see Figure 44)
(421)
Structural Reliability Analysis
MSc Thesis A Rippi 43
The main disadvantage of LRFD is the fact that there is a set of load and resistance factors that
need to be calibrated in order to cover the specific cases that are likely to occur This is because
of the lack of sufficient statistical data that are necessary for this calibration procedure
Consequently it cannot reassure that all the designs to which the set of factors is applied result
in the reliability level that was aimed for As a result in majority of the cases a conservative
calibration that gives the required reliability is chosen which can lead to ldquoover-designrdquo
structures
Figure 44 Design with partial factors for the load and resistance parameters (source
Schweckendiek 2006)
The evaluation of an elementrsquos reliability starts in principle with calculating the probability of
failure and subsequently the reliability for the given strength and load In practise the problem
is often that the strength is unknown but it has to be determined for a given reliability The
determination of the required reliability can be estimated with the help of Level II and III
methods by iteratively adjusting the strength in the calculation until a sufficiently small
probability of failure is found
In the design domain regulations and guidelines follow the standard that the characteristic
value14 of the strength is divided by a factor and that the characteristic value of the load is
multiplied by a factor as it is described in Eq 421
The link between Level I and the estimation of the failure probability has been achieved through
Level II methods The design point that results from a Level II method calculation is the point
with the greatest joint probability density of the strength and the load and it is therefore
possible that when failure occurs the strength and load values will be close to that point The
design values can be specified as
(422)
(423)
14 The characteristic value of a parameter is its main representative value and its value can be specified as a mean value an upper or lower value or a nominal value according to EN 1990 definition (2002)
Structural Reliability Analysis
MSc Thesis A Rippi 44
where are the coefficient of variation for the resistance and the solicitation respectively
15 and are the importance factor and β is the reliability factor These can be also expressed
in terms of characteristic values as follows
(424)
By substituting function 423 with 421 and 422 the partial safety factors are defined as
(425)
(426)
where 16 and are the values for load or resistance respectively to which a probability of
(non-)exceedance of 95 corresponds (non-exceedance for the load and exceedance for the
resistance) (for a standard normal distributed parameter this is 1645)
424 Response Surface Techniques (RS)
In case that the models presented above tend to be time consuming for the limit state function
evaluation or convergence problems of the optimization algorithms occur a better methodology
is recommended known as response surface The RS is mainly used when (unknown) response is
only available from experiments or complex FEM computations (for example large highly-non-
linear FEM models) Indeed once a RS has been built up the various standard methods may be
applied at a negligible cost An analytical limit state function replaces the real response function
The main idea is that the response consisting of a complex function of input variables is
approximated by a simple function of the input variables
A list of possible response surfaces techniques is given below
Linear and quadratic Taylor expansions
Polynomial RS based on least squares
Kriging method
Polynomial chaos expansion
A detailed explanation of the aforementioned methods can be found in OpenTURNS 15 (2015a)
The first method is associated with the approximation of the model response around a specific
set of input parameters while the rest seek a global approximation of the model over its whole
domain of definition The most sophisticated one is the polynomial chaos expansion but one the
15 α values are negative for solicitation (ie RV for which increasing their values lead to adversely effects on the reliability) and positive for resistance (ie RV for which increasing their values implies a beneficial effect on the reliability) 16 is negative and can be negative or positive
Structural Reliability Analysis
MSc Thesis A Rippi 45
other hand its complexity does not make it attractive In this research a polynomial RS is
applied based on least squares
In combination with FEA the standard procedure is as follows
1 Select the most important random variables on the basis of engineering basis
2 A SA is carried out in combination with FEM
3 Reduce the stochastic variables if needed according to the SA
4 A RS is constructed through the response data
5 A reliability calculation is carried out using the RS instead of the real response
A polynomial RS is generally constructed by fitting a quadratic function to the sampling points
whose general expression is
sum
sum
sum sum
(427)
The type of the reliability method to be used is of little importance since the time consuming
LSFE (using FEM) are replaced by analytical expressions However of main importance is now
the accuracy of the RS compared to the real response There are several parameters in order to
evaluate the goodness of fit of the RS to the real model among which is the well-known R-
squared
In Figure 45 an example of a quadratic response surface is showed that is fitted on the data
response indicated with the black circles This was a 6-dimensional problem that means 6
different variables were included in the system (non-visualized dimensions are kept constant at
their mean value) In this thesis the response of the system is to be acquired after the coupling
of FEM with a reliability method and that would give the ldquoblackrdquo circles that are shown in Figure
45 The response surface technique applies a curve fitting on the real response of the system
(ldquoblackrdquo circles) whose analytical formula can later be used for the reliability analysis of similar
type of systems with negligible computation time
Structural Reliability Analysis
MSc Thesis A Rippi 46
Figure 45 Quadratic response surface with 28 terms (ie ) for 6 dimensions ( ie x
=[ x1 hellip x6])
43 Coupling Reliability Analysis with FEM The limit state function evaluations (LSFE) will be carried out with the software Plaxis 2D 2015
which is a special two-dimensional finite element software used to perform deformations and
stability analysis for various types of geotechnical applications Moreover considering the case
study under investigation in this thesis Plaxis in contrast with other FEM such as Abacus
Comsol DIANA etc offers several techniques to realistically simulate structural elements such
as sheet pile walls and anchors and their interaction with soil while the variety of the
constitutive models for the soil body that are available and the ability to include the history of
the construction phases can lead to a better analysis of the systemrsquos behavior in terms of the
stress level and the deformations It is essential at that point to mention that using FEM for this
purpose means that the limit state formulation is implicit and can only be solved numerically
The reliability analysis is carried out through an uncertainty package In this section firstly a
description of the reliability package and its possibilities are given Finally an explanation of the
coupling procedure between the reliability tool and FEM is given together with the calculation
process that was followed for the parameters manipulation
431 The functionality and possibilities of OT
The reliability analysis tool to be utilized in this research is OpenTURNS (OT) which is a
scientific library usable as a Python module dedicated to the treatment of uncertainties and it is
still under development during the work of this thesis Several reliability packages are available
such as OpenEarth (applied in Matlab) Prob2b (applied in Probox software of TNO) and
Probabilistic Toolkit (software of Deltares) Plaxis 2015 features Python connection possibility
and thus coupling was decided to be carried out with an uncertainty package in Python and OT
is the most developed one Moreover until now only OT is an open source package and can be
Structural Reliability Analysis
MSc Thesis A Rippi 47
used easily from anyone as there is a wide community of experienced people supporting with
their knowledge while many related manual reports are already available for starting learning
A list of the current available reliability methods in OT is given in Figure 46 In this figure apart
from the standard methods the alternative method of the Adaptive Response surfaces is
introduced that is mainly used when (unknown) response is only available from complex FE
computations
Figure 46 Reliability methods available in OpenTURNS
In this thesis the methods to be basically applied are the FORM and the DS analysis while MC is
also used mainly for confirming the application of FORM Additionally the RS technique has
been mainly implemented for enhancing the performance of FORM and DS analysis
OT handles 47 types of distributions amongst which the ones used in this research are Normal
Lognormal Uniform and Truncated Normal distribution Regarding the joint distributions that
are available to be selected for random vectors 11 types of copulas are existing in OT amongst
which the most known ones are the Independent the Gumbel and the Normal copula In this
thesis the independent copula has been used during the reliability analysis which means that
the variables were assumed to be independent among each other However it is strongly
advised a further elaboration and research considering correlation matrices for specific soil
parameters
432 Coupling OpenTURNS-Plaxis
The finite element simulation will be carried out in Plaxis 2015 Plaxis is a finite element
software for plane strain and axi-symmetric modelling of soil and rock behaviour Moreover it
supports a fully automatic mesh generation allowing for a virtually infinite number of 6-node
and 15-node elements
Structural Reliability Analysis
MSc Thesis A Rippi 48
The coupling of reliability analysis and FEA requires an interface for the communication
between each other When a reliability tool is coupled with another software program the
reliability program carries out the whole reliability analysis and it uses the other program only
for the evaluation of the limit state function More precisely OT should be able to read and
amend Plaxis output values for important variables such as material parameters pore pressures
generation and stresses development and corresponding deformations inside the dike
Respectively Plaxis has to be also capable of obtaining the new values that has been set by OT
for the variables that are treated as stochastic during an iterative process according to the
reliability assessment In Figure 47 an illustration of the coupling methodology and its function
is shown
Figure 47 Coupling scheme OpenTURNS-Plaxis
In principle an input file is firstly required where the user set the preferable reliability method
to be used the stochastic input parameters and their probability distributions the joint
probability distribution and the corresponding correlation matrix and finally the limit state
function is formed depending on the situation In Appendix D an example of the input files that
were used for soil sheet pile wall and anchor analysis is shown However such input files
should be interpreted so as to be readable by both Plaxis and OT Therefore an input interpreter
was created which is actually a python script that helps OT to start up the reliability analysis
according to the assigned method variables distributions and LSF As it was mentioned before
the evaluation of the limit state function is conducted by Plaxis For that purpose the input
interpreter should be also able to send the next set of input parameters to Plaxis However an
additional means of connecting the interpreter with Plaxis is also needed This is can be
Structural Reliability Analysis
MSc Thesis A Rippi 49
achieved via a Plaxis interface which actually call Plaxis to make the calculations while it also
transfers the required value of the limit state function to input interpreter and this in turn to OT
Eventually the probability of failure is obtained as a model result However it is also essential
that Plaxis simulation procedure converges to the desirable criteria and under the physical
boundary conditions that have been determined Likewise the convergence criteria of the
reliability methods shall be manipulated so as the optimization algorithms to able to converge
efficiently (see Appendix A2 for further explanation)
44 Overview In section 42 a summary of the principles of the main reliability methods was made From
Level III methods MC and DS were presented whereas from Level II the basics of a FORM
analysis were introduced Besides the concept of the Level I method was explained Even
though this method was not used in this research for the evaluation of the probability of failure
it was applied in section 643 for the deterministic analysis of the system under investigation
and the preliminary calculations of the dimensions In Appendix F more reliability methods are
introduced and they are further described In the sequence their implementation in conjunction
with Plaxis was discussed through the coupling of FEM with OT For more information
concerning OT special features (ie SA optimization algorithms and probability distributions) a
reference is made to Appendices A and E
Before any of these reliability methods was applied on the case study of this thesis their
performance and their compatibility with Plaxis were tested with simple examples At first the
probability of failure was evaluated with FORM SORM Importance Sampling (IS) crude MC and
DS for simple artificial linear and non-linear limit state functions
and
respectively where R B and F are the random variables distributed normally
or lognormally The crude MC technique has been applied in order to validate the failure
probability in case the various methods give different results Secondly a flexible circular
footing over a one-layer soil material was used from Plaxis tutorial in order to test the coupling
between Plaxis and the reliability method and reassure the functionality of the interface In this
pilot example the objective was to determine the probability that the settlements of the footing
exceed a certain threshold of settlements and thus The respective
components that were taken into account for the failure mechanism was the cohesion (c) the
friction angle (φ) and the specific weight (γ) of the soil
From a qualitative evaluation of the reliability methods that has been done and according to the
experience gained from the aforementioned applications of some methods valuable conclusions
were drawn in terms of their efficiency More precisely regarding the sampling methods it does
not required previous knowledge of the failure domain apart from the IS method in which a
starting point inside the failure domain shall be provided Moreover the calculation effort
depends on the magnitude of the failure probability and the required accuracy whereas for the
IS it depends also on the selection of the sampling region In Waarts (2000) it has been proved
that DS tends to be more efficient than MC for low dimensional problems as it is indicated in
Figure 48 (number of random variables nlt100)
Structural Reliability Analysis
MSc Thesis A Rippi 50
Figure 48 Required number of samples for MC and DS as a function of the random variables
(source Waarts 2000)
As far as the approximation reliability methods (FORM and SORM) are concerned it is not
necessary to be aware of the failure region in advance Furthermore it has been observed that
the required iterations and the calculation time is quite lower in comparison with the sampling
methods However the accuracy of the method is highly dependent on the shape of the LSF
In Table 41 a summary of the most applied reliability methods evaluation is presented
According to this evaluation it was decided that the methods of FORM and DS as well as the
combination of them with RS techniques are going to be tested and evaluated in terms of their
efficiency and robustness These methods were chosen from both Level II and Level III methods
due to their expected reduced computational time the non-requirement of previous knowledge
about the LSF and their satisfactory performance in similar case studies In Chapter 5 the
different failure mechanisms and the relevant LSF that are going to be utilized in this thesis are
presented whereas in Chapter 7 the results deduced from the reliability analysis are elaborated
In Chapter 7 there is also an indication of which reliability method was used for each systemrsquos
component the selection of which was based on the aforementioned evaluation
Table 41Evaluation of reliability methods
Method Previous knowledge Accuracy Calculation Effort
FORM not required full accuracy for
Gaussian variables and linear LSF
depends on LSF linearity and the number of random
variables
SORM not required exact up to 2nd order
LSF error dependent on the shape of LSF
depends on LSF linearity and the number of random
variables
Structural Reliability Analysis
MSc Thesis A Rippi 51
Crude MC not required can be controlled by convergence criteria
depends on the magnitude of failure probability and the
required accuracy
IS required can be controlled by convergence criteria
depends on the choices made for the sampling region
DS not required can be controlled by convergence criteria
for low dimensional problems (nlt100) DS is more efficient
than MC
Especially RS were created based on a SA sampling in order basically to increase the efficiency
of FROM and DS analysis in case of FEM numerical errors Precisely an analytical formula or a
response surface was needed that can approximately represent the performance of the structure
under investigation and that is called in case of Plaxis errors This formula was essential in
order to keep the reliability analysis running by providing a response value for the LSF when
Plaxis calculation is unsuccessful and thus incapable of returning a result It should be
mentioned that in case a RS is implemented for the reliability analysis Figure 47 can be
reduced to a graph consisting of only the input file OT and the input interpreter as no Plaxis
calculations would be necessary any more The LSFE will then be performed through the
response surface that is provided by the user Due to this limited use of the RS method it will
not be further treated in this thesis
Structural Reliability Analysis
MSc Thesis A Rippi 52
MSc Thesis A Rippi 53
5 Failure Mechanisms and Limit State Functions
The reliability analysis of a structure requires the definition of the different failure modes that
are relevant to the corresponding structural elements In this chapter the possible failure
mechanisms of the systemrsquos elements expressed in LSF are identified and their role in the
system reliability is explained Based on these LSF the reliability of the system components was
evaluated These LSFs focus on the ULS whose exceedance implies failure of the corresponding
component
51 Introduction to the system analysis and the limit states For system reliability analysis all the possible failure mechanisms have to be identified and
summarized in a fault tree In Figures 51 and 52 the fault trees of a dike section without
structural elements and a retaining structure with sheet piles are depicted respectively In this
project the combination of these two separate structures ie a reinforced dike section with an
anchored sheet pile wall is to be simulated and studied in terms of its failure modes
As it is described later in this chapter failure can be expressed in different ways depending on
the structure and stakeholders demands and safety standards This research is mainly focused
on the ULS of failure and Mohr Coulomb was chosen as the soil failure criterion due to its
efficiency on detecting failure Such a failure criterion is actually defining the stress strain
relationship and the gradual weakening of the soil (ie the gradual reduction of the strength
parameters such as the friction angle and the cohesion due to the stresses development until
soil collapses) under a stress condition In Plaxis new version there is the ability of gradually
reducing structural properties such the strength of steel elements However the results
retrieved of such a safety analysis have not been yet investigated and for that reason such
calculations are not included in this project
As far as the dike section is concerned the failure mechanism to be investigated in this project is
the macro instability of the slope rather than the other mechanisms depicted in Figure 51 On
the other hand because of Plaxis limitations to distinguish among the different structural
elements failure mechanisms each element was considered as a singular case for investigating
its failure mode In the next section a distinction is also made between the ULS and the SLS
For a reinforced dike with retaining walls there are basically four classes of structural elements
(see Figures 12 and 21)
∙ Retaining wall (ie sheet piles)
∙ Anchors
∙ Walings
∙ Soil structure (ie dike section)
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 54
For the reliability analysis of each of the elements the probability of failure is to be determined
In the next sections of this chapter the respective LSF that are considered for each class are
further elaborated according to the ULS criterion while in the end the general combined fault
tree is to be schematized
Figure 51 Fault tree of a dike section
Figure 52 Fault tree for retaining structures using anchored sheet piles (source Schweckendiek 2006)
The system failure in this research is considered as a serial system of the anchor sheet pile
wall and soil body LSFs The fault tree of the system failure is presented in Figure 53 together
with the LSF of each component as they are formulated according to sections 522 and 523
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 55
Figure 53 Fault tree of the system failure
It should be mentioned at that point that the probability of failure of the system differs from the
overall probability of flooding which takes into account any potential failure mechanism of the
system combined with the uncertainties of the water elevation More precisely as far as the dike
safety is concerned a probability of failure (where ) under a certain
water level is estimated from the fault tree of Figure 51 which in sequence is multiplied
with the occurrence probability of the corresponding water level in order to estimate the
overall probability of failure for the specific water elevation Then that product is accumulated
over a required range of water levels that can jeopardise the overall stability of the structure in
order to calculate the overall probability of failure or else the probability of flooding This
probability is then compared to the one established from the safety standards so as to reassure
the safety or not of the structure can be estimated as follows
int
sum (51)
In this thesis the probability is to be estimated where is a specific water level
and especially the design water level as it is considered to be the most challenging part of the
procedure described above gives the probability of occurrence of the water level (ie it is
the PDF of ) After setting up the steps and implementing them successfully for the estimation
of the calculation of the overall probability of failure is just a repetition of the same
procedure for more water levels It should be mentioned that in this thesis failure consists only
of the macro-instability failure mode whereas the rest are excluded for the time being A
simplified alternative for the overcoming the integral of Eq51 is to repeat the calculations for
certain water levels and then sum their products instead of integrating the full range of the
them
52 Limit State Functions
521 Serviceability Limit State
The SLS is evaluated in design calculations in order to avoid excessive deformations of a
structure that could lead to the loss of its service requirements and its functionality In some
cases the SLS criteria are used to prevent damage to adjacent structures in the surroundings
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 56
For example concerning the design of a dike section attention should be paid to deformations
caused to buildings that are located in the inland part of the dike in case of an extreme event
In Figure 54 an example of a deformed dike is depicted In this figure some of the potential
locations that excessive deformations might be experienced are illustrated such as
∙ the vertical settlement of top of dike (arrow 1)
∙ the horizontal displacement of the top of the sheet pile (it might also be the middle part
of the sheet pile that ends up with the largest deformation according to the loading
conditions and the specific soil structure different deformations might occur) (arrow 2)
∙ Uplift upstream of the dike (arrow 3) special attention shall be paid to this type of
deformation not only due to the direct effect on inland structures but also as a sign of
developing piping mechanism
(a)
(b)
Figure 54 Figure (a) shows the initial undeformed structure while (b) illustrates the deformed mesh (not to scale) (the plotted length of the anchor does not have any physical meaning) In (b) the possible locations and directions (black arrows) of potentially excessive deformations
According to the current Dutch regulations (Larsen et al 2013) the SLS criteria for designing
reinforced dikes with sheet pile walls require that
∙ the settlement on the top of dike does not exceed 10 cm and
∙ the horizontal displacements of the sheet pile wall does not overcome 10 cm
In this research the SLS criteria and especially the dike settlements were used roughly for the
preliminary determination of the structures characteristics that are discussed in section 643
Dike Anchor
Sheet pile wall
1
2 3
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 57
More precisely displacements at the top of the dike were also taken into account as
complementary to the safety factor in order to determine the required length of the sheet pile
wall As far as the reliability analysis in concerned the LSF were formulated based on the ULS of
the different systemrsquos components that are presented in the next section
522 ULS for Structural Members
In the design process one is most interested in the ULS of a failure mechanism This state
describes the situation wherein the acting extreme loads are just balanced by the strength of the
construction If that limit state is exceeded the construction will lose its functionality and thus
collapse or fail In general most attention is paid to the behaviour of the structure after
completion However during construction there are also periods in which the construction may
fail The different phases of construction are listed in section 642 In the present section the
analytical LSF of the structural elements are given as they are going to be used in the reliability
analysis
Sheet pile wall
The most relevant failure mode for the sheet pile wall is the exceedance of the yield strength
which corresponds to the ultimate steel strength The response of the structure is mainly due to
bending moments and the axial forces (shear forces are considered to be negligible) Where an
axial force is present allowance should be made for its effect on the moment resistance
Figure 55 The water and earth pressures as well as the anchor forces lead to bending moments and axial forces on the sheet pile wall
In Figure 55 an example of the axial forces and the bending moments that can be developed in
the sheet pile wall with one anchor layer are illustrated after an extreme water level loading
Accordingly the maximum stresses on the sheet pile wall are composed of a bending moment
and a normal force component17
[
] (52)
where [kNm] and [kN] are the bending moment and the axial normal force
respectively that depend on the depth level where they are calculated over the sheet pile length
[m3] is the elastic section modulus and [m2] the cross-sectional area of the sheet pile
wall
17 the vertical anchor force component is reducing by its interaction with the soil over depth
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 58
Bending moment and axial force can be variable over the depth and that is why they are
expressed as a function of z-depth FEM has the advantage to take into account second order
effects ie a stiffer structure will experience higher bending moments than a more flexible one
Taking into account the above the limit state function can be formed as the difference between
the maximum developed stress and the yield stress
[
] (53)
where and can be characterized as the load variables while and can be considered
as the resistance variables and are assumed to be constant over depth
Concerning the permissible displacements for the ULS an upper limit for the maximum
horizontal displacements of the top of the sheet pile has been set to 150L where L is the
vertical length of the sheet pile wall (Larsen et al 2013) This requirement was mainly used in
section 643 where the structural properties were determined after the deterministic analysis
Anchors
Anchors are loaded by their reaction to the horizontal loads on the retaining walls The failure of
the anchor element is actually represented by the failure of the steel members of the anchor
(tubes bars cables etc) that are loaded by traction forces
As it is illustrated in Figure 56 the anchor experiences an axial tensile N that is nearly constant
over its length
Figure 56 Axial loading of anchor inside a dike
The elastic behaviour of an anchor involves only a relationship between axial force N and
displacement (elongation) u of the form
[ ] (54)
where EA [kN] is the anchor stiffness consisting of the steel Youngrsquos modulus E [kNm2] and the
anchor cross section A [m2] and L [m] in the length of the anchor
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 59
Similarly to the sheet pile wall the limit state function of the anchor involves the certain yield or
ultimate strength of the steel members and the maximum stress that the anchor experiences
during its loading Consequently the LSF is as following
(55)
where [kN] is the calculated anchor force and [m2] is the cross sectional area of the
anchor (both of them considered to be constant over the depth) It is essential to mention that
the anchor is also subjected to bending moments due to soil settlements (that are implicitly
illustrated via the uniformly distributed load q over the tie rod) that should be taken into
account in order to investigate the displacements of the tie rod itself However in this thesis
only the axial forces on the anchor are considered without taking into account the individual
deformations and its reaction with the surrounding soil
Walings
The waling is the element that transfers the loads from the retaining wall to the anchors (see
Figure 21) The loading of the walings can be schematized as a continuous beam on several
supports as it is depicted in Figure 57
Figure 57 Loading of walings
If Nα [kNm] is the calculated anchor force per m then the developed bending in the support Ms
and in the opening Mo can be approximated as follows
[ ] (56)
(57)
where La is the mutual anchor distance Considering as the design moment of the waling to be
the limit state function can be formulated as follows
Waling Sheet pile Anchor
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 60
(58)
For a conservative design of the waling the limit state function will give the same or lower
failure probability than the anchor itself Therefore it will actually not be necessary to carry out
this analysis The anchor failure will be the determinant mechanism (Schweckendiek 2006)
523 ULS for Soil Failure
In this thesis emphasis is given on the dike global instability which actually consists of several
failure modes Figure 58 indicates the possible failure mechanisms of flood defences The
combination if these can lead to the overall instability of the dike and thus to soil body failure
However Plaxis is not capable of accounting for all of them More precisely Plaxis assumes the
soil to be a continuous body and thus it can model movements in the scale of soil bodies (using a
relative fine mesh in relation to the moving soil body) but not on a soil particle level Plaxis can
simulate the groundwater flow in a soil body and from this someone can deduce input for a
piping analysis (see failure mechanism G in Figure 58) (using other softwaremethods) but as
it has been mentioned above it cannot determine movement of soil particles due to
groundwater flow Moreover Plaxis cannot deal with the flow and waves occurring in ldquoopen
waterrdquo ie water outside the soil in a canal a lake or sea for instance
Figure 58 Possible failure mechanisms of flood defences (source Flood Defences 2015)
Therefore Plaxis can handle the failure modes (C) (D) (E) (J) (K) and (L) of those illustrated in
Figure 57 In this thesis loads due to extreme temperature conditions or ships are not
considered while the settlements are taken into account implicitly though the mechanisms (C)
(D) and (E) These mechanisms are thus investigated in this thesis and they mainly refer to the
macro-instability of the inner and outer slope of the dike or the overall dike body (ie horizontal
sliding (mechanism D))
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 61
For these failure modes soil failure is defined as shear failure In Figure 59 the most relevant
patterns of the macro-instability regarding the shear strength of the soil are illustrated
Applying FEA the most critical failure mode is determined automatically However it is not
always straightforward what the trigger mechanism of the failure was and that is why more
investigation and FEA tests are needed in order to obtain a clear view
(a) Outer slope failure (b) ldquoActive siderdquo failure
(c) ldquoPassive siderdquo failure (d) Overall failure
(e) ldquoKranzrdquo stability
Figure 59 Macro-instability failure modes in a dike with an anchored sheet pile wall
In Figure 59 failure modes (a)-(d) follows Mohr-Coulomb failure criterion In principle the
total stress state inside a dike section consists of the effective stresses and the pore pressures
according to Terzaghi principle
[
] (59)
Mohr-Coulomb failure criterion basically considers as limit for the failure to occur the
maximum shear stress that the soil can withstand This is determined from the friction
angle the cohesion of the soil and the current stress state The drained shear strength
(similarly in for the undrained shear strength) in the soil according to Mohr-Coulomb is defined
as follows
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 62
[
] (510)
where is the effective friction angle In Appendix B1 a detailed explanation of this failure
criterion is attached
Taking into account the above mentioned failure mechanism (a) is mainly triggered by the
increase of the pore pressures in the outer slope which subsequently causes a decrease in the
effective stresses ( ) which at the same time leads to a reduction of the shear strength
(considering Mohr-Coulomb failure criterion) When the shear stresses in the soil exceed the
shear strength a slip plane forms and a soil wedge collapses
Failure modes (b) and (c) are primarily determined from the active and passive effective
stresses respectively As far as the ldquoactive siderdquo failure is concerned displacements of the sheet
pile wall towards the downstream side of the dike lead to development of active stresses
where is the active pressure coefficient During the ldquoactiverdquo failure the
retaining soil is allowed to relax which leads to a decrease of horizontal stresses and
simultaneously to a decrease in shear strength Similarly to failure mode (a) a soil wedge
collapses The result is an increased earth pressure on the wall for which it is not designed
However this type of failure in dikes is not such determinant for the soil and the sheet pile wall
failure because there is always the resistance of the passive side of the dike that keeps the active
side stable enough Such a mechanism is thus more relevant for deep excavations with retaining
walls
In mechanism (c) the passive soil resistance is exceeded by the horizontal loads In this case the
wall moves inland and a wedge of soil compresses The shear strength is larger due to the
deformation of the wall and the horizontal stresses increase stresses where is
the passive pressure coefficient ( ) This failure mechanism usually occurs due to an
underestimation of the sheet pile length or due to the presence of a weak soil layer in the
passive side
Failure mode (d) is actually a combination of (b) and (c) possibly accompanied also with the
development of excess pore pressures as it was discussed in failure mechanism (a)
Moreover too short anchors could cause a soil block to collapse as it is showed in Figure 59(e)
(ldquoKranzrdquo stability)
As it has been already mentioned for mechanism (d) failure modes can be correlated and
combined with each other in order to lead to a final failure state Therefore the common failure
probability would be smaller than the sum of the singular probabilities In this thesis the total
probability of failure of the soil body due to global instability is to be determined This is due to
the fact that the application of FEM is capable of simulating the combination of the failure
modes depicted in Figure 58 and thus the overall failure (related to (C) (D) and (E) modes of
Figure 58) can be evaluated However with the classical engineering approach for the stresses
calculation each one of the failure modes depicted in Figure 59 should be separately evaluated
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 63
Below a description of the available methods to formulate the LSF of the soil failure is given
After the evaluation of these alternatives a selection was made regarding the most suitable LSF
for the current case study
Limit State functions for soil failure
Soil mechanical failure can be analysed in several ways (Schweckendiek 2006)
1 Excessive Deformations
2 φ-c Reduction
3 Relative Shear Resistance
4 Plaxis definition of soil collapse
The possibilities and limitations of the aforementioned methods are briefly discussed below
1 Excessive Deformations
Similarly to the SLS a limit state function can be formulated by deformations that are
unacceptable As it is illustrated in Figure 54(b) there are some points (the top of the dike the
sheet pile top and the inland soil level behind the dike) the displacements of which shall be
limited to the minimum possible during the structurersquos lifetime in order for the structure to
meet the required service standards Therefore the calculated deformations at locations 1 2
and 3 could give a clear indication of failure given the maximum acceptable deformation
In case that there is a limit value for each location then the limit state function can be formed as
follows
[ ] (511)
with this criterion designer can control the developed deformation on the system and improve
the structure so as to limit the displacements to the required level and subsequently increase
the reliability
On the other hand such an approach might exhibit several problems during the reliability
analysis that are listed below
Many failure mechanisms suddenly happen and as a result no significant displacements
are observed before moving very close to failure In sequence this may cause problems
for the iterative procedures of some reliability methods such as FORM and DS
The determination of suitable maximum admissible deformations is not
straightforward They have to be large enough to serve as failure criterion and
simultaneously they shall not be larger than the values that can be calculated within
the limits of equilibrium in FEM-calculations This requires previous knowledge on the
analysed system as well as on the feasibility of FEM-calculations on the specific subject
Last but not least the location of the undesirable displacements in a dike section is also
under investigation as there can be multiple vulnerable spots on the dike that should
be kept under a certain range of deformations in order not to cause a collapse and that
they are not always known in advance
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 64
2 φ-c Reduction
An available type of calculation in Plaxis is the so-called ldquoSafety Calculationrdquo with which global
safety factors can be computed In Appendix B2 a more detailed elaboration of this method is
exhibited This feature can also be used in reliability analysis by formulating the LSF as follows
(512)
where is the Multiplier Safety Factor obtained by the φ-c reduction The main concept is
that if the safety factor is smaller than 1 then it is considered as failure This method can provide
us with the probability of failure for a general soil body collapse However even if this method
seems to be quite simple there are some issues during its implementation in a reliability
analysis
For complex limit state functions there can be convergence problems for some
reliability methods with iterative procedures (ie FORM) This can be explained from the
fact that in a φ-c reduction soil strength properties follows a certain path as it is
described in Appendix B2 independently of the values for the random variables that
have been set from the reliability method
The safety factor is a general safety factor regarding the failure of the system
Therefore sometimes it is not straightforward what the ldquotriggerrdquo factors are that led to a
certain type of failure
A safety calculation in Plaxis is time-consuming and the outcome can be unstable (Plaxis
calculationsrsquo convergence depends on the number of calculation steps)
Plaxis cannot handle safety factors below 1 Therefore a new limit of the safety factor
shall be introduced in this limit state function
3 Relative Shear Resistance
In this method the basic idea is to define soil failure according to the failure criterion that Plaxis
is set to use for the analysis In this research the Mohr Coulomb failure criterion is applied in
combination with a linear ndashelastic perfectly-plastic stress-strain relationship (that is the so-
called Mohr-Coulomb model) For this model the initiation of plasticity is considered as failure
Therefore the maximum shear resistance is defined just before plastic yielding occur for any
given stress state Of course the occurrence of plasticity does not directly indicate the failure of
the soil structure however this model can give a first estimate for the stress state and the
deformations
According to this method the relative shear resistance is defined as the ratio between the
mobilized shear stress and the maximum shear resistance In Figure 510 the Mohr-Coulomb
model is illustrated in a Cartesian stress-plane In the graph the principle stresses σ1 and σ3 are
indicated More precisely the principle stresses are defined as follows
radic
(513)
(514)
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 65
radic
(515)
Figure 510 Mohrs stress circle and Mohr-Coulomb failure criterion
Hence the mobilized shear stress is
radic
(516)
And the maximum shear resistance which is the distance of the mean stress point to the yield
surface is defined as
(517)
Then the relative shear resistance is a measure for load-resistance ratio that can be determined
in any integration point in the soil continuum
(518)
However this criterion requires prior knowledge of the possible relevant failure mechanisms
and their potential locations on the soil structure This is necessary in this method in order to
choose a suitable cluster of integration points where the average value of the relative shear
strength is to be determined Therefore the limit state function to be considered is not
straightforward and it needs problem investigation in advance
4 Plaxis definition of soil collapse
In Plaxis the construction stages are analysed by performing a Load advancement ultimate level
procedure which is controlled by a total multiplier
(519)
where is the load that is applied by Plaxis is the load at the beginning of the
calculation phase (ie the load that has been reached at the end of the previous calculation
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 66
phase) and is the defined load configuration (ie water level soil weight vertical load
on the top of the dike etc) Thus the main idea is that Plaxis applies an incremental load until it
reaches the defined one When has reached an ultimate level (which by default is 1)
the current phase is finished However if a staged construction calculation has not properly
finished the multiplier fails to converge to the desired ultimate level at the end of the
phase and a warning of ldquoSoil collapsesrdquo appears in the Log info box
In other words a collapse load has been reached In this case the total specified load has not
been applied In physical terms this means that the current value of the stiffness parameter CSP
is less than 0015 (Plaxis 2D 2015b) It is a measure for the amount of plasticity that occurs
during the calculation When the solution is fully elastic CSP is equal to unity whereas at failure
it approaches zero Therefore soil reaches its upper limit of plasticity and it collapses which can
be visualized as a settlement of the dike body Such a warning in the analysis is thus assumed to
be a possible failure situation
53 Overview In this chapter an overview of the relevant possible LSF of the structural elements and the dike
body respectively has been given The different LSFs are formulated and discussed whereas the
failure of each component is going to be evaluated in the ULS of each component
The system failure in this research is considered as the combination of the anchor sheet pile
wall and soil body LSF Particularly for the anchor and the sheet pile failure their yield stress
was chosen as a limitation for their failure by adopting the next LSFs
[(
)] sheet pile wall LSF (520)
anchor LSF (521)
As far as the soil body is concerned in this research method 4 is to be used in order to identify
soil failure The limit state function is then formed as a single value in case of a successful and an
unsuccessful computation Therefore the limit state function was chosen to be equal to 1
(actually the value of is retrieved from Plaxis that in case of a successful calculation is
1) in case of a successful Plaxis analysis and -1 in case that soil collapses in any phase (error in
Plaxis Log file ldquoSoil Collapsesrdquo) This formula can be written as follows
(522)
This accounts for the global instability of the dike body in contrast with the other methods that
are related to a certain dike area (methods 1 and 3) or they are cannot handle overall failure
due to numerical issues (method 2) More precisely methods 1 and 3 were excluded due to the
prior knowledge that is required in method 3 while in method 1 a definition of the maximum
admissible deformations shall be first introduced However attention shall be paid to the type
of the warning that Plaxis gives because it might also be that the maximum number of load steps
was not sufficient In that case the phase must be recalculated using a larger number of steps
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 67
Someone could argue that such a failure can be considered as a system failure rather than a soil
failure alone Partially this is true as the incapability of the sheet pile wall or the anchor to keep
the upstream part of the dike stable can lead to non-convergence and thus according to Plaxis
definition to failure This incapability can be translated into various scenarios of failures that
stem from either the structural elements or the soil body weakness and that are illustrated in
Figure 511 This picture shows the different scenarios that can take place due to macro
instability that was discussed in 523 and the failure of the retaining wallrsquos elements
Figure 511 Combined failure scenarios (initial and developed mechanisms) (source Larsen et al 2013)
However this does not mean that the sheet pile wall or the anchor fail because of exceeding
their yield stress but rather due to possible unacceptable deformations This is also an
advantage of the inherent residual strength of the structural elements that allows a large
deformation before they fail However the residual strength of the elements was not taken into
account in this thesis as they were considered as elastic Moreover the message for the ldquosoil
collapsesrdquo can appear in a phase prior to the installation of the wall which means that this type
of failure is exclusively attributed to the deficiency of the soil strength Therefore it is not
always straightforward what is failing first and lead to the system failure Consequently by
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 68
considering such an analysis representative of a system failure it is likely to underestimate the
overall probability of failure whereas if it is considered as a soil failure only it might
overestimate the total probability of failure since it is later combined with the rest LSF of the
anchor and the sheet pile wall In this research it will be considered only for the soil failure that
brings the results to the safe side even if it does not lead to the best optimization In this thesis a
serial system of the anchor the sheet pile and the soil body LSF was considered the fault tree of
which is depicted in Figure 53
Another way to evaluate the reliability of the different elements could be the reach of a
threshold displacement according to the SLS criteria Such a threshold could be also considered
for the system as a whole rather than the individual elements However in that case someone
should be cautious regarding the choice of the value for the maximum deformations as well as
the location of their occurrence
Concluding in this section the possible failure mechanisms of a flood defence (Figure 58) were
shown Then it was clarified which of these mechanisms can be modelled in Plaxis and they
were further explained and schematized in Figure 58 with the presence of the retaining wall as
they would have to be handled in case of hand calculations Plaxis can incorporate these
mechanisms automatically and thus no distinguish among the different patterns of these
specific modes is necessary Last but not least in Figure 511 the possible failure modes for the
system of the dike and the retaining wall that can be simulated in Plaxis are depicted In this
thesis the plasticity of the structural elements is not considered and thus possible plastic hinges
in the anchor and the sheet pile are not taken into account Therefore scenario 3 is excluded
In Chapter 7 the results after the reliability analysis of the aforementioned LSFs are given It
was essential though before proceeding with the reliability analysis to get a better
understanding into the system under investigation For this purpose in the next chapter a
description of the case study as it has been simulated in Plaxis is given while also the different
structural properties are defined for the following reliability analysis The soil variables to be
considered as stochastic are presented and a first insight into of the systemrsquos behaviour is
obtained through the mean values calculation
MSc Thesis A Rippi 69
6 Case Study-Dike with an anchored sheet pile wall
In this chapter a description of the case study is presented together with the relevant soil and
structural parameters In sequence a deterministic analysis follows in order to specify the basic
structural characteristics and obtain a first sense of the stresses magnitude developed on the
structure This was carried out based on the recommendations given in the CUR 166 (2005)
The aim of this deterministic analysis is to redefine the structural properties in order to avoid
having an overdesigned structure understand the current design procedure and see if there are
any possibilities of improvement by applying the proposed reliability analysis
61 Case Description The case study concerns a dike section with an anchored sheet pile wall which was initially used
to showcase the current methodology of designing dikes with sheet piles using partial factors
For more information a reference is made to Breedeveld (2011)
In this research this case study has been modified and simplified for the needs of the thesis and
it is illustrated in Figure 61 while in Table 61 the soil materials are identified Details for the
properties of the soil layers are given in section 62 We assume the structure to be
homogeneous to the third dimension and therefore a plane-strain model is applied The dike
soil materials (layers 3 and 4 in Figure 61) were adopted from the case study used in
Breedeveld (2011) that was presented in Chapter 2 Especially in Figure 25 the original case
study as it was modeled in Plaxis is shown
Table 61 Soil materials that corresponds to Figure 61
1 Aquifer Sand layer 2 Aquitard Clay layer 3 Dike new material Clay layer 4 Dike old material Clay layer
The geometrical properties the groundwater level and the design water level were taken
identical to the original case study as well as the soil properties of the dike materials The soil
layers under the dike were modified for the need of the thesis and were modeled using random
average quantities for which the statistics were chosen arbitrarily but in a realistic range
according to the NEN6740- Table 1 (see Appendix C) included in the Dutch code for
geotechnical structures Moreover the vertical load was removed in order to examine the
impact of the water elevation on the dike stability
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 70
Figure 61 Case study Dike with anchored sheet pile wall Dimensions and Soil materials
62 Soil Parameters The soil parameters for the dike material were obtained from the report of Breedeveld (2011)
where for some parameters the characteristic values were available whereas for some others
the mean values were given For the soil layers below the dike Table 1 of NEN6740 was used as
it was mentioned before which gives characteristic values For those parameters whose value
was a characteristic 18 one a transformation was made according to their probability
distributions in order to acquire the mean values that are required for the reliability analysis
Only the properties required for the use of the Mohr-Coulomb model are presented
The parameters of each soil layer are presented in Tables 63(a) and (b) and the front number of
each layer indicates the corresponding soil layer as showed in Figure 61
The distribution types and the coefficients of variation for each parameter were chosen
according to the knowledge that has been obtained until now about the physically possible
ranges of such parameters and the recommendations that have been given in several researches
up to now (Baecheramp Christian 2003 Schweckendiek 2006 GriffithsampFenton 2007 Phoon
2008 Wolters 2012 Bach 2014) The distributions and the CoV of all parameters are displayed
in Table 62 They were considered to be the same for all soil layers Particularly for the
saturated and unsaturated volumetric weight γsat and γunsat respectively a relationship was
established in order to derive the one from the other In general γunsat varies between the real
dry weight and γsat depending each time on the degree of saturation An estimated maximum
difference between γunsat and γsat is approximately 4 kNm3 (Brinkgreve 2015) In this thesis γsat
was expressed as the summation of γunsat and a variable with uniform distribution in the range
of [0 2]19 Initially a deterministic relationship was defined between the two variables but later
on it was proved the variable was important for the soil failure and thus it was decided to
consider it as random
18 for the definition of the characteristic value see Appendix E 19 The range of [0 4] was not chosen in order to avoid extreme conditions
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 71
Table 62 Soil parameter distributions
Soil parameter Symbol DistributionRelation COV Unit
Unsaturated
Volumetric weight γunsat Normal (microσ) 5 [kNm3]
Saturated Volumetric
weight γsat γsat= γunsat+U(02) 5 [kNm3]
Cohesion c Lognormal (microσ0) 20 [kPa]
Friction angle φ Truncated normal (microσ045) 10 [ ˚]
Youngrsquos Modulus E Lognormal (microσ0) 25 [kPa]
Poissonrsquos ratio ν Truncated normal (microσ005) 10 [-]
Interface strength Rinter Truncated normal (microσ0099) 20 [-]
Table 63 (a) Soil statistical properties under the dike and (b) soil statistical properties in the dike
(a) 1 SAND (very silty)
Soil property Symbol Characteristic
value
Mean
value STD Unit
Saturated
Volumetric
weight
γsat 20 22 11 [kNm3]
Unsaturated
Volumetric
weight
γunsat 19 21 103 [kNm3]
Cohesion c 0 0 0 [kPa]
Friction angle φ 30 36 36 [ ˚]
Dilatancy angle ψ 0 0 0 [ ˚]
Youngrsquos Modulus E 20000 30769 7692 [kPa]
Poissonrsquos ratio ν 025 03 003 [-]
Interface strength Rinter 044 066 013 [-]
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 72
2 CLAY (clean medium)
Soil property Symbol Characteristic
value
Mean
value STD Unit
Saturated
Volumetric
weight
γsat 17 185 093 [kNm3]
Unsaturated
Volumetric
weight
γunsat 17 185 093 [kNm3]
Cohesion c 10 14 282 [kPa]
Friction angle φ 175 21 21 [ ˚]
Dilatancy angle ψ 0 0 0 [ ˚]
Youngrsquos Modulus E 2000 3077 769 [kPa]
Poissonrsquos ratio ν 029 035 0035 [-]
Interface strength Rinter 034 05 01 [-]
(b) 3 DIKE NEW (very sandy clay)
Soil property Symbol Characteristic
value
Mean
value STD Unit
Saturated
Volumetric
weight
γsat 17 185 093 [kNm3]
Unsaturated
Volumetric
weight
γunsat 17 185 093 [kNm3]
Cohesion c 4 564 113 [kPa]
Friction angle φ 29 347 347 [ ˚]
Dilatancy angle ψ 0 0 0 [ ˚]
Youngrsquos Modulus E 1625 2500 625 [kPa]
Poissonrsquos ratio ν 029 035 0035 [-]
Interface strength Rinter 034 05 01 [-]
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 73
4 DIKE OLD (little sandy medium clay)
Soil property Symbol Characteristic
value
Mean
value STD Unit
Saturated
Volumetric
weight
γsat 195 212 106 [kNm3]
Unsaturated
Volumetric
weight
γunsat 19 207 103 [kNm3]
Cohesion c 13 183 367 [kPa]
Friction angle φ 28 335 335 [ ˚]
Dilatancy angle ψ 0 0 0 [ ˚]
Youngrsquos Modulus E 2925 4500 1125 [kPa]
Poissonrsquos ratio ν 029 035 0035 [-]
Interface strength Rinter 034 05 01 [-]
63 Finite Element Model The structure has been modelled with the finite element software Plaxis 2D 2015 In Figure 62
the model as it was created and the automatically generated mesh are presented The ldquoboxrdquo
around the dike section was used to refine the mesh in this area The mesh in the interfaces
between the sheet pile and the soil was also refined for a better representation of the potential
high stresses that can develop in that area
Figure 62 FEM model and generated mesh for case study
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 74
The sheet pile has been modelled with elastic20 plate elements and the free anchor length with
an elastic fixed-end-anchor element The grout body of the anchor was not explicitly modelled
but it is implied in the end of the anchor as its final point is considered to be fixed in the soil A
fixed-end-anchor was chosen because even if it is a simplification of an anchorage system it can
still give reliable results in terms of the anchor axial force which is actually included in the limit
state function However the interaction of the soil and the grout body in that case cannot be
simulated The interaction and the modelling of a grouted body in the soil are still under
investigation due to several difficulties in modelling the real conditions
For all the steel members a Youngrsquos modulus of E=210 GPa was used while the specific
structural parameters are to be specified after the deterministic analysis that is described in the
next section
64 Deterministic Analysis The design concept of dikes with sheet piles modelled in FEM is based on the combination of
several partial factors in order to evaluate the overall required safety factor Additionally the
design recommendations of CUR 166 (2005) are considered for using the design or the
characteristic values of the structural parameters and the loads The design procedure (see
section 22) that has been initially followed in this structure is reported in Breedeveld (2011)
and the following structural elements have been defined accordingly
Cross sectional area of the sheet pile (per meter)
Length of the sheet pile wall
Steel quality of the piles
Cross sectional area of the anchor (per meter)
Free length of the anchor (without the grounded body)
Steel quality of the anchorrsquos reinforcement
Profile of waling
The required overall safety factor was calculated up to 18 according to the current design
procedure described in section 22 according which the above mentioned structural parameters
were defined Moreover apart from the overall safety factor in Breedeveld (2011) additional
partial factors are applied on the deduced forces and bending moments of the anchor and the
sheet pile wall as they are defined in CUR 166 (2005) In particular in Table 64 the different
partial factors that are applied on the several structural developed actions are given
Table 64 Partial factors for the different structural actions according to CUR 166 (2005)
Structural action Partial factor γ
Sheet pile normal force Fn 115
Sheet pile bending moment M 115
Anchor normal force FA 125
Waling bending moment Mw 110
20 The elements were considered as elastic in order to avoid numerical problems in FORM analysis For a better explanation a reference is made to the evaluation of the optimization algorithms in Appendix A2
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 75
However as it is mentioned in section 62 alterations have been made in the soil layers below
the dike while also the vertical load has been removed On the top of that in section 22 it was
shown that the large applied partial factors led to an overestimation of the design values for the
moments of the sheet pile Therefore a redesign of the sheet pile and the anchor is carried out
in this section in order to determine the new required characteristics of the structural elements
according to the new loading situation with less strict partial factors In particular in this thesis
the required overall safety factor was considered 12 according to CUR 166 (2005) calculation
scheme that is described in the next section The reason for using a lower safety factor was to
avoid an overestimation of the design forces and moments of the sheet pile wall that would
render a reliability analysis meaningless as the structure would be quite safe Such an analysis is
also important in order to show the discrepancy between the two different design procedures
and indicate the optimal one
However the reliability analysis was carried out for the mean values of the soil parameters and
for a safety factor of 10 in order to account only for soil properties uncertainties A safety factor
of 10 was not directly used as the design procedure followed in section 643 accounted only for
the macro-instability of the system and thus other failure mechanisms such as piping
overtopping infiltration etc were not taken into account In that way we prevent coming up
with a retaining wall that would be volatile under other crucial failure mechanisms
In section 641 the calculation scheme to be used in this thesis for the deterministic analysis is
further described In the sequence in section 642 the construction phases that were followed
in Plaxis simulation are listed and illustrated and finally in section 643 the results of the
preliminary design are displayed
641 Calculation Scheme and Design Values
In this research only the design philosophy suggested in CUR 166 is applied and the desired
safety factor of the structure is considered to be 12 for structures classified in ldquosafety class IIIrdquo
in order to avoid an overestimation of the structurersquos dimensions as it happened in the original
case study At this point it should be also clarified that the redesign was conducted considering
only the overall stability of the structure and no other failure mechanisms such as piping
internal erosion or overtopping
The prescribed calculation schemes in CUR 166 (2005) for designing structures with retaining
walls are basically the following
Calculation scheme A Calculations with design values The calculations are executed
using the design values for the soil parameters retaining height water levels and
stiffness of the structure It uses two different soil stiffnesses When using a high soil
stiffness the anchor force is generally relatively high The advantage of this scheme is
that it requires relative little effort The disadvantage is that the deformations may be
overestimated due to the use of design values in every stage However a designer is not
interested in deformations when performing a ULS calculation
Calculation scheme B Calculations with characteristic values The calculations are
executed using characteristic values of the soil strength (crsquo and φrsquo) Design values are
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 76
used for the retaining height water levels external loads and stiffness of the soil Here
again two different soil stiffnesses are used In the end of the governing stage a phi-c
reduction should be carried out As the characteristic values for the soil parameters are
used the deformations are most probably smaller than in case of calculation scheme A
The safety margin lies in the fact that phi-c reduction is imposed to obtain a safety factor
of 115 (class II) or 12 (class III)
For the soil stiffness parameters two options are given in CUR 166 a low and a high value
(61)
where is the (low and the high) design value the (low and the high) characteristic value
and is the mean value of the stiffness parameter In Woltersrsquos master thesis (2012) both
have been applied and the appeared to be governing for the bending moment in the wall
and the anchor force
In this research the Calculation scheme B was used and the as the soil stiffness
parameter It should be mentioned that regarding the soil parameters the characteristic values
were available with an exemption for the stiffness for which the mean values were given The
groundwater level was manually schematized according to TAW (2004) because of lack of data
while the maximum water level in the river side was taken equal to the design water level
according the Dutch regulations for the specific dike region
In Tables 65 and 66 the partial factors according to CUR 166 and the corresponding design
values for the different soil properties are presented As someone can notice in these tables
only the strength parameters (phi and c) and the stiffness parameter (E) are exhibited For the
rest parameters the design value is identical to the characteristic value (see Tables 63(a) and
(b)) as a partial factor of 1 is used
Table 65 Partial factors and design values for the soil layers under the dike
1 SAND (very silty)
Soil property Symbol Partial factor γk Design value Unit
Cohesion c 11 0 [kPa]
Friction angle φ 12 2569 [ ˚]
Youngrsquos Modulus E 13 1538462 [kPa]
2 CLAY (medium)
Soil property Symbol Partial factor γk Design value Unit
Cohesion c 11 91 [kPa]
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 77
Friction angle φ 12 1472 [ ˚]
Youngrsquos Modulus E 13 153846 [kPa]
Table 66 Partial factors and design values for the soil materials in the dike
3 DIKE NEW (medium clay)
Soil property Symbol Partial factor γk Design value Unit
Cohesion c 11 364 [kPa]
Friction angle φ 12 2479 [ ˚]
Youngrsquos Modulus E 13 1250 [kPa]
4 DIKE OLD (stiff clay)
Soil property Symbol Partial factor γk Design value Unit
Cohesion c 11 1182 [kPa]
Friction angle φ 12 239 [ ˚]
Youngrsquos Modulus E 13 2250 [kPa]
642 Construction Stages
The construction and the gradual loading of the dike ware modelled as follows (see Figure 63)
1 K0-procedure for the generation of the initial stresses under horizontal groundwater
level
2 Dike self-weight under horizontal groundwater level
3 Rise groundwater level to the phreatic water elevation
4 Interpolation21 of the clay layer under the dike and activation of the sheet pile wall and
the anchor
5 Apply extreme water level conditions on the dike structure
6 (φ-c reduction for the determination of the safety factor)
21 Due to different permeability between the clay and the sand layer an approximation of the developed pore pressures is carried out through interpolation between the two layers
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 78
1
2
3
4
5
Figure 63 Construction and loading stages
For the soil modelling the Mohr-Coulomb model was used with a non-associative flow rule
(φneψ) For phases 1-4 the calculations were conducted under drained conditions whereas in
phases 5 and 6 the undrained behaviour of clay layers was considered
643 Determination of the structural elementsrsquo characteristics
According to the calculation scheme that was described above the design characteristics of the
structural elements were obtained from phase 6 In particular the target overall safety factor is
12 with which initially the length of the sheet pile wall was calculated
+370 NAP
+700 NAP
+1050 NAP
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 79
Firstly a length of 10 m and a profile of AZ 12 with a steel quality of 240 kNm2 yield stress
were used while for the anchor a cross-section of 628 mm2 and S355 were chosen for the first
calculation However even if the safety factor reached 12 it seemed to be quite unstable during
the calculation steps while the developed anchor stress was above the ultimate yield stress
What is more the deformations resulted in that case were more than 30 cm Therefore in order
to stabilize the safety factor and to decrease the displacements the length of the sheet pile was
increased 1 m (so 11 m in total) while also a better steel quality for the sheet pile was chosen
S355 Furthermore the anchor cross section was increased to 933 mm2 with a steel quality of
MW450 Last but not least the waling was designed based on the anchor force
In the following paragraphs the design procedure of each element is described providing also
the results of moments and forces that were retrieved from Plaxis
Sheet Pile wall
As it was partially mentioned in the previous paragraph the length of the sheet piles is
considered to be sufficient when the safety factor of the system22 reaches at least 12 For the
reasons discussed above the sheet pile length was decided to be 11 m and with a profile of AZ
12 S355
With this configuration the safety factor reached is 13 as it is depicted at the right graph of
Figure 64 Someone can notice that the safety factor from 117 in the situation without a sheet
pile structure increased to 13 with the structure
Figure 64 Safety factor deduced from the phi-c reduction calculation without the anchored sheet pile wall at the left (safety factor=117) and with the wall at the right (safety factor=13) The steps on the x-axis represent the calculation steps have been carried out in order to reach an equilibrium after incrementally reducing the strength parameters ie phi and c
The design moment is to be taken for a safety factor of 12 that is reached in step 34 where it
was found to be 1519 kNmm and the corresponding axial force 9133 kNm Therefore
considering Eq 52 for checking the exceedance of the ultimate yield stress and applying the
prescribed partial factor of 115 for both the moment and the axial force we get
22 Plaxis always gives the safety factor of the system which is under investigation
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 80
In Figure 65 the maximum bending moments with a safety factor 10 and 12 are respectively
showed From the values being shown in that figure someone can easily distinguish the
difference between the two moments It must be also mentioned in that point that the required
safety factor defined in Chapter 2 was estimated at 18 (and not 12 that is considered here)
which gives excessive design moments (almost 7 times larger) For that reason in this research
a lower safety factor was assumed avoiding thus the design of a quite conservative structure
Figure 65 Bending moments on the sheet pile for a safety factor safety factor=10 at the left and
safety factor=12 at the right
The explanation behind this discrepancy is that in case of the phi-c reduction procedure the
artificially decreasing strength of the soil and on the other hand the non-reduced strength and
stiffness23 parameters of the structural elements tends to lead in an increased arching
phenomenon that makes the structure attract the most of the load
In this research the reliability of the structure will be evaluated for a safety factor of 1 or else
during phase 5 and by considering the mean values of the soil parameters As someone can
notice from Figure 65 the magnitude of the maximum bending moment for a safety factor =10
is quite low which implies a possible minor contribution to the overall failure
At this point it is essential to mention that Mohr-Coulomb model is capable of simulating the
elastic behaviour and the state at failure Therefore in case that plasticity of the soil or the
structural elements plays an important role a different constitutive model is recommended to
be applied
Anchor
For a safety factor 12 the anchor force was found to be 10387 kNm Considering a mutual
anchor distance of 3 m the total axial force that is undertaken by the rod multiplied also with
the prescribed design partial factor is
23 ldquoStrengthrdquo and ldquostiffnessrdquo are usually confused terms Strength is a measure of the maximum load that can be placed on a material before it permanently deforms or breaks and it is represented by the yield stress σy in metals and by φ and c in soil Stiffness is a measure of the amount of deflection that a load causes in a material and it is represented by the Youngrsquos modulus E and the Poisson ratio ν both in soil and metals
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 81
Therefore choosing an anchor with a cross sectional area of A = 933 mm2 and a steel quality of
MW450 ( ) the design stress level is
In comparison with the sheet piles the anchor does develop large stresses as it was concluded
from the results in the deterministic calculations This is probably attributed to the high
stiffness of the anchor that apparently attracts the load and thus the anchor develops large axial
force
Waling
The type of the waling system is defined according to the anchor force as it was described in
section 522 Therefore by choosing a waling type of 2UPE 200 the developed bending
moment can be estimated as follows
where is the design bending moment of the walling system according to the chosen profile
In Table 67 the characteristics of each structural element that was described above are
displayed in detail
65 Overview In this chapter an introduction of the case study as it has been used in this research is made The
soil materials and their properties are summarized and their mean and design values are
defined The probability distributions of the soil parameters to be considered as stochastic were
proposed according to the literature study and the engineering judgement Moreover the mesh
generation and the modelling features are discussed
Last but not least due the observed overestimation of the retaining wallrsquos design moment and
axial force according to the current design regulation a new deterministic analysis was carried
out in order to redefine the structural properties and come up with a less conservative structure
This makes the reliability analysis of the system meaningful in a sense that an extremely safe
structure would give a failure probability of zero Besides the scope of a reliability analysis is
also the optimization of the construction that corresponds to a financially attractive and
simultaneously safe structure This can be made by conducting a risk analysis whose first step
would be the analysis that is carried out in this research Therefore the aim is to start with the
marginal required structural properties according to the safety standards In a risk analysis that
should conducted for several dimensions of the structure until the most optimized one is
identified However this is out of the scope of this thesis
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 82
In Table 67 the structural properties of the sheet pile wall the anchor and the waling are
summarized as they were found in the deterministic analysis
Table 67 Design parameters of structural elements
SHEET PILES
Property Symbol Value Unit
Profile - AZ 12 [-]
Steel quality - S355 [-]
Length L 11 [m]
Moment of inertia I 21430 [cm4m]
Elastic section modulus Wel 1245 [cm3m]
Mass w 943 [kgm]
Sheet pile thickness d 85 [mm]
Cross sectional area A 1201 [cm2m]
ANCHOR
Property Symbol Value Unit
Steel quality - MW450 [-]
Free length Lafree 104 [m]
angle φ 30 [deg]
Cross sectional area A 933 [mm2]
Mutual anchor distance s 3 [m]
WALING
Property Symbol Value Unit
Profile - 2UPE200 [-]
Steel quality - S355 [-]
Elastic section modulus Wel 191 [cm3m]
In the original case study the sheet pile wall is 162 m long and it has a profile of AZ 28 (S240)
while the anchor is an MW450 with a cross-section of 1134 mm2 As it is noticed both sheet
pilesrsquo and anchorrsquos properties have been reduced after the new deterministic analysis However
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 83
it should be mentioned that in the original case there was a vertical load on the top of the dike
that was displaced while the clay layers under the dike was merged into one in the new case
study These changes might lead to less heavy structure However the difference among the
design properties between the new and the original case study is mainly due to the required
safety factor that was assumed In the new case study the safety factor was set to 12 whereas in
the original cases study a safety factor of 18 was considered as it was calculated according to
the partial safety factors
In the next chapter the results after the reliability analysis of the just designed case study are
presented
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 84
MSc Thesis A Rippi 85
7 Reliability analysis results with stochastic soil properties
In this chapter the probability of failure of the different structural elements that were discussed
in Chapter 4 will be estimated by varying the soil properties of the several soil layers under and
inside the dike At the beginning a better insight into the system behavior is gained by
inspecting the deterministic Plaxis calculations for the mean values of the soil variables After
that a global SA is carried out in order to evaluate the most important soil parameters and
inspect their impact on the different structural elements in combination with each other and
individually Finally the results regarding the probability of failure of the anchor the sheet pile
wall and the soil are presented as they deduced by the reliability analysis (concerning the ULS of
the elements) Finally the systemrsquos reliability is evaluated The methodology that was followed
during the probabilistic analysis is described below and it is mainly dedicated to the DS
rationale that was used in order to obtain the influence factors and to evaluate the results of
Plaxis calculations
71 Method description For the calculation of the failure probability of the systemrsquos different components FORM and DS
were used for the retaining wall and the soil body respectively From these methods the
probability of failure and the important factors are going to be deduced for each component (ie
anchor sheet pile wall and soil body) However before starting with the reliability analysis it is
essential to carry out a sensitivity analysis in order to obtain a first impression of what are the
most influencing variables and thus determine the input random parameters for the reliability
analysis In the figure below the steps for the reliability analysis as they were followed in this
thesis are presented
Figure 71 Steps of the reliability analysis
Below the methodology of both the sensitivity and the reliability analysis are discussed and
explained in detail
Sensitivity analysis
In this thesis the FAST method was utilized for conducting the sensitivity analysis which is
further elaborated in Appendix A1 Here an explanation is given for the quantitative measure of
sensitivity which is represented by Sensitivity Indices The first-order sensitivity index of an
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 86
input variable pi is the measure of the main (direct) effect of pi on the output variance
(where i ne j) the second-order sensitivity indices measures the interaction effect of pi and pj on
the output variance Other higher-order indices are defined in the same manner The total
sensitivity index is the sum of all sensitivity indices involving factor pi (Homma amp Saltelli
1996) For example the total sensitivity index of factor 1 for a model with 3 input factors is
given as
The first-order (Si) and total ( ) indices can be interpreted as following (Pandya et al 2008)
high pi is an influent parameter
and both small pi is not an influent parameter (neither alone nor in interaction
with other parameters)
and nearly the same no interaction of pi with the other parameters
and very different high interaction of pi with the other parameters
Reliability analysis
As far as the FORM analysis is concerned section 422 gives sufficient information about the
methodology that is followed in order to obtain the failure probability and the influence factors
(ie α-values) The main concept and principles of DS are briefly explained in section 421
whereas here a more precise description of the implementation of DS and the how the α2-
values can be deduced is given
In particular DS is a method in which ldquodirectionsrdquo are sampled The concept of this method is
schematically depicted in Figure 72(a) For each sampled direction it is evaluated whereas
along this line the LSF Z equals to zero or not This procedure is repeated for a number of
directions and this method is applied in the standard normal space (u-space see also section
422) using equations to transform the standard normal variables u1 hellip un to their
corresponding ldquorealrdquo variables (in the x-space) X1 hellipXn and vice versa
(a) (b)
Figure 72 (a) Schematic view of the DS method in the u-space The numbers 0 1 2 3 4 show successive LSF evaluations along one direction (b) Schematic view of the line search algorithm β is the distance from the origin in the u-space along one direction The numbers 0 1 2 3 4 correspond to the ones shown in figure (a) (source Hydra-Ring 12 2015)
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 87
The search procedure for the location on the line where Z=0 is often referred to as the ldquoline
search methodrdquo and it is depicted in Figure 72 (a) and (b) In both figures the numbers 0-4 refer
to successive Z-function evaluations along one direction Figure 72(a) shows 3 evaluated
directions in the u-space whereas Figure 72(b) shows the (fictional) value of the Z-function
along one direction (where ||u|| is the distance from the origin along the line) The maximum
length of the direction line as well as the step size along the direction (for example the distance
between 0 and 1 along one line) can be steered in order to increase the efficiency of the method
and presumably decrease the computational time
In the sequence a method was developed in this thesis in order to estimate the α2-values of the
random variables out of a DS analysis as OT does not have an available method so far For that
purpose a transformation of the output samples to u-space was firstly carried out Then the
distance to the origin of all samples (in u-space) that are located on a direction where failure
(Z=0) was detected is calculated as follows
radicsum
i=1hellip (71)
where n is the number of random variables and uij is the ith ldquofailurerdquo sample of the jth random
variable is equivalent with the reliability index of this direction βi The α-values for each
random variable can then be calculated as (for more information about these values see
sections 422- Eq 417 and 423)
(72)
After that three different methods were created in order to evaluate the influence of the
random variables on the response of the LSF ie the α2-values (square of Eq 72)
Shortest distance ldquoβminrdquo
Average 10
Average all
In the Shortest distance method the sample with the smallest distance to the origin is
considered to be the design point and thus the corresponding influence factors are estimated
according to that point The smallest distance is also equivalent to the reliability index βmin This
method is well-known for estimating the design point out of MC techniques (Vrijling amp Van
Gelder 2002) However for a soil failure modeled in FEM this might not be always indicative of
the real influence that the soil properties can have on the system failure Therefore the last two
techniques were used as additional in order to get a better insight into the dominant soil
properties In Figure 73 a general scheme of the methodology that was followed for obtaining
the α2-values is depicted where also the different techniques are presented
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 88
Figure 73 DS methodology for calculating α2-values
The second method Average 10 takes into account the α2-values of the samples in the failure
domain whose distance to the origin lies within the 10 higher than βmin and averages them
Therefore it is firstly required the calculation of βmin as it was described for the Shortest distance
method Such a technique helps to identify the other directions that are close to the one which
gives the shortest distance to the origin and thus inspect if the important variables that were
indicated according to the first method are influential for other failure mechanisms as well
Last but not least the third method Average all averages the α2-values of all the samples
located in the failure domain This shows the overall contribution of the random variables to the
response of the LSF under investigation Concerning the soil failure LSF that is described in
section 73 such a method can be more suitable for distinguishing the soil properties that are
involved in the majority of the failure mechanisms rather than only in the one that the first
method indicates
72 Mean values calculations The results of the deterministic analysis that are presented in the previous chapter included a
degree of conservatism due to the load and material factors (partial safety factors) In order to
overcome this limitation structural reliability calculations were carried out and are discussed in
the next sections by considering relevant soil parameters as random However before starting
the probabilistic calculations by coupling Plaxis with OT that cannot be fully interpreted and
verified to a certain extent it is essential to first carry out some deterministic calculation Such
calculations can help to get a better insight into the behavior of the system and show
qualitatively what kind of response to wait later in the analyses These calculations were taken
place based on the mean values of the soil properties
Deformations
The analysis of this case study focuses on the ULS of the structure For the stresses and
deformation calculations the Mohr-Coulomb model is used that is generally accepted for the
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 89
simulation of the soil behavior until failure and for detecting the failure modes (a more detailed
explanation of this model is provided in Appendix B1) However this model is not so suitable
for analyzing the deformations of soft soils or the settlements around the wall
The deformation pattern that would more likely occur due to primarily the soil body failure can
be illustrated by conducting a ϕ-c reduction calculation In Figure 74 (a)-(c) the displacement
pattern after the ϕ-c reduction of the mean values of the friction angle and the cohesion
respectively As it can be seen a slip surface in the inner side of the dike has been created that
rotates towards the inland
(a) deformed mesh
(b) shadings
(c) arrows
Figure 74 Deformations after the ϕ-c reduction (mean values)
Stresses on the structural elements and the soil body
The anchor and the sheet pile wall are installed in phase 4 In phase 5 the anchor normal force
increases and especially for the mean values it amounts to Na = 7983kNm
As far as the sheet pile wall is concerned the bending moments and the normal forces that are
developed on the wall are indicated in Figure 75 as they were deduced from phase 5 Similarly
to the anchor the maximum stresses occur in phase 5 and are mainly located in the area of the
wall which is surrounded by the clay layer (brownish soil layer in Figure 74(a)) As someone
can observe the bending moments are not significant taking into account that the maximum one
observed reaches 2132 kNm
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 90
Figure 75 Bending moments and normal forces for the mean values calculation
For the interpretation of the stress field of the soil body the effective principle stresses and the
relative shear stress τrel (see Eq 517) were illustrated in Figure 76 (a) and (b) respectively In
Figure 76 (a) the crosses indicate the direction and the magnitude of the principle stresses σ1rsquo
and σ3rsquo and from which an active soil behavior (
) can be noticed on the river side of
the retaining wall whereas a passive behavior (
) is dominant in the inland side
(a) Effective principle stresses
(b) Relative
shear stress τrel
Figure 76 Effective principle stresses and relative shear stress in mean values calculation
In Figure 76 (b) the distribution of the relative shear strength shows the potential developing
plastic area in case of failure that expands from the river side of the wall and around it It can
also betoken the possible shape of formation of a slip surface that starts from the outer side of
the dike and goes around the wall
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 91
73 Sensitivity Analysis Results In section 312 different types of SA are presented and that can be applied through OT In this
research due to the time limit and the amount of parameters FAST method was used for
carrying out the SA as it allows the calculation of the first order sensitivity indices as well as the
total order indices and the computational time required is less than the other available methods
In this section the results from the conducted SA are presented separately for the anchor the
sheet pile and the soil performance Regarding the sheet pile wall and the anchor the sensitivity
of the total developed stresses towards the soil properties was evaluated according to Eq 53
and 55 respectively The influence of the soil parameters on the soil behaviour was evaluated
based on the calculation of the overall safety factor after a safety calculation (ϕ-c reduction
method) The total order indices are considered for distinguishing the most important
parameters while their difference with the first order indices Si indicates the level of interaction
effect amongst the parameters on the output variance
As far as the anchor and the sheet pile wall are concerned the SA was carried out in two steps
(1) SA with all the soil parameters (listed in Table 62) and (2) SA with the most important
parameters found in step 1 In Figure 77(a) and (b) the sensitivity indices from step (1) are
presented as they were found for the anchor and the sheet pile respectively Step 1 mainly
provides an overview of the contribution of the variablesrsquo variance to the total variance of the
limit state response and it helps to identify the most influencing parameters Step 2 was
basically carried out in order to build the response surface based on the SA output that was
later utilized in the FORM analysis as a complementary function in case of Plaxis errors At that
point it should be mentioned that the Dikenew material was not considered neither for the SA
nor for the reliability analysis of the structural elements as its influence was assumed to be
insignificant relative to the other soil layers However Dikenew was considered in case of the
soil body investigation
From Figure 77(a) it can be concluded that as far as the anchor is concerned the shear stiffness
G of the clay layer seems to be the most crucial soil property In particular the anchorrsquos stress
level seems also to be sensitive towards the Poisson ratio of the clay layer ν the friction angle φ
of the sand layer and the strength parameters φ c of the Dikeold material Moreover the
interface strength Rinter between the Dikeold material and the sheet pile wall seems to display
an additional notable influence as well The unit soil weight γ does not appear to affect the
anchor performance significantly
As far as the sheet pile wall is concerned the shear stiffness of the clay layer and the Dikeold
material seem to be dominant according to Figure 77(b) From these two soil layers it appears
that also the strength properties (φ c and Rinter) as well as the unit soil weight γ have a relative
impact on the stress level of the wall Furthermore the sand layer contributes via especially its
friction angle and shear stiffness
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 92
(a)
(b)
Figure 77 Total and First order sensitivity indices for the (a) anchor and the (b) sheet pile in step 1
As it was mentioned before step 2 was used for building the response surfaces for the anchor
and the sheet pile limit states The input random variables in this step are the most influencing
ones that were found in step 1 In Table 71 the soil properties that were considered as random
during step 2 are presented for both anchor and sheet pile It is essential to mention that the RS
was constructed by fitting a quadratic polynomial function to the sampling points of the SA
whose general expression is given in Eq 425 In Figure 78(a) and (b) an illustration of these
response surfaces is made However the figures depict the response of the anchor and the sheet
pile stress level over the fluctuation of only two variables amongst the 14 that are presented in
Table 71
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 93
Table 71 Soil properties considered for the RS construction of the anchor and the sheet pile respectively
Soil layer Anchor Sheet pile
Clay Sand
Dikeold
The linear regression of all the parameters was based on least squares and the R2 parameter
was found 0989 for the sheet pile and 0996 for the anchor Therefore the regression was quite
representative for the domain of the samples over which it was adjusted However there were
many residuals (ie difference between the observed and the predicted value) in the order of
103 and 104 which means that a possible expansion of this RS to an extended domain of the
input variables may lead to an underestimation of the limit state response
(a)
(b)
Figure 78 Polynomial response surface adjusted for the (a) anchor and (b) sheet pile stress level For the anchor the friction angle of the sand layer and the shear stiffness of the clay layer were accounted for this illustration whereas for the sheet pile wall the shear stiffness of the sand and the cohesion of the clay layer were used
The SA for the soil performance was also conducted in two steps However in that case there
was no need of a response surface construction since the LSF for the reliability analysis (Eq
522) takes the value of -1 in case of Plaxis error due to soil collapse Eq 522 though takes only
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 94
two different values -1 and 1 for soil collapse or not respectively which cannot be used for
carrying a SA For that reason the φ-c reduction method was used and the sensitivity of the
value of the safety factor (Msf) was investigated instead The second step of the SA was carried
out in order to divide the soil properties into two different groups of variables and thus
investigate the sensitivity of the soil in a more efficient way rather than including all the
parameters at one SA It should be mentioned that in the soil analysis soil properties of the
Dikenew material were added In Figure 79(a) and (b) the total and the first order indices are
depicted for step 1 and 2 respectively Step 1 includes all the stiffness and strength parameters
of all the soil layers whereas in step 2 the most important variables that were found in step 1
together with the unit soil weight γ are included Subsequently step 1 detects the most
influencing stiffness and strength properties and in step 2 the impact of γ is investigated
As it is observed in Figure 79(a) and (b) the safety factor appeared to be sensitive towards the
sandrsquos friction angle while also the strength properties (c and φ) and the shear stiffness (G) of
the clay layer play an important role Moreover the cohesion of the Dikeold material and the
unit soil weight of the clay sand and Dikenew layer seem to be influencing as well As it has
been mentioned above for this SA the LSF was different than the one considered in the
reliability analysis Therefore even if the SA at that point helps to evaluate in general the most
dominant soil properties the engineering judgement should also be included for the selection of
random variables that are going to be utilized in the reliability analysis
(a)
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 95
(b)
Figure 79 Sensitivity indices for the soil in step 1 and 2
It is also advisable that the difference between the total and the first order sensitivity indices be
investigated in order to have an insight into the most interactive variables The magnitude of the
interaction effect of a variable is crucial for the system behaviour because a variable might not
be important as a singularity but it is possible its combination with another variable to have a
considerable effect on the limit state under investigation Therefore this can be considered as
an additional criterion for choosing the set of the random variables to be used in the reliability
analysis In Figure 710(a) and (b) the difference between the two indices is presented for the
anchor and the sheet pile in (a) and the soil in (b) as it was deduced from the first step of the SA
Regarding the sheet pile wall the most interactive variables are the shear stiffness and the unit
weight of the clay layer the friction angle of the sand and the cohesion shear stiffness interface
strength and unit weight of the Dikeold material Concerning the anchorrsquos limit state the
cohesion of the Dikeold material seems to have the highest interaction with the rest of the
variables Last but not least regarding the soil performance the stiffness of all the soil layers
and the friction angle of the clay and the sand layer have a higher interaction effect amongst the
other parameters The importance of the aforementioned variables can be enhanced
considering both total indices and the difference that they display between their first and total
indices
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 96
(a)
(b)
Figure 710 Interaction effect amongst the parameters in (a) anchor and sheet pile and in (b) soil This effect is expressed as the difference between the total order indices T and the first order indices S as they were deduced from the first step of the SA for both anchorsheet pile and soil respectively
Considering the discussion above some preliminary conclusions can be drawn regarding the
influence of the several variables to the different components of the system
1 both anchor and sheet pile wallrsquos limit states are sensitive mostly towards the shear
stiffness of the soil
2 sheet pile wall stress level is also affected by the unit soil weight more than the anchor
does
3 soil body is mostly influenced by the strength properties and the unit soil weight
4 the interaction effect among the soil properties is more profound for the soil and the
sheet pile limit states rather than the anchorrsquos
It must be mentioned that this SA is a global one that gives an overview of the impact of the soil
parameters in terms of the system behaviour However near the design point a local sensitivity
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 97
can possibly give a better estimation of the variablesrsquo impact Near the design point the
structural elements as well as the soil body usually reaches plasticity that implies a non-linear
behaviour Such a behaviour is unlikely to be initiated during a SA that usually samples around
the mean values This can be counteracted by conducting more iterations so as to have a larger
amount of available samplings but this can make the SA to be quite time consuming and thus to
lose its benefit to give a fast preliminary estimation of the most influencing soil layers and
properties
Besides the purpose of a preliminary global SA is to reduce the number of the stochastic
parameters to a reasonable amount so as the initial reliability analysis to be more efficient and
computationally accessible The parameters that were eventually considered as stochastic in
each LSF are summarized in the table below and they were chosen according to both the SA and
the engineering judgement In the last column the stochastic variables used also for the
systemrsquos analysis are presented
Table 72 Stochastic variables for each LSF
Soil
Failure Anchor Failure
Sheet Pile failure
System failure
Sa
nd
γunsat radic
c
φ radic radic radic radic
G radic radic radic radic
v
radic
Rinter radic
Cla
y
γunsat radic
radic radic
c radic radic radic radic
φ radic radic radic radic
G radic radic radic radic
v
radic
Rinter radic radic radic
Dik
e_n
ew
γunsat radic
c radic
φ radic
G
v
Rinter
Dik
e_o
ld
γunsat
radic radic
c radic radic radic radic
φ radic radic radic radic
G
radic radic radic
v
radic
Rinter radic radic radic
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 98
74 Soil Shear Failure As far as the soil failure is concerned the limit state function that was implemented is given
below (see also sections 52 and 53) It was formulated as a single value depending on the
success of computation in Plaxis
(71)
The reliability method that was chosen in that case is DS due to the formulation of the limit state
function and the convergence issues of FORM There were used 12 input stochastic soil
parameters as it is indicated in Table 72 (see also input file of Soil Failure in Appendix D) The
variables were chosen based on the SA that was presented in the previous section For the soil
failure Poisson ratio was initially included in the analysis but it was noticed that values close to
04 during Plaxis calculations were leading to unrealistic behaviour of the dike More precisely
it has been observed that Poisson ratio equal to 04 can lead to swelling effects on the dike
which consequently lead to a raise of the dike crest instead of collapsing as it would be expected
in a soil failure analysis Therefore Poisson ratio was excluded from the reliability analysis of
the soil failure
The results of the reliability analysis according to DS are presented below In particular 300
iterations were carried out and 3 different failure directions were found Specifically in Table
73 the probability of failure the reliability index the number of iterations and the duration of
the analysis are firstly shown In the sequence the points from each failing direction with the
lowest reliability index are shown
Table 73 Reliability results of DS for the soil failure
DS Pf 1310-8 β 55 Number of LSF calls 1840 Number of iterations (or directions)
300
Elapsed time (hr) 34
Failure points (in real space) Parameter 1st direction 2nd direction 3rd direction Dikenew_phi [deg] 2277 2543 3021 Dikenew_c [kPa] 529 557 444 Dikenew_gammaunsat [kNm3]
1748 1848 1973
Dikenew_gammasat [kNm3]
1948 1984 1997
Sand_phi [deg] 2782 2039 3899 Sand_G [kPa] 808437 715473 715785 Sand_gammaunsat [kNm3]
2183 2122 2097
Sand_gammasat [kNm3] 2352 2145 2283 Clay_G [kPa] 56934 90063 80134 Clay_c [kPa] 704 1593 794
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 99
Clay_phi [deg] 1039 1629 1077 Clay_gammaunsat [kNm3]
1772 1413 1612
Clay_gammasat [kNm3] 1833 1433 1719 Dikeold_c [kPa] 1621 2978 915 Dikeold_phi [deg] 2474 3149 3281
According to the Shortest distance method the failure point from the 2nd direction turned out to
be the design point the importance factors of which are depicted in Figure 714 (first pie chart
at the left) However before concluding for the final design point an assessment of the above
failing directions was made
In particular the combination of the variables from each direction was later implemented in
Plaxis in order to illustrate the failure This is also necessary for verifying and assessing the
failure that Plaxis is giving and inspecting the error in the logfile in order to check if it is a ldquorealrdquo
(ie error=Soil collapses) or a ldquofakerdquo failure (ie error=Load advancement failed or Not enough
load steps) In Figure 711 the ldquorealrdquo failure is defined
Moreover a demonstration of the displacements is also essential in order to certify the failure
shape that normally has the dike crest settled down A different output of the displacements
formation (ie dike crest goes up) would probably imply a ldquofakerdquo failure However it should be
mentioned that a ldquofakerdquo failure in Mohr Coulomb model can be a ldquorealrdquo failure in another model
that can more realistically simulate the deformation patterns of the structure In general even
though Mohr Coulomb can be successfully applied for detecting failure it is not very suitable for
the determination of the displacements and especially for soft soils such that the dike under
investigation is constructed by It should be also mentioned at that point that in this case study
the system is subjected to unloading conditions24 (ie decrease of effective stresses) in soft soil
layers after phase 4 which cannot be modelled successfully by the Mohr-Coulomb model In that
case other constitutive models such as the Soft Soil Model and the Hardening Soil model are
presumably more reliable for simulating such situations For further information for this model
a reference is made to Plaxis (2015) Therefore someone should be always critical and verify
the failure points based on hisher engineering judgement
Figure 711 Prerequisites for real failure in Plaxis
24 From phase 4 to 5 there is an increase in water pressures and thus due to constant total stresses there is a decrease of the effective stresses This can be characterised as an unloading situation
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 100
In order to make it more understandable the different failure points were illustrated in Plaxis
and their deformed mesh and total displacements are presented in Figure 712 (a)-(c) for each
failure direction As one can observe in the first two directions the dike crest settles down
whereas in the 3rd one the dike body seems to swell and move to the right (the blue lines that
surrounds the dike soil layers indicate the initial dike geometry) However even if the 2nd
direction gave a normal pattern of displacements the error message was ldquoNot enough load
stepsrdquo The combination of the variables of the 3rd direction was then tested again with the
Hardening Soil model (HS) in order to illustrate the deformations with a more suitable model
and it was proved that the deformations seem to be realistic as it is depicted in Figure 713 With
Mohr Coulomb only the 1st direction seems to comply with the diagram in Figure 711 and that
would properly be considered as the real design point
(a) Deformed mesh
(b) Total displacements |u| in m (shadings)
(c) Total displacements direction (arrows)
1st direction 2nd direction 3rd direction
Figure 712 Illustration of the design point in soil body failure
(a) Deformed mesh (b) Total displacements |u| in m (shadings)
(c) Total displacements direction (arrows)
Figure 713 Displacements pattern of the 3rd failure direction according to HS
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 101
However taking into account that the other directions showed characteristics of failing behavior
they should also be accounted for the determination of the importance factors Excluding them
from the analysis would not be wise as the error message or the deformation schematization can
be a limitation of FEM or Mohr Coulomb model respectively In any case the points that lead to
failure shall be treated critically in order to decide if they should be included or not In this case
the points from the three directions seem to be realistic as both strength and stiffness
parameters are low enough in order to lead to a potential failure Therefore they were finally
considered for the analysis and the importance factors were averaged over all the points as it is
depicted in Figure 714 (last graph at the right)
Figure 714 Importance factors α2 for soil failure
Another option is also to average over the 10 range of close to the lowest reliability index that
was calculated with the shortest distance method (ie Average 10 method) in order to come
up with the α2 values However such methodology can lead to averaging over points of the same
direction that are not very different with each other Therefore the rest of the failure points are
automatically omitted from the consideration This can be also verified from Figure 714 where
the pie chart of the Shortest Distance and the Average 10 are quite similar whereas the Average
all chart introduces the importance of other soil variables as well
From a general perspective the soil properties of the clay layer seems to be determinant
whereas also sand and Dikeold material play an important role to the soil failure More precisely
according to the Average all method the unit weight and the friction angle of the clay layer
turned out to be the most influencing while also the cohesion of the Dikeold material and
friction angle of the sand layer contribute to the failure domain From a qualitative point of view
the weakening and consequently the settlement of the Dikeold material pushes the subsoil
creating an additional surcharge for the underlying clay and sand layer which act like the
foundation soil the incapability then of the clay layer to withstand the overlying load due to the
low unit weight in combination with the low friction angle of the clay as well as the low stiffness
and strength of the sand layer can lead to the creation of an inner slip surface and thus failure
Since the clay layer can be considered as a foundation soil the importance of the friction angle
and the unit weight can be also explained from the Terzaghirsquos bearing capacity theory (Terzaghi
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 102
1943) where these two factors are crucial for the determination of the ultimate bearing capacity
of the soil However these can be better verified by conducting large scale experiments in dikes
and inspecting the sensitivity of the dikersquos stability towards the soil properties
75 Anchor Failure The reliability of the support system is estimated based on the anchor and the waling probability
of failure However the load on the waling is proportional to the anchor force as it appeared in Eq
57 Therefore after the estimation of the anchor reliability by calculating the probability of
exceedance of a certain admissible anchor force the waling can be designed deterministically
Besides the failure probability of the waling must be lower than the anchorrsquos (given that no
uncertainties of the strength and the structural properties are considered) For the anchor failure
the LSF was formulated as follows
(72)
where [kN] is the anchor force is the yield stress and the cross
section area The reliability method to be utilized in this analysis is FORM and the selection of the
random variables to be used was made according to the sensitivity analysis results and the
engineering judgment and they are listed in Table 72 In particular the variables from the second
step of the sensitivity analysis were considered which will also help at a later stage to compare
FORM sensitivity method and FAST In Table 74 the results of the reliability analysis with FORM
are listed Additionally the design point together with the importance factors of each variable is
presented It should be mentioned at that point that the sensitivity analysis helped to define a
starting point closer to the design point and thus reduce the computational time required for the
analysis More precisely the analysis lasted approximately 16 hours whereas by starting from
mean values from which by default FORM starts the iterations it would take almost a day
Table 74 Reliability results for the anchor failure with FORM
FORM Pf 28410-6 β 45 Number of LSF calls 119 Maximum number of iterations
100
Elapsed time (hr) 16
Design point Parameter value α-value Clay_phi [deg] 1814 -0059 Clay_nu [-] 033 0086 Clay_c [kPa] 1487 -0089 Clay_G [kPa] 41739 0911 Clay_Rinter [-] 051 -0021 Sand_Rinter [-] 069 -0051 Sand_nu [-] 030 0007 Sand_phi [deg] 3259 -0174 Sand_G [kPa] 1130770 0048 Dikeold_phi [deg] 2842 -0020 Dikeold_nu [-] 036 -0045
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 103
Dikeold_c [kPa] 1873 -0045 Dikeold_G [kPa] 243183 -0326 Dikeold_Rinter [-] 053 -0072
Figure 715 Importance factors α2 in anchor failure from FORM analysis
The influence coefficients in Figure 715 indicate that this limit state is governed by the shear
stiffness of the clay layer G Furthermore the shear stiffness of the Dikeold material seems to
contribute significantly to the anchor yield stress exceedance It can be concluded that the
problem is still in the elastic domain as far as the soil behaviour is concerned In case of
predominately plastic behaviour the strength properties of the soil (φ and c) become more
important In Figure 716 a demonstration of the design point as it was acquired by the FORM
analysis is presented The figures indicate that presumably due to the clay layer withdrawal in the
passive side as it is depicted in Figure 716(b) with the red shadings sheet pile moves to the right
and subsequently anchor is tensioned enough to reach its yield stress25
(a) Deformed mesh (b) Total displacements |u| in m (shadings)
(c) Total displacements direction (arrows)
Figure 716 Design point illustration from the anchor failure
25 A better representation of the soil-anchor interaction could be possibly obtained if the anchor was modelled as a node-to-node element and the anchorage body as an embedded element
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 104
The calculations were repeated with the 6 most influencing parameters both with FORM and DS
in order to investigate the effect of reducing the number of random variables and in order to
validate the results of FORM analysis In Table 75 the reliability results of respectively FORM and
DS are presented while in Figure 717 the importance factors of both methods are juxtaposed The
design point and the importance factors of DS were estimated based on the Average 10 method
(see section 71)
Table 75 Reliability results of FORM and DS with reduced variables in anchor failure
FORM Pf 8510-6 DS Pf 1510-5 β 43 β 42 Number of LSF calls
28 Number of LSF calls
776
Maximum number of iterations
100 Number of iterations
100
Elapsed time (hr) 034 (starting point close to the design point)
Elapsed time (hr)
115
Design point Design point Parameter value α-value value α-value Clay_nu [-] 033 0119 033 0242 Clay_c [kPa] 1404 -0026 1535 -0212 Clay_G [kPa] 43236 0928 44548 0807 Sand_phi [deg] 3163 -0110 2954 0203 Dikeold_G [kPa] 239465 -0330 251723 -0366 Dikeold_Rinter [-] 053 -0064 060 -0263
Figure 717 Importance factors α2 with FORM and DS analysis respectively
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 105
As it can be noticed comparing FORM results from Table 74 and 75 the reliability index changes
from β=45 with 14 parameters to β=43 with the 6 most influencing parameters This verifies the
importance of the most influencing parameters that it was found in the analysis with the 14
parameters The small discrepancy though between the two reliability indexes might be
attributed to the interaction effect of the soil variables on the output performance
The estimation of failure probability with DS is almost the same with FORM As far as the
importance factors are concerned DS also came up with the conclusion that the shear stiffness of
the clay and the Dikeold material are the most crucial soil properties for the anchor stress level
However according to DS the stiffness of clay ν and the interface strength Rinter between the
Dikeold material and the sheet pile wall seem to also contribute to the limit state of the anchor
The reason why the strength of the clay layer under the dike and the Dikeold material (the part of
the dike that comes in contact with the anchor and the sheet pile wall) are the most important
variables can be explained by the principle of soil arching According to this principle the soil
columns on both sides of the rigid sheet pile wall are more compressive than the soil columns on
the top of the wall (ie overlying dike body) because of the higher stiffness of the wall when
compared with soils As such soil columns on both sides tend to settle more than the soils on top
of the rigid wall and this differential settlement causes a downward shear force acting along the
sides of soil columns on top of the wall As such the vertical load on the wall becomes larger than
the sole weight of soil columns on its top and the anchor that actually receives the most of this
vertical load reaches eventually its yielding stress
Last but not least it should be mentioned that the response surface which was constructed based
on the SA worked efficiently in most of the Plaxis errors cases in a sense that it gave reasonable
results of the anchor stress level
76 Sheet pile wall failure For the reliability analysis of the sheet pile wall the strength property ie the yield stress was
considered constant and equal to 355 kPa Therefore the LSF that was applied for this case was
formulated as follows
[
] (73)
where [kNm] and [kN] are the maximum bending moment and the axial force
that are exerted on the wall respectively is the elastic section modulus ( =1245 cm3m)
and is the cross sectional area ( =1201 cm2m for a AZ12 profile)
In sections 643 and 72 the maximum bending moments of the sheet pile wall were calculated
with the design values of the soil properties and the mean values respectively It has been
observed that the bending moments were not so significant and determinant in comparison with
the anchor force and the safety factor of the system that were also considered (in section 643) in
order to determine the structural properties
In a later stage a SA was carried out for the sheet pile wall that has been further discussed in
section 73 Figure 718 shows the stress level of the wall that was developed after 3150 iterations
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 106
by considering as random variables the ones that are defined in Table 72 As someone can
observe the stress level is generally quite low and even an order lower than the yield stress (ie
355105 Pa)
That small magnitude of bending moments can be probably attributed to the counteraction of the
active earth pressures from the passive ones that due to the homogeneity of the soil in both sides
of the wall are quite similar
Figure 718 Stress level [in Pa] of the sheet pile wall after 3150 iterations of the SA
The reliability analysis has been carried out with FORM method and considering as random
variables those listed in Table 72 However due to the low level of stresses that are developed
and the simultaneously high yield stress the method converged after almost 35 days of
calculations As it can be noticed in Table 76 the analysis came up with an extremely low
probability of failure It should be mentioned at that point that the convergence of the analysis
was successful thanks to the RS that was fitted This means that a Plaxis error of soil collapse
preceded and it activated the RS which implies that the soil failure is more likely to occur before
the sheet pile failure
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 107
Table 76 Reliability results for the sheet pile wall failure with FORM
FORM Pf 7210-38 β 13 Number of LSF calls 2619 Maximum number of iterations
100
Elapsed time (hr) 84 (35 days)
Design point Parameter value α-value Clay_phi [deg] 1164 0261 Clay_c [kPa] 586 0336 Clay_G [kPa] 14275 0663 Clay_Rinter [-] 052 -0016 Clay_gammaunsat [kNm3]
1848 0000
Clay_gammasat [kNm3] 1948 0000 Sand_phi [deg] 3020 0000 Sand_G [kPa] 1193650 0000 Dikeold_phi [deg] 1360 0406 Dikeold_c [kPa] 592 0438 Dikeold_G [kPa] 168798 0000 Dikeold_Rinter [-] 053 -0024 Dikeold_gammaunsat [kNm3]
2275 -0150
Dikeold_gammasat [kNm3]
2375 0000
This can be also verified from the importance factors that are depicted in Figure 719 where apart
from the shear stiffness of the clay layer which seems to be the predominant soil property for the
sheet pile LSF the strength parameters (φ and c) of the clay and the Dikeold materials also play
an important role This implies that the soil behaviour has already passed into its plastic domain
where the governing properties are the strength parameters before the sheet pile exceeds its
yield stress
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 108
Figure 719 Importance factors α2 of the sheet pile stress level in the plastic domain of soil
Nevertheless in order to investigate the potential most influencing soil properties for the sheet
pile LSF in the elastic domain as well the available output data of the FORM analysis have been
assessed The highest observed value of the sheet pile stress before its failure was around
235000 kPa After tracing the results it seems that when FORM tries to reduce further the
stiffness and strength properties of the soil layers soil collapses before the sheet pile failure The
vectors of the random variable that led to a stress level on the sheet pile wall in the range of
230000-235000 kPa were chosen in order to evaluate the importance factors The method that
was followed in that case is the Shortest Distance and the results are shown in Figure 720
Figure 720 Importance factors α2 of the sheet pile stress level in the elastic domain of soil
As it can be noticed from Figure 720 the results in the elastic domain regarding the influence of
the stochastic parameters are comparable with anchorrsquos analysis The stiffness parameter of the
clay layer (69) is the most dominant variable for the system behaviour whereas the stiffness of
the sand (15) and the Dikeold (4) layers seems also to have a significant influence This is not
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 109
surprising as the nature of the loads that generate high stresses on the anchor as well as the sheet
pile is the same namely the horizontal earth and water forces
Concluding in this case study the sheet pile wall is not expected to show a decisive behaviour to
the systemrsquos reliability according to the reliability analysis and the deterministic analysis (see
section 643) It should be stressed the fact that the fitted RS in a lot of calculations seemed not to
perform suitably for the sheet pile and this can be attributed to the fact that the it was created
based on the SA which was far from the plastic domain In that case a better fit of a RS is
recommended However such an action is still under investigation that for future considerations
in the design concept will be contemplated as a valuable technique for complex structures
modelled in FEM whose reliability analysis with conventional methods is time consuming and the
coupling may lead to numerical problems In the next section the analysis of the systemrsquos
reliability is presented
77 System Failure The estimation of the systemrsquos reliability was based on the systemrsquos fault tree of Figure 53
according to which the general LSF was formulated as the minimum of the three LSF that were
presented above for the soil the sheet pile and the anchor and it is given below
[ ]
[(
)]
(74)
According to this LSF the failure that occurs first is counted as a system failure The analysis was
carried out with DS reliability method and the variables that were taken as stochastic are listed in
Table 72 As it was discussed in section 51 the systemrsquos probability of failure differs from the
probability of flooding for which a combination between the systems reliability and the water
level uncertainty shall be made
In Table 77 the reliability analysis results are presented as they were deduced from DS Similarly
to the soil failure in Figure 721 the importance factors according to the Shortest Distance the 10
Average and the Average all methods are illustrated
Table 77 Reliability analysis results from the system failure
DS Pf 2410-3 β 30 Number of LSF calls 1807 Number of iterations (or directions)
300
Elapsed time (hr) 39
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 110
Figure 721 Importance factors α2 of the system failure
The analysis took almost one and a half days and it came with indicative valuable though results
about the system behaviour According to the averaged importance factors in Figure 721
(deduced by the Average all method) we can notice features from both soil and anchor failure
For example the unit weight (14) the friction angle (7) and the cohesion (8) of the clay
layer as well as the friction angle (8) and the stiffness (7) of the sand layer that played an
essential role in the soil failure analysis Moreover the stiffness of the clay (11) and the Dikeold
(9) layers that appeared to be important for the anchor and sheet pile LSF They seem to be also
essential in the system analysis
However the other two methods (Shortest Distance and Average 10) came up with the cohesion
and the interface strength of the Dikeold material to be the most influencing soil properties for
the system For that reason an investigation was conducted regarding the output results of the
system analysis in order to obtain an insight into the different reasons of the system failure More
precisely the failure points were divided into those that come from Plaxis Error and those that
come from the yield stress exceedance of the structural elements It was then observed that many
failure points originate from Plaxis errors that are related to ldquoNot enough load stepsrdquo (Error codes
102 and 112) rather than ldquoSoil seems to collapserdquo (Error code 101)
Therefore in order to estimate the actual important soil properties the vectors of the random
variables that led to errors of ldquoSoil seems to collapserdquo and to anchorsheet pile (SP) failure were
studied separately The importance factors are depicted in Figure 722 which actually verifies the
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 111
importance of the variables that were also deduced from the Averaged all method in Figure 721
(which are basically the unit weight and the stiffness of the clay layer and the friction angle of the
sand layer) It should be mentioned that the failure of the structural elements contributes 65
whereas the soil failure contributes 35 on the total probability of failure The latter would be
possibly lower than 2410-3 that was initially found as several failure points are not attributed to
real failure but to Plaxis numerical errors
Figure 722 Influence factors α2 for the structural elements (left) and the soil LSF (right) respectively
The system analysis can lead to valuable conclusions for the system behaviour as a whole and
individually for the different elements given though that an inspection of the results is followed
Moreover such an analysis is taking into account correlations among the three sub-failure
mechanisms implicitly For example the different LSFs may be affected by the same soil variable
such as the shear stiffness of the clay layer for the anchor and the sheet pile LSF In that case the
most dominant failure mechanism will occur first and the rest are excluded26 However during
the system analysis someone is not aware of what failed or not and therefore heshe is not
capable of distinguishing the weaknesses of the system that need further improvement Besides
with this technique is not also easy to inspect the failure mechanisms that occur in order to get
some knowledge about the systemrsquos behaviour Therefore the reliability analysis of the
components is recommended to be carried out separately if time is available
78 Comparison between Global Sensitivity Analysis and Local Sensitivity
Analysis (FORM)
The aim of this section is a comparison between the importance factors that were deduced in
section 75 according to a FORM analysis and the results of the global SA that were presented in
section 73 for the anchor failure
The output of the global SA is actually the total indices that show the influence of the random
variables The importance factors of FORM and the total indices of global SA cannot be directly
26 This method is not capable though of accounting for the simultaneous occurrence of different failure mechanisms as the system was considered a serial system
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 112
compared as they are different values More precisely the total index of a random variable is the
ratio between the total variance and the portion of the total variance that stems from the
uncertainty of the specific variable and its interactions with the rest (see Eq A10 in Appendix A)
On the other hand importance factors of FORM express the ratio between the normalized value of
the variable in the design point and the reliability index (see Eq 72) Therefore the comparison
between these values is carried out qualitatively in terms of the variables that are distinguished
as the most influencing
In particular according to global SA the shear stiffness G of the clay layer and the cohesion c of
the Dikeold material seem to be the most determinant soil properties Moreover the strength
reduction parameters Rinter of the Dikeold and the clay layers and the friction angle φ of the
Dikeold and the sand layer appear to have a relatively notable contribution to the anchors limit
state
On the other hand FORM analysis has concluded to the same results as the global SA with the
difference that Dikeold material affects the anchorrsquos stress level more with its shear stiffness
rather than its cohesion In Table 78 the results of both FORM and global SA are presented
separately for each soil layer and property (in blue for the global SA and in black for the FORM
analysis)
Table 78 Important soil properties according to global SA and FORM analysis for the anchor failure (radicradic=very important radic=less important but still considerable)
FORM global SA Sand Clay Dikeold
ν
radicradic
ϕ radicradic
radic
c radic radic radicradic
G radicradicradicradic radicradic
Rinter
radic radicradic
The validation of the most important parameters was carried out with the repetition of the
analysis with DS as it was showed in section 75 The small difference of the probability of failure
as well as the same tendency of the important factors of both DS and FORM verified the reliability
of FORM analysis Therefore the similarity of the results deduced according to global SA with the
ones given by FORM indicates the fact that in this case study the global SA should be considered
as a trustable way of SA for obtaining a first overview of the dominant variables that affect the
limit state under investigation Furthermore it is capable of giving a general insight into the
outputrsquos (in this case anchor stress level) behavior and the soil layers that can potentially stand as
predominant during the reliability analysis For example in order to obtain a first sense of the
reliability and the magnitude of the expected probability of failure someone can inspect the
output samples of the LSF deduced from the global SA In Figure 723 the samples of the anchor
stress level are presented as they were deduced after 3150 iterations during the global SA
(computation time =75 hr = 3 days) The fact that after 3150 iterations and variables
combinations the anchor did not reach its ultimate stress level (ie 450 kPa) implies that the
probability of failure might be small This can also help for qualitative decision making concerning
a systemrsquos behavior without conducting a reliability analysis Moreover global SA proved the high
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 113
interaction (large difference between first and total order indices) among the parameters that can
affect the reliability analysis and can help to decide the amount of the random variables
Figure 723 Anchor stress level during global SA (step 1)
However as someone can observe from Table 78 there are also some discrepancies concerning
the dominance of some variables This can be attributed to the fact that the influence of all
parameters in specific points such as the design point cannot be identified by the global SA
Moreover it is possible that the samples used in the SA could not reach the field where the
system has a plastic behavior and where presumably other variables may also affect the limit
state
As general conclusions for global SA some possibilities and limitations are listed below
Possibilities
bull Indicates the level of interaction effect on the output variance
bull Indicate the amount of the random variables that should be considered in the reliability
analysis
bull Gives a general insight into the output behavior
bull Valuable for future use in the construction of response surfaces
Limitations
bull Cannot rely on global SA for the evaluation of the most important parameters near the
design point
bull Difficult to reach plastic zone (large amount of samples are probably needed)
79 Discussion
In this chapter the reliability analysis results were presented for the soil the anchor the sheet
pile and eventually the system failure taking soil properties of the several soil layers as stochastic
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 114
Initially a SA was carried out for each system component in order to identify the major soil
properties and reduce the number of the stochastic parameters The reliability analysis was
performed either with FORM or DS and in both cases the design point and the importance factors
were estimated Eventually the list of subquestions that is proposed in section 11 was answered
The way they were treated and the corresponding conclusions are elaborated below
Which reliability methods are computationally tractable in combination with FEM
Before starting analyzing the reliability of the system it was essential to evaluate and
assess the performance of the available reliability methods in order to be aware of their
performance in terms of computational time and flexibility in different kind of LSFs After
a literature review and the implementation of some of them in simple LSF problems a
general overview has been obtained for the most applied reliability methods concerning
the previous knowledge required for the system behavior in order to provide the right
input parameters its accuracy and the computational effort Eventually FORM and DS
were chosen to be employed for this case study considering the no preliminary knowledge
of the LSF is required their reliable accuracy and the limited computational effort that
they have in comparison with other methods
The anchorrsquos and sheet pilersquos reliability has been evaluated with FORM method whereas
soil and systemrsquos reliability have been assessed by DS Carrying out a FORM analysis and
steering the algorithmrsquos parameters in order to converge proved to be more difficult than
it was expected However an efficient performance was finally achieved and the failure
probability of the anchor failure was successfully estimated together with the importance
factors (or else the α values) The analysis was repeated with the most important variables
and they were validated with DS The FORM analysis took approximately one and half
hour to converge (with 14 random variables) after providing a starting point that it was
estimated to be close to the design point whereas for the validation it took almost half an
hour On the other hand in case of the sheet pile wall FORM analysis took almost 4 days
The computational time also depends on the LSF to be approximated the number of
variable and the complexity of the problem that each time is analyzed As far as the soil
analysis is concerned DS took approximately one and a half day to converge for 300
iterations and 14 random variables
Which limit states are relevant for the envisaged application of retaining walls in dikes and
how can they be formulated using FEM analysis outcomes
The relevant LSFs of a dike with an anchored sheet pile wall were investigated For each
element ie the sheet pile the anchor and the soil body a separate LSF has been set up
The focus of this research was on the ULS that in case of the structural elements it was
represented by the ultimate yield stress of the anchor and the sheet pile respectively as
they were expressed in Eq 72 and 73 The determination of the soil failure has been also
elaborated and the available alternatives that can be considered in that case were
elaborated After a short evaluation the Plaxis definition of soil failure was selected that it
was formulated as it is shown in Eq 71 Finally the systemrsquos LSF was formed as a
combination of the aforementioned limit states More precisely the system was
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 115
considered as a serial and thus system failure occurs when one of the elements fails first
(see Eq 74)
How robust (convergence) are the tractable methods
The robustness of the different reliability methods can be controlled by the convergence
criteria that are defined by the user and depend on the required accuracy In case of FORM
the convergence criteria consist of a set of errors that has mostly to do with the
approximation of the reliability index and the limit state threshold whereas in DS the
convergence is controlled through the coefficient of variation (CoV)
In order to reassure the robustness of FORM the method has been validated with DS The
validation showed that the FORM method complies with the result of DS with a small
difference though in the design point and the importance factors FORM analysis seemed
to converge sufficiently and faster than DS However the choice of the optimization
algorithm plays a key role to the methodrsquos efficiency In this research the convergence
errors were manipulated for increasing the efficiency of the method More precisely it has
been observed that by relaxing the relative error and the constraint error (the explanation
of these errors and an evaluation of the different optimization algorithms are available in
Appendix A2) FORM convergence can be accelerated
The reliability analysis of the soil body and the system was performed with DS Similarly
to FORM analysis in order to improve the performance of the searching algorithm some
of the methodrsquos parameters should be stipulated for improving its efficiency and
reliability Therefore the maximum step size of the algorithms was tripled and it was
proved that for the same time period double iterations were feasible to be carried out
This does not mean that the larger the step size the better it is The choice of the step size
must be made wisely depending on the reliability index that it is expected and
respectively it is allowed to change for achieving a better performance In this case study
it was proved that after 300 iterations the CoV of DS tends to be high and it can therefore
still be improved by carrying out more iterations In Figure 724 an example of a DS
convergence graph concerning the system analysis is shown The middle red line shows
the probability estimation whereas the green lines show the boundaries At that example
the CoV was 040
DS seems to perform successfully both in soil failure with one LSF and in system failure
where 3 different LSF were included In such cases performing an approximation method
such as FORM would not be wise as the output from each iteration does not provide
information for going to the next that a FORM analysis normally needs In contrast it just
gives an indication of failure or not This means that only a random sampling method can
handle it and converge successfully as it was proved
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 116
Figure 724 Directional Sampling convergence graph (at confidence level 095) of the system LSF
What is the contribution of different uncertainties in the failure mechanisms of the system
According to the global SA and the reliability analysis of FORM and DS the friction angle
the cohesion and the shear stiffness properties appeared to have an influence on the
different LSFs This is not surprising as they all have large CoV of 10 20 and 25
respectively Thus their uncertainty generally showed an impact on the systemrsquos
behaviour It is also essential to mention that the global SA has shown that the uncertainty
of the interface strength Rinter (CoV=20) has also a notable contribution to the structural
elements LSFs However the FORM analysis which is more representative close to the
design point came up with a less important impact of that variable
The reliability analysis showed that as far as the structural elements (anchor and sheet
pile) are concerned the shear stiffness of the clay and the Dikeold material is determinant
and especially for the sheet pile the strength parameters of the soft soils seem also to play
a crucial role In contrast with the anchor where the soil body seems still to behave in an
elastic manner during the sheet pile failure soil appears to reach plasticity in more
locations where the strength properties (ϕ and c) are also influential The most dominant
failure mode appeared to be the anchor failure whereas the soil failure is always preceded
the sheet pile failure The dominance of the anchor failure can be explained from the fact
that the most influent parameter in this LSF is the stiffness that is also the variable with
the highest uncertainty (CoV=25) Therefore the plastic limit of the anchor is more
likely to be reached
Regarding the soil body reliability the analysis detected 3 different failure directions in
which the clay and the sand layer below the dike seem to be the most influent More
specifically the friction angle φ the cohesion c the unit weight γ of the clay layer the
friction angle of the sand layer and the cohesion of the Dikeold material turned out to play
an important role The decrease of the clay layerrsquos unit weight lead to the incapability of
the layer to withstand the overlying load of the dike body and at the same time the
weakening of the sand layer contribute to the creation of an inner sliding surface Even if
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 117
the uncertainty of the unit weight is small (CoV=5) it is still important for the stability
of the system as the maximum deformations appear in that layer (see Figure 712) Last
but not least the illustration of the several failure points in Plaxis allows someone to
inspect the failure mode and realize if it is a real failure of the structure or if it is a
weakness of the modelling
Finally during the system analysis multiple failure directions were detected whereas the
pie charts of the importance factors include soil properties that affect both soil and
structural elements as it was expected However due to Plaxis numerical errors several
failure points were found to be not realistic and they should be excluded in order to come
up with a more reliable failure probability which will be presumably lower The most
suitable method for estimating the importance factors though seems to be the Average all
method which is not distorted from the Plaxis numerical errors However it is
recommended that a further research be done on how realistic are the failure points
coming from Plaxis warnings and to verify if they comply with Figure 711 failure
definition
In Table 79 a synopsis of the most important soil properties for the different LSFs is
presented It is also apparent the common contribution that some of them have in
different LSFs In the last two columns of the table the results of the influencing soil
variables for the system LSF are shown with the Average 10 and Average all method
respectively For the reasons that were explained in the previous paragraph the Average
all method is considered to be the most suitable in that case and as it can be noticed from
the table it gives the most representative picture of the governing soil variables of each
LSF
Table 79 Important soil properties for the different LSFs (radicradic=very important radic=less important but still considerable the most important parameters for each LSF is painted in orange)
Soil material
Soil property
Anchor Sheet pile
Soil body
System Average
10
System Average
all
Clay
G radicradic radicradic
radic radicradic
φ
radic radicradic
c
radic radic radic
Rinter γ
radicradic
radicradic
Sand
G
radic radic
φ
radicradic radicradic
c
Rinter γ
Dikeold
G radicradic
radic
φ
radic
c
radicradic radic radicradic radic
Rinter radic
radicradic radic
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 118
γ
radic radic
Rinter was taken from the validation of FORM with DS Figure 717 (right graph)
taken from the sensitivity of the sheet pile in the soil elastic domain Figure 720
Can response surface techniques help to increase the efficiency and robustness of the
reliability model
In this research RS techniques were used only as a source for providing a reasonable
value to the FORM analysis in order to continue running in case of Plaxis errors These
response surfaces were constructed based on the SA results for both the anchor and the
sheet pile wall Apparently the domain of the SA is limited and most of the times it is
incapable of approaching the design point However RS worked efficiently in many cases
where Plaxis failed to give an output result such as in case of the sheet pile LSF Moreover
the RS could be used individually for the estimation of the failure probability (instead of
Plaxis calculations) but a better fit is advised to be firstly achieved that would be
representative in a wider plastic domain of the structural elements Therefore the RS
technique can be used to enhance the efficiency of the reliability analysis of a complex
system in the sense that it can be considered as solution for FEM errors What is more RS
techniques are a quite promising method for the reliability analysis of a system as it can
also be used alone without coupling with FEM and eliminate the computational effort
However this requires more investigation of such a technique
How can the current design approach be improved
In section 22 the current design concept of the case study that is under investigation in
this research is described At that section the design values of the moments and the forces
on the structural elements are presented as well as the required total safety factor FEM
=18 is given as it was initially calculated by Breedeveld (2011) It has been observed that
a significant overestimation is made in terms of the design values that might later lead to
heavy and cost inefficient structures For that purpose alterations was made on the
original case study in terms of the structural elements and some soil properties and a new
deterministic design was made in order to come up with a less heavy structure This
would also render it possible to investigate the reliability of the system rather than taking
zero probabilities or non-convergence Moreover it would finally help to see if there are
any potentials for optimizing the design by applying a full probabilistic analysis
According to the soil reliability analysis the probability of failure was found 1310-8 and
the reliability index β=55 As far as the anchor failure is concerned the probability of
failure found to be 28410-6 with a reliability index β=45 according to the FORM analysis
The sheet pile failure seem not be significant as it will more likely occur after the soil
failure Eventually the system analysis was performed the probability of failure was
found to be 2410-3 and the reliability index β=30 which is high enough to consider the
system safe
The systemrsquos reliability index that came out of this analysis cannot be directly compared
with the required one (ie β =40) because they stem from different probabilities The
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 119
former refers to the probability of failure due to the global instability given a design water
level whereas the latter refers to the probability of flooding that incorporates also the
water level uncertainty It must be mentioned that in reality there is no sheet pile wall
inside the dike section in the dike ring 43 but this section is under investigation for
reinforcement purposes as it does not comply with the safety standards after a recent
inspection that has been carried out Moreover the original case study taken from
Breedeveld (2011) was subjected to multiple alterations for the needs of the current
research Therefore a comparison of the reliability index with the required one from the
regulations would not be meaningful to make However what it can be compared in that
case is the order of magnitude of the reliability index After the analysis a β of 30 was
estimated for the probability of failure whereas for the overall probability of flooding a β
of 40 is required whose magnitude is the comparable with the former This can firstly be
considered a sort of the probability ldquovalidationrdquo in the sense that the analysis gave
realistic results and secondly it can imply that the structure that has been investigated in
this research is likely to be safe towards macro-instability as its coupling with the water
level uncertainty would certainly give a reliability index above 40 Furthermore a
reliability index of 30 indicates that there is room for optimizing the design concept of
such a system if someone considers also the fact that the soil strength of the subsoil was
further reduced for the needs of the current study
Due to the limitations of the modeling and the alterations of the original case study these
reliability indices should not be taken into account as an indicative number for the safety
of this dike section but rather as an order of magnitude in order to realize what the
potentials for improving the design procedure are This research can also contribute to
the way of designing not only the system but also each structural element as the reliability
of each can be calculated in advance According to the methodology developed in this
thesis a comparison can later on be made for a real case situation between the results of a
full reliability analysis and the current design procedure In that sense a tangible
improvement can be suggested for the design concepts and even to introduce a new way
for the future design of dikes with retaining walls
It should be stressed that the conclusions are based on the results of this case study with its
geometry and set of material parameters and the variation coefficients Changes in the geometry
the material parameters or the statistical properties could lead to different results The
generalization of these conclusions should thus be treated carefully
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 120
MSc Thesis A Rippi 121
8 Conclusions and Recommendations
81 Conclusions The advances in structural reliability analysis have made a large variety of methods available The
applicability and efficiency of these methods depends on the problem that is analyzed and on the
models that are used One the other hand the necessity for optimizing the design procedure of
several structures is becoming more and more intense A solution to this demand can be the
reliability analysis of a system for which rare information about its application on real-world
problems is available This work can contribute to making probabilistic analysis more accessible
for dikes reinforced with a sheet pile wall and at the same time to analyzing such a complex
system in terms of the factors that mainly influence the system behavior In particular the main
research question of this research was
How can the probability of failure due to global instability of a dike with a sheet pile wall modeled
by a Finite Element Model be analyzed
For that purpose in this research a reliability analysis was conducted by making use of a soil-
structure model that was simulated in FEM The output of the FEM calculations was utilized as an
input in the probabilistic model that was used in this thesis in order to assess the reliability of a
dike with an anchored sheet pile wall In that case the load was basically the earth and water
pressure whose impact was investigated on the anchor (LSF 1) the sheet pile (LSF 2) and the soil
body (LSF 3) respectively as well as on the system as a whole
It was demonstrated that the calculation effort of this kind of analysis does not necessarily have to
be extremely high Especially when previous knowledge about the system behavior and the
reliability models used is available This can be achieved by carrying out a sensitivity analysis In
this thesis a global SA was conducted that was proved capable of giving a representative
estimation of the most influencing variables and give a general overview of the systemrsquos
performance However close to the design point local SA such as FORM becomes more suitable
As far as the reliability of the structural elements is concerned their limit state was evaluated in
terms of the exceedance of their yielding stress The reliability assessment was carried out with
FORM from which the probability of failure and the influence factors were concluded More
precisely the anchor failure mode seemed to be the most determinant one with the shear
stiffness of the soft soils to be the most important soil properties The sheet pile failure is quite
rare and it is unlikely to occur before the soil failure since the soil body has already entered its
plastic domain
The soil and the system failure were challenging to assess as there are multiple criteria that
someone can choose In this thesis regarding the soil failure the Plaxis definition of soil collapse
was used and it was evaluated with DS In that case the unit weight and the strength parameters
of the soft soil layers seemed to be influencing The detected failure directions were investigated
Conclusions amp Recommendations
MSc Thesis A Rippi 122
in terms of their validity More precisely the failure points were illustrated in Plaxis in order to
visualize the failure and see if it is real failure or not (real if dike crest settles) That procedure
revealed the possibilities and limitations of FEM modeling The advantage is that a potential
failure can be detected and its visualization is possible On the other hand one should be aware of
the limitations related with the constitutive model that is used In this study the Mohr Coulomb
method was applied that although it can efficiently identify possible failure mechanisms the
representation of the deformations is not so realistic
The failure of the system was considered as a serial system of LSF 1 2 and 3 and it was also
assessed with DS The importance factors stemmed from the reliability analysis indicated soil
features that were important in all the LSFs However it should be stressed that the frequency of
the Plaxis numerical errors was such that quite a few failure points seem to be unrealistic In this
thesis the identification of those errors and their partially exclusion from the analysis was
achieved The latter was attained by the construction of the so-called Response Surfaces (RS)
based on a preliminary sensitivity analysis Such a technique found to be valuable in order to
overcome such a problem however a study for fitting more suitable RS in each problem is
recommended Nevertheless the further elimination of those numerical errors is strongly
recommended to be investigated in future research given the results from the current study
Retrieving the importance factors from a DS analysis was also a challenging part In this study
additional methods were developed in order to evaluate the importance factors and define the
design point after a DS analysis It revealed that alternative possible approaches of the
importance factors (Average 10 and Average all methods) can presumably provide a better
understanding of the systemrsquos behaviour Such methods seemed to be valuable for excluding the
ldquonoiserdquo of the possible fake failure points that originated from FEM numerical errors However
due to the inherent randomness of DS in terms of the selection of the directions the design point
shall be inspected and compared with the other failure points that were found Then someone can
contemplate if it is necessary to proceed with an averaging of the importance factors in order to
get a clearer view the variables influence Of course such a procedure takes time that is not always
available but according to the requirements of each research the most preferable assumptions
should be made
Last but not least the reliability analysis came up with relatively low probabilities of failure which
implies that there is still room for optimizing the design concept of a dike with retaining wall
More precisely the original case study of Breedeveld (2011) has been altered by reducing the
strength properties of the subsoil and installing a less heavy sheet pile wall after following a less
strict deterministic design It was then observed that even then the reliability analysis came up
with a respectively high reliability index It should be stressed that the failure probability of the
system should be also coupled with the uncertainty of the water level in order to obtain an overall
picture of the total probability of failure Nevertheless this research proves the potentials to
optimize the design of such a system which presumably will lead to a less expensive construction
However additional failure mechanisms such as piping internal erosion etc should be also
investigated and the design properties of the system (such as the length and cross section area of
the sheet pile wall) shall be based on them
Conclusions amp Recommendations
MSc Thesis A Rippi 123
82 Recommendations The following topics are recommended for further research
Additional research is required on FEM modeling both for the structural elements and the
soil body More in particular in this research it has been proved that Mohr Coulomb
model cannot realistically estimate the deformations in a soft soil In that case there are
other more advanced models such Hardening Soil model and Soft Soil model that are
capable of simulating the soil displacements and the stress-strain relationship closer to
the actual ones However these methods require more parameters to be specified and
their coupling with reliability packages might not be so robust and functional
As far as the structural elements modeling is concerned there are two major topics that
future research should focus on Firstly the anchor rod can be simulated in Plaxis as a
node-to-node element and at the bottom an embedded beam can be installed in order to
represent the anchorage body This would provide a better reaction with the surrounding
soil body and it could probably simulate the anchor-soil interaction more realistically
Another alternative would be to simulate it as a plate element in order to be able to test
the moments that are also developed on the anchor separately Secondly it is
recommended that the elements be investigated close to their plastic area and a reliability
analysis to be carried out with plastic elements instead In this research an attempt was
made for such an analysis but it appeared not to be functional in cooperation with FORM
analysis
Another issue that is recommended to be investigated is the inspection and the validation
of the failure that Plaxis is giving During the soil failure analysis different types of failure
were observed from which some might not be realistic but due to the modeling limitations
Therefore an attempt should be made in order to further eliminate the effect of Plaxis
numerical errors on the reliability analysis that can lead to fake failure modes A check of
the failure points can be also conducted by using a more suitable constitutive model for
the deformations rather than Mohr Coulomb However this is not always feasible for each
individual case Therefore the illustration of the failure mechanisms in large scale
experiments can give a better insight into the systemrsquos behavior and valuable knowledge
can be acquired of the expected failure modes In this way also 3D effects can be
investigated see what is their effect to the dike instability and at what extend they shall be
considered
Apart from the soil body the stresses and the displacements of the structural elements
shall be also verified It is thus recommended that field tests and measurements take place
in embedded elements like sheet piles and anchor in order to validate Plaxis results and
learn what is the most suitable manner for the their simulation (fixed-end-anchor node-
to-node element plate etc)
Concerning the reliability analysis of the specific case study a research on the potential
relevant LSFs is recommended for the different elements Specifically in this thesis the
LSF were chosen in order to represent the ULS of the element There is the possibility that
the SLS is used and certain acceptable deformations can be settled as a threshold for the
Conclusions amp Recommendations
MSc Thesis A Rippi 124
limit states However the choice of this displacement limit is not straightforward and a
qualitative research shall also be conducted in order to identify the impacts of different
values Besides it is not only the value that should be selected but also the location on the
structure that is going to be checked
It is also advised that the potential correlations among the soil parameters to be included
In Wolters (2012) and Bach (2014) correlation matrices are introduced for specific soil
properties that can be included in a future research and reveal what is their effect on the
reliability analysis The influence of correlations depends on what kind of variables are
correlated and what is the influence of those variables on the system behavior However a
preliminary analysis with independent variables is advisable to be carried out at the
beginning so as for someone to be able to analyze the effect of the variables individually
and then evaluate the correlation effect more wisely
It would be also advisable a comparison to be made between the design of a real case
according to a full probabilistic analysis and partial safety factors In that way a direct
insight into the advantages and disadvantages of each method can be gained and
fundamental improvements of the design concept can be made for dikes with retaining
walls
At this moment the most reliable way of conducting a reliability analysis is with sampling
methods like MC or DS However such methods are usually unattractive due to their large
computational time Thus the tolerance of sampling methods accuracy is lowered in order
also to reduce the number of iterations or approximating methods are used like FORM
analysis However even if approximating methods are used they are usually preferred to
be validated with sampling methods in order for their results to be acceptable A real
validation of the failure probability would mean that a structure should be experimentally
tested in different positions and under the same conditions in order to see if the
probability of failure that is calculated in each case is the same Therefore a real
validation of the failure probability cannot yet be initiated but the computational effort of
the sampling methods can be eliminated with the development of technology and the
improvement of the computer science
MSc Thesis A Rippi 125
References
Abdo T amp Rackwitz R (1990) A new beta-point algorithm for large time-invariant and time
variant reliability problems Proceedings of the 3rd IFIP WG 75 Conference Berkeley pp1-12
California USA
Ang AHS and Tang W (1984) Probability Concepts in Engineering Planning and Design Volume
IBasic Principles John Wiley and Sons New York USA
Baecher GB and Christian JT (2003) Reliability and Statistics in Geotechnical Engineering
Chichester West Sussex John Wiley amp Sons Ltd
Behaviour of Soils and Rocks GEO-ENGINEERING - CIE4361 (Vol 2015) Retrieved from
httpsblackboardtudelftnlbbcswebdavpid-2386735-dt-content-rid-
7926659_2courses32281-14150221-possibilities26limitationspdf
Breedeveld J (2011) Technisch Rapport Analyse Macrostabiliteit Dijken met de Eindige Elementen
Methode Bijlage Voorbeeldcase Deltares Delft the Netherlands
Brinkgreve RBJ (2015) Personal discussion TU Delft the Netherlands
Brinkgreve RBJ and Bakker HL (1991) Non-linear finite element analysis of safety factors In
G Beer JR Booker and J P Carter (Eds) Proceedings of the Seventh International Conference on
computer methods and advances in Geomechanics Rotterdam A A Balkema Brookfield
Cukier RI Fortuin M Shuler KE Petschek AG and Schaibly JH (1973) A study of the
sensitivity of coupled reaction systems to uncertainties in rate coefficients I Theory California
University USA
Cundall PA (2002) A discontinuous future for numerical modeling in soil and rock Published in
Proceedings of the third international conference Discrete Element Methods Numerical Modeling
of Discontinua Edited by Cook and Jensen Santa Fe pp 3-4
CUR (2005) Handbook Quay Walls Gouda the Netherlands
CUR-publication 190 (1997) Probabilities in civil engineering Part 1 Probabilistic design in theory
Stichting CUR Gouda
CUR 166 ndash Damwandconstructies (2005) Technical Recommendation 4e herziene uitgave 2005
The Netherlands
Deltaprogramma (2014) Synthesedocument Veiligheid Achtergronddocument B1 Ministrie van
Infrastructuur en Milieu
EN 1990 2002 Basis of Structural Design European Committee for Standardization
References
MSc Thesis A Rippi 126
Farrokh N (2007) Tools and Strategies for Dealing with Uncertainty in Geotechnics In DV
Griffiths and GA Fenton (Ed) Probabilistic methods in Geotechnical Engineering (pp 71-95) New
York NY Springer Vienna
Fenton GA and Vanmarcke EH (1990) Simulation of random fields via local average
subdivision J Eng Mech ASCE 116(8)1733ndash1749
Flood Defences HYDRAULIC ENGINEERING ndash CIE5314 (Vol 2015) Retrieved from
httpsblackboardtudelftnlbbcswebdavpid-2125084-dt-content-rid-
7621002_2courses29759-
131404CIE5314_FloodDefences_05_failure_mechanisms_stability_Schweckendiekpdf
Gemeentewerken-Rotterdam (2003) Alle triaxiaal proeven rotterdam Technical report Gemeentewerken Rotterdam Rotterdam the Netherlands
Ghanem R and Spanos P (1991) Stochastic Finite Elements ndashA Spectral Approach
Springer New York
Gonzaacutelez A Cuadrado A Josa A and Olivella S (2013) Safety assessment of limit equilibrium
methods for the design of cantilever and anchored sheet pile walls Escola de Camins Barcelona
Griffiths DV and Fenton GA (Eds) (2007) Probabilistic Methods in Geotechnical Engineering Lecture notes monographs edited works and proceedings in the field of Mechanics Engineering Computer Science and Applied Mathematics Italy Hasofer AM and Lind NC (1974) An exact and invariant first-order reliability format Journal of
Engineering Mechanics Division ASCE 100(1) 111-121
Hoek E Read J Karzulovic A and Chen ZY (2000) Rock slopes in civil and mining engineering
Published in Proceedings of the International Conference on Geotechnical and Geological
Engineering GeoEng2000 19-24 November Melbourne
Homma T and Saltelli A (1996) Importance measures in in global sensitivity analysis of
nonlinear models Reliab Eng Syst Saf 52(1) pp1-17
Hydra-Ring 12 (2015) Probabilistic toolbox for the WTI2017 Technical Reference Manual Version
12 Deltares Delft
Jaky J (1944) ldquoThe coefficient of earth pressure at rest In Hungarian (A nyugalmi nyomas tenyezoje)rdquo J Soc Hung Eng Arch (Magyar Mernok es Epitesz-Egylet Kozlonye) 355ndash358
Johansson E amp Sandeman E (2014) Modelling of Deep Excavation in Soft Clay A comparison of different calculation methods to in-situ measurements Chalmers University of Technology Gothenburg Sweden
Joint Committee on Structural Safety (1981) General principles on reliability for structural design
International Association for Bridge and Structural Engineering
References
MSc Thesis A Rippi 127
Larsen H Lubking P and Breedeveld J (2013) Ontwerp stabiliteitsschermen (type II) in primaire
waterkeringen (groene versie) Deltares Delft
Melchers R E (2002) Safety and risk in structural engineering Prog Struct Engng Mater
4 193ndash202
Moumlllman A amp Vermeer P (2008) Reliability analysis of a dike failure case study Elbe river
University of Stuttgart Germany
Nelson RB (1999) Lecture Notes in Statistics An Introduction to Copulas Springer New York
OpenTURNS 15 (2015a) Reference Guide OpenTURNSrsquo methods for Step C uncertainty
propagation pp155-161
OpenTURNS 15 (2015b) Use Cases Guide for the Textual User Interface pp190-192
Pandya N Marsdea E Floquet P and Gabas N (2008) Sensitivity analysis of a model for
atmospheric dispersion of toxic gases In B Braunschweig and X Joulia (Ed) 18th European
Symposium on Computer Aided Process Engineering-ESCAPE18 (pp1143-1148) Elsevier BV
Phoon K (Ed) (2008) Reliability-based Design in Geotechnical Engineering Computations and Applications Taylor amp Francis Britain
Plaxis 2D (2015a) Tutorial manual Delft the Netherlands
Plaxis 2D (2015b) Reference Manual The Netherlands
Plaxis (2015) Material Models Manual The Netherlands
Powell MJD (1994) A direct search optimization method that models the objective and
constraint functions by linear interpolation In S Gomez amp JP Hennart (Ed) Advances in
Optimization and Numerical Analysis (pp51-67) Dordrecht Kluwer Academic
Rankine W (1857) On the stability of loose earth Philosophical Transactions of the Royal Society
of London Vol 147
Saltelli A (2004) What is Sensitivity Analysis In A Saltelli K Chan and EM Scott (Ed)
Sensitivity Analysis (pp 3-13) John Wiley amp Sons Publication
Saltelli A S Tarantola F Campolongo and M Ratto (2004) Sensitivity Analysis in Practice
A Guide to Assessing Scientific Models John Wiley amp Sons Ltd
Schittkowski K (1985) NLPQL A Fortran subroutine solving constrained nonlinear
programming problems Annals of Operations Research 5 485-500
Schweckendiek T (2006) Structural Reliability applied to Deep Excavations Coupling Reliability
Methods with Finite Elements TU Delft Delft
References
MSc Thesis A Rippi 128
Schweckendiek T Vrouwenvelder ACWM Calle EOF Jongejan RB and Kanning W (2013)
Partial Factors for Flood Defenses in the Netherlands Modern Geotechnical Codes of Practice ndash
Development and Calibration Fenton GA et al (eds) Special Geotechnical Publication Taylor amp
Francis
Seed RB Bea R G Abdelmalak R I Athanasopoulos A G Boutwell G P Bray J D Briaud J-L Cheung C Cobos-Roa D Cohen-Waeber J Collins B D Ehrensing L Farber D Hanemann M Harder L F Inkabi K S Kammerer A M Karadeniz D Kayen RE Moss R E S Nicks J Nimmala S Pestana J M Porter J Rhee K Riemer M F Roberts K Rogers J D Storesund R Govindasamy A V Vera-Grunauer X Wartman J E Watkins C M Wenk Jr E and Yim SC (2006) Investigation of the Performance of the New Orleans Flood Protection Systems in Hurricane Katrina on August 29 2005 Volume I Main Text and Executive Summary New Orleans USA
Steenbergen RDJM (2013) Lecture notes Probabilistic Design Delft the Netherlands TU Delft
TAW (Technische Adviescommissie voor de Waterkeringen) (2004)Technisch Rapport Waterspanningen bij dijken The Netherlands
Terzaghi K (1942) Theoretical Soil Mechanics New York Wiley ISBN 978-0-471-85305-3 Teunissen JAM (2009) SBW Analyse macrostabiliteit van dijken met Eindige Elementen
Modellen Achtergronden bij Activiteit EEM 04a Opstellen stappenplan Deltares the Netherlands
Terzaghi K (1943) Theoretical Soil Mechanics John Wiley and Sons New York
Turaacutengi T amp Rabitz H (2004) Local Methods In A Saltelli K Chan and EM Scott (Ed)
Sensitivity Analysis (pp 81-99) John Wiley amp Sons Publication
Vergouwe R Huting RJM and van der Scheer P (2014) Veiligheid Nederland in Kaart 2 Overstromingsrisico dijkring 43 Betuwe Tieler- en Culemborgerwaarden The Netherlands
Vrijling JK Schweckendiek T and Kanning W (2011) Safety Standards of Flood Defenses Keynote at ISGSR 2011 (International Symposium on Geotechnical Safety and Risk) Munich Vogt Schuppener Straub amp Braumlu (eds)
Vrijling JK and Gelder PHAJM van (2002) Probabilistic Design in Hydraulic Engineering Lecture notes TUDelft Faculty of Civil Engineering and Geosciences Delft Vrouwenvelder ACWM Steenbergen HMGM and Slijkhuis KAH (1999) Theoriehandleiding PC-Ring Delft the Netherlands TNO
Waarts PH (2000) Structural reliability using Finite Element Analysis An appraisal of DARS
(Directional Adaptive Response surface Sampling) Delft University of Technology Delft the
Netherlands
Wyllie DC amp May CW (2004) Rock Slope Engineering (4th edition) Spon Press London
MSc Thesis A Rippi 129
Appendix A OpenTURNS features
In this Appendix an overview of the most relevant of OT is given Firstly in section 112 an
introduction of the applicable methods in SA is made while in section 52 the results from such an
analysis are presented In Appendix A1 a more detailed elaboration is made of how the FAST
sensitivity method that was applied in this project works
In Appendix A2 an explanation of the different optimization algorithms available for FORM
analysis is given The theory behind these algorithms and their convergence criteria are discussed
while finally an evaluation of their performance after applying them in case of the anchor failure
is presented This evaluation has been conducted in order to decide about the most relatively
reliable algorithm that should be adopted for the reliability analysis
Finally in Appendix A3 the types of probability distributions that have been used in this study
for the random variables and which are available in OT are displayed
A1 Fourier Amplitude Sensitivity Test (FAST) In this section the extended FAST sensitivity method is discussed as it is also described in the
Reference Guide of OpenTURNS 15 (2015) FAST is based on the Fourier decomposition of the
variance of the model response ( ) the latter being represented by its Fourier expansion
is an input random vector of independent components Its key idea is to
recast this representation as a function of a scalar parameter by defining
exploring the support of the input random vector
For each input the same procedure is realized in three steps
1) Sampling
Deterministic space-filling paths with random starting points are defined ie each input Xi is
transformed as follows
( ( )) (A1)
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 130
Figure A1 Search curves in the input space as they are defined from the transformation (the blue
dot indicates the direction of the current search path)
where is the number of input variables N is the length of the discretization of the s-space with
s varying in (-π π) by step 2πN is a random phase-shift chosen uniformly in [0 2π]
is a set of integer frequencies assigned to each input Xi The frequency
associated with the input of interest is set to the maximum admissible frequency satisfying the
Nyquist criterion (which ensures to avoid aliasing effects)
(A2)
with M the interference factor usually equal to 4 or higher It corresponds to the truncation level
of the Fourier series ie the number of harmonics that are retained in the decomposition realised
in the third step of the procedure And the maximum frequency of the complementary set of
frequencies is
(A3)
with the index ldquo-irdquo which meaning ldquoall but irdquo
2) Simulations
Output is computed such as
Then is expanded onto a Fourier series
sum [ ] (A4)
where and are Fourier coefficients defined as follows
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 131
int
(A5)
int
(A6)
These coefficients are estimated thanks to the following discrete formulation
sum
(A7)
sum
(A8)
3) Estimations by frequency analysis
The quantitative measure of sensitivity is represented by Sensitivity Indices The first-order
sensitivity index of an input pi is the measure of the main (direct) effect of pi on the output
variance (where i ne j) the second-order sensitivity indices measures the interaction effect of
pi and pj on the output variance Other higher-order indices are defined in the same manner The
total sensitivity index is the sum of all sensitivity indices involving factor pi (Homma amp Saltelli
1996) For example the total sensitivity index of factor 1 for a model with 3 input factors is
given as
The first-order (Si) and total ( ) indices can be interpreted as following (Pandya et al 2008)
high pi is an influent parameter
and both small pi is not an influent parameter (neither alone nor in interaction with
other parameters)
and nearly the same no interaction of pi with the other parameters
and very different high interaction of pi with the other parameters
Total indices are especially suited to apportion the model output variation to the input factors in a
comprehensive manner The FAST method calculates the first-order and the total sensitivity
indices whereas Sobolrsquo method in addition to these also provides all higher-order sensitivity
indices to determine quantitatively the interaction between parameters However the
computational cost and calculation time of Sobolrsquos method tends to be higher than that of the
FAST method
Combining equations A1-A8 the first-order indices are estimated as follows
sum (
)
sum (
)
(A9)
where is the total variance the portion of D arising from the uncertainty of the ith input and N
the size of the sample using to compute the Fourier series
Subsequently the total-order indices are estimated as follows
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 132
sum (
)
sum (
)
(A10)
where is the part of the variance due to all inputs except the ith input
A2 Optimization Algorithms in FORM
Principles of optimization algorithms
The possible optimization algorithms in OT are
Abdo-Rackwitz (ARF)
Cobyla
Sequential Quadratic Programming (SQP)
Below a description of the main principles of the optimization methods mentioned above is
quoted ARF and SQP methods are elaborated according to Abdo and Rackwitz (1990) while
Cobylarsquos principles according to Powell (1994) Finally an evaluation of the performance of the
different methods follows after applying them in case of the anchor failure
The general optimization problem (objective function) can be written in the classical form
subject to inequality constraints (constraint function) (A11)
The Lagrangian function of the general problem is defined by
sum (A12)
where are the so-called Lagrangian multipliers The SQP method (developed by Schittkowski
(1985)) replaces the original problem by a sequence of quadratic programming problems which
are exactly solvable and which approximate the original one This is done by approximating the
Lagrangian function by its second order Taylor expansion in an initial point u0
(A13)
where
sum
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 133
sum
sum
in which is the gradient operator and represents the Hessian matrix27 of the function f in
the point u0 The optimality conditions for any iteration point k of the sequence of quadratic
expansions are
sum [ sum
] (A14)
(A15)
The exact calculation of the Hessian matrix is generally too expensive and cannot be efficiently
implemented for a general case Therefore the gradient information obtained in each point
during iteration is used to build up an approximation of this matrix using one of the known
update formulas
The new iteration point is defined by
(A16)
where is the step length and is a direction in which a line search is performed The process
stops when the optimality conditions of the original problem are satisfied
The most time consuming part in this algorithm is the updating of the Hessian matrix and the
solution of the system of equations A fair approximation of the Hessian of non-quadratic
functions is also obtained with about n updates of the matrix This means that the approximation
used in the few (say ten) iterations to reach convergence cannot be very good when the problem
has large number of variables The rounding errors during the updating process in large problems
can make the approximate Hessian to become singular Close to singularity the search direction
can be significantly distorted In this case the algorithm has to restart the iteration with a unit
Hessian matrix in the point where singularity occurred
The posterior algorithm ARF developed by Abdo and Rackwitz (1990) overcomes this problem
by obtaining a constant approximation of the true Hessian matrix Only the contribution of the
objective function to the Hessian is considered The numerical decomposition of the matrix
contains the scalar product of the gradients of the constraints as elements in each iteration
27 Hessian is a square matrix of second-order partial derivatives of a scalar-valued function or scalar field It describes the local curvature of a function of many variables
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 134
Cobyla algorithm has been developed by Powell (1994) and it constructs linear polynomial
approximation to the objective and constraint functions by interpolation at the vertices of
simplices (a simplex in n dimensions is the convex hull of n+1 points n being the number of
variables) It generates the next vector of variables from function values at the vertices
of a nondegenerate simplex in In this case there are unique linear
functions and say that interpolate f and at the vertices
and the optimization problem A15 by the linear programming problem
( ) (A18)
( )
The iterative use of expression A18 puts this method in the class of ldquosequential linear
programming algorithmsrdquo It is also essential to notice that in this method the gradients are
derived by interpolation at the vertices of simplices and not analytically as it happens in SQP and
ARF
Additionally Cobyla algorithm makes use of a trust region bound ρ so that the trust region
condition on the new vector of variables is
(A19)
Such a region gives the user some control over the steps that are taken automatically and which
respond satisfactorily to the fact that there may be no finite solution to the linear programming
problem The algorithm also employs a merit function of the form
( ) ( ) [ ( ) ] (A20)
in order to compare the goodness of two different vectors of variables Here is a parameter that
is adjusted automatically depending on the how close to optimizing the objective function the
analysis is and the subscript ldquo+rdquo means that the expression in square brackets is replaced by zero
if and only if its value is negative so we have ( ) ( ) whenever is feasible Parameters
and are changing ( only reduces) automatically according to the improvement of the
optimization problem that sequential vectors might cause Therefore if the change in the merit
function does not happen to improve the optimization then these values are changing Especially
regarding the parameter user gives the initial and the final values of it namely and (it
is recommended that be a reasonable change to make the variables for a coarse exploration
of the calculation while should be approximately the required distance from the final vector
of variables to the solution of the optimization problem)
Convergence criteria
The convergence of the aforementioned algorithms is controlled by the evaluation of the
following errors expressed in the standard space (for the relationship between the real space (x-
space) and the standard space (u-space) see section 422) (OpenTURNS 15 (b) 2015 pp 190-
192)
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 135
The absolute error which is the distance between two successive iterates
(A21)
The constraint error which is the absolute value of the limit state function minus the
threshold
(A22)
The relative error which is the relative distance between two successive iterates (with
regards to the second iterate)
(A23)
The residual error which is the orthogonality error (lack of orthogonality between the
vector linking the center and the iterate and the limit state function)
( ) (A24)
The algorithm converges if all the final error values are less than the maximum value specified by
the user The convergence can be also controlled by altering the maximum number of iterations
which should be higher in case of stricter required errors What is more there is the possibility to
manipulate the starting point that the algorithm uses for its first iteration Usually the default
starting point in a FORM analysis consists of the mean values of the random variables However if
there is already some knowledge about the potential design point it can be set as the starting
point of the algorithm in order to accelerate the convergence and save a lot of time
The aforementioned errors the maximum number of iterations and the starting point were
manipulated during the analysis in order to understand the functionality of the different
algorithms and to improve their efficiency
Evaluation of the algorithms performance
An evaluation of the above optimization algorithms has been made in order to make a decision of
the most suitable for the current project In principle a literature overview regarding the
performance of the algorithms is presented Then the 3 algorithms were tested considering the
anchor failure limit state function (see Eq 55) with a stress limit of 450000 Pa and their results
are discussed below
According to Abdo and Rackwitz (1990) the SQP method seems to have better convergence
behavior for highly curved constraint functions which is especially true when very expensive
structural state functions are involved such as finite element structural analyses However the
mentioned algorithm fails to reach convergence at a problem dimension (amount of random
variables) of around n=50 Therefore for increased dimensionality SQP is less efficient in terms
of storage and CPU time compared with other methods due to singularity of the updated Hessians
matrices In OT a warning is introduced since an analysis has started that a default
implementation of Hessian is being used and that the computation can be severely wrong
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 136
Abdo and Rackwitz (1990) also proved that ARF algorithm can handle problems of up to 2000
variables while the storage requirements and the CPU time are much smaller than with SQP
Therefore optimization problems of high dimensionality can be evaluated
In OT a warning appears at the beginning of a FORM calculation saying that a default
implementation of the gradient is being used and that the computation can be wrong A similar
message also appears for the SQP algorithm saying that a default implementation of the Hessian
matrix is used and that the computationrsquos results might be unreliable Moreover in both
algorithms it is advisable to check the values of the random variables that were used during the
analysis in order to ensure that they fluctuate within acceptable boundaries (Figure A4 is an
example of a variable that surpassed the minimum boundary at some stage)
Firstly the Cobyla algorithm was applied on the case study in combination with a FORM analysis
Cobyla does not require a gradient evaluation of the limit state function and does not use a default
implementation of the Hessians matrix as ARF and SQP does while the computational time is
quite small comparable to the aforementioned methods
The Cobyla algorithm was tested with (i) 3 (ii) 10 and (iii) 19 parameters respectively in order to
check its efficiency28 As it has been mentioned in the previous section the convergence criteria to
be steered are the errors (absolute relative residual and constraint) the maximum number of
iterations and the starting point At first the default errors were used with a value of 10-3 but it
was realised that such an order of errors lead to non-convergence of the algorithm Therefore the
algorithm was then checked after relaxing the maximum errors to values of 01 05 10 50 and
100 Finally the combination of the errors that led to convergence was 01 05 01 02 for the
maximum absolute relative residual and constraint error respectively After the analysis the
number of LSF calls in case (i) (ii) and (iii) was 25 even if the maximum one was set to 100 and
the probability of failure was around 035 The probability of failure was quite high and that is
why an inspection of the output file was carried out In Figures A2 and A3 the values of the
anchor limit state function are presented in the right graph during the iterations It was then
observed that the threshold of 450 kPa was never surpassed during the iterations
28 All of the algorithms were first checked with very simple LSF (linear and quadratic) and their results were verified by MC and DS
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 137
Figure A2 Friction angle and cohesion progress during a FORM analysis with Cobyla and with 3 random variables The figure at the right illustrates the limit state function evolution
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 138
Figure A3 Friction angle and cohesion progress during a FORM analysis with Cobyla and with 19 input parameters The last figure at the right illustrates the limit state function evolution
An additional action was then taken by relaxing the maximum number of iterations from 100 to
200 but no improvement was noticed (the LSF calls were again 25) It was also given extreme
values to the errors and especially to the absolute constraint and relative error in order to check
the response of the algorithm but that did not lead to better results Furthermore the
parameter (see theory part for explanation) was manipulated but even then no further
improvement was noticed in terms of the results In Figures A2 and A3 the performance of the
algorithm during the analysis can be seen through the values of the variables that were assigned
during the iterations
As someone can notice from the graphs above it is obvious that Cobyla does not perform very
well in high dimensionality problems as the variables do not seem to vary a lot More specifically
in Figure A3 with the 19 parameters it can be noticed that the random variable does not change
at all apart from two peaks at the beginning and at the end of the analysis respectively The
incapability of the algorithm to change the parameters in order to reach the limit of the anchor
stress ie 450 kPa (while here a maximum of approximately 267 kPa is reached) leads to a poor
approximation of both the design point and the probability of failure In Figure A2 with only 3
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 139
parameters the variability of the parameter is higher during the analysis which leads to a better
approach of the limit (around 370 kPa which is larger than with the 19 parmaters) but even then
a failure probability of 035 was not a realistic value On the other hand Cobyla is easy to use only
for small number of variables (ie less than 9) with linear LSF otherwise the linear
approximations can be highly inefficient Apparently the limit state function of the anchor stress
level as a function of the soil parameters is much more complex in order to be approximated by
this algorithm
Therefore the other available optimization algorithms ie ARF and SQP were examined as well
Because of the similarity of the two algorithms only the ARF performance is presented below and
a link with SQP is made when it is needed
The ARF algorithm was tested initially with 19 parameters The settings for the errors that were
used were 01 for the absolute relative residual and constraint error respectively and the
maximum number of iterations was set to 100
Figure A4 Friction angle and cohesion progress during a FORM analysis with ARF algorithm and with 19 input parameters The figure at the right illustrates the limit state function evolution while the red spots indicate the points where the 450 kPa threshold was surpassed
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 140
The algorithm worked efficiently which means that the algorithm managed to converge towards
the threshold of 450 kPa of the limit state function as it is depicted in Figure A4 (see red dots) It
was also noticed that the random variables fluctuated a lot during the iteration steps even if the
amount was significant (19 parameters) The capability of the algorithm to vary all the 19
variables evenly and in a wide range makes it possible to find the real design point and thus to
come up with a more reliable failure probability The resulted probability of failure was 3810-9
and the reliability index was ϐ=58
However the analysis took almost 4 days to finish which is a long duration and the number of LSF
calls counted at 4151 Moreover even if the results were promising after tracing all the 19
parameters in order to inspect the algorithmrsquos performance it was found out that some variables
took unacceptable values29 during the analysis As an example in Figure A5 Poissonrsquos ratio
received negative values (see red circle) at the beginning of the iterations There were also other
such examples in the same analysis such as the friction angle which took sky-high values and the
interface strength which took negative values as well as values above 1 which is unacceptable It
should be mentioned that the design point of the algorithm was completely reasonable but even
then such a performance can prevent the algorithm from converging to the right design point
quickly and increase the computational effort
Figure A5 Poisson ratio progress during the FORM analysis with ARF algorithm
Taking into account the above implications there are two main concerns at that stage (a) the
high computational time and (b) the unreasonable values that the variables took Firstly in order
to reduce the computational time the maximum errors were relaxed More precisely it was
noticed that after relaxing the maximum relative error a lot which is more related to the
29 In some cases a value of ldquoinfrdquo was given to the variables that led to termination of the analysis without results
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 141
relaxation factor (in this case an extreme value of 1000030 was tried) and increasing the absolute
error to 05 the number of LSF calls was reduced to 724 and the analysis was finished after
almost 10 hours In Figure A6 the values that the cohesion and the friction angle took during the
iterative procedure while the red spots in the right figure indicate the points that are included in
the failure space Figure A4 shows the progress of friction angle after relaxing the errors
Figure A6 Progress of the friction angle during the FORM analysis with ARF algorithm with errors
05 10000 01 01 for the absolute relative residual and constraint error respectively
As a general conclusion after the test of different combination of maximum errors by relaxing the
constraint absolute and relative error too much successive iterations of a certain variablersquos value
seem to have a wide difference which can detain the algorithm from converging Therefore the
wisest in this case study is to keep the maximum errors below or around 10 except for the
relative and the constraint error that can relax more than 1 (depending on the case) in order to
reduce the computational time to the minimum possible
To prevent having unacceptable values of the parameters and in order to increase the efficiency of
the algorithm the random variables were reduced to 11 after a SA and some of them were
truncated in a certain range according to observations and experience until now However the
algorithm then started to give an error due to zero gradient of the objective function at a specific
point (similarly SQP gave an error due to zero hessian matrix at a specific point) while the upper
limit of the LSFs could not be reached As a countermeasure the errors were relaxed (05 10 05
10 for the absolute relative residual and constraint error respectively) in order for the algorithm
not to be trapped in a specific point while also the step of the algorithm was manipulated It
was finally proved that this kind of error stemmed from the fact that the structural elements were
considered as elastoplastic and thus a limit of the stress level was considered that misled the
algorithm The elements were eventually considered as elastic and the numerical problem was
solved
30 Practically such a value has the meaning that there is no limitation of the relative error during the optimization In the standard space (where these errors are expressed) a value also greater than 10 can be also quite high However here extreme values were given due to the time limit to test multiple values
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 142
Before figuring out that the elastoplastic behaviour was giving the problem Cobyla as a gradient
free algorithm was tested again as a potential solution but with changing the starting point close
to the design one that was found with ARF It is worthwhile to take a look into the performance of
the algorithm with 19 random variables in Figure A7 In that way we boost Cobyla to converge
and it eventually performed quite differently than in Figure A3 and without giving convergence
problems The computational time was estimated to be 9 hours and the probability of failure was
very small in the order of 10-120 However the validation of the failure probability with DS
showed that the result of Cobyla was far from correct and thus ARF was preferred for continuing
with the rest of the analysis For educational reasons a FORM analysis with ARF was again
performed after changing the starting point and it was found that the computational time was
reduced to 4 hours Therefore it is still worthy to be aware of the fact that by changing the
starting point the performance of the algorithm can improve a lot However that requires a prior
knowledge of the potential design point that usually does not exist
Figure A 7 Cobyla performance after changing the starting point closer to the failure point
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 143
It must be also mentioned that by reducing the number of the variables to 11 as it was mentioned
before the algorithms could not converge to the errors that has been set and the computational
time increased compared with the case of 19 variables Taking into account the high interaction
among the variables that was proved in section 73 through the SA it can be possible that the
combination of the soil variables could lead to failure rather than the variables individually
Therefore an analysis was tried again with the complete set of the parameters and it was realised
that the algorithm did converge to the limit value
A3 Distribution Types
Uniform Distribution
If the probability of a result is between a lower limit a and an upper limit b and if every result
between these limits is equally probable a uniform distribution should be used This distribution
is defined by the two limit parameters and the probability density function is written
(A25)
and the CDF is
(A26)
Figure A8 Probability density function of a uniform distribution with a=-2 and b=2
The mean and the variance of the distribution are also defined by the limit parameters
(A27)
(A28)
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 144
Normal Distribution
One of the most commonly used distribution types is the normal distribution or Gauss-
distribution and it is defined by its first two central moments the mean micro and the standard
deviation σ The PDF is given by
radic
(A29)
and its CDF is
int
radic
(A30)
Figure A9 Probability density function of a normal distribution with different mean and standard
deviation values (notice that the curve with micro=0 and σ2=1 belongs to the standard normal distribution)
The normal distribution with mean and standard deviation is called the standard
normal distribution In the most reliability methods such as FORM SORM DS etc the analysis is
carried out in the standard normal space which actually means that the stochastic parameters are
transformed into standard normal variables
Lognormal Distribution
If a variable X has a lognormal distribution the variable Y=ln(X) has a normal distribution The
probability density function is defined as
radic (
) (A31)
where and the mean and the standard deviation of the normally distributed random
variable Y Sometimes it is useful to work with the Yrsquos parameters rather than Xrsquos These
parameters can be expressed as follows
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 145
radic
(A32)
(A33)
where and are the mean and the standard deviation of Y In Figure A10 an example is given
of a log-normally distributed variable with different values of mean and standard deviation
Figure A10 Probability density function of a log-normal distribution with different mean and standard deviation values
Truncated Normal Distribution
The truncated normal distribution is the probability distribution of a normally distributed
random variable whose value is either bounded below or above or both
Suppose has a normal distribution and lies within the interval
Then conditional on has a truncated normal distribution whose probability
density function is
(
) (
)
(A34)
and otherwise and is the probability density function and the CDF of the standard
normal distribution similar to Eq (1) and (2) If
and
then the CDF is
defined as
(A35)
Then the mean value and the variance of the conditional will be respectively
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 146
(A36)
[
(
) ] (A37)
In the next figure the PDF histograms of a variable with a mean of 0 and a standard deviation 1
were plotted for two cases (a) variable is normally distributed (b) variable is truncated-normally
distributed with =-2 and b=2
Figure A 11 Truncated normal distribution in comparison with normal distribution
MSc Thesis A Rippi 147
Appendix B Plaxis 2D (2015) features
In this Appendix the basic features of Plaxis 2D that were treated in this thesis are discussed First
of all an introduction over the Mohr Coulomb model is made as the soil failure criteria stem from
that model Then the initial stress generation in Plaxis is discussed and how that was applied in
the current master thesis while also the φ-c reduction technique is described which was used in
the SA of the soil failure Last but not least a more detailed explanation of the interface strength
Rinter is given which represents the modeling of the soil-structure interaction in Plaxis and it has
been considered as a random variable in the current project
B1 Mohr Coulomb failure criterion Soil tends to behave in a highly non-linear way under load This non-linear stress-strain behaviour
can be modelled at several levels of sophistication There are different constitutive models
available for simulating the behaviour of the soil whose parameters can change depending on the
level of sophistication that is required in each case
As someone can notice in Table B1 different models can be applied in different cases of soil
behaviour Hardening Soil model is becoming more and more interesting in soil structures as it
has been proved to approach the plastic behaviour of the soil and the non-linear stress-strain
relationship in a reliable and promising level Moreover UBCSAND31 and the Hypoplastic model
are more sophisticated models capable of representing several aspects of soil behaviour
however the amount of the model parameters is large and for that reason they are recommended
only in case that enough soil data are available for determining these parameters
Table B 1 Features of material models (source Behaviour of Soils and Rocks 2015)
ModelFeature Elasticity Failure Hardening Softening Small strain stiffness
Cyclic loading
Liquefaction Creep
Mohr-Coulomb x x
Duncan-Chang x x
Hardening Soil x x x
HSsmall x x x
x
Modified Cam-Clay
x x x x
Soft Soil x x x
Soft Soil Creep x x x
x
UBCSAND x x x
x x
Hypoplastic (x) x x x
x (x)
31 UBC are the initials for University of British Columbia in Vancouver in Canada
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 148
In this master thesis the Mohr-Coulomb model is going to be applied which is simple but an
efficient model to describe the soilrsquos failures condition Below a description of that model is made
The Mohr-Coulomb model is a linear elastic perfectly plastic model (see Figure B1) The linear
elastic part is based on Hookersquos law while the perfectly plastic part is based on the Mohr-Coulomb
failure criterion formulated in a non-associated plasticity framework
The main principle of elastoplasticity is that total strains and strain rates consist of elastic and
plastic strain components
(B1)
(B2)
in which the elastic strain rate component abides by the Hookersquos law and thus it can be expressed
as follows
(B3)
where is the elastic stress-strain matrix and the effective stress rate The plastic strain
component is given by
(B4)
where is a scalar defining the magnitude of the plastic strains called plastic multiplier and
is
a vector describing their direction with to be the so-called plastic potential function
Figure B1 Basic principle of elastoplasticity
The occurrence of (elasto)plasticity can be indicated by the yield function which is
illustrated in Figure B2 in principle stress space and it is derived according to the Mohr-Coulomb
failure criterion expressed in the next formula
(B5)
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 149
Considering the sign of this yield function the different soil states can be recognised as follows
if flt0 (inside yield contour) pure elastic behaviour
if f=0 and dflt0 unloading from aplastic state (=elastic behaviour)
if f=0 and df=0 elastoplastic behaviour
if fgt0 (outside yield contour) impossible stress state
Figure B2 Mohr-Coulomb yield surface in principle stress space
Finally in Mohr-Coulomb model the calculation of the plastic strains is achieved by the so-called
non-associated flow rule where with the dilatancy angle
This makes a difference between the friction and the dilatancy angle which prevents the model
from overestimating potential uplift
Concluding this model uses 5 input parameters
Stiffness parameters E (Youngrsquos modulus) and v (Poissonrsquos ratio) as they are included in
the elastic stress-strain matrix according to Hookersquos law
Strength parameters c (cohesion) and (friction angle)
Dilative behaviour ψ (dilatancy angle)
The author is aware of the possibilities and limitations of this model In particular the model is
capable of
a first order approach of the soil behaviour in general
describing quite well the (drained) failure behaviour
It should be mentioned at that point that the Youngrsquos modulus E cannot be steered via the
Python interface (as it appears to be a readonly parameter) and thus the shear stiffness G was
considered instead that is defined as follows
[kPa]
f=0 (yield contour)
flt0
fgt0
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 150
On the other hand the model has also limitations that are referring to
the linear elastic behaviour until failure (no strain- or stress path-dependent stiffness
behaviour)
overestimation of the shear strength in case of an undrained behaviour
However this model is considered reliable for detecting failure and giving realistic stresses
magnitudes in the soil and structural elements
Moreover in this project a non-associative behaviour was assumed by considering a zero
dilatancy angle However it has been proved by Teunissen (2009) that in case of deep failure
mechanisms in a dike structure with soil materials of a friction angle above 30deg problems arise in
the numerical analysis in finite element calculations Particularly the strength in the sliding
surface decreases due to rotation and the ground stresses are relieving around it by releasing
energy which might lead to calculation implications These phenomena are not treated in this
thesis but it is advisable to check and review the results after such an analysis
B2 φ-c Reduction Technique In this section the ldquoSafety Calculationrdquo is applied as it is described in Plaxis (Plaxis 2D (a) 2015)
In the Safety Calculation the φ-c reduction technique (Brinkgreve amp Bakker 1991) is adopted
where the shear strength parameters tanφ and c of the soil as well as the tensile strength are
successively reduced until failure of the structure occurs The dilatancy angle ψ is in principle not
affected by the phic reduction procedure However the dilatancy angle can never be larger than
the friction angle When the friction angle φ has reduced so much that it becomes equal to the
(given) dilatancy angle any further reduction of the friction angle will lead to the same reduction
of the dilatancy angle The strength of interfaces if used is reduced in the same way
The total multiplier is used to define the value of the soil strength parameters at a given
stage in the analysis
(B6)
where the strength parameters with the subscript ldquoinputrdquo refer to the properties entered in the
material sets and the parameters with the subscript ldquoreducedrdquo refer to the reduced values used in
the analysis is set to 10 at the beginning of a calculation to set all material strengths to
their input values
The incremental multiplier is used to specify the increment of the strength reduction of the
first calculation step (this increment is by default 01 but user can change it) The strength
parameters are successively reduced automatically until all required steps have been performed
(by default these steps are set to 100 but this can be subjected to changes depending on the
situation) It must always be checked whether the final step has resulted in a fully developed
failure mechanism In other words if has reached an almost constant value while the
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 151
deformation is continuing that indicates an equilibrium state In that case the factor of safety is
given by
(B7)
B3 Initial Stress Generation The first construction phase in Plaxis was the initial stress generation before the dike has been
constructed In that phase for horizontal soil surfaces with any soil layers parallel to this surface
and horizontal phreatic level32 a K0-procedure can be applied where K0 is the ratio between
horizontal and vertical stresses
(B8)
In practice the value of for a normally consolidated soil is often assumed to be related to the
friction angle by Jakyrsquos empirical expression (Jaky 1944)
Using very low or very high K0-values in the K0 procedure may lead to stresses that violate the
Mohr-Coulomb failure criterion In this case Plaxis automatically reduces the lateral stresses such
that the failure condition is obeyed Hence these stresses points are in a plastic state and are thus
indicated as plastic points Although the corrected stress state obeys the failure condition it may
result in a stress field which is not in equilibrium
B4 Interface Strength In FEM for a proper modelling of soil-structure interaction interface elements can be added to
plates in order to simulate the thin zone of intensely shearing material at the contact between the
plate and the soil
The strength properties of the interface are linked to the strength properties of the adjacent soil
layer and they are reduced by a factor of as follows
(B9)
(B10)
(B11)
where index ldquoirdquo refers to the interfacersquos parameters whereas index ldquosoilrdquo refers to the adjacent soil
parameters The behaviour of the interfaces is defined according to the Coulomb criterion which
distinguishes between elastic behaviour where small displacements can occur within the
interface and plastic interface behaviour when permanent slip may occur
32 For non-horizontal surfaces a Gravity loading is recommended
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 152
When the interface is elastic then both slipping (relative movement parallel to the interface) and
gapping (relative displacements perpendicular to the interface) can occur The magnitudes of the
interface displacements are
(B12)
(B13)
where is the shear modulus of the interface is the 1-D compression modulus of the
interface is the virtual thickness of the interface (generated automatically from Plaxis) is
the elastic interface normal stiffness and is the elastic interface shear stiffness
The shear and compression moduli are related as follows
(B14)
(B15)
Taking into account the above relations the displacements can be highly influenced by the elastic
parameters and thus the strength reduction factor can have a considerable impact on the
systemrsquos behaviour In the Reference manual of Plaxis 2D (2015) a value of 23 is generally
recommended while a value of 1 would not normally be used It should be stated that in the
framework of this thesis there are no available measurements in order to define the sensitivity to
this parameter and hence it is considered uncertain variable following a truncated normal
distribution
Appendix C ndash NEN 6740 ndash Table 1
MSc Thesis A Rippi 153
Appendix C NEN 6740 - Table 1
MSc Thesis A Rippi 154
MSc Thesis A Rippi 155
Appendix D Input Files for the Reliability Analysis
Below an example of the input files that were used for the anchor the sheet pile the soil body and
the system failure is presented in Figures D1 D2 D3 and D4 respectively In particular
regarding the input file of the anchor failure the meaning and the scope of each line is briefly
explained in Table D1
Figure D 1 Anchor analysis input file (with FORM method)
APPENIDX D ndash Input Files for the Reliability Analysis
MSc Thesis A Rippi 156
Table D1 Explanation of input components
A Reliability method and its parameters eg FORM with a maximum number of iterations 100 a maximum absolute error of 05 a relative error of 1000 a residual error of 05 and a constraint error of 05
B Number of random variables
C Soil material as it is specified in Plaxis eg AKleiVenigOnder is the clay layer under the dike
D Soil parameter eg Rinter is the interface strength of the clay layer
E Probability distribution and its parameters eg Truncated normal distribution with mean 066 standard deviation 013 and bounds from 00 to 10
F Copula type
G I Limit state function I is the threshold of the function
H Response surface function that is activated in case of Plaxis errors
Figure D 2 Sheet pile wall analysis input file (with FORM method)
APPENIDX D ndash Input Files for the Reliability Analysis
MSc Thesis A Rippi 157
Figure D 3 Soil analysis input file (with DS method)
APPENIDX D ndash Input Files for the Reliability Analysis
MSc Thesis A Rippi 158
Figure D 4 System analysis input file (with DS method)
MSc Thesis A Rippi 159
Appendix E Characteristic and mean values
The characteristic value of a parameter implies that there is 5 probability that the real value is
higher (load) or lower (resistance) than the characteristic In Figure 22 the probability density
function of a standard normal distributed parameter (mean value equals to zero and standard
deviation 1) is depicted In case of such a distribution the characteristic value is 164σ higher than
the mean value
Figure E1 Characteristic value illustration for a normal distribution (source Wolters 2012)
In this study full probabilistic calculations are taking place in order to estimate the corresponding
structural reliability According to the probabilistic design concept the mean values and the
standard deviations are used instead of characteristic values In that case characteristic values
and shall be converted to their mean values and as follows
where
(E1)
(E2)
and are the values for load or resistance respectively to which a probability of (non-
)exceedance of 95 corresponds (non-exceedance for the load and exceedance for the resistance)
(for a standard normal distributed parameter this is 1645) The definition of the rest parameters
is given in section 423
MSc Thesis A Rippi 160
MSc Thesis A Rippi 161
Appendix F Reliability methods
In this appendix an overview of different reliability methods is given More precisely in F1 an
explanation is given concerning the generation of random samples that is applied in MC as well as
in other sampling methods In Appendix F2 more advanced sampling methods than crude MC are
presented that can be used as accelerating reliability methods in the sense that the computational
time can be reduced Finally in Appendix F3 some basic principles of the FOSM are elaborated
that can help to better understand the concept of FORM as well
F1 Generation of random samples in Monte Carlo
The non-exceedance probability of an arbitrary random variable is uniformly distributed between
zero and one regardless of the distribution of the variable In formula (CUR ndashpublication 190
1997)
(F1)
is the uniformly distributed variable between zero and one and is the non-exceedance
probability Thus for the variable X
(F2)
in which is the inverse of the PDF of X Using this formula a random number X can be
generated from an arbitrary distribution by drawing a number of from the uniform
distribution between zero and one
More or less the same way base variables of a statistical vector can be drawn from a known joint
probability distribution function However the joint PDF must then be formulated as the product
of the conditional probability distributions of the base variables of the vector In formula this is
( )
By taking m realizations of the uniform probability distribution between zero and one a value can
be determined for every
(F3)
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 162
If the variables are statistically independent this can be simplified to
(F4)
By inserting the values for the LSF one can check whether the obtained vector ( ) is
located in the safe area
F2 Other Sampling Methods
Quasi-Monte Carlo Simulation
As an alternative to pseudo-random number generation (used in Crude MC) the quasi-random
number sequences are introduced from where the method Quasi-MC took its name Quasi-MC is a
technique which approximates the probability of failure using low discrepancy sequences33
x1hellipxN instead of randomly generated sequences
If the domain of failure is the goal is to estimate the following
probability
int ( )
(F5)
The main principle of this method is the integral of the function on [ ] can be
approximated by using some low discrepancy sequence x1hellipxN as follows
int ( )
sum
(F6)
The low discrepancy sequence is generated in Δ according to Lebesgue measure and then may be
transformed to any measure micro thanks to the inverse CDF technique in order to approximate the
integral
int ( )
sum
(F7)
In Figure 4 a comparison is made between quasi-random sample and pseudo-random sample in
order to illustrate the difference in sampling for the two methods
33 The discrepancy of a sequence is low if the number of points in the sequence falling into an arbitrary set B is close to proportional to the measure of B as would happen on average (but not for particular samples) in the case of a uniform distribution
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 163
Figure F1 Example of Quasi random Sample (left) and Pseudo Random Sample (right) (source
Waarts 2000)
This method a priori provides an asymptotically higher convergence rate than traditional MC but
no general rule can guarantee a better efficiency of the Quasi-MC sampling than the classical MC
sampling The advantages of Quasi-MC sampling tend to disappear with the increase of the
number of dimensions of the input variables It is recommended to use the Quasi-MC technique
with high sampling sizes or with very low dimensionality (in OT this method is valid only if the
input random variables are independent) (OpenTURNS 15 2015)
Importance Simulation
The main idea of the Importance Sampling method is to replace the initial probability distribution
of the input variables by a more efficient one Efficient means that more events will be counted
in the failure domain and thus reduce the variance of the estimator of the probability of
exceeding a threshold In other words importance sampling starts with the observation that if we
are going to sample randomly we should distribute the points to put the most of them in the
region that contains information (ie the failure region) and to waste as few as possible
In this method a sampling density is used instead of the actual probability density function
For a MC simulation the probability of failure is calculated by
sum ( ( )) ( )
( )
(F8)
where ( ( )) ( )
( ( )) ( )
The problem is that prior knowledge is needed of the failure area which in most cases is not
available In Figure 5 examples of the different sampling methods are shown
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 164
(a) (b)
(c)
Figure F2 Comparison of Importance Sampling (a) MC (b) and Numerical integration (c) (source
Latin Hypercube Simulation
Latin Hypercube Sampling is a sampling method enabling to better cover the domain of variations
of the input variables thanks to a stratified sampling strategy (in OT this method is applicable in
the case of independent input variables) Stratified sampling tries to place the sampling points so
that more of them will be found in regions where the variance of the limit state function g(x) is
largest In Figure F3 a comparison is made between the LHS and MC in terms of their sampling
technique The sampling procedure according to Baecher and Christian (2003) is based on the
next steps
1 Divide the range of each variable into several intervals of equal probability
2 If there are D variables to be sampled each in k regions then for each variable the k
sampling points are placed in independent random order
3 K samplings are made by choosing the first randomized value for each variable the second
randomized value for each variable and so on until k sets of randomized variables have
been chosen
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 165
Figure F 3 Comparison between Latin Hypercube sampling and MC (source OpenTurns 15 2015a)
F3 First Order Second Moment (FOSM) Method The First-Order Second-Moment (FOSM) approach (Ang amp tang 1984) provides analytical
approximation for the mean and the standard deviation of a parameter of interest as a function of
the mean and standard deviations of the various input factors and their correlations
Consider Z to be a function of random variables x1 x2 hellip xn that is
(F9)
In general case are correlated with covariance matrix [C] ie [C]=[σ][R][σ] where [σ]
is a diagonal matrix of standard deviations and [R] is the (positive-definite and symmetric)
correlation matrix with diagonal elements Rii=1 and non-diagonal elements Rij=ρij (ρij is the
correlation coefficient between variables i and j) In scalar notation Cij= σiσjRij (Farrokh 2007)
The first order approximation to the mean variance and standard deviation of the Z function is
based on the first terms of a Taylor series expansion of Z The following first-order estimates of
mean and variance are obtained
[ ]
(F10)
[ ] (F11)
Where the vector b denotes evaluated at the mean values of ie
(F12)
If there is no correlation among the variables Eq 223 can be written as
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 166
sum
(F13)
Then the reliability index can be calculated in 6 steps (Baecher amp Christian 2003)
1 Identify all variables that affect the mechanism that is researched
2 Determine the best estimate (usually the mean value) of each variable E[ ] and use these
to calculate the best estimate of the function E[Z]
3 Estimate the uncertainty in each variable and in particular its variance Var[ ]
4 Perform sensitivity analyses by calculating the partial derivatives of Z with respect to each
of the uncertain variables or by approximating each derivative by the divided difference
5 Use the equation of Var[Z] to obtain the variance of the function Z
6 Calculate the reliability index
[ ]
(F14)
MSc Thesis A Rippi 167
MSc Thesis A Rippi iii
MSc Thesis A Rippi iv
Abstract
Some dike sections in the Netherlands failed to comply with the safety standards and one of the
most promising countermeasures is the construction of retaining walls inside the dike The
Dutch design codes for dikes with retaining walls rely on Finite Element Analysis (FEM) in
combination with partial safety factors However this can lead to conservative designs For this
reason in this research a reliability analysis is carried out with FEM calculations aiming to
demonstrate the feasibility of reliability analysis for such a soil-structure interaction problem
The case study concerns a (river) dike with an anchored sheet pile wall modelled in PLAXIS The
sensitivity and reliability analyses were enabled by coupling the uncertainty software package
OpenTURNS and PLAXIS through a Python interface The most relevant (ultimate) limit states
concern the anchor the sheet pile wall and global instability (soil body failure) The case was
used to investigate the applicability of the First Order Reliability Method (FORM) and
Directional Sampling (DS) to analysing these limit states Finally also the system reliability was
evaluated using sampling-based methods (DS)
Due to the considerable number of random variables before starting the reliability analysis a
sensitivity analysis was conducted for each limit state This indicated the most important soil
layers to be accounted for and the variables to be considered as stochastic The sensitivity
analysis and later on the reliability analysis were based on analytical formulations of the limit
state functions The anchor and the sheet pile limit states were formulated in terms of their
yield stress for global instability loss of equilibrium in the FEM analysis was used to define
failure Moreover the systemrsquos reliability was evaluated by taking into account all the three limit
states that were mentioned previously
The goal is to implement the coupling between FEM and reliability methods in order to analyse
the components of such a system (ie anchor sheet pile wall and dikersquos soil body) estimate the
probability of failure and identify the most important soil properties that affect the behaviour of
each component and the system as a whole The results of this research can be used to assess
and optimize the current design procedure for dikes with retaining walls
MSc Thesis A Rippi v
MSc Thesis A Rippi vi
Table of Contents
Preface ii
Abstract iv
List of abbreviations x
1 Introduction 1
11 Project objective and main research questions 2
12 Research approach and outline 3
2 System description and current design concept 7
21 System description and forces configuration 7
22 Current design concept 12
23 Safety standards 17
3 Literature study 21
31 Background 21
311 Finite Element Modeling 21
312 Uncertainties and Sensitivity analysis 24
32 Previous Studies 29
33 Overview 34
4 Structural Reliability Analysis 35
41 Basics of Reliability Analysis 35
42 Overview of Reliability Analysis Methods 36
421 Level III Methods 36
422 Level II Methods 40
423 Level I Methods (semi-probabilistic) 42
424 Response Surface Techniques (RS) 44
43 Coupling Reliability Analysis with FEM 46
431 The functionality and possibilities of OT 46
432 Coupling OpenTURNS-Plaxis 47
44 Overview 49
5 Failure Mechanisms and Limit State Functions 53
51 Introduction to the system analysis and the limit states 53
52 Limit State Functions 55
521 Serviceability Limit State 55
522 ULS for Structural Members 57
MSc Thesis A Rippi vii
523 ULS for Soil Failure 60
53 Overview 66
6 Case Study-Dike with an anchored sheet pile wall 69
61 Case Description 69
62 Soil Parameters 70
63 Finite Element Model 73
64 Deterministic Analysis 74
641 Calculation Scheme and Design Values 75
642 Construction Stages 77
643 Determination of the structural elementsrsquo characteristics 78
65 Overview 81
7 Reliability analysis results with stochastic soil properties 85
71 Method description 85
72 Mean values calculations 88
73 Sensitivity Analysis Results 91
74 Soil Shear Failure 98
75 Anchor Failure 102
76 Sheet pile wall failure 105
8 Conclusions and Recommendations 121
81 Conclusions 121
82 Recommendations 123
References 125
Appendix A 129
OpenTURNS features 129
A1 Fourier Amplitude Sensitivity Test (FAST) 129
A2 Optimization Algorithms in FORM 132
Principles of optimization algorithms 132
Convergence criteria 134
Evaluation of the algorithms performance 135
A3 Distribution Types 143
Uniform Distribution 143
Normal Distribution 144
Lognormal Distribution 144
Truncated Normal Distribution 145
Appendix B 147
MSc Thesis A Rippi viii
Plaxis 2D (2015) features 147
B1 Mohr Coulomb failure criterion 147
B2 φ-c Reduction Technique 150
B3 Initial Stress Generation 151
B4 Interface Strength 151
Appendix C NEN 6740 - Table 1 153
Appendix D 155
Input Files for the Reliability Analysis 155
Appendix E 159
Characteristic and mean values 159
Appendix F 161
Reliability methods 161
F1 Generation of random samples in Monte Carlo 161
F2 Other Sampling Methods 162
F3 First Order Second Moment (FOSM) Method 165
MSc Thesis A Rippi ix
MSc Thesis A Rippi x
List of abbreviations
CDF Cumulative Distribution Function
CoV Coefficient of Variation
CUR Civieltechnisch Centrum Uitvoering Research en Regelgeving
DS Directional Sampling
FAST Fourier Amplitude Sensitivity Analysis
FEA Finite Element Analysis
FEM Finite Element Model
FERM Finite Element Reliability Method
FORM First Order Reliability Method
FOSM First Order Second Moment
LEM Limit Equilibrium Method
LRFD Load and Resistance Factor Design
LSF Limit State Function
MC Monte Carlo
OT OpenTURNS
PDF Probability Distribution Function
RFEM Random Finite Element Method
RS Response Surface
SA Sensitivity Analysis
SLS Serviceability Limit State
SORM Second Order Reliability Method
ULS Ultimate Limit State
VNK Veiligheid Nederland in Kaart
MSc Thesis A Rippi xi
The roots of education are bitter but the fruit is sweet
Aristotle
MSc Thesis A Rippi 1
1 Introduction
In the Netherlands according to the Flood Protection Program (Hoogwater
Beschermingsprogramma1) and the Room for the River program (Ruimte voor de Rivier2)
alternative structural techniques for the reinforcement of existing dikes or for future dike
constructions additional to the conventional ones have been introduced and tend to be
attractive The heightening of the dike crest and the construction of a stability berm are some of
the most common current actions for dike strengthening Additionally filter layers geotextiles
and drainage systems can be applied in advance in order to prevent several failure mechanisms
Recently cantilever or anchored sheet pile walls and diaphragm walls tend to be used as an
alternative procedure for embankment reinforcing (see Figure 11) Such an alternative can be
chosen for an improvement of the macro stability of the dikersquos inner slope and additionally for
saving space in the land area that needs to be habited
Figure 11 Design options for dike reinforcement (source Flood Defences 2015)
For that purpose engineers need a concrete and unequivocal design methodology for such
combined structures In the meantime Deltares in cooperation with the Water Board of
Rivierenland is working on a design guideline which will be applicable to design the so-called
Type II stability fences3 A draft of such a guideline is elaborated in the report of Larsen et al
(2013) In this report suggestions and recommendations are outlined for the design of such
structures with Finite Element Modeling (FEM) and partial factors Remarks are also made for
the advantages and the limitations of such models and how they should be handled so that
reasonable and trustworthy results can be retrieved from the analysis
1 httpwwwhoogwaterbeschermingsprogrammanldefaultaspx 2 httpalbertawatercomhow-is-water-governedwhat-is-room-for-the-river 3 Type II structures are structures which in combination with a soil construction fulfil water retaining functionalities
Introduction
MSc Thesis A Rippi 1
The concept behind the recommended design criteria (Larsen et al 2013) is the definition and
the evaluation of several partial factors with the view to determine overall safety factors These
factors are further discussed in section 22 However validation of these factors shall be carried
out before being used in any case otherwise either the safety of the structure is jeopardized or
the structure will be overdesigned and cost inefficient
For improving the design criteria researchers for a long time focused on enhancing structural
models (beams shells etc) and constitutive laws (elasticity plasticity damage theories etc)
With the development of computer science a great amount of work has been devoted to
numerically evaluate approximated solutions of the boundary value problems describing the
mechanical system FEM is probably nowadays the most applied approach for the solution of
these problems
However the increasing level of detail of the constitutive models and the constant enhancement
of the computational tools do not solve the problem of identification of the model parameters
and the inherent physical and modelling uncertainties Moreover in most civil engineering
applications the intrinsic randomness of materials (soil rock concrete etc) or loads (water
elevation wind earthquake motion etc) is such that deterministic models are using average or
later on characteristic values of the properties at best lead to rough representations of the
reality
As a counteraction a semi-probabilistic methodology has been developed that was based on the
application of characteristic and design values by using partial factors Current design codes
such as Eurocode provide target reliabilities for different types of structures and structural
elements according to the potential consequences of failure However these partial factors are
not always equally suitable and efficient for all types of structural applications since they have
been calibrated under specific conditions Besides the consequences of failure of flood defences
such as dikes can be comparable to the investments in increasing the reliability of such systems
are For that reason it would be advisable for these systems to define target reliabilities based
on a risk assessment (ie tailor-made solution) rather than using the standard partial factors
coming from general geotechnical design codes which may be either too low or too high for a
given flood defence system
One step of such a risk assessment is accounting for randomness and spatial variability of the
mechanical properties of materials is one of the tasks of stochastic or probabilistic mechanics
which has developed fast in the last decade In this master thesis project the uncertainty of soil
properties is going to be treated in terms of its contribution to failure For that purpose
probabilistic methods are going to be implemented for a dike with an anchored sheet pile wall
(see Figure 12) simulated in a FEM software that is specialized in soil mechanics The successful
implementation and in future research the verification of such methods can be considered as
the most preferable and cost efficient way to design structures with high safety requirements
and not only for the validation of the partial factors Of course such a procedure tends to be
time consuming However the gradual improvement of the current probabilistic methods in
combination with the state-of-the-art computer capabilities as well as the scientific knowledge
gained in terms of different systems behaviour and failure modes can introduce a more
optimized way of designing structures with considerable investments
Introduction
MSc Thesis A Rippi 2
Figure 12 Reinforced dike section with an anchored sheet pile wall
11 Project objective and main research questions At the beginning of this research it has been observed that the design procedure being followed
for dikes with retaining walls modeled in FEM can lead to an overestimation of the critical loads
and thus to a potential cost inefficient final structure An example of the magnitude of the design
values in such a case study is given in the next chapter where also a more detailed explanation
of the current design concept is described
The objective of this study is to implement a full probabilistic analysis for evaluating the
reliability of a dike reinforced with an anchored sheet pile wall Such a research analysis can
subsequently come up with valuable recommendations for the improvement of the present
design approach Soilrsquos and structural elementsrsquo behavior should be taken into account both
separately and as a system Some of the reliability methods can deal with system reliability
problems such as Monte Carlo (MC) and Directional Sampling (DS) whereas for others like
FORM and SORM additional methods should be applied that use reliability information for each
individual limit state function to obtain the systemrsquos reliability
In principle the probability of failure of different limit states is to be computed individually
while in the sequence the system reliability is going to be estimated As far as the soil failure
mechanisms are concerned this research will focus mainly on the global stability of the dike
slope while for the sheet pile wall and the anchorrsquos failure the exceedance of the steel yield
stress is going to be considered
The scope of this thesis is not only to test different reliability models robustnessrsquo in conjunction
with FEM simulations but also to get a better insight into the specific system behavior (ie of a
dike with a retaining wall) analyze each component separately and investigate its response
under certain load conditions Furthermore the minimization of the computational effort and
time could also be carried out meaningfully under the constraint of sufficient accuracy The
accuracy should be such that the probability of failure is acceptable for the ultimate limit state
(ULS) in a normal design process and according to the current safety standards Finally the
Sheet pile
wall
Anchor
Dike section
Introduction
MSc Thesis A Rippi 3
robustness of the coupling between the reliability model and the FEM will be tested via their
capability of adapting to new input parameters without encountering convergence errors
during execution
The main research question of this master thesis project is thus formed as follows
How can the probability of failure of a dike with a sheet pile wall due to global instability modeled
by a Finite Element Model be analyzed
Essential questions regarding the soil models and failure criteria as well as the reliability
methods and the systems behavior are generated which are listed below
Subquestion 1 Which reliability methods are computationally tractable in
combination with FEM
Subquestion 2 How robust (convergence) are the tractable methods
Subquestion 3 Which limit states are relevant for the envisaged application of
retaining walls in dikes and how can they be formulated using FEM
analysis outcomes
Subquestion 4 What is the contribution of different uncertainties in the failure
mechanisms of the system
Subquestion 5 Can response surface techniques help to increase the efficiency and
robustness of the reliability model
Subquestion 6 How can the current design approach for dikes with sheet piles be
improved
In the next chapters the above research question and the related subquestions are going to be
answered by following the methodology that is described in the next section
12 Research approach and outline In this section an overview is given of how the aforementioned objective and sub-questions are
approached The thesis is mainly divided into 8 chapters and 6 appendices In Figure 13 the
thesis outline is illustrated and an indication of which subquestions (Sq1-6) are treated to
which chapter is given In Chapter 1 an introduction into this research content is made and a
first illustration of the system under investigation is presented together with the main research
questions
Chapter 2 has been devoted to the description of the system (ie a dike with an anchored sheet
pile wall) and to the case study that has been adopted for being analyzed in terms of its
reliability The forces configuration is also illustrated and the results of the case study according
to the current design concept are presented and evaluated Last but not least the new
recommended safety standards are included and the current required reliability of the specific
dike section is presented
Introduction
MSc Thesis A Rippi 4
In Chapter 3 a literature study related to FEM uncertainties and sensitivity analysis is
presented Moreover previous studies that are associated with the objective of this research are
mentioned whereas some of their results were also taken into account for proceeding with this
research
The main scope of this thesis is the implementation of different reliability methods on a specific
case study with the view to investigate and analyze its behavior Some of these methods are
continuously mentioned through the test and thus in Chapter 4 an overview is given of the
most well-known reliability methods Eventually an evaluation of these methods is made based
on literature and preliminary testing with simple case studies and the procedure of their
coupling with FEM is discussed The evaluation helped to qualitatively answer subquestion 1
while also an introduction of how special reliability methods such Response Surfaces were used
in this thesis that partially covers subquestion 5 In this chapter an introduction into the Limit
State Functions (LSF) concept is made and how they are considered for the reliability analysis of
a system An LSF actually represents a failure mode that can be detected in a structural or soil
element and it is expressed as a function of several variables In Chapter 5 the failure
mechanisms and the corresponding LSF that are related to the specific case study are identified
and formulated which answers subquestion 3
In Chapter 6 the case study whose reliability is to be evaluated is presented as it was modeled
in FEM The boundary conditions are specified and the soil and structural properties are
indicated and illustrated Moreover a deterministic design is also taking place in order to
roughly estimate the structural elementsrsquo properties This will help to make a qualitative
comparison between the current design procedure and the design according to a fully
probabilistic approach by referring thus to subquestion 6
Eventually the results of the aforementioned analysis are presented in Chapter 7 The results
mainly include the estimated probability of failure of each of the system components as well as
of the system as a whole by considering the soil properties as stochastic Moreover an
interpretation of the failure points is made and an assessment of the level of impact of the
random variables on the systemrsquos behavior is carried out At that chapter subquestions 2 4 and
5 are mainly treated
Last but not least in Chapter 8 the general conclusions are presented together with some
valuable recommendations for future consideration and research Furthermore a reflection on
the methodology and how the different research questions were approached is made
Introduction
MSc Thesis A Rippi 5
Figure 13 Thesis outline
Introduction
MSc Thesis A Rippi 6
MSc Thesis A Rippi 7
2 System description and current design concept
In this section the system to be analyzed is described more in detail and the individual
components are identified Moreover the forcesrsquo configuration is explained and illustrated as
they would have been calculated with the conventional way in parallel with a qualitative
evaluation of the potential expected deformations Finally the current design concept and the
safety standards that are referred to such a system are introduced
21 System description and forces configuration The system to be investigated in this thesis consists of a dike section reinforced with a one-layer
anchored sheet pile wall The system has been simulated by FEM and later on it was coupled
with a reliability package for carrying out a reliability analysis In the figure below an
illustration of the system is depicted and the several elements of an anchored sheet pile wall are
showed
Figure 21 System layout and different components of the anchored sheet pile wall
In particular an anchored sheet pile wall consists of the sheet piles which are anchored in the
soil via a tie rod The rod is grouted most of the times in a sandy layer with a cemented grout
body and it both alleviates the sheet piles from the axial forces exerted by the upper structure
and keeps the wall stable in case of excessive developed moments due to the lateral earth
pressure Last but not least a waling system is applied mainly downstream of the sheet piles in
order to transfer the loads from the piles to the anchors in such a manner so as to avoid
excessive local stresses on the intersection between the sheet piles and the tie rod
Cross-Section
Top View
System description and current design concept
MSc Thesis A Rippi 8
The main load configuration acting on a dike section as well as a sheet pile wall is depicted in
Figures 22-24 together with the possible displacements As far as the soil body stability is
concerned most of the slope stability analysis computer programs are based on the limit
equilibrium concept according which a soil mass tends to slide down under the influence of
gravity The transitional or rotational movement is considered on an assumed or known slip
surface below the soil There an equilibrium should be achieved among the driving and the
resisting forces In that case the driving moments consist of the soil weight the water pressures
and the loads around the center of the slip surface such as a possible vertical load on the top of
the dike crest (ie traffic load) The magnitude of the water pressures is controlled by the water
elevation on the river side
In Figure 22(a) the external and internal forces acting on a slice of the slip surface are
illustrated The driving forces are the soil weight W the water pressure U and any additional
load that can contribute to the rotation of the slip surface which in turn activate the lateral
active earth pressure The resisting forces consist of the lateral passive earth pressure
(Rankine 1857) and the shear strength (Terzaghi 1943) Their expressions are written as
follows
[kN] (21)
[kN] (22)
[kPa] (23)
where is the active lateral earth pressure coefficient and the passive lateral earth
pressure [kNm3] is the unit weight of the corresponding soil layer and H [m] is the thickness
of the soil layer that the lateral earth pressure comes from (in case of multiple soil layers there
should be a distinction among the different forces) and [kPa] is the effective cohesion of the
soil layer The safety factor that is written in the Figure 22(a) as FS is defined by the ratio of the
shear strength (excluding the pore water pressure) divided by the weight of the earth body
(including the pore water pressure)
The loss of equilibrium between the driving and the resisting moments lead to the rotation and
instability of a slip surface as it is illustrated in Figure 22(b) A slide surface can take various
shapes according to the Limit Equilibrium Method (LEM) that is followed According to Bishop
method this surface tends to be circular in Uplift-Van method a horizontally compressed zone
can be also considered whereas in Spencer method the shape of the slide body can be arbitrary
The slip surface can be located either in the landside or in the river side of the dike depending
on the load the soil characteristics and the design of the dike (ie inner or outer berm
reinforcing revetment on the outer slope etc) Most of the times the landside area of the dike is
jeopardized due to different failure mechanisms that are described in section 523 The
instability of the outer slope is mainly relevant in case of sea dikes and a potential damage can
be usually reconstructed until the next flooding
System description and current design concept
MSc Thesis A Rippi 9
(a)
(b) Figure 22 (a) External and internal forces acting on a slip surface of a dike section and (b) possible deformation pattern
As a counteraction for the slope instability a sheet pile wall can be installed inside the dike The
special thing about these structures is that they make possible a greater freedom in form and
functionality than a traditional dike design This structure derives its strength from the
materials used such as steel which are able to withstand higher pressures than clay for instance
The general stability is due to friction and wedging in the bottom
After the reinforcement of the inner side of the dike with a retaining wall the strength capacity
of the wall is also important for the global stability of the system The forces to be taken into
account for the sheet pile wall stability are the active and the passive earth pressures (effective
pressures) and the water pressure as they are illustrated in Figure 23(a) In that figure the
System description and current design concept
MSc Thesis A Rippi 10
forces were simplified in a singular triangular shape as in a homogeneous soil body in order to
explain and depict the overall picture of the acting forces However the stresses distribution
over depth can be more complex depending on the variety of the soil layers that are present In
Figure 23(a) it can be noticed that a sheet pile wall which is installed close to the inner berm
might not have significant instability issues as there is the passive side that contributes to
resisting forces However in the passive side the soil can differ and be weaker than this of the
active side Therefore the passive force in that case might not be very supportive and thus the
wall shall be designed cautiously
Additionally the anchor resistance shall be reassured for the sake of the stability of both the
sheet pile wall and the overlying soil body In Figure 23(a) the forces acting on the anchorage
are displayed The distributed load qs comes from the earth pressure of the upper soil layer (it
might not be uniformly distributed as it is depicted in the figure but more in a trapezoidally type
of load) whereas the tension force Ft4 stems both from the earth pressure and the displacement
of the sheet pile wall The bending and the tension capacity of the anchor are determinant for its
stability In Figure 23(b) a possible displacement pattern of the retaining wall and the
anchorage is illustrated
Furthermore corrosion is an additional weakening impact on the wall that depends on the
water level and the pore water pressures near the structure The thickness of the wall and the
material properties are playing a key role to the resistance towards corrosion This will not be a
subject of this thesis however it should be taken into account in case of design purposes
(a)
4 where O is the average circumference of the pile P [kPa] is the shaft friction at depth z and L is the length of the pile
System description and current design concept
MSc Thesis A Rippi 11
(b) Figure 23 (a) Mobilizing forces acting on the sheet pile wall and the anchorage and (b) possible deformation pattern
It should be mentioned that as far as the pore pressures are concerned in Figure 23(a) only the
hydrostatic pore pressures are depicted Generally the pore pressures (or active pore pressures)
are the sum of the steady state pore pressures and the excess pore pressures
[kNm2] (24)
Steady state pore pressures can be hydrostatic (based on a horizontal phreatic line) or non-
hydrostatic (based on steady state groundwater flow) Excess pore pressures are based on
loading of undrained soils In FEM these various pore pressures are taken into account
automatically In Figure 24 an example of a potential distribution of the active pressures
nearby the sheet pile wall is depicted as it was deduced from Plaxis Therefore as it can be
noticed from the figure the distribution can indeed sometimes approximated as triangular
System description and current design concept
MSc Thesis A Rippi 12
Figure 24 Example distribution of the active pore pressures acting nearby the sheet pile wall
The knowledge of the type of forces that are exerted on both the structural components and the
soil body as well as of the most expected form of displacements is valuable in order for
someone to be able to evaluate the results deduced from FEM and detect potential modelling
errors Therefore this validation was necessary before starting with the reliability analysis
In the next sections the current design approach of the system described above is elaborated
Additionally the Dutch recommended safety standards are presented for this type of structures
which show the need of carrying out a reliability analysis For that purpose a case study was
adopted that it is presented first as it was designed according to the current regulations while
finally a comparison is made between the original and the new case study as it was found to be
according to the reliability analysis
22 Current design concept In the Netherlands dike reinforcement is becoming a major issue concerning the design of flood
defenses Special reinforcing structures such as cofferdams anchored sheet piles and diaphragm
walls are used for strengthening the dike Especially in the context of the Flood Protection
Program and the Room for the River program many dike reinforcement projects have been
suggested
In the report of Larsen et al (2013) the design procedure of a dike reinforced with sheet pile
wall is prescribed using FEM The safety philosophy that is followed in this technical report
actually composes the current procedure for designing dikes with sheet pile walls and it is
linked with the usage of partial safety factors Particularly the required overall safety factor
FEM that should be compared with the one from FEM calculations is determined as follows
SFEM
b d m n
(25)
where
System description and current design concept
MSc Thesis A Rippi 13
Partial safety factor indicating the uncertainties of the soil composition and
the water pressures (also called schematization factor)
Partial safety factor which is related to the calculation model and the way the
calculations have been carried out (also called model factor)
Partial safety factor which is related to the material parameters (also called
material factor)
Partial safety factor associated with damage caused during the soil tests (also
called loss factor)
Safety factor of load
Each of these partial factors shall be at first calibrated by conducting a full probabilistic analysis
of the system under consideration In most of the cases these partial factors have been already
calibrated on previous similar projects and then they are reused for any similar case In
Schweckendiek et al (2013) a new approach of using partial factors for flood defences is
proposed whose application is still under consideration At the moment the aforementioned
partial factors for dike reinforced with sheet pile walls modelled in FEM are estimated
according to mostly engineering judgement and political negotiations
In sequence the FEM design procedure that is applied for sheet pile walls inside soil structures
is based on a so-called ldquosafety calculationrdquo (for information for this type of calculation see
Appendix B2) In this calculation an artificial reduction to the soil strength parameters (friction
angle and cohesion) is applied in order to evaluate the overall safety factor and confirm if this is
lower or higher than the required one that was estimated according to Eq 25 Furthermore the
resulting moments and forces developed on the structural elements during the specific
calculation are used for their design
This is a quite helpful and indicative calculation type available in Plaxis that can give a sense of
the structurersquos safety factor and the possible failure mechanisms that can occur under the
predefined load configuration However such a method of calculating the developed stresses
can be misleading in terms of the moments and forces acting on the structural elements and the
total deformations of the system
In this thesis the reliability of a dike with an anchored sheet pile wall is going to be evaluated
and in particular the case study of Breedeveld (2011) is to be used for that purpose In Figure
26 the location of the existing dike section is presented while in Figure 25 the structure as it
was modelled in FEM is illustrated It is essential to mention that the retaining wall does not
exist in reality but it was placed so as to implement and demonstrate the current design
regulations in the report of Breedeveld (2011) The anchorrsquos angle has been taken equal to 30deg
from the vertical which is generally a steep anchor inclination That is expected to reduce the
part of the horizontal force that is undertaken by the anchor and thus the most of it exerts on
the sheet pile wall The angle is usually taken 30deg-45deg from the horizontal axis (CUR 2005) but
this is of course depended on the load conditions and the construction requirements for each
System description and current design concept
MSc Thesis A Rippi 14
case The influence of the anchor orientation and the construction methods that are generally
followed for an anchored sheet pile wall are not considered in this thesis The case study was
taken as it was modelled in the aforementioned project and the reader should be aware that this
serves the proceedings of a reliability analysis and not of a perfectly constructed retaining wall
Figure 25 Original dike section in Varik-Heesselt as it is presented in Breedeveld (2011)
In particular based on this case study and trying to understand the design criteria a
comparison was carried out between the ldquosafety calculationrdquo and a normal ldquoplastic calculationrdquo
(ie elastoplastic drained or undrained analysis) In both cases the boundary conditions the
external loads and the input parameters are the same The difference between them is the
calculation procedure followed in order to come up with final stresses which in the case of the
ldquoplastic calculationrdquo does not include a reduction of the soil strength properties In Figures 27
and 28 the results from both calculations are presented In Figure 27 the total deformations
are exhibited while in Figure 28 the moments and the axial forces on the sheet pile are shown
Figure 26 Location of the Varik-Heeselt dike section (section number TG-078) in the Dike ring 43
BetuweTieler-en Culemborgerwaarden west (source Breedeveld 2011)
System description and current design concept
MSc Thesis A Rippi 15
As it is realised from this figure the discrepancy between the two results is quite high while as
far as the deformations are concerned no realistic conclusions can be drawn from a ldquosafety
calculationrdquo Moreover taking as design moments the moments that are deduced from the
ldquosafety calculationrdquo will lead to a much heavier and more expensive sheet pile cross section than
the ldquoplastic calculationrdquo
It is also essential to mention that the required safety factor was estimated up to FEM =18
according to the partial-factor procedure that is defined in Eq 25 Such a safety factor as
prerequisite for coming up with the design values of the moments and forces of the sheet pile
and the procedure of the ldquosafety calculationrdquo in FEM can lead to a conservative and financially
inaccessible design
(a) ldquoPlastic calculationrdquo Deformed mesh and maximum deformations (umax= 008 m)
(b) ldquoSafety calculationrdquo Deformed mesh and maximum deformations (umax= 9041103 m)
Figure 27 Deformed mesh and maximum deformations in the end of (a) plastic calculation and
(b) safety calculation
30deg
System description and current design concept
MSc Thesis A Rippi 16
(a) ldquoSafety
calculationrdquo
Mmax = 9392 kNmm Nmax= -4402 kNm
(b) ldquoPlastic
calculationrdquo
Mmax = 6833 kNmm Nmax= -1372 kNm
Figure 28 Bending moments and axial forces exerted on the sheet pile during the (a) safety
calculation and (b) plastic calculation
On the other hand the aforementioned design procedure guaranties a strong structure capable
of probably undertaking more than the expected loads and thus ensuring the safety against
flooding of the landside However the knowledge on fully probabilistic methods that has been
obtained as well as the development of the technology can introduce a reliability analysis of the
system as a preferable way for the future design concept Such a procedure should be
accompanied also by field test that would be able to validate FEM results as well as in-situ
measurements and inspections of the soil properties for calibrating their statistical
characteristics
Due to the conservative results that the aforementioned procedure came up with and because of
the simplifications that this case study was later subjected to in order to be used in this thesis a
new design was carried out in Chapter 6 and the properties of the structural elements were
redefined
The inclination towards probabilistic methods has not only been created due to the incomplete
design regulations for the system under consideration but also the current proposed safety
standards which are discussed in the section below recommends a risk based safety assessment
of the primary flood defenses which in principle implies the evaluation of the failure probability
In the next section an overview of the new recommended safety standards related to primary
flood defenses is given and the required reliability for the stability of the aforementioned case
study is elaborated
System description and current design concept
MSc Thesis A Rippi 17
23 Safety standards An important factor for the evaluation of a structurersquos reliability is the safety standards that
have been set and according which the acceptance or not of the failure probability is made After
the safety assessment5 of 2011 it has been declared that 33 (1225 km) of the primary flood
defenses (3767 km) in the Netherlands does not comply with the safety standards For this
purpose a new session of investigations has commenced in order to assess the reliability of the
existing dikes and the potential amendment of the safety standards for future constructions
Taking into account this latest information and within the framework of the Flood Risk in the
Netherlands 2 project6 (VNK2) a new Delta program was founded whose one of the main
decisions was the ldquoDelta Safety Decisionrdquo This decision contains a proposal for new safety
standards of the primary dikes as they are depicted in Figure 29 which are stricter than the
previous ones and formulated in terms of failure probability
According to this project a new policy regarding the safety assessment of the flood defenses has
been settled and proposed for the future design criteria which is based on the evaluation of the
acceptable flood risk (probability of failure (x) consequences) of each dike ring rather than the
probability of exceedance in order to achieve a level of protection that is in balance with the
societal value (Cost Benefit Analysis and life loss)
VNK has been already using reliability evaluation techniques in order to calibrate partial safety
factors Especially in geotechnical engineering the high inherent uncertainty of the soil
properties renders the evaluation of the structural reliability essential for the safety assessment
of the structure Subsequently this implies the investigation of the response of the different
reliability methods applied in a real case study of a dike The techniques that are used to
evaluate the probability of failure are discussed thoroughly in Chapter 4 and in Appendix F In
case of a complicated structure such as a dike with a retaining wall both the reliability of the
dike body and the reliability of the wall shall be evaluated and eventually a common standard
for the reliability of the system shall be defined
5 A safety assessment is an evaluation of the function of the flood protection structure and it is being conducted every approximately 6 years after the reassessment of the existing hydraulic boundary conditions 6 in Dutch Veiligheid Nederland in Kaart 2 or VNK2 It has been initiated by the Ministry of Infrastructure and the Environment (IampM) the Association of Regional Water Authorities (UvW) and the Association of Provincial Authorities (IPO)
System description and current design concept
MSc Thesis A Rippi 18
Figure 29 Maximum admissible flooding probability for primary defences according to Delta
programme 2015 (Deltaprogramma 2014)
As far as the retaining walls are concerned CUR (2005) design procedure distinguishes the
following three safety classes for retaining walls with corresponding reliability indexes
Table 21 Safety classes and corresponding reliability indices (CUR 2005)
Class I Relatively simple constructions no personal safety risks and relatively minor
damage in the case of overall failure ϐ=25
Class II Considerable damage in the case of overall failure minor personal safety
risks ϐ=34
Class III Major damage in the case of overall failure andor considerable personal
safety risks ϐ=42
System description and current design concept
MSc Thesis A Rippi 19
Additionally as far as a dike section is concerned the required overall reliability index shall be
concluded according to the new norms that are shown in Figure 29 However the reliability
index of the different failure mechanisms of a dike shall be estimated separately
According to Vergouwe et al (2014) that is part of the VNK2 report the major failure
mechanisms in dike ring 43 where the case study under investigation is located (see Figure 26)
is uplift and piping and macro-instability of the inner slope Especially in Figure 210 the
estimated percentage of the contribution of each failure mechanism to the overall probability of
flooding7 is indicated In Vergouwe et al (2014) the overall flooding probability of the specific
dike section that the case study concerns was evaluated to 17800 Therefore the macro-
instability on which this thesis is concentrated with a contribution percentage of 226 has a
probability of about 310-5 (
) and thus a target reliability index β of
approximately 4 This reliability level can be considered as a benchmark for evaluating the one
that will be entailed in the end of this thesis
Figure 210 Percentage contribution of the different failure mechanisms to the probability of flooding (source Vergouwe et al 2014)
It should be mentioned though that the aforementioned safety standards concerning the dike
ring 43 were defined according to the probability of exceedance of a certain water level that the
dike has to retain Therefore this complies with the old norms and not with those appeared in
Figure 29 However for large engineered systems such as flood defence systems with large
potential consequences and substantial investments it is worthwhile to assign target reliability
levels based on a risk assessment of the area surrounded by the specific dike ring There are
three widely used types of criteria for evaluating the risks related to floods and major industrial
hazards (Vrijling et al 2011)
Economic criteria
Individual Risk Criteria
Societal Risk criteria
7 The dependencies among the failure mechanisms were not taken into account Therefore this probability stated in Vergouwe et al (2014) is not exactly equal to the probability of flooding
System description and current design concept
MSc Thesis A Rippi 20
In order to check then if a certain system abides by the target reliability that is required the
failure probability of each failure mechanism shall be calculated based on a LSF The
aggregation of the failure probabilities of all the modes gives the failure probability of the
system which is supposed to be compared with the required in order to reassure the reliability
of the structure Therefore this thesis is dealing with the calculation of this failure probability of
a certain system modelled in FEM analysis and understanding of the systemrsquos behaviour and
finally the investigation of the robustness of coupling FEM with reliability methods for
evaluating the reliability of a structure In the next chapter an overview over the research has
been done until now concerning coupling and reliability methods applications is presented and
a discussion over the most important is made
MSc Thesis A Rippi 21
3 Literature study
In this chapter an overview of some principles related to FEM and the uncertainties in
geotechnical engineering is presented More precisely an introduction into FEM concept and a
discussion over the different types of FEM is made Moreover the uncertainties in geotechnical
engineering are stressed and the general framework according to which they are handled is
described Finally previous studies that are associated with the application of FEM on soil
structures and the reliability analysis are discussed
31 Background
311 Finite Element Modeling
FEM is a numerical method whose essence is to convert a problem described by partial
differential equations over space and time into one by dividing the space-time continuum into a
set of discrete elements and assuming that the unknowns vary over each element FEM solution
process is as follows
1 Divide structure into pieces (elements with nodes) (discretizationmeshing)
2 Connect the elements at the nodes to form an approximate system of equations for the
whole structure (forming element matrices)
3 Solve the system of equations involving unknown quantities at the nodes (eg
displacements)
4 Calculate desired quantities (ie strains and stresses) at selected elements
The properties of each element are set the same as the material properties that have been
defined by the user Then the Deterministic Finite Element Method (the finite element method
that was formulated with deterministic variables) can be used in conjunction with means and
standard deviations of the input variables to obtain reliability estimates
In the simple coupling of FEM with reliability analysis each parameter that is considered as
stochastic is given a particular probability density function estimated either by field tests or by
engineering judgement This approach is also called ldquosingle random variablerdquo and it assumes
that the spatial correlation length is infinite In other words the soil is considered to be
homogeneous and the stochastic property assigned to the soil is taken at random from a certain
probability distribution
Nevertheless a more realistic model should take into account the spatial correlation within
smaller regions where the stochastic property is allowed to vary For that purpose the Random
Finite Element Method (RFEM) was introduced by Fenton and Vanmarcke (1990) in which the
random variables are correlated to one another using auto-correlation functions
The most sophisticated method of FEM is the Stochastic Finite Element Method introduced by
Ghanem and Spanos (1991) which accounts for the uncertainties in the input parameters
Literature study
MSc Thesis A Rippi 22
implicitly within the finite element calculation This aims at representing the complete response
probability distribution function (PDF) in an intrinsic way Two steps are basically applied for
that purpose
Discretization of the problem and definition of the random vector with an unknown
joint PDF and
Expansion of the response in a particular basis of random vectors with a finite variance
called the polynomial chaos
In this thesis the Deterministic Finite Element Method is to be used coupled with reliability
analysis According to Waarts (2000) coupling FEM with reliability methods (FERM) would lead
to the following advantages
In comparison to standard finite element analysis it gives direct insight into the
structural reliability and decisive parameters
Structures designed using FERM will either be safer orand more economically built in
comparison to structures designed using safety factors and classical constitutive models
FERM is expected to be useful for the determination of LSF that cannot be explicitly formulated
and that differ in each case such as soil limit state in different structural schematizations
Moreover it is likely to be valuable in areas where little knowledge exists on the systems
reliability of structures with multiple components (ie a dike with a sheet pile wall)
FEM in practice
The conventional method for stability analysis in a soil body is represented by LEM although
FEM is increasingly used by designersresearchers The latter has been proved to be quite
realistic for the progressive behaviour (ie stress-strain development in different construction
phases) of a soil system under the effect of stress redistribution in comparison with classical
models Especially in their master thesis Johansson amp Sandeman (2014) compared the
deformations and the forces measured at a deep excavation supported by anchored sheet pile
wall in a railway tunnel located in Gothenburg in Sweden with 1D finite element software the
2D finite element software (Plaxis) and hand calculations They proved that Plaxis produces
reliable results for horizontal deformations in the sheet pile wall and anchor forces when
compared to in-situ measurements
Moreover in Gonzaacutelez et al (2013) it was proved that the simplification of reality done by the
classical methods such as Blumrsquos Engelrsquos Kreyrsquos methods etc (for further information for these
methods a reference is made to Gonzaacutelez et al (2013)) allows us to generally understand the
behaviour of the system wall-soil Nonetheless the results that came out of this analysis were
found to be quite conservative whereas FEM managed to give a more realistic interpretation of
the wallrsquos movement
Seed et al (2006) investigated the performance of the New Orleans Flood Protection Systems
during hurricane Katrina in August 2005 by simulating the different levee sections with FEM As
an example in Figure 31 pictures of the failure of the Lower Ninth ward IHNC East Bank in
New Orleans after the hurricane Katrina are presented FEM was proved to be an efficient tool
to realistically interpret the shape and the triggers of the failure mechanism
Literature study
MSc Thesis A Rippi 23
Figure 31 Failure of the Lower Ninth ward IHNC East Bank in New Orleans modelled in FEM with Plaxis (source Seed et al 2006) At the bottom right the shear strain contours are depicted that are developed through the embankment and at the bottom left the deformed mesh is presented illustrating the failure mode of the levee
In recent implementation of design concepts and technical recommendations such as CUR 166
ndash Damwandconstructies (2005) which refers to the design of retaining walls it can be observed
that guidelines for Finite Element Analysis (FEA) have been set parallel to the existing analytical
or empirical calculation methods Therefore FEA tends to become more and more accepted as
an alternative for Serviceability Limit State (SLS) as well as for ULS design
According to Wyllie and Mah (2004) LEM give an estimation of the factor of safety with no
information on deformation of the slope However in numerical analysis the failure surface can
evolve during the calculation in a way that is representative of the natural evolution of the
physical failure plane in the slope In that way a better insight into the evolution of failure
mechanisms can be gained
Cundall (2002) compared the characteristics of numerical solutions and LEM in solving the
factor of safety of slopes and concluded that continuum mechanics-based numerical methods ie
FEM have the following advantages
No pre-defined slip surface is needed
The slip surface can be of any shape (in contrast with Bishop and Kranz stability)
Multiple failure surfaces are possible
No static assumptions are needed
Literature study
MSc Thesis A Rippi 24
Structures (such as footings embankments etc) and structural elements (such as
beams cables sheet piles etc) and interface can be included without concern about
compatibility
It is also important to recognize that LEM only identifies the onset of failure whereas FEM
includes the effect of stress redistribution and progressive failure after its initiation Numerical
models can also be used to determine the factor of safety of a slope in which a number of failure
mechanisms can exist simultaneously or where the mechanism of failure may change as
progressive failure occurs (Hoek et al 2000)
On the other hand the accuracy of FEM is dependent on usersrsquo settings such as the refinement
of the mesh generation the loading steps etc Moreover even if FEM can be functional and easy
to use it is essential that the user has a deep knowledge of the subject under investigation (soil
and structural mechanics) in order to be sceptical and critical with the FEM results able to
interpret the output behaviour of the structure and changecorrect everything that seems
peculiar Moreover the usage of FEM implies also some basic knowledge on numerical methods
and the general background of the FEM so as to solve possible numerical errors
An ideal way of modelling in FEM would be a prior calibration of the program according to data
related with the structurersquos properties and performance (ie stress generation) However this
means that enough field measurements should be carried out and under specific conditions in
order to be considered as a base for FEM calibration Unfortunately such field tests are not
always available and if so their reliability is on doubt For example measurements related with
the developed moments or displacements of a retaining wall inside a dike section are rare to be
found but even if there are some available they cannot represent the situation near the failure
domain where most of the engineers are worried about
312 Uncertainties and Sensitivity analysis
Uncertainty in geotechnical engineering can be categorized into aleatoric epistemic and
decision model uncertainty as they are depicted in Figure 32 (Baecher amp Christian (2003))
Aleatoric uncertainty is associated with the natural variability which is defined as the intrinsic
randomness of natural processes Such variability can be expressed by changes of the soil
properties over time at a certain location (temporal variability) or over space at a single time
(spatial variability) Epistemic uncertainty can be divided into the site characterization model
and parameters uncertainty and it is attributed to lack of information about events or lack of
understanding the physical laws that prohibits the ability of creating a realistic model Finally
the decision uncertainty describes the difficulty of being aware of social objectives defining
social values such as discount rates and predicting the planning horizon
Literature study
MSc Thesis A Rippi 25
Figure 32 Categories of uncertainty in geotechnical engineering (modified from Baecher amp
Christian (2003))
Below the general steps of an uncertainty study are described and an introduction into the
sensitivity analysis concept is made
Global methodology of an uncertainty study
A first step of an uncertainty study8 can be described as ldquothe definition of the problemrdquo Initially
the variables of interest (or else the output variables) of which the uncertainty is to be
quantified shall be specified In sequence given several input variables for which the user may
have data andor expertengineering judgment a model denoted usually by a mathematical
function should be introduced that enables the computation of the set variable of interest
After the general context has been staged we should choose the criteria with which the
uncertainty can be evaluated The most complete measure of uncertainty when dealing with a
random vector is the probability distribution In order to assess the value of this distribution
function the following criteria can be followed
Probability of exceeding a threshold the aim is to assess the probability that the variable
of interest exceeds a threshold important for the goals at stake
Quantiles the aim is to assess the threshold that a variable of interest may exceed with a
probability equal to a given value
Central dispersion the aim is to obtain an ldquoorder of magnituderdquo of the uncertainty by
specifying the average value and the variance of a variable of interest
8 In the engineering world this is also called reliability analysis However in this section the general uncertainty study is described In particular reliability describes the ability of a system or component to function under stated conditions for a specified time period
Literature study
MSc Thesis A Rippi 26
The next step is to define a model to represent and quantify the uncertainties of the input
variables One shall investigate each variable as a singularity and come up with the most
suitable probability density function (mostly depending on available data) Besides it is
essential to assess also the potential correlations among the variables that can be included in a
joint probability density function which is discussed later in this section
Eventually the uncertainties of the input variables shall be translated in terms of uncertainty on
the variables of interest This procedure is called uncertainty propagation and can be carried
out via several reliability methods (approximation methods or sampling methods) that are
extensively elaborated in Chapter 4 In Figure 33 a schematization of the different steps during
an uncertainty study is given
Figure 33 Methodology of uncertainty study
Last but not least a better understanding of uncertaintiesrsquo influence can be achieved by
analyzing the contribution of the different uncertainty sources to the uncertainty of the
variables of interest via a sensitivity analysis Such an analysis aims at identifying important
parameters for the system response besides it gives a better insight into the model used to
quantify the uncertainties Before conducting an uncertainty analysis it is advisable to filter out
parameters of less significance in order to reduce the modeling computational effort Below
different methods for sensitivity analysis are described
Sensitivity analysis
Sensitivity analysis (SA) is the study of how the variation in the output of a model can be
apportioned quantitatively and qualitatively to the variation in the model parameters (Saltelli
et al 2004) Saltelli (2004) proposes one possible way of grouping these methods into three
classes (i) screening methods (ii) global SA methods and (iii) local SA methods
i Screening methods
Screening is a particular instance of a sampling-based method The objective here is to identify
which input variables are contributing significantly to the output uncertainty in high-
dimensionality models rather than exactly quantifying sensitivity (ie in terms of variance)
Screening tends to have a relatively low computational cost when compared to other
Literature study
MSc Thesis A Rippi 27
approaches and can be used in a preliminary analysis to weed out less influential variables
before applying a more informative analysis to the remaining set One of the most commonly
used screening methods is the elementary effect method
ii Global SA methods
Global SA techniques incorporate the whole range of variation and the probability density
function of the input parameters to calculate their influence on the output Many global
sensitivity analysis techniques are now available such as
Variance-based methods
o Sobolrsquos method9
o Fourier Amplitude Sensitivity Test (FAST)10
o Analysis of Covariance (ANCOVA)
Regression-based methods11
Both FAST and Sobolrsquos method rely on the assumption of parameter independence while
ANCOVA can also handle correlated input parameters The main principles of FAST method are
presented in Appendix A1 For more information about the other methods the reader can
betake himself to the related references
iii Local SA methods
Local SA methods provide the slope of the calculated model output in the parameter space at a
given set of values (Turanyi and Rabitz 2004) This basically means that local methods involve
taking the partial derivative of the output Y with respect to an input factor Xi |
|
where the
subscript indicates that the derivative is taken at some point in the space of the input
Examples for these are (Schweckendiek 2006)
the α-values in a FORM calculation (further discussion in section 422)
Local Probabilistic Sensitivity Measure
In reliability analysis the local sensitivities will be more important than the global ones in the
sense that the influence of all variables in specific points such as the design point cannot be
identified by the global methods However local SA can only inspect one point at a time and the
sensitivity index of a specific parameter is dependent on the central values of the other
parameters
In this thesis the influence of the different parameters on the response of the system is
evaluated according to local sensitivity indices However before the main part of the reliability
analysis starts it was necessary to filter out the less influencing variables in order to reduce the
number of the major variables and make the analysis more efficient and computationally
affordable For this purpose the global sensitivity method FAST was used whose results are
presented in Chapter 7
9 created by Sobolrsquo (1993) 10 introduced by Cukier et al (1973) 11 introduced by Helton et al (2005)
Literature study
MSc Thesis A Rippi 28
Probabilities as a measure of uncertainties
Most engineers today use the concept of probabilities as the standard way to talk about
uncertainty One way to quantify a probability of a random variable is the calculation of the
cumulative probability function (CDF) of non-exceedance which can be obtained by
the probability distribution function For a random vector this reads
( ) (11)
From this the joint probability density function of this random vector can be determined as
( )
( )
(12)
This function is depicted in Figure 34 by means of contour levels The joint probability density
function has the shape of a hill as it is showed in the 3D figure and the inclination of the ellipses
reveals a correlation between the two variables and
Figure 34 Joint probability density function in 2D and 3D
The functions that join or couple multivariate distribution functions to their one-dimensional
marginal distribution functions are called copulas These are going to be used in order to define
the joint density probability distribution from which the random variables are taking their
values However no detailed explanation is made in this report regarding copulas and its
characteristics as it is automatically implemented in OT whereas the user has to define only the
type of copula that heshe considers the most appropriate (a reference is made to Nelson (1999)
for further information) According to the type of copula that is selected the order of the values
of the variables can be chosen within an iterative reliability method In this thesis an
independent copula was used considering that the input variables are independent the ones
from the others due to the time limit
Literature study
MSc Thesis A Rippi 29
32 Previous Studies In this section a summary of the most relevant researches on the topic of FEM applied in soil
structures and coupling with reliability methods is given
To begin with an attempt to estimate the reliability of a structure modeled in FEM has been
made by Waarts (2000) by introducing an optimized reliability method in terms of
computational effort and efficiency
In particular in Waarts (2000) two adopted reliability methods are introduced both making
use of a response surface (a detailed explanation of the reliability methods and the response
surface techniques is made in Chapter 4 and Appendix F) These adaptive response surfaces are
used in combination with FORM and DS respectively The accuracy and the effectiveness of
these methods are investigated on the basis of artificial LSFs and a comparison is made with the
existing standard reliability methods The most efficient combinations of response surface
techniques and reliability methods were with FORM (FORM-ARS) and DS (DARS) Comparing
these two methods DARS predominated over FORM-ARS as it can cope with a much wider
range of limit state functions
In a later stage the above best performing reliability method (ie DARS) is further investigated
in terms of its efficiency on the basis of complex structures reliability In Figure 35 a couple of
case studies used to verify the performance of DARS are given
Figure 35 Left Two-story two-bay frame structure Right overview of a sheet pile installation
(source Waarts 2000)
The conclusions of this research showed that DARS serves its purpose and that the benefit from
using it increases with increasing number of random variables In Figure 36 the Limit State
Function Evaluations (LSFE) that are carried out as a function of the number of variables are
presented
Literature study
MSc Thesis A Rippi 30
Figure 36 Number of LSFE as a function of the number of variables n The upper line depicts the
required number of LSFE N=160 (it is supposed that on average 3 LSFE per sample are required
and consequently the number of LSFE equals 3x160n=480n) The lower line shows the
performance of DARS (source Waarts 2000)
More recently a methodology for assessing the reliability of geotechnical structures and in
particular deep excavations has been developed by Schweckendiek (2006) More precisely this
study deals with the application of different reliability methods in combination with FEM which
carries out the LSFE The reliability methods are applied initially on simple examples in order to
be calibrated and eventually the most efficient methods are used for the reliability assessment
of a deep excavation with a sheet pile wall The several failure mechanisms are treated in detail
concerning the structural elements of the excavation as well as the soil medium
Finally the failure probability of the system is estimated according to a fault tree
schematization of a deep excavation with one layer anchored sheet pile wall Uncertainties in
the soil properties the phreatic levels and also the corrosion of the sheet pile wall were taken
into account In Figure 37 a picture of the case study is shown together with a FEA illustration
and the importance factors of different soil properties in terms of the probability of failure after
a FORM analysis
Literature study
MSc Thesis A Rippi 31
Figure 37 Finite Element simulation and contribution of several structural and soil properties in
the failure mechanism of an anchored retaining wall in layered soil (source Schweckendiek 2006)
For a similar case study as in Schweckendiek (2006) a cantilever and an anchored sheet pile
wall the safety of the structure based on LEM is evaluated in a research conducted by Gonzaacutelez
et al (2013) Moreover in order to compare the efficiency of the classical methods with FEM
the finite element software Plaxis has been used This research shows the limitation of the
classical models and the efficiency of FEM as far as an anchored sheet pile wall in concerned
More precisely LEM results found to be conservative as they do not take into account the
confinement around the free length of the anchor rod that increases the passive pressure After
comparing numerical results with the classical methods in anchored walls the failure
mechanism given by the FEM suggests an intermediate failure state where the sheet pile wall
describes a translation movement on the deep zone and at the same time a rotation movement
around the anchor point that is closer to the reality
Furthermore Moumlllmann and Vermeer (2008) conducted a reliability analysis of a dike failure A
case study at river Elbe in Saxony Germany has been used and the failure probabilities of
different dike cross sections involving different failure modes were compared More precisely
overflowwave overtopping upliftpiping slope stability and damage of the revetment on the
waterside were taken into account For each failure mode the particular failure probability was
determined using the software package PC-Ring (Vrouwenvelder et al 1999) while the
reliability method used for the assessment of the failure probability was FORM That was
coupled with a slope stability software where Bishoprsquos approach was performed
However this method is limited to circular slip surfaces and prescribed pore pressure
distributions within the dike In order to overcome this limitation FEM was used for further
analysis In combination with FEM an adaptive response surface technique was used in order to
represent the limit state function of each failure mode In Figure 38 the simulated structure
Literature study
MSc Thesis A Rippi 32
together with the response surface of the slope instability failure mode is showed Note that this
structure is approached with a single soil layer (same soil within and under the dike) and that
no structural element is implemented inside the dike
Figure 38 Dike model for the probabilistic FEA of dike stability and the corresponding response
surface of the limit state function for 3 stochastic input parameters (source Moumlllmann and
Vermeer 2008)
The failure probability of each of the aforementioned failure mechanisms were finally calculated
and compared with those recommended in each case while the performance of the coupling
between FEM and response surface method was assessed
Additionally a more recent investigation in terms of dikes with sheet piles employing FEA was
carried out by Breedeveld (2011) The main objective of this study was to display the
implementation of partial factors for design purposes using FEM and especially the software
Plaxis in simulating dikes with a sheet pile wall The dike was firstly simulated as a single
structure without reinforcement while in sequence the occurring stresses and pore pressures
were calculated with an existing sheet pile wall The results of the distribution of the effective
stresses within and below the dike are illustrated in Figure 39 His case study has been
introduced in Chapter 2 while in Chapter 6 a more detailed explanation of how it was modelled
in FEM is given This case study will be used as an example case in this thesis
(a)
Literature study
MSc Thesis A Rippi 33
(b)
Figure 39 Effective stress distribution before (a) and after (b) the installation of the sheet pile wall (source Breedeveld 2011)
Last but not least Bach (2014) has been occupied with the reliability analysis of bored piles and
the case study that he examined is presented in Figure 310 The major objectives of this
research were (1) Propose models to calibrate resistance factors for the Load and Resistance
Factor Design (LRFD) (see section 423 for further explanation of this method) approach and
find a suitable model aiming to directly determine reliability of a bored pile considering some
types of defect that may occur in the bored pile (2) Select a quality control method and evaluate
its reliability when applied to bored piles
For that purpose he made a coupling calculation between the finite element software Plaxis
(version 90) and the numerical probabilistic toolbox Prob2B in order to design bored pile
foundations in light of the reliability-based design Two failure modes the geotechnical failure
mode and the structural failure mode were proposed in order to comprehensively assess the
reliability of an axially loaded pile The parameter uncertainty was considered through the use
of statistical parameters and probability distributions for material parameters in soil models
The soil parameters were treated as random variables The geometry parameters of pile were
used as deterministic quantities because a change in pile shape in the calculation process
requires establishing a new mesh which is now impossible with regard to the Plaxis software
Figure 310 Nine calculation cases considered with different pile toe levels (source Bach 2014)
Literature study
MSc Thesis A Rippi 34
For the reliability of the pile the ULS of the pile was investigated and finally a displacement
criterion of a percentage of the pile diameter used depending on the soil type and
characteristics that the structure is founded on
33 Overview In this chapter concepts that are related to FEM probabilities uncertainty and sensitivity
analysis were discussed The introduction of FEM applications and the principles of how to
handle uncertainties will help the reader to better understand and follow the process of this
thesis
Furthermore preliminary researches related to several aspects of this thesis are presented that
mainly concern
Reliability methods and their applications [Waarts 2000 Schweckendiek 2006 Bach
2014]
Coupling FEM with reliability methods [Waarts 2000 Schweckendiek 2006 Bach
2014]
(Reliability) Analysis of geotechnical structures and especially dikes reinforced with
retaining walls [Waarts 2000 Schweckendiek 2006 Moumlllmann and Vermeer 2008
Breedeveld 2011 Gonzaacutelez et al 2013 Bach 2014]
As someone might have noticed from the overview of the existing research on the topic both the
performance of different reliability methods and the simulation of a dike with a sheet pile wall
in FEM have been carried out These references were quite helpful in order to get an idea of
coupling FEM with reliability methods as well as FEM and reliability methods individually
Moreover a first dive into the system under investigation and the variables to be handled was
made However noone has analysed the reliability of such a complicated system that is more
and more often seemed to be applied in the Netherlands Thus it would be essential to evaluate
the implementation and the results of the reliability analysis of a dike with sheet piles simulated
in FEM The conclusions of such a research are going to contribute to design and optimization
concepts and hopefully to a better understanding of the system behaviour
MSc Thesis A Rippi 35
4 Structural Reliability Analysis
The aim of this chapter is to introduce the basic theory of a reliability analysis and to summarize
the principles of different reliability methods that are widely used for the uncertainty
evaluation of a system Particularly three of them (MC DS and FORM)12 were used in this thesis
and are described in this chapter Finally a brief evaluation of the selected reliability methods
and the way that they are applied on this thesis by coupling them with FEM are discussed
41 Basics of Reliability Analysis In the consideration of the reliability of an element the determination of the probability of
failure is the central issue The limit between failure and non-failure is defined as a limit state
and the reliability is the probability that this limit state is not exceeded The limit states are
interpreted through the so-called limit state functions (LSF) whose general form is
(41)
in which R is the strength or more general the resistance to failure and S is the load or that
which is conducive to failure (ldquosolicitationrdquo) The basic principle of structural design is that the
resistance needs to be higher than the load or in other words that the LSF is larger than zero
(Zgt0) The main objective of the design is to ensure that this performance criterion is valid
throughout the lifetime of a structure Nevertheless the majority of the quantities that both
resistance and load consists of are uncertain Therefore a probability of satisfying the preceding
criterion is estimated and expresses the reliability of the structure P(Zgt0) The probability of
failure is
(42)
Hence the probability of failure is complement to the reliability to the sense that
(43)
The LSF is plotted in Figure 41 indicating the failure space and the limit state function by Z=0
12 Besides RS techniques were used for proceeding with the analysis but most for solving numerical problems rather than evaluating the failure probability The RS techniques are still under investigation and need to be tested before being applied in real cases However it is highly possible that this would be a potential way of evaluating the failure probability in the future
Structural Reliability Analysis
MSc Thesis A Rippi 36
Figure 41 Limit Sate function
In this case the design concept is based on the evaluation of the so-called design point which is
the point in the failure space with the greatest probability density Generally this point is
located on the border between the safe and the unsafe areas
In the structural domain the Joint committee on structural safety (1981) proposed a level-
classification of the calculation methods This classification includes the following three levels
Level III these methods calculate the probability of failure by considering the probability
density functions of all strength and load variables The reliability of an element is linked
directly to the probability of failure
Level II this level comprises a number of methods for determining the probability of
failure and thus the reliability It entails a linearization of the LSF in a carefully selected
point These methods approximate the probability distribution of each variable by a
standard normal distribution
Level I at this level no failure probabilities are calculated The level I calculation is a
design method according to the standards which consider an element sufficiently reliable
if a certain margin is present between the representative values of the strength and the
loads This margin is created by taking the so-called partial safety factors into account in
the design
In the next section the different reliability methods to be used in this thesis are further
elaborated
42 Overview of Reliability Analysis Methods
421 Level III Methods
The foundation of the Level III failure probability calculation is the mathematical formulation of
the subset of the probability space which involves failure (Zlt0) Level III reliability methods
Structural Reliability Analysis
MSc Thesis A Rippi 37
(also known as fully probabilistic methods) compute the probability of failure based on the
exact probability density function and the exact limit state functions Therefore such methods
are considered to be the most accurate Well-known sampling methods are (Quasi-) MC
simulation Importance Simulation DS and the Latin Hybercube Simulation Below the former is
discussed as it represents the fundamental theory on which the sampling methods are based on
while DS is also further explained as it is continuously used in this thesis For the rest a brief
explanation is provided in Appendix F There are also other sampling methods such as the
Numerical and the Directional integration which will not be described in this study However
for further information about them a reference though is made to Waarts (2000) In Level III
methods errors can only occur by a too low number of simulations
Crude Monte Carlo Simulation
The MC method uses the possibility of drawing random numbers from a uniform probability
density function between zero and one If and are the
vectors for the resistance and the load respectively that consists of n variables then by taking
multiple realizations of the uniform probability distribution a value can be determined for
every and
By repeating this procedure a large number of times the failure probability can be estimated
(44)
where is the total number of simulations (nm draws from the uniform distribution in which
m is the number of base variables) and is the number of simulations for which Zlt0
The simulationrsquos relative error is
(45)
If the expected value of the relative error is zero the standard deviation is
radic
(46)
Based on the central limit theorem the error is normally distributed provided is sufficiently
large The probability that the relative error is smaller than the given value E is then
(47)
Thus for the reliability of the relative error is smaller than For the wanted k and
E the required number of simulations can be determined with
Structural Reliability Analysis
MSc Thesis A Rippi 38
(48)
If for instance a reliability of 95 is required for a maximum relative error E=01 the required
number of simulations amounts to
(49)
The number of simulations is therefore still dependent on the probability of failure
MC simulation is applicable whatever the complexity of the deterministic model is However the
number of samples required to get an accurate estimation of may be dissuasive especially
when the value of is small (10-k) Thus crude MC is not applicable when small values of are
sought andor when the CPU cost of each run of the model is non-negligible
In Appendix F and especially in sections F1 and F2 a further elaboration of the MC simulation
is presented Moreover the methods of Quasi-MC Importance sampling and Latin Hybercube
are described
Directional Sampling
The directional simulation method is an accelerated sampling method It implies a preliminary
iso-probabilistic transformation as for FORM and SORM methods however it remains based on
sampling and is thus not an approximation method Below the method as it is described in Open
TURNS 15 Reference Guide (2015 pp190-193) is presented
Each simulation of the directional simulation algorithm is made of three steps Using the
probability distribution of a random vector we seek to evaluate the following probability
( ) (410)
Here is a deterministic vector and ( ) the limit state function which enables the
definition of the event For the iteration these steps are the
following
Let A point is drawn randomly on S according to a uniform
distribution
In the direction starting from the origin and the passing through solutions of the
equation ( ) (ie limits of ) The set of values of that belong to is
deduced for these solutions it is a subset of
Then one calculates the probability By property of independent
standard variable
is a random variable distributed according to chi-square
distribution which makes the computation effortless
Structural Reliability Analysis
MSc Thesis A Rippi 39
Finally the estimate of the probability after N simulations is the following
sum
(411)
The main idea is that for each direction we go along the direction by step of a specified length
from the origin to the maximum distant point and we check if there is a sign change on each
segment so formed (see Figure 42) We go until the maximum distant point Then for all the
segments where a sign change is detected we research the root λ in the segment (there are
also other strategies available in OT however all of them follow the same principle of
searching) The following figure illustrates the principle of an iteration in dimension 2
Figure 42 DS of a 2-dimensional problem
Similar to MC method the outcome of the estimated probability of failure is a random variable
and the error in the estimate can be made as small as possible by taking a sufficient number of
samples For DS the standard deviation σ of the estimated failure probability can be quantified
as follows (Melchers 2002)
radic
sum
(412)
From this equation relative errors and the confidence intervals can be estimated As can be
seen the error in the estimated probability of failure will decrease with increasing number of
sampled directions Equation (412) can be used to determine the number of sampled directions
that is required for a reliable estimate of the failure probability The convergence criterion in
this method is usually the coefficient of variation (CoV) which is defined as
Structural Reliability Analysis
MSc Thesis A Rippi 40
422 Level II Methods
Level II methods (known also as fully probabilistic methods with approximations) can take all
the probabilistic properties of the random variables into account but they include
approximations of the limit state function and therefore their use and outcomes should be
inspected and evaluated in order to be considered as reliable However experience in that
methods has shown that the computational effort is profoundly decreasing in comparison with
Level III methods and that the application of them can provide important parameters such as
the reliability index (β) and the influence factors (α) Some of the most known Level II methods
are First Order Second Moment (FOSM) Method First and Second Order Reliability Methods
(FORM and SORM) and the Point Estimate method (PEM) In the next paragraphs FORM is
elaborated as it will be mainly used in this thesis while in Appendix F3 the principles of FOSM
are discussed
First Order Reliability Method (FORM)
The principles of FORM are based on the First-Order Second-Moment (FOSM) approach which
first introduced the reliability index concept This method is presented in Appendix F3 in detail
However the reliability index estimated using the FOSM approach is not ldquoinvariantrdquo which
means that the index value depends on the limit state function format (Farrokh 2007) Hasofer
and Lind (1974) proposed an invariant definition of the reliability index that leads to a
geometric interpretation known as first-order reliability method (FORM)13 The first step of this
method is the transformation of the random variables to equivalent standard normally
distributed variables and the whole procedure is carried out in u-space (or else standard space)
For variables that are normally distributed this step is as follows
(413)
For other types of distributions there are procedures available for carrying out this
transformation such as Generalised Nataf and Rosenblatt transformations In this study these
transformations are not discussed however for further explanation a reference is made to the
Reference Guide of OpenTURNS 15 (2015)
In the sequence the limit state function Z is expressed in terms of
The second step is the approximation of the function with the first two terms of the Taylor-
polynomial The approximation reads (CUR 1997)
( ) ( ) sum
(414)
where is the point where the linearization takes place This approximation of Z is linear and
according to the central limit theorem it is normally distributed The expected value of the LSF
can be approximated by the expected value of the linearized function
13 VNK2 uses FORM in view of the calculation time required In case of inaccuracies VNK2 switches to more time-consuming techniques such as DS whereby a large number of draws are made from the probability density functions of the various stochastic variables (uncertain quantities)
Structural Reliability Analysis
MSc Thesis A Rippi 41
( ) sum
(415)
While the standard deviation (sum of contributions of each variable to the variance of Z) is
defined as
radicsum
( )
(416)
Using the definition of FOSM for the reliability index [ ]
where [ ] is the mean value of
the limit state function and substituting Eq 415 and Eq 416 in it the reliability index can be
approximated However linearization in different points can lead to different values of the
reliability index Hasofer and Lind definition of the reliability index overcomes this limitation
and renders it equal to the minimum distance from the origin to the design point
(417)
where radic
Looking for the design point is basically an optimization problem Many analytical and
numerical approaches can be used for that purpose A relatively straightforward method to do
this is by firstly assuming that the design point is the mean value (the starting point can also be
another point according to an engineering judgement in order to accelerate the optimization
procedure) The obtained β-value is used to determine a new point in which the LSF is
linearized In this case the importance factors αi are calculated as
( )
radicsum (
( ) )
(418)
where is the design point and is the number of variables expresses the
contribution of the variance of each variable to the total variance of Z in the design point Figure
43 illustrates the linear approximation of the limit state function and the aforementioned
parameters for a two dimensional problem The probability that Zlt0 can be determined using
the standard normal CDF
(
) (419)
Structural Reliability Analysis
MSc Thesis A Rippi 42
The new calculation point is determined by
(420)
After some iterations the chosen optimization algorithm finds the final design point and the
new reliability index In Appendix A2 an overview of the different optimization algorithms
available in OT is presented together with an evaluation of their performance in the specific
case study
Figure 43 Two-dimensional illustration of u-space LSF and design point
423 Level I Methods (semi-probabilistic)
At the beginning of the probabilistic concepts incorporation in the field of structural
engineering the most notable development was the implementation of LFRD At the past single
factors of safety on the ratio of total resistance to total load were used in order to ensure that
the stresses developed from an applied load on a structure were lower than the allowable ones
LRFD replaces those factors by introducing a set of partial safety factors on the individual
components of load and resistance These partial factors can be selected such as they account
for uncertainties associated with the resistance and the load respectively
The current design philosophy in CUR 166 and Eurocode is characterized by the use of
characteristic values of the parameters (in Appendix E a further explanation of the
characteristic values is given together with the way that they are converted into the mean
values for being used in the reliability analysis later on) The values of the partial safety factors
are additional factors to the characteristic values Therefore the characteristic values are
multiplied with the aforementioned safety factors ( ) and the following criterion has
to be satisfied for a reliable structural performance (see Figure 44)
(421)
Structural Reliability Analysis
MSc Thesis A Rippi 43
The main disadvantage of LRFD is the fact that there is a set of load and resistance factors that
need to be calibrated in order to cover the specific cases that are likely to occur This is because
of the lack of sufficient statistical data that are necessary for this calibration procedure
Consequently it cannot reassure that all the designs to which the set of factors is applied result
in the reliability level that was aimed for As a result in majority of the cases a conservative
calibration that gives the required reliability is chosen which can lead to ldquoover-designrdquo
structures
Figure 44 Design with partial factors for the load and resistance parameters (source
Schweckendiek 2006)
The evaluation of an elementrsquos reliability starts in principle with calculating the probability of
failure and subsequently the reliability for the given strength and load In practise the problem
is often that the strength is unknown but it has to be determined for a given reliability The
determination of the required reliability can be estimated with the help of Level II and III
methods by iteratively adjusting the strength in the calculation until a sufficiently small
probability of failure is found
In the design domain regulations and guidelines follow the standard that the characteristic
value14 of the strength is divided by a factor and that the characteristic value of the load is
multiplied by a factor as it is described in Eq 421
The link between Level I and the estimation of the failure probability has been achieved through
Level II methods The design point that results from a Level II method calculation is the point
with the greatest joint probability density of the strength and the load and it is therefore
possible that when failure occurs the strength and load values will be close to that point The
design values can be specified as
(422)
(423)
14 The characteristic value of a parameter is its main representative value and its value can be specified as a mean value an upper or lower value or a nominal value according to EN 1990 definition (2002)
Structural Reliability Analysis
MSc Thesis A Rippi 44
where are the coefficient of variation for the resistance and the solicitation respectively
15 and are the importance factor and β is the reliability factor These can be also expressed
in terms of characteristic values as follows
(424)
By substituting function 423 with 421 and 422 the partial safety factors are defined as
(425)
(426)
where 16 and are the values for load or resistance respectively to which a probability of
(non-)exceedance of 95 corresponds (non-exceedance for the load and exceedance for the
resistance) (for a standard normal distributed parameter this is 1645)
424 Response Surface Techniques (RS)
In case that the models presented above tend to be time consuming for the limit state function
evaluation or convergence problems of the optimization algorithms occur a better methodology
is recommended known as response surface The RS is mainly used when (unknown) response is
only available from experiments or complex FEM computations (for example large highly-non-
linear FEM models) Indeed once a RS has been built up the various standard methods may be
applied at a negligible cost An analytical limit state function replaces the real response function
The main idea is that the response consisting of a complex function of input variables is
approximated by a simple function of the input variables
A list of possible response surfaces techniques is given below
Linear and quadratic Taylor expansions
Polynomial RS based on least squares
Kriging method
Polynomial chaos expansion
A detailed explanation of the aforementioned methods can be found in OpenTURNS 15 (2015a)
The first method is associated with the approximation of the model response around a specific
set of input parameters while the rest seek a global approximation of the model over its whole
domain of definition The most sophisticated one is the polynomial chaos expansion but one the
15 α values are negative for solicitation (ie RV for which increasing their values lead to adversely effects on the reliability) and positive for resistance (ie RV for which increasing their values implies a beneficial effect on the reliability) 16 is negative and can be negative or positive
Structural Reliability Analysis
MSc Thesis A Rippi 45
other hand its complexity does not make it attractive In this research a polynomial RS is
applied based on least squares
In combination with FEA the standard procedure is as follows
1 Select the most important random variables on the basis of engineering basis
2 A SA is carried out in combination with FEM
3 Reduce the stochastic variables if needed according to the SA
4 A RS is constructed through the response data
5 A reliability calculation is carried out using the RS instead of the real response
A polynomial RS is generally constructed by fitting a quadratic function to the sampling points
whose general expression is
sum
sum
sum sum
(427)
The type of the reliability method to be used is of little importance since the time consuming
LSFE (using FEM) are replaced by analytical expressions However of main importance is now
the accuracy of the RS compared to the real response There are several parameters in order to
evaluate the goodness of fit of the RS to the real model among which is the well-known R-
squared
In Figure 45 an example of a quadratic response surface is showed that is fitted on the data
response indicated with the black circles This was a 6-dimensional problem that means 6
different variables were included in the system (non-visualized dimensions are kept constant at
their mean value) In this thesis the response of the system is to be acquired after the coupling
of FEM with a reliability method and that would give the ldquoblackrdquo circles that are shown in Figure
45 The response surface technique applies a curve fitting on the real response of the system
(ldquoblackrdquo circles) whose analytical formula can later be used for the reliability analysis of similar
type of systems with negligible computation time
Structural Reliability Analysis
MSc Thesis A Rippi 46
Figure 45 Quadratic response surface with 28 terms (ie ) for 6 dimensions ( ie x
=[ x1 hellip x6])
43 Coupling Reliability Analysis with FEM The limit state function evaluations (LSFE) will be carried out with the software Plaxis 2D 2015
which is a special two-dimensional finite element software used to perform deformations and
stability analysis for various types of geotechnical applications Moreover considering the case
study under investigation in this thesis Plaxis in contrast with other FEM such as Abacus
Comsol DIANA etc offers several techniques to realistically simulate structural elements such
as sheet pile walls and anchors and their interaction with soil while the variety of the
constitutive models for the soil body that are available and the ability to include the history of
the construction phases can lead to a better analysis of the systemrsquos behavior in terms of the
stress level and the deformations It is essential at that point to mention that using FEM for this
purpose means that the limit state formulation is implicit and can only be solved numerically
The reliability analysis is carried out through an uncertainty package In this section firstly a
description of the reliability package and its possibilities are given Finally an explanation of the
coupling procedure between the reliability tool and FEM is given together with the calculation
process that was followed for the parameters manipulation
431 The functionality and possibilities of OT
The reliability analysis tool to be utilized in this research is OpenTURNS (OT) which is a
scientific library usable as a Python module dedicated to the treatment of uncertainties and it is
still under development during the work of this thesis Several reliability packages are available
such as OpenEarth (applied in Matlab) Prob2b (applied in Probox software of TNO) and
Probabilistic Toolkit (software of Deltares) Plaxis 2015 features Python connection possibility
and thus coupling was decided to be carried out with an uncertainty package in Python and OT
is the most developed one Moreover until now only OT is an open source package and can be
Structural Reliability Analysis
MSc Thesis A Rippi 47
used easily from anyone as there is a wide community of experienced people supporting with
their knowledge while many related manual reports are already available for starting learning
A list of the current available reliability methods in OT is given in Figure 46 In this figure apart
from the standard methods the alternative method of the Adaptive Response surfaces is
introduced that is mainly used when (unknown) response is only available from complex FE
computations
Figure 46 Reliability methods available in OpenTURNS
In this thesis the methods to be basically applied are the FORM and the DS analysis while MC is
also used mainly for confirming the application of FORM Additionally the RS technique has
been mainly implemented for enhancing the performance of FORM and DS analysis
OT handles 47 types of distributions amongst which the ones used in this research are Normal
Lognormal Uniform and Truncated Normal distribution Regarding the joint distributions that
are available to be selected for random vectors 11 types of copulas are existing in OT amongst
which the most known ones are the Independent the Gumbel and the Normal copula In this
thesis the independent copula has been used during the reliability analysis which means that
the variables were assumed to be independent among each other However it is strongly
advised a further elaboration and research considering correlation matrices for specific soil
parameters
432 Coupling OpenTURNS-Plaxis
The finite element simulation will be carried out in Plaxis 2015 Plaxis is a finite element
software for plane strain and axi-symmetric modelling of soil and rock behaviour Moreover it
supports a fully automatic mesh generation allowing for a virtually infinite number of 6-node
and 15-node elements
Structural Reliability Analysis
MSc Thesis A Rippi 48
The coupling of reliability analysis and FEA requires an interface for the communication
between each other When a reliability tool is coupled with another software program the
reliability program carries out the whole reliability analysis and it uses the other program only
for the evaluation of the limit state function More precisely OT should be able to read and
amend Plaxis output values for important variables such as material parameters pore pressures
generation and stresses development and corresponding deformations inside the dike
Respectively Plaxis has to be also capable of obtaining the new values that has been set by OT
for the variables that are treated as stochastic during an iterative process according to the
reliability assessment In Figure 47 an illustration of the coupling methodology and its function
is shown
Figure 47 Coupling scheme OpenTURNS-Plaxis
In principle an input file is firstly required where the user set the preferable reliability method
to be used the stochastic input parameters and their probability distributions the joint
probability distribution and the corresponding correlation matrix and finally the limit state
function is formed depending on the situation In Appendix D an example of the input files that
were used for soil sheet pile wall and anchor analysis is shown However such input files
should be interpreted so as to be readable by both Plaxis and OT Therefore an input interpreter
was created which is actually a python script that helps OT to start up the reliability analysis
according to the assigned method variables distributions and LSF As it was mentioned before
the evaluation of the limit state function is conducted by Plaxis For that purpose the input
interpreter should be also able to send the next set of input parameters to Plaxis However an
additional means of connecting the interpreter with Plaxis is also needed This is can be
Structural Reliability Analysis
MSc Thesis A Rippi 49
achieved via a Plaxis interface which actually call Plaxis to make the calculations while it also
transfers the required value of the limit state function to input interpreter and this in turn to OT
Eventually the probability of failure is obtained as a model result However it is also essential
that Plaxis simulation procedure converges to the desirable criteria and under the physical
boundary conditions that have been determined Likewise the convergence criteria of the
reliability methods shall be manipulated so as the optimization algorithms to able to converge
efficiently (see Appendix A2 for further explanation)
44 Overview In section 42 a summary of the principles of the main reliability methods was made From
Level III methods MC and DS were presented whereas from Level II the basics of a FORM
analysis were introduced Besides the concept of the Level I method was explained Even
though this method was not used in this research for the evaluation of the probability of failure
it was applied in section 643 for the deterministic analysis of the system under investigation
and the preliminary calculations of the dimensions In Appendix F more reliability methods are
introduced and they are further described In the sequence their implementation in conjunction
with Plaxis was discussed through the coupling of FEM with OT For more information
concerning OT special features (ie SA optimization algorithms and probability distributions) a
reference is made to Appendices A and E
Before any of these reliability methods was applied on the case study of this thesis their
performance and their compatibility with Plaxis were tested with simple examples At first the
probability of failure was evaluated with FORM SORM Importance Sampling (IS) crude MC and
DS for simple artificial linear and non-linear limit state functions
and
respectively where R B and F are the random variables distributed normally
or lognormally The crude MC technique has been applied in order to validate the failure
probability in case the various methods give different results Secondly a flexible circular
footing over a one-layer soil material was used from Plaxis tutorial in order to test the coupling
between Plaxis and the reliability method and reassure the functionality of the interface In this
pilot example the objective was to determine the probability that the settlements of the footing
exceed a certain threshold of settlements and thus The respective
components that were taken into account for the failure mechanism was the cohesion (c) the
friction angle (φ) and the specific weight (γ) of the soil
From a qualitative evaluation of the reliability methods that has been done and according to the
experience gained from the aforementioned applications of some methods valuable conclusions
were drawn in terms of their efficiency More precisely regarding the sampling methods it does
not required previous knowledge of the failure domain apart from the IS method in which a
starting point inside the failure domain shall be provided Moreover the calculation effort
depends on the magnitude of the failure probability and the required accuracy whereas for the
IS it depends also on the selection of the sampling region In Waarts (2000) it has been proved
that DS tends to be more efficient than MC for low dimensional problems as it is indicated in
Figure 48 (number of random variables nlt100)
Structural Reliability Analysis
MSc Thesis A Rippi 50
Figure 48 Required number of samples for MC and DS as a function of the random variables
(source Waarts 2000)
As far as the approximation reliability methods (FORM and SORM) are concerned it is not
necessary to be aware of the failure region in advance Furthermore it has been observed that
the required iterations and the calculation time is quite lower in comparison with the sampling
methods However the accuracy of the method is highly dependent on the shape of the LSF
In Table 41 a summary of the most applied reliability methods evaluation is presented
According to this evaluation it was decided that the methods of FORM and DS as well as the
combination of them with RS techniques are going to be tested and evaluated in terms of their
efficiency and robustness These methods were chosen from both Level II and Level III methods
due to their expected reduced computational time the non-requirement of previous knowledge
about the LSF and their satisfactory performance in similar case studies In Chapter 5 the
different failure mechanisms and the relevant LSF that are going to be utilized in this thesis are
presented whereas in Chapter 7 the results deduced from the reliability analysis are elaborated
In Chapter 7 there is also an indication of which reliability method was used for each systemrsquos
component the selection of which was based on the aforementioned evaluation
Table 41Evaluation of reliability methods
Method Previous knowledge Accuracy Calculation Effort
FORM not required full accuracy for
Gaussian variables and linear LSF
depends on LSF linearity and the number of random
variables
SORM not required exact up to 2nd order
LSF error dependent on the shape of LSF
depends on LSF linearity and the number of random
variables
Structural Reliability Analysis
MSc Thesis A Rippi 51
Crude MC not required can be controlled by convergence criteria
depends on the magnitude of failure probability and the
required accuracy
IS required can be controlled by convergence criteria
depends on the choices made for the sampling region
DS not required can be controlled by convergence criteria
for low dimensional problems (nlt100) DS is more efficient
than MC
Especially RS were created based on a SA sampling in order basically to increase the efficiency
of FROM and DS analysis in case of FEM numerical errors Precisely an analytical formula or a
response surface was needed that can approximately represent the performance of the structure
under investigation and that is called in case of Plaxis errors This formula was essential in
order to keep the reliability analysis running by providing a response value for the LSF when
Plaxis calculation is unsuccessful and thus incapable of returning a result It should be
mentioned that in case a RS is implemented for the reliability analysis Figure 47 can be
reduced to a graph consisting of only the input file OT and the input interpreter as no Plaxis
calculations would be necessary any more The LSFE will then be performed through the
response surface that is provided by the user Due to this limited use of the RS method it will
not be further treated in this thesis
Structural Reliability Analysis
MSc Thesis A Rippi 52
MSc Thesis A Rippi 53
5 Failure Mechanisms and Limit State Functions
The reliability analysis of a structure requires the definition of the different failure modes that
are relevant to the corresponding structural elements In this chapter the possible failure
mechanisms of the systemrsquos elements expressed in LSF are identified and their role in the
system reliability is explained Based on these LSF the reliability of the system components was
evaluated These LSFs focus on the ULS whose exceedance implies failure of the corresponding
component
51 Introduction to the system analysis and the limit states For system reliability analysis all the possible failure mechanisms have to be identified and
summarized in a fault tree In Figures 51 and 52 the fault trees of a dike section without
structural elements and a retaining structure with sheet piles are depicted respectively In this
project the combination of these two separate structures ie a reinforced dike section with an
anchored sheet pile wall is to be simulated and studied in terms of its failure modes
As it is described later in this chapter failure can be expressed in different ways depending on
the structure and stakeholders demands and safety standards This research is mainly focused
on the ULS of failure and Mohr Coulomb was chosen as the soil failure criterion due to its
efficiency on detecting failure Such a failure criterion is actually defining the stress strain
relationship and the gradual weakening of the soil (ie the gradual reduction of the strength
parameters such as the friction angle and the cohesion due to the stresses development until
soil collapses) under a stress condition In Plaxis new version there is the ability of gradually
reducing structural properties such the strength of steel elements However the results
retrieved of such a safety analysis have not been yet investigated and for that reason such
calculations are not included in this project
As far as the dike section is concerned the failure mechanism to be investigated in this project is
the macro instability of the slope rather than the other mechanisms depicted in Figure 51 On
the other hand because of Plaxis limitations to distinguish among the different structural
elements failure mechanisms each element was considered as a singular case for investigating
its failure mode In the next section a distinction is also made between the ULS and the SLS
For a reinforced dike with retaining walls there are basically four classes of structural elements
(see Figures 12 and 21)
∙ Retaining wall (ie sheet piles)
∙ Anchors
∙ Walings
∙ Soil structure (ie dike section)
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 54
For the reliability analysis of each of the elements the probability of failure is to be determined
In the next sections of this chapter the respective LSF that are considered for each class are
further elaborated according to the ULS criterion while in the end the general combined fault
tree is to be schematized
Figure 51 Fault tree of a dike section
Figure 52 Fault tree for retaining structures using anchored sheet piles (source Schweckendiek 2006)
The system failure in this research is considered as a serial system of the anchor sheet pile
wall and soil body LSFs The fault tree of the system failure is presented in Figure 53 together
with the LSF of each component as they are formulated according to sections 522 and 523
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 55
Figure 53 Fault tree of the system failure
It should be mentioned at that point that the probability of failure of the system differs from the
overall probability of flooding which takes into account any potential failure mechanism of the
system combined with the uncertainties of the water elevation More precisely as far as the dike
safety is concerned a probability of failure (where ) under a certain
water level is estimated from the fault tree of Figure 51 which in sequence is multiplied
with the occurrence probability of the corresponding water level in order to estimate the
overall probability of failure for the specific water elevation Then that product is accumulated
over a required range of water levels that can jeopardise the overall stability of the structure in
order to calculate the overall probability of failure or else the probability of flooding This
probability is then compared to the one established from the safety standards so as to reassure
the safety or not of the structure can be estimated as follows
int
sum (51)
In this thesis the probability is to be estimated where is a specific water level
and especially the design water level as it is considered to be the most challenging part of the
procedure described above gives the probability of occurrence of the water level (ie it is
the PDF of ) After setting up the steps and implementing them successfully for the estimation
of the calculation of the overall probability of failure is just a repetition of the same
procedure for more water levels It should be mentioned that in this thesis failure consists only
of the macro-instability failure mode whereas the rest are excluded for the time being A
simplified alternative for the overcoming the integral of Eq51 is to repeat the calculations for
certain water levels and then sum their products instead of integrating the full range of the
them
52 Limit State Functions
521 Serviceability Limit State
The SLS is evaluated in design calculations in order to avoid excessive deformations of a
structure that could lead to the loss of its service requirements and its functionality In some
cases the SLS criteria are used to prevent damage to adjacent structures in the surroundings
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 56
For example concerning the design of a dike section attention should be paid to deformations
caused to buildings that are located in the inland part of the dike in case of an extreme event
In Figure 54 an example of a deformed dike is depicted In this figure some of the potential
locations that excessive deformations might be experienced are illustrated such as
∙ the vertical settlement of top of dike (arrow 1)
∙ the horizontal displacement of the top of the sheet pile (it might also be the middle part
of the sheet pile that ends up with the largest deformation according to the loading
conditions and the specific soil structure different deformations might occur) (arrow 2)
∙ Uplift upstream of the dike (arrow 3) special attention shall be paid to this type of
deformation not only due to the direct effect on inland structures but also as a sign of
developing piping mechanism
(a)
(b)
Figure 54 Figure (a) shows the initial undeformed structure while (b) illustrates the deformed mesh (not to scale) (the plotted length of the anchor does not have any physical meaning) In (b) the possible locations and directions (black arrows) of potentially excessive deformations
According to the current Dutch regulations (Larsen et al 2013) the SLS criteria for designing
reinforced dikes with sheet pile walls require that
∙ the settlement on the top of dike does not exceed 10 cm and
∙ the horizontal displacements of the sheet pile wall does not overcome 10 cm
In this research the SLS criteria and especially the dike settlements were used roughly for the
preliminary determination of the structures characteristics that are discussed in section 643
Dike Anchor
Sheet pile wall
1
2 3
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 57
More precisely displacements at the top of the dike were also taken into account as
complementary to the safety factor in order to determine the required length of the sheet pile
wall As far as the reliability analysis in concerned the LSF were formulated based on the ULS of
the different systemrsquos components that are presented in the next section
522 ULS for Structural Members
In the design process one is most interested in the ULS of a failure mechanism This state
describes the situation wherein the acting extreme loads are just balanced by the strength of the
construction If that limit state is exceeded the construction will lose its functionality and thus
collapse or fail In general most attention is paid to the behaviour of the structure after
completion However during construction there are also periods in which the construction may
fail The different phases of construction are listed in section 642 In the present section the
analytical LSF of the structural elements are given as they are going to be used in the reliability
analysis
Sheet pile wall
The most relevant failure mode for the sheet pile wall is the exceedance of the yield strength
which corresponds to the ultimate steel strength The response of the structure is mainly due to
bending moments and the axial forces (shear forces are considered to be negligible) Where an
axial force is present allowance should be made for its effect on the moment resistance
Figure 55 The water and earth pressures as well as the anchor forces lead to bending moments and axial forces on the sheet pile wall
In Figure 55 an example of the axial forces and the bending moments that can be developed in
the sheet pile wall with one anchor layer are illustrated after an extreme water level loading
Accordingly the maximum stresses on the sheet pile wall are composed of a bending moment
and a normal force component17
[
] (52)
where [kNm] and [kN] are the bending moment and the axial normal force
respectively that depend on the depth level where they are calculated over the sheet pile length
[m3] is the elastic section modulus and [m2] the cross-sectional area of the sheet pile
wall
17 the vertical anchor force component is reducing by its interaction with the soil over depth
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 58
Bending moment and axial force can be variable over the depth and that is why they are
expressed as a function of z-depth FEM has the advantage to take into account second order
effects ie a stiffer structure will experience higher bending moments than a more flexible one
Taking into account the above the limit state function can be formed as the difference between
the maximum developed stress and the yield stress
[
] (53)
where and can be characterized as the load variables while and can be considered
as the resistance variables and are assumed to be constant over depth
Concerning the permissible displacements for the ULS an upper limit for the maximum
horizontal displacements of the top of the sheet pile has been set to 150L where L is the
vertical length of the sheet pile wall (Larsen et al 2013) This requirement was mainly used in
section 643 where the structural properties were determined after the deterministic analysis
Anchors
Anchors are loaded by their reaction to the horizontal loads on the retaining walls The failure of
the anchor element is actually represented by the failure of the steel members of the anchor
(tubes bars cables etc) that are loaded by traction forces
As it is illustrated in Figure 56 the anchor experiences an axial tensile N that is nearly constant
over its length
Figure 56 Axial loading of anchor inside a dike
The elastic behaviour of an anchor involves only a relationship between axial force N and
displacement (elongation) u of the form
[ ] (54)
where EA [kN] is the anchor stiffness consisting of the steel Youngrsquos modulus E [kNm2] and the
anchor cross section A [m2] and L [m] in the length of the anchor
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 59
Similarly to the sheet pile wall the limit state function of the anchor involves the certain yield or
ultimate strength of the steel members and the maximum stress that the anchor experiences
during its loading Consequently the LSF is as following
(55)
where [kN] is the calculated anchor force and [m2] is the cross sectional area of the
anchor (both of them considered to be constant over the depth) It is essential to mention that
the anchor is also subjected to bending moments due to soil settlements (that are implicitly
illustrated via the uniformly distributed load q over the tie rod) that should be taken into
account in order to investigate the displacements of the tie rod itself However in this thesis
only the axial forces on the anchor are considered without taking into account the individual
deformations and its reaction with the surrounding soil
Walings
The waling is the element that transfers the loads from the retaining wall to the anchors (see
Figure 21) The loading of the walings can be schematized as a continuous beam on several
supports as it is depicted in Figure 57
Figure 57 Loading of walings
If Nα [kNm] is the calculated anchor force per m then the developed bending in the support Ms
and in the opening Mo can be approximated as follows
[ ] (56)
(57)
where La is the mutual anchor distance Considering as the design moment of the waling to be
the limit state function can be formulated as follows
Waling Sheet pile Anchor
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 60
(58)
For a conservative design of the waling the limit state function will give the same or lower
failure probability than the anchor itself Therefore it will actually not be necessary to carry out
this analysis The anchor failure will be the determinant mechanism (Schweckendiek 2006)
523 ULS for Soil Failure
In this thesis emphasis is given on the dike global instability which actually consists of several
failure modes Figure 58 indicates the possible failure mechanisms of flood defences The
combination if these can lead to the overall instability of the dike and thus to soil body failure
However Plaxis is not capable of accounting for all of them More precisely Plaxis assumes the
soil to be a continuous body and thus it can model movements in the scale of soil bodies (using a
relative fine mesh in relation to the moving soil body) but not on a soil particle level Plaxis can
simulate the groundwater flow in a soil body and from this someone can deduce input for a
piping analysis (see failure mechanism G in Figure 58) (using other softwaremethods) but as
it has been mentioned above it cannot determine movement of soil particles due to
groundwater flow Moreover Plaxis cannot deal with the flow and waves occurring in ldquoopen
waterrdquo ie water outside the soil in a canal a lake or sea for instance
Figure 58 Possible failure mechanisms of flood defences (source Flood Defences 2015)
Therefore Plaxis can handle the failure modes (C) (D) (E) (J) (K) and (L) of those illustrated in
Figure 57 In this thesis loads due to extreme temperature conditions or ships are not
considered while the settlements are taken into account implicitly though the mechanisms (C)
(D) and (E) These mechanisms are thus investigated in this thesis and they mainly refer to the
macro-instability of the inner and outer slope of the dike or the overall dike body (ie horizontal
sliding (mechanism D))
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 61
For these failure modes soil failure is defined as shear failure In Figure 59 the most relevant
patterns of the macro-instability regarding the shear strength of the soil are illustrated
Applying FEA the most critical failure mode is determined automatically However it is not
always straightforward what the trigger mechanism of the failure was and that is why more
investigation and FEA tests are needed in order to obtain a clear view
(a) Outer slope failure (b) ldquoActive siderdquo failure
(c) ldquoPassive siderdquo failure (d) Overall failure
(e) ldquoKranzrdquo stability
Figure 59 Macro-instability failure modes in a dike with an anchored sheet pile wall
In Figure 59 failure modes (a)-(d) follows Mohr-Coulomb failure criterion In principle the
total stress state inside a dike section consists of the effective stresses and the pore pressures
according to Terzaghi principle
[
] (59)
Mohr-Coulomb failure criterion basically considers as limit for the failure to occur the
maximum shear stress that the soil can withstand This is determined from the friction
angle the cohesion of the soil and the current stress state The drained shear strength
(similarly in for the undrained shear strength) in the soil according to Mohr-Coulomb is defined
as follows
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 62
[
] (510)
where is the effective friction angle In Appendix B1 a detailed explanation of this failure
criterion is attached
Taking into account the above mentioned failure mechanism (a) is mainly triggered by the
increase of the pore pressures in the outer slope which subsequently causes a decrease in the
effective stresses ( ) which at the same time leads to a reduction of the shear strength
(considering Mohr-Coulomb failure criterion) When the shear stresses in the soil exceed the
shear strength a slip plane forms and a soil wedge collapses
Failure modes (b) and (c) are primarily determined from the active and passive effective
stresses respectively As far as the ldquoactive siderdquo failure is concerned displacements of the sheet
pile wall towards the downstream side of the dike lead to development of active stresses
where is the active pressure coefficient During the ldquoactiverdquo failure the
retaining soil is allowed to relax which leads to a decrease of horizontal stresses and
simultaneously to a decrease in shear strength Similarly to failure mode (a) a soil wedge
collapses The result is an increased earth pressure on the wall for which it is not designed
However this type of failure in dikes is not such determinant for the soil and the sheet pile wall
failure because there is always the resistance of the passive side of the dike that keeps the active
side stable enough Such a mechanism is thus more relevant for deep excavations with retaining
walls
In mechanism (c) the passive soil resistance is exceeded by the horizontal loads In this case the
wall moves inland and a wedge of soil compresses The shear strength is larger due to the
deformation of the wall and the horizontal stresses increase stresses where is
the passive pressure coefficient ( ) This failure mechanism usually occurs due to an
underestimation of the sheet pile length or due to the presence of a weak soil layer in the
passive side
Failure mode (d) is actually a combination of (b) and (c) possibly accompanied also with the
development of excess pore pressures as it was discussed in failure mechanism (a)
Moreover too short anchors could cause a soil block to collapse as it is showed in Figure 59(e)
(ldquoKranzrdquo stability)
As it has been already mentioned for mechanism (d) failure modes can be correlated and
combined with each other in order to lead to a final failure state Therefore the common failure
probability would be smaller than the sum of the singular probabilities In this thesis the total
probability of failure of the soil body due to global instability is to be determined This is due to
the fact that the application of FEM is capable of simulating the combination of the failure
modes depicted in Figure 58 and thus the overall failure (related to (C) (D) and (E) modes of
Figure 58) can be evaluated However with the classical engineering approach for the stresses
calculation each one of the failure modes depicted in Figure 59 should be separately evaluated
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 63
Below a description of the available methods to formulate the LSF of the soil failure is given
After the evaluation of these alternatives a selection was made regarding the most suitable LSF
for the current case study
Limit State functions for soil failure
Soil mechanical failure can be analysed in several ways (Schweckendiek 2006)
1 Excessive Deformations
2 φ-c Reduction
3 Relative Shear Resistance
4 Plaxis definition of soil collapse
The possibilities and limitations of the aforementioned methods are briefly discussed below
1 Excessive Deformations
Similarly to the SLS a limit state function can be formulated by deformations that are
unacceptable As it is illustrated in Figure 54(b) there are some points (the top of the dike the
sheet pile top and the inland soil level behind the dike) the displacements of which shall be
limited to the minimum possible during the structurersquos lifetime in order for the structure to
meet the required service standards Therefore the calculated deformations at locations 1 2
and 3 could give a clear indication of failure given the maximum acceptable deformation
In case that there is a limit value for each location then the limit state function can be formed as
follows
[ ] (511)
with this criterion designer can control the developed deformation on the system and improve
the structure so as to limit the displacements to the required level and subsequently increase
the reliability
On the other hand such an approach might exhibit several problems during the reliability
analysis that are listed below
Many failure mechanisms suddenly happen and as a result no significant displacements
are observed before moving very close to failure In sequence this may cause problems
for the iterative procedures of some reliability methods such as FORM and DS
The determination of suitable maximum admissible deformations is not
straightforward They have to be large enough to serve as failure criterion and
simultaneously they shall not be larger than the values that can be calculated within
the limits of equilibrium in FEM-calculations This requires previous knowledge on the
analysed system as well as on the feasibility of FEM-calculations on the specific subject
Last but not least the location of the undesirable displacements in a dike section is also
under investigation as there can be multiple vulnerable spots on the dike that should
be kept under a certain range of deformations in order not to cause a collapse and that
they are not always known in advance
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 64
2 φ-c Reduction
An available type of calculation in Plaxis is the so-called ldquoSafety Calculationrdquo with which global
safety factors can be computed In Appendix B2 a more detailed elaboration of this method is
exhibited This feature can also be used in reliability analysis by formulating the LSF as follows
(512)
where is the Multiplier Safety Factor obtained by the φ-c reduction The main concept is
that if the safety factor is smaller than 1 then it is considered as failure This method can provide
us with the probability of failure for a general soil body collapse However even if this method
seems to be quite simple there are some issues during its implementation in a reliability
analysis
For complex limit state functions there can be convergence problems for some
reliability methods with iterative procedures (ie FORM) This can be explained from the
fact that in a φ-c reduction soil strength properties follows a certain path as it is
described in Appendix B2 independently of the values for the random variables that
have been set from the reliability method
The safety factor is a general safety factor regarding the failure of the system
Therefore sometimes it is not straightforward what the ldquotriggerrdquo factors are that led to a
certain type of failure
A safety calculation in Plaxis is time-consuming and the outcome can be unstable (Plaxis
calculationsrsquo convergence depends on the number of calculation steps)
Plaxis cannot handle safety factors below 1 Therefore a new limit of the safety factor
shall be introduced in this limit state function
3 Relative Shear Resistance
In this method the basic idea is to define soil failure according to the failure criterion that Plaxis
is set to use for the analysis In this research the Mohr Coulomb failure criterion is applied in
combination with a linear ndashelastic perfectly-plastic stress-strain relationship (that is the so-
called Mohr-Coulomb model) For this model the initiation of plasticity is considered as failure
Therefore the maximum shear resistance is defined just before plastic yielding occur for any
given stress state Of course the occurrence of plasticity does not directly indicate the failure of
the soil structure however this model can give a first estimate for the stress state and the
deformations
According to this method the relative shear resistance is defined as the ratio between the
mobilized shear stress and the maximum shear resistance In Figure 510 the Mohr-Coulomb
model is illustrated in a Cartesian stress-plane In the graph the principle stresses σ1 and σ3 are
indicated More precisely the principle stresses are defined as follows
radic
(513)
(514)
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 65
radic
(515)
Figure 510 Mohrs stress circle and Mohr-Coulomb failure criterion
Hence the mobilized shear stress is
radic
(516)
And the maximum shear resistance which is the distance of the mean stress point to the yield
surface is defined as
(517)
Then the relative shear resistance is a measure for load-resistance ratio that can be determined
in any integration point in the soil continuum
(518)
However this criterion requires prior knowledge of the possible relevant failure mechanisms
and their potential locations on the soil structure This is necessary in this method in order to
choose a suitable cluster of integration points where the average value of the relative shear
strength is to be determined Therefore the limit state function to be considered is not
straightforward and it needs problem investigation in advance
4 Plaxis definition of soil collapse
In Plaxis the construction stages are analysed by performing a Load advancement ultimate level
procedure which is controlled by a total multiplier
(519)
where is the load that is applied by Plaxis is the load at the beginning of the
calculation phase (ie the load that has been reached at the end of the previous calculation
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 66
phase) and is the defined load configuration (ie water level soil weight vertical load
on the top of the dike etc) Thus the main idea is that Plaxis applies an incremental load until it
reaches the defined one When has reached an ultimate level (which by default is 1)
the current phase is finished However if a staged construction calculation has not properly
finished the multiplier fails to converge to the desired ultimate level at the end of the
phase and a warning of ldquoSoil collapsesrdquo appears in the Log info box
In other words a collapse load has been reached In this case the total specified load has not
been applied In physical terms this means that the current value of the stiffness parameter CSP
is less than 0015 (Plaxis 2D 2015b) It is a measure for the amount of plasticity that occurs
during the calculation When the solution is fully elastic CSP is equal to unity whereas at failure
it approaches zero Therefore soil reaches its upper limit of plasticity and it collapses which can
be visualized as a settlement of the dike body Such a warning in the analysis is thus assumed to
be a possible failure situation
53 Overview In this chapter an overview of the relevant possible LSF of the structural elements and the dike
body respectively has been given The different LSFs are formulated and discussed whereas the
failure of each component is going to be evaluated in the ULS of each component
The system failure in this research is considered as the combination of the anchor sheet pile
wall and soil body LSF Particularly for the anchor and the sheet pile failure their yield stress
was chosen as a limitation for their failure by adopting the next LSFs
[(
)] sheet pile wall LSF (520)
anchor LSF (521)
As far as the soil body is concerned in this research method 4 is to be used in order to identify
soil failure The limit state function is then formed as a single value in case of a successful and an
unsuccessful computation Therefore the limit state function was chosen to be equal to 1
(actually the value of is retrieved from Plaxis that in case of a successful calculation is
1) in case of a successful Plaxis analysis and -1 in case that soil collapses in any phase (error in
Plaxis Log file ldquoSoil Collapsesrdquo) This formula can be written as follows
(522)
This accounts for the global instability of the dike body in contrast with the other methods that
are related to a certain dike area (methods 1 and 3) or they are cannot handle overall failure
due to numerical issues (method 2) More precisely methods 1 and 3 were excluded due to the
prior knowledge that is required in method 3 while in method 1 a definition of the maximum
admissible deformations shall be first introduced However attention shall be paid to the type
of the warning that Plaxis gives because it might also be that the maximum number of load steps
was not sufficient In that case the phase must be recalculated using a larger number of steps
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 67
Someone could argue that such a failure can be considered as a system failure rather than a soil
failure alone Partially this is true as the incapability of the sheet pile wall or the anchor to keep
the upstream part of the dike stable can lead to non-convergence and thus according to Plaxis
definition to failure This incapability can be translated into various scenarios of failures that
stem from either the structural elements or the soil body weakness and that are illustrated in
Figure 511 This picture shows the different scenarios that can take place due to macro
instability that was discussed in 523 and the failure of the retaining wallrsquos elements
Figure 511 Combined failure scenarios (initial and developed mechanisms) (source Larsen et al 2013)
However this does not mean that the sheet pile wall or the anchor fail because of exceeding
their yield stress but rather due to possible unacceptable deformations This is also an
advantage of the inherent residual strength of the structural elements that allows a large
deformation before they fail However the residual strength of the elements was not taken into
account in this thesis as they were considered as elastic Moreover the message for the ldquosoil
collapsesrdquo can appear in a phase prior to the installation of the wall which means that this type
of failure is exclusively attributed to the deficiency of the soil strength Therefore it is not
always straightforward what is failing first and lead to the system failure Consequently by
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 68
considering such an analysis representative of a system failure it is likely to underestimate the
overall probability of failure whereas if it is considered as a soil failure only it might
overestimate the total probability of failure since it is later combined with the rest LSF of the
anchor and the sheet pile wall In this research it will be considered only for the soil failure that
brings the results to the safe side even if it does not lead to the best optimization In this thesis a
serial system of the anchor the sheet pile and the soil body LSF was considered the fault tree of
which is depicted in Figure 53
Another way to evaluate the reliability of the different elements could be the reach of a
threshold displacement according to the SLS criteria Such a threshold could be also considered
for the system as a whole rather than the individual elements However in that case someone
should be cautious regarding the choice of the value for the maximum deformations as well as
the location of their occurrence
Concluding in this section the possible failure mechanisms of a flood defence (Figure 58) were
shown Then it was clarified which of these mechanisms can be modelled in Plaxis and they
were further explained and schematized in Figure 58 with the presence of the retaining wall as
they would have to be handled in case of hand calculations Plaxis can incorporate these
mechanisms automatically and thus no distinguish among the different patterns of these
specific modes is necessary Last but not least in Figure 511 the possible failure modes for the
system of the dike and the retaining wall that can be simulated in Plaxis are depicted In this
thesis the plasticity of the structural elements is not considered and thus possible plastic hinges
in the anchor and the sheet pile are not taken into account Therefore scenario 3 is excluded
In Chapter 7 the results after the reliability analysis of the aforementioned LSFs are given It
was essential though before proceeding with the reliability analysis to get a better
understanding into the system under investigation For this purpose in the next chapter a
description of the case study as it has been simulated in Plaxis is given while also the different
structural properties are defined for the following reliability analysis The soil variables to be
considered as stochastic are presented and a first insight into of the systemrsquos behaviour is
obtained through the mean values calculation
MSc Thesis A Rippi 69
6 Case Study-Dike with an anchored sheet pile wall
In this chapter a description of the case study is presented together with the relevant soil and
structural parameters In sequence a deterministic analysis follows in order to specify the basic
structural characteristics and obtain a first sense of the stresses magnitude developed on the
structure This was carried out based on the recommendations given in the CUR 166 (2005)
The aim of this deterministic analysis is to redefine the structural properties in order to avoid
having an overdesigned structure understand the current design procedure and see if there are
any possibilities of improvement by applying the proposed reliability analysis
61 Case Description The case study concerns a dike section with an anchored sheet pile wall which was initially used
to showcase the current methodology of designing dikes with sheet piles using partial factors
For more information a reference is made to Breedeveld (2011)
In this research this case study has been modified and simplified for the needs of the thesis and
it is illustrated in Figure 61 while in Table 61 the soil materials are identified Details for the
properties of the soil layers are given in section 62 We assume the structure to be
homogeneous to the third dimension and therefore a plane-strain model is applied The dike
soil materials (layers 3 and 4 in Figure 61) were adopted from the case study used in
Breedeveld (2011) that was presented in Chapter 2 Especially in Figure 25 the original case
study as it was modeled in Plaxis is shown
Table 61 Soil materials that corresponds to Figure 61
1 Aquifer Sand layer 2 Aquitard Clay layer 3 Dike new material Clay layer 4 Dike old material Clay layer
The geometrical properties the groundwater level and the design water level were taken
identical to the original case study as well as the soil properties of the dike materials The soil
layers under the dike were modified for the need of the thesis and were modeled using random
average quantities for which the statistics were chosen arbitrarily but in a realistic range
according to the NEN6740- Table 1 (see Appendix C) included in the Dutch code for
geotechnical structures Moreover the vertical load was removed in order to examine the
impact of the water elevation on the dike stability
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 70
Figure 61 Case study Dike with anchored sheet pile wall Dimensions and Soil materials
62 Soil Parameters The soil parameters for the dike material were obtained from the report of Breedeveld (2011)
where for some parameters the characteristic values were available whereas for some others
the mean values were given For the soil layers below the dike Table 1 of NEN6740 was used as
it was mentioned before which gives characteristic values For those parameters whose value
was a characteristic 18 one a transformation was made according to their probability
distributions in order to acquire the mean values that are required for the reliability analysis
Only the properties required for the use of the Mohr-Coulomb model are presented
The parameters of each soil layer are presented in Tables 63(a) and (b) and the front number of
each layer indicates the corresponding soil layer as showed in Figure 61
The distribution types and the coefficients of variation for each parameter were chosen
according to the knowledge that has been obtained until now about the physically possible
ranges of such parameters and the recommendations that have been given in several researches
up to now (Baecheramp Christian 2003 Schweckendiek 2006 GriffithsampFenton 2007 Phoon
2008 Wolters 2012 Bach 2014) The distributions and the CoV of all parameters are displayed
in Table 62 They were considered to be the same for all soil layers Particularly for the
saturated and unsaturated volumetric weight γsat and γunsat respectively a relationship was
established in order to derive the one from the other In general γunsat varies between the real
dry weight and γsat depending each time on the degree of saturation An estimated maximum
difference between γunsat and γsat is approximately 4 kNm3 (Brinkgreve 2015) In this thesis γsat
was expressed as the summation of γunsat and a variable with uniform distribution in the range
of [0 2]19 Initially a deterministic relationship was defined between the two variables but later
on it was proved the variable was important for the soil failure and thus it was decided to
consider it as random
18 for the definition of the characteristic value see Appendix E 19 The range of [0 4] was not chosen in order to avoid extreme conditions
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 71
Table 62 Soil parameter distributions
Soil parameter Symbol DistributionRelation COV Unit
Unsaturated
Volumetric weight γunsat Normal (microσ) 5 [kNm3]
Saturated Volumetric
weight γsat γsat= γunsat+U(02) 5 [kNm3]
Cohesion c Lognormal (microσ0) 20 [kPa]
Friction angle φ Truncated normal (microσ045) 10 [ ˚]
Youngrsquos Modulus E Lognormal (microσ0) 25 [kPa]
Poissonrsquos ratio ν Truncated normal (microσ005) 10 [-]
Interface strength Rinter Truncated normal (microσ0099) 20 [-]
Table 63 (a) Soil statistical properties under the dike and (b) soil statistical properties in the dike
(a) 1 SAND (very silty)
Soil property Symbol Characteristic
value
Mean
value STD Unit
Saturated
Volumetric
weight
γsat 20 22 11 [kNm3]
Unsaturated
Volumetric
weight
γunsat 19 21 103 [kNm3]
Cohesion c 0 0 0 [kPa]
Friction angle φ 30 36 36 [ ˚]
Dilatancy angle ψ 0 0 0 [ ˚]
Youngrsquos Modulus E 20000 30769 7692 [kPa]
Poissonrsquos ratio ν 025 03 003 [-]
Interface strength Rinter 044 066 013 [-]
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 72
2 CLAY (clean medium)
Soil property Symbol Characteristic
value
Mean
value STD Unit
Saturated
Volumetric
weight
γsat 17 185 093 [kNm3]
Unsaturated
Volumetric
weight
γunsat 17 185 093 [kNm3]
Cohesion c 10 14 282 [kPa]
Friction angle φ 175 21 21 [ ˚]
Dilatancy angle ψ 0 0 0 [ ˚]
Youngrsquos Modulus E 2000 3077 769 [kPa]
Poissonrsquos ratio ν 029 035 0035 [-]
Interface strength Rinter 034 05 01 [-]
(b) 3 DIKE NEW (very sandy clay)
Soil property Symbol Characteristic
value
Mean
value STD Unit
Saturated
Volumetric
weight
γsat 17 185 093 [kNm3]
Unsaturated
Volumetric
weight
γunsat 17 185 093 [kNm3]
Cohesion c 4 564 113 [kPa]
Friction angle φ 29 347 347 [ ˚]
Dilatancy angle ψ 0 0 0 [ ˚]
Youngrsquos Modulus E 1625 2500 625 [kPa]
Poissonrsquos ratio ν 029 035 0035 [-]
Interface strength Rinter 034 05 01 [-]
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 73
4 DIKE OLD (little sandy medium clay)
Soil property Symbol Characteristic
value
Mean
value STD Unit
Saturated
Volumetric
weight
γsat 195 212 106 [kNm3]
Unsaturated
Volumetric
weight
γunsat 19 207 103 [kNm3]
Cohesion c 13 183 367 [kPa]
Friction angle φ 28 335 335 [ ˚]
Dilatancy angle ψ 0 0 0 [ ˚]
Youngrsquos Modulus E 2925 4500 1125 [kPa]
Poissonrsquos ratio ν 029 035 0035 [-]
Interface strength Rinter 034 05 01 [-]
63 Finite Element Model The structure has been modelled with the finite element software Plaxis 2D 2015 In Figure 62
the model as it was created and the automatically generated mesh are presented The ldquoboxrdquo
around the dike section was used to refine the mesh in this area The mesh in the interfaces
between the sheet pile and the soil was also refined for a better representation of the potential
high stresses that can develop in that area
Figure 62 FEM model and generated mesh for case study
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 74
The sheet pile has been modelled with elastic20 plate elements and the free anchor length with
an elastic fixed-end-anchor element The grout body of the anchor was not explicitly modelled
but it is implied in the end of the anchor as its final point is considered to be fixed in the soil A
fixed-end-anchor was chosen because even if it is a simplification of an anchorage system it can
still give reliable results in terms of the anchor axial force which is actually included in the limit
state function However the interaction of the soil and the grout body in that case cannot be
simulated The interaction and the modelling of a grouted body in the soil are still under
investigation due to several difficulties in modelling the real conditions
For all the steel members a Youngrsquos modulus of E=210 GPa was used while the specific
structural parameters are to be specified after the deterministic analysis that is described in the
next section
64 Deterministic Analysis The design concept of dikes with sheet piles modelled in FEM is based on the combination of
several partial factors in order to evaluate the overall required safety factor Additionally the
design recommendations of CUR 166 (2005) are considered for using the design or the
characteristic values of the structural parameters and the loads The design procedure (see
section 22) that has been initially followed in this structure is reported in Breedeveld (2011)
and the following structural elements have been defined accordingly
Cross sectional area of the sheet pile (per meter)
Length of the sheet pile wall
Steel quality of the piles
Cross sectional area of the anchor (per meter)
Free length of the anchor (without the grounded body)
Steel quality of the anchorrsquos reinforcement
Profile of waling
The required overall safety factor was calculated up to 18 according to the current design
procedure described in section 22 according which the above mentioned structural parameters
were defined Moreover apart from the overall safety factor in Breedeveld (2011) additional
partial factors are applied on the deduced forces and bending moments of the anchor and the
sheet pile wall as they are defined in CUR 166 (2005) In particular in Table 64 the different
partial factors that are applied on the several structural developed actions are given
Table 64 Partial factors for the different structural actions according to CUR 166 (2005)
Structural action Partial factor γ
Sheet pile normal force Fn 115
Sheet pile bending moment M 115
Anchor normal force FA 125
Waling bending moment Mw 110
20 The elements were considered as elastic in order to avoid numerical problems in FORM analysis For a better explanation a reference is made to the evaluation of the optimization algorithms in Appendix A2
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 75
However as it is mentioned in section 62 alterations have been made in the soil layers below
the dike while also the vertical load has been removed On the top of that in section 22 it was
shown that the large applied partial factors led to an overestimation of the design values for the
moments of the sheet pile Therefore a redesign of the sheet pile and the anchor is carried out
in this section in order to determine the new required characteristics of the structural elements
according to the new loading situation with less strict partial factors In particular in this thesis
the required overall safety factor was considered 12 according to CUR 166 (2005) calculation
scheme that is described in the next section The reason for using a lower safety factor was to
avoid an overestimation of the design forces and moments of the sheet pile wall that would
render a reliability analysis meaningless as the structure would be quite safe Such an analysis is
also important in order to show the discrepancy between the two different design procedures
and indicate the optimal one
However the reliability analysis was carried out for the mean values of the soil parameters and
for a safety factor of 10 in order to account only for soil properties uncertainties A safety factor
of 10 was not directly used as the design procedure followed in section 643 accounted only for
the macro-instability of the system and thus other failure mechanisms such as piping
overtopping infiltration etc were not taken into account In that way we prevent coming up
with a retaining wall that would be volatile under other crucial failure mechanisms
In section 641 the calculation scheme to be used in this thesis for the deterministic analysis is
further described In the sequence in section 642 the construction phases that were followed
in Plaxis simulation are listed and illustrated and finally in section 643 the results of the
preliminary design are displayed
641 Calculation Scheme and Design Values
In this research only the design philosophy suggested in CUR 166 is applied and the desired
safety factor of the structure is considered to be 12 for structures classified in ldquosafety class IIIrdquo
in order to avoid an overestimation of the structurersquos dimensions as it happened in the original
case study At this point it should be also clarified that the redesign was conducted considering
only the overall stability of the structure and no other failure mechanisms such as piping
internal erosion or overtopping
The prescribed calculation schemes in CUR 166 (2005) for designing structures with retaining
walls are basically the following
Calculation scheme A Calculations with design values The calculations are executed
using the design values for the soil parameters retaining height water levels and
stiffness of the structure It uses two different soil stiffnesses When using a high soil
stiffness the anchor force is generally relatively high The advantage of this scheme is
that it requires relative little effort The disadvantage is that the deformations may be
overestimated due to the use of design values in every stage However a designer is not
interested in deformations when performing a ULS calculation
Calculation scheme B Calculations with characteristic values The calculations are
executed using characteristic values of the soil strength (crsquo and φrsquo) Design values are
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 76
used for the retaining height water levels external loads and stiffness of the soil Here
again two different soil stiffnesses are used In the end of the governing stage a phi-c
reduction should be carried out As the characteristic values for the soil parameters are
used the deformations are most probably smaller than in case of calculation scheme A
The safety margin lies in the fact that phi-c reduction is imposed to obtain a safety factor
of 115 (class II) or 12 (class III)
For the soil stiffness parameters two options are given in CUR 166 a low and a high value
(61)
where is the (low and the high) design value the (low and the high) characteristic value
and is the mean value of the stiffness parameter In Woltersrsquos master thesis (2012) both
have been applied and the appeared to be governing for the bending moment in the wall
and the anchor force
In this research the Calculation scheme B was used and the as the soil stiffness
parameter It should be mentioned that regarding the soil parameters the characteristic values
were available with an exemption for the stiffness for which the mean values were given The
groundwater level was manually schematized according to TAW (2004) because of lack of data
while the maximum water level in the river side was taken equal to the design water level
according the Dutch regulations for the specific dike region
In Tables 65 and 66 the partial factors according to CUR 166 and the corresponding design
values for the different soil properties are presented As someone can notice in these tables
only the strength parameters (phi and c) and the stiffness parameter (E) are exhibited For the
rest parameters the design value is identical to the characteristic value (see Tables 63(a) and
(b)) as a partial factor of 1 is used
Table 65 Partial factors and design values for the soil layers under the dike
1 SAND (very silty)
Soil property Symbol Partial factor γk Design value Unit
Cohesion c 11 0 [kPa]
Friction angle φ 12 2569 [ ˚]
Youngrsquos Modulus E 13 1538462 [kPa]
2 CLAY (medium)
Soil property Symbol Partial factor γk Design value Unit
Cohesion c 11 91 [kPa]
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 77
Friction angle φ 12 1472 [ ˚]
Youngrsquos Modulus E 13 153846 [kPa]
Table 66 Partial factors and design values for the soil materials in the dike
3 DIKE NEW (medium clay)
Soil property Symbol Partial factor γk Design value Unit
Cohesion c 11 364 [kPa]
Friction angle φ 12 2479 [ ˚]
Youngrsquos Modulus E 13 1250 [kPa]
4 DIKE OLD (stiff clay)
Soil property Symbol Partial factor γk Design value Unit
Cohesion c 11 1182 [kPa]
Friction angle φ 12 239 [ ˚]
Youngrsquos Modulus E 13 2250 [kPa]
642 Construction Stages
The construction and the gradual loading of the dike ware modelled as follows (see Figure 63)
1 K0-procedure for the generation of the initial stresses under horizontal groundwater
level
2 Dike self-weight under horizontal groundwater level
3 Rise groundwater level to the phreatic water elevation
4 Interpolation21 of the clay layer under the dike and activation of the sheet pile wall and
the anchor
5 Apply extreme water level conditions on the dike structure
6 (φ-c reduction for the determination of the safety factor)
21 Due to different permeability between the clay and the sand layer an approximation of the developed pore pressures is carried out through interpolation between the two layers
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 78
1
2
3
4
5
Figure 63 Construction and loading stages
For the soil modelling the Mohr-Coulomb model was used with a non-associative flow rule
(φneψ) For phases 1-4 the calculations were conducted under drained conditions whereas in
phases 5 and 6 the undrained behaviour of clay layers was considered
643 Determination of the structural elementsrsquo characteristics
According to the calculation scheme that was described above the design characteristics of the
structural elements were obtained from phase 6 In particular the target overall safety factor is
12 with which initially the length of the sheet pile wall was calculated
+370 NAP
+700 NAP
+1050 NAP
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 79
Firstly a length of 10 m and a profile of AZ 12 with a steel quality of 240 kNm2 yield stress
were used while for the anchor a cross-section of 628 mm2 and S355 were chosen for the first
calculation However even if the safety factor reached 12 it seemed to be quite unstable during
the calculation steps while the developed anchor stress was above the ultimate yield stress
What is more the deformations resulted in that case were more than 30 cm Therefore in order
to stabilize the safety factor and to decrease the displacements the length of the sheet pile was
increased 1 m (so 11 m in total) while also a better steel quality for the sheet pile was chosen
S355 Furthermore the anchor cross section was increased to 933 mm2 with a steel quality of
MW450 Last but not least the waling was designed based on the anchor force
In the following paragraphs the design procedure of each element is described providing also
the results of moments and forces that were retrieved from Plaxis
Sheet Pile wall
As it was partially mentioned in the previous paragraph the length of the sheet piles is
considered to be sufficient when the safety factor of the system22 reaches at least 12 For the
reasons discussed above the sheet pile length was decided to be 11 m and with a profile of AZ
12 S355
With this configuration the safety factor reached is 13 as it is depicted at the right graph of
Figure 64 Someone can notice that the safety factor from 117 in the situation without a sheet
pile structure increased to 13 with the structure
Figure 64 Safety factor deduced from the phi-c reduction calculation without the anchored sheet pile wall at the left (safety factor=117) and with the wall at the right (safety factor=13) The steps on the x-axis represent the calculation steps have been carried out in order to reach an equilibrium after incrementally reducing the strength parameters ie phi and c
The design moment is to be taken for a safety factor of 12 that is reached in step 34 where it
was found to be 1519 kNmm and the corresponding axial force 9133 kNm Therefore
considering Eq 52 for checking the exceedance of the ultimate yield stress and applying the
prescribed partial factor of 115 for both the moment and the axial force we get
22 Plaxis always gives the safety factor of the system which is under investigation
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 80
In Figure 65 the maximum bending moments with a safety factor 10 and 12 are respectively
showed From the values being shown in that figure someone can easily distinguish the
difference between the two moments It must be also mentioned in that point that the required
safety factor defined in Chapter 2 was estimated at 18 (and not 12 that is considered here)
which gives excessive design moments (almost 7 times larger) For that reason in this research
a lower safety factor was assumed avoiding thus the design of a quite conservative structure
Figure 65 Bending moments on the sheet pile for a safety factor safety factor=10 at the left and
safety factor=12 at the right
The explanation behind this discrepancy is that in case of the phi-c reduction procedure the
artificially decreasing strength of the soil and on the other hand the non-reduced strength and
stiffness23 parameters of the structural elements tends to lead in an increased arching
phenomenon that makes the structure attract the most of the load
In this research the reliability of the structure will be evaluated for a safety factor of 1 or else
during phase 5 and by considering the mean values of the soil parameters As someone can
notice from Figure 65 the magnitude of the maximum bending moment for a safety factor =10
is quite low which implies a possible minor contribution to the overall failure
At this point it is essential to mention that Mohr-Coulomb model is capable of simulating the
elastic behaviour and the state at failure Therefore in case that plasticity of the soil or the
structural elements plays an important role a different constitutive model is recommended to
be applied
Anchor
For a safety factor 12 the anchor force was found to be 10387 kNm Considering a mutual
anchor distance of 3 m the total axial force that is undertaken by the rod multiplied also with
the prescribed design partial factor is
23 ldquoStrengthrdquo and ldquostiffnessrdquo are usually confused terms Strength is a measure of the maximum load that can be placed on a material before it permanently deforms or breaks and it is represented by the yield stress σy in metals and by φ and c in soil Stiffness is a measure of the amount of deflection that a load causes in a material and it is represented by the Youngrsquos modulus E and the Poisson ratio ν both in soil and metals
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 81
Therefore choosing an anchor with a cross sectional area of A = 933 mm2 and a steel quality of
MW450 ( ) the design stress level is
In comparison with the sheet piles the anchor does develop large stresses as it was concluded
from the results in the deterministic calculations This is probably attributed to the high
stiffness of the anchor that apparently attracts the load and thus the anchor develops large axial
force
Waling
The type of the waling system is defined according to the anchor force as it was described in
section 522 Therefore by choosing a waling type of 2UPE 200 the developed bending
moment can be estimated as follows
where is the design bending moment of the walling system according to the chosen profile
In Table 67 the characteristics of each structural element that was described above are
displayed in detail
65 Overview In this chapter an introduction of the case study as it has been used in this research is made The
soil materials and their properties are summarized and their mean and design values are
defined The probability distributions of the soil parameters to be considered as stochastic were
proposed according to the literature study and the engineering judgement Moreover the mesh
generation and the modelling features are discussed
Last but not least due the observed overestimation of the retaining wallrsquos design moment and
axial force according to the current design regulation a new deterministic analysis was carried
out in order to redefine the structural properties and come up with a less conservative structure
This makes the reliability analysis of the system meaningful in a sense that an extremely safe
structure would give a failure probability of zero Besides the scope of a reliability analysis is
also the optimization of the construction that corresponds to a financially attractive and
simultaneously safe structure This can be made by conducting a risk analysis whose first step
would be the analysis that is carried out in this research Therefore the aim is to start with the
marginal required structural properties according to the safety standards In a risk analysis that
should conducted for several dimensions of the structure until the most optimized one is
identified However this is out of the scope of this thesis
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 82
In Table 67 the structural properties of the sheet pile wall the anchor and the waling are
summarized as they were found in the deterministic analysis
Table 67 Design parameters of structural elements
SHEET PILES
Property Symbol Value Unit
Profile - AZ 12 [-]
Steel quality - S355 [-]
Length L 11 [m]
Moment of inertia I 21430 [cm4m]
Elastic section modulus Wel 1245 [cm3m]
Mass w 943 [kgm]
Sheet pile thickness d 85 [mm]
Cross sectional area A 1201 [cm2m]
ANCHOR
Property Symbol Value Unit
Steel quality - MW450 [-]
Free length Lafree 104 [m]
angle φ 30 [deg]
Cross sectional area A 933 [mm2]
Mutual anchor distance s 3 [m]
WALING
Property Symbol Value Unit
Profile - 2UPE200 [-]
Steel quality - S355 [-]
Elastic section modulus Wel 191 [cm3m]
In the original case study the sheet pile wall is 162 m long and it has a profile of AZ 28 (S240)
while the anchor is an MW450 with a cross-section of 1134 mm2 As it is noticed both sheet
pilesrsquo and anchorrsquos properties have been reduced after the new deterministic analysis However
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 83
it should be mentioned that in the original case there was a vertical load on the top of the dike
that was displaced while the clay layers under the dike was merged into one in the new case
study These changes might lead to less heavy structure However the difference among the
design properties between the new and the original case study is mainly due to the required
safety factor that was assumed In the new case study the safety factor was set to 12 whereas in
the original cases study a safety factor of 18 was considered as it was calculated according to
the partial safety factors
In the next chapter the results after the reliability analysis of the just designed case study are
presented
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 84
MSc Thesis A Rippi 85
7 Reliability analysis results with stochastic soil properties
In this chapter the probability of failure of the different structural elements that were discussed
in Chapter 4 will be estimated by varying the soil properties of the several soil layers under and
inside the dike At the beginning a better insight into the system behavior is gained by
inspecting the deterministic Plaxis calculations for the mean values of the soil variables After
that a global SA is carried out in order to evaluate the most important soil parameters and
inspect their impact on the different structural elements in combination with each other and
individually Finally the results regarding the probability of failure of the anchor the sheet pile
wall and the soil are presented as they deduced by the reliability analysis (concerning the ULS of
the elements) Finally the systemrsquos reliability is evaluated The methodology that was followed
during the probabilistic analysis is described below and it is mainly dedicated to the DS
rationale that was used in order to obtain the influence factors and to evaluate the results of
Plaxis calculations
71 Method description For the calculation of the failure probability of the systemrsquos different components FORM and DS
were used for the retaining wall and the soil body respectively From these methods the
probability of failure and the important factors are going to be deduced for each component (ie
anchor sheet pile wall and soil body) However before starting with the reliability analysis it is
essential to carry out a sensitivity analysis in order to obtain a first impression of what are the
most influencing variables and thus determine the input random parameters for the reliability
analysis In the figure below the steps for the reliability analysis as they were followed in this
thesis are presented
Figure 71 Steps of the reliability analysis
Below the methodology of both the sensitivity and the reliability analysis are discussed and
explained in detail
Sensitivity analysis
In this thesis the FAST method was utilized for conducting the sensitivity analysis which is
further elaborated in Appendix A1 Here an explanation is given for the quantitative measure of
sensitivity which is represented by Sensitivity Indices The first-order sensitivity index of an
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 86
input variable pi is the measure of the main (direct) effect of pi on the output variance
(where i ne j) the second-order sensitivity indices measures the interaction effect of pi and pj on
the output variance Other higher-order indices are defined in the same manner The total
sensitivity index is the sum of all sensitivity indices involving factor pi (Homma amp Saltelli
1996) For example the total sensitivity index of factor 1 for a model with 3 input factors is
given as
The first-order (Si) and total ( ) indices can be interpreted as following (Pandya et al 2008)
high pi is an influent parameter
and both small pi is not an influent parameter (neither alone nor in interaction
with other parameters)
and nearly the same no interaction of pi with the other parameters
and very different high interaction of pi with the other parameters
Reliability analysis
As far as the FORM analysis is concerned section 422 gives sufficient information about the
methodology that is followed in order to obtain the failure probability and the influence factors
(ie α-values) The main concept and principles of DS are briefly explained in section 421
whereas here a more precise description of the implementation of DS and the how the α2-
values can be deduced is given
In particular DS is a method in which ldquodirectionsrdquo are sampled The concept of this method is
schematically depicted in Figure 72(a) For each sampled direction it is evaluated whereas
along this line the LSF Z equals to zero or not This procedure is repeated for a number of
directions and this method is applied in the standard normal space (u-space see also section
422) using equations to transform the standard normal variables u1 hellip un to their
corresponding ldquorealrdquo variables (in the x-space) X1 hellipXn and vice versa
(a) (b)
Figure 72 (a) Schematic view of the DS method in the u-space The numbers 0 1 2 3 4 show successive LSF evaluations along one direction (b) Schematic view of the line search algorithm β is the distance from the origin in the u-space along one direction The numbers 0 1 2 3 4 correspond to the ones shown in figure (a) (source Hydra-Ring 12 2015)
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 87
The search procedure for the location on the line where Z=0 is often referred to as the ldquoline
search methodrdquo and it is depicted in Figure 72 (a) and (b) In both figures the numbers 0-4 refer
to successive Z-function evaluations along one direction Figure 72(a) shows 3 evaluated
directions in the u-space whereas Figure 72(b) shows the (fictional) value of the Z-function
along one direction (where ||u|| is the distance from the origin along the line) The maximum
length of the direction line as well as the step size along the direction (for example the distance
between 0 and 1 along one line) can be steered in order to increase the efficiency of the method
and presumably decrease the computational time
In the sequence a method was developed in this thesis in order to estimate the α2-values of the
random variables out of a DS analysis as OT does not have an available method so far For that
purpose a transformation of the output samples to u-space was firstly carried out Then the
distance to the origin of all samples (in u-space) that are located on a direction where failure
(Z=0) was detected is calculated as follows
radicsum
i=1hellip (71)
where n is the number of random variables and uij is the ith ldquofailurerdquo sample of the jth random
variable is equivalent with the reliability index of this direction βi The α-values for each
random variable can then be calculated as (for more information about these values see
sections 422- Eq 417 and 423)
(72)
After that three different methods were created in order to evaluate the influence of the
random variables on the response of the LSF ie the α2-values (square of Eq 72)
Shortest distance ldquoβminrdquo
Average 10
Average all
In the Shortest distance method the sample with the smallest distance to the origin is
considered to be the design point and thus the corresponding influence factors are estimated
according to that point The smallest distance is also equivalent to the reliability index βmin This
method is well-known for estimating the design point out of MC techniques (Vrijling amp Van
Gelder 2002) However for a soil failure modeled in FEM this might not be always indicative of
the real influence that the soil properties can have on the system failure Therefore the last two
techniques were used as additional in order to get a better insight into the dominant soil
properties In Figure 73 a general scheme of the methodology that was followed for obtaining
the α2-values is depicted where also the different techniques are presented
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 88
Figure 73 DS methodology for calculating α2-values
The second method Average 10 takes into account the α2-values of the samples in the failure
domain whose distance to the origin lies within the 10 higher than βmin and averages them
Therefore it is firstly required the calculation of βmin as it was described for the Shortest distance
method Such a technique helps to identify the other directions that are close to the one which
gives the shortest distance to the origin and thus inspect if the important variables that were
indicated according to the first method are influential for other failure mechanisms as well
Last but not least the third method Average all averages the α2-values of all the samples
located in the failure domain This shows the overall contribution of the random variables to the
response of the LSF under investigation Concerning the soil failure LSF that is described in
section 73 such a method can be more suitable for distinguishing the soil properties that are
involved in the majority of the failure mechanisms rather than only in the one that the first
method indicates
72 Mean values calculations The results of the deterministic analysis that are presented in the previous chapter included a
degree of conservatism due to the load and material factors (partial safety factors) In order to
overcome this limitation structural reliability calculations were carried out and are discussed in
the next sections by considering relevant soil parameters as random However before starting
the probabilistic calculations by coupling Plaxis with OT that cannot be fully interpreted and
verified to a certain extent it is essential to first carry out some deterministic calculation Such
calculations can help to get a better insight into the behavior of the system and show
qualitatively what kind of response to wait later in the analyses These calculations were taken
place based on the mean values of the soil properties
Deformations
The analysis of this case study focuses on the ULS of the structure For the stresses and
deformation calculations the Mohr-Coulomb model is used that is generally accepted for the
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 89
simulation of the soil behavior until failure and for detecting the failure modes (a more detailed
explanation of this model is provided in Appendix B1) However this model is not so suitable
for analyzing the deformations of soft soils or the settlements around the wall
The deformation pattern that would more likely occur due to primarily the soil body failure can
be illustrated by conducting a ϕ-c reduction calculation In Figure 74 (a)-(c) the displacement
pattern after the ϕ-c reduction of the mean values of the friction angle and the cohesion
respectively As it can be seen a slip surface in the inner side of the dike has been created that
rotates towards the inland
(a) deformed mesh
(b) shadings
(c) arrows
Figure 74 Deformations after the ϕ-c reduction (mean values)
Stresses on the structural elements and the soil body
The anchor and the sheet pile wall are installed in phase 4 In phase 5 the anchor normal force
increases and especially for the mean values it amounts to Na = 7983kNm
As far as the sheet pile wall is concerned the bending moments and the normal forces that are
developed on the wall are indicated in Figure 75 as they were deduced from phase 5 Similarly
to the anchor the maximum stresses occur in phase 5 and are mainly located in the area of the
wall which is surrounded by the clay layer (brownish soil layer in Figure 74(a)) As someone
can observe the bending moments are not significant taking into account that the maximum one
observed reaches 2132 kNm
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 90
Figure 75 Bending moments and normal forces for the mean values calculation
For the interpretation of the stress field of the soil body the effective principle stresses and the
relative shear stress τrel (see Eq 517) were illustrated in Figure 76 (a) and (b) respectively In
Figure 76 (a) the crosses indicate the direction and the magnitude of the principle stresses σ1rsquo
and σ3rsquo and from which an active soil behavior (
) can be noticed on the river side of
the retaining wall whereas a passive behavior (
) is dominant in the inland side
(a) Effective principle stresses
(b) Relative
shear stress τrel
Figure 76 Effective principle stresses and relative shear stress in mean values calculation
In Figure 76 (b) the distribution of the relative shear strength shows the potential developing
plastic area in case of failure that expands from the river side of the wall and around it It can
also betoken the possible shape of formation of a slip surface that starts from the outer side of
the dike and goes around the wall
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 91
73 Sensitivity Analysis Results In section 312 different types of SA are presented and that can be applied through OT In this
research due to the time limit and the amount of parameters FAST method was used for
carrying out the SA as it allows the calculation of the first order sensitivity indices as well as the
total order indices and the computational time required is less than the other available methods
In this section the results from the conducted SA are presented separately for the anchor the
sheet pile and the soil performance Regarding the sheet pile wall and the anchor the sensitivity
of the total developed stresses towards the soil properties was evaluated according to Eq 53
and 55 respectively The influence of the soil parameters on the soil behaviour was evaluated
based on the calculation of the overall safety factor after a safety calculation (ϕ-c reduction
method) The total order indices are considered for distinguishing the most important
parameters while their difference with the first order indices Si indicates the level of interaction
effect amongst the parameters on the output variance
As far as the anchor and the sheet pile wall are concerned the SA was carried out in two steps
(1) SA with all the soil parameters (listed in Table 62) and (2) SA with the most important
parameters found in step 1 In Figure 77(a) and (b) the sensitivity indices from step (1) are
presented as they were found for the anchor and the sheet pile respectively Step 1 mainly
provides an overview of the contribution of the variablesrsquo variance to the total variance of the
limit state response and it helps to identify the most influencing parameters Step 2 was
basically carried out in order to build the response surface based on the SA output that was
later utilized in the FORM analysis as a complementary function in case of Plaxis errors At that
point it should be mentioned that the Dikenew material was not considered neither for the SA
nor for the reliability analysis of the structural elements as its influence was assumed to be
insignificant relative to the other soil layers However Dikenew was considered in case of the
soil body investigation
From Figure 77(a) it can be concluded that as far as the anchor is concerned the shear stiffness
G of the clay layer seems to be the most crucial soil property In particular the anchorrsquos stress
level seems also to be sensitive towards the Poisson ratio of the clay layer ν the friction angle φ
of the sand layer and the strength parameters φ c of the Dikeold material Moreover the
interface strength Rinter between the Dikeold material and the sheet pile wall seems to display
an additional notable influence as well The unit soil weight γ does not appear to affect the
anchor performance significantly
As far as the sheet pile wall is concerned the shear stiffness of the clay layer and the Dikeold
material seem to be dominant according to Figure 77(b) From these two soil layers it appears
that also the strength properties (φ c and Rinter) as well as the unit soil weight γ have a relative
impact on the stress level of the wall Furthermore the sand layer contributes via especially its
friction angle and shear stiffness
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 92
(a)
(b)
Figure 77 Total and First order sensitivity indices for the (a) anchor and the (b) sheet pile in step 1
As it was mentioned before step 2 was used for building the response surfaces for the anchor
and the sheet pile limit states The input random variables in this step are the most influencing
ones that were found in step 1 In Table 71 the soil properties that were considered as random
during step 2 are presented for both anchor and sheet pile It is essential to mention that the RS
was constructed by fitting a quadratic polynomial function to the sampling points of the SA
whose general expression is given in Eq 425 In Figure 78(a) and (b) an illustration of these
response surfaces is made However the figures depict the response of the anchor and the sheet
pile stress level over the fluctuation of only two variables amongst the 14 that are presented in
Table 71
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 93
Table 71 Soil properties considered for the RS construction of the anchor and the sheet pile respectively
Soil layer Anchor Sheet pile
Clay Sand
Dikeold
The linear regression of all the parameters was based on least squares and the R2 parameter
was found 0989 for the sheet pile and 0996 for the anchor Therefore the regression was quite
representative for the domain of the samples over which it was adjusted However there were
many residuals (ie difference between the observed and the predicted value) in the order of
103 and 104 which means that a possible expansion of this RS to an extended domain of the
input variables may lead to an underestimation of the limit state response
(a)
(b)
Figure 78 Polynomial response surface adjusted for the (a) anchor and (b) sheet pile stress level For the anchor the friction angle of the sand layer and the shear stiffness of the clay layer were accounted for this illustration whereas for the sheet pile wall the shear stiffness of the sand and the cohesion of the clay layer were used
The SA for the soil performance was also conducted in two steps However in that case there
was no need of a response surface construction since the LSF for the reliability analysis (Eq
522) takes the value of -1 in case of Plaxis error due to soil collapse Eq 522 though takes only
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 94
two different values -1 and 1 for soil collapse or not respectively which cannot be used for
carrying a SA For that reason the φ-c reduction method was used and the sensitivity of the
value of the safety factor (Msf) was investigated instead The second step of the SA was carried
out in order to divide the soil properties into two different groups of variables and thus
investigate the sensitivity of the soil in a more efficient way rather than including all the
parameters at one SA It should be mentioned that in the soil analysis soil properties of the
Dikenew material were added In Figure 79(a) and (b) the total and the first order indices are
depicted for step 1 and 2 respectively Step 1 includes all the stiffness and strength parameters
of all the soil layers whereas in step 2 the most important variables that were found in step 1
together with the unit soil weight γ are included Subsequently step 1 detects the most
influencing stiffness and strength properties and in step 2 the impact of γ is investigated
As it is observed in Figure 79(a) and (b) the safety factor appeared to be sensitive towards the
sandrsquos friction angle while also the strength properties (c and φ) and the shear stiffness (G) of
the clay layer play an important role Moreover the cohesion of the Dikeold material and the
unit soil weight of the clay sand and Dikenew layer seem to be influencing as well As it has
been mentioned above for this SA the LSF was different than the one considered in the
reliability analysis Therefore even if the SA at that point helps to evaluate in general the most
dominant soil properties the engineering judgement should also be included for the selection of
random variables that are going to be utilized in the reliability analysis
(a)
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 95
(b)
Figure 79 Sensitivity indices for the soil in step 1 and 2
It is also advisable that the difference between the total and the first order sensitivity indices be
investigated in order to have an insight into the most interactive variables The magnitude of the
interaction effect of a variable is crucial for the system behaviour because a variable might not
be important as a singularity but it is possible its combination with another variable to have a
considerable effect on the limit state under investigation Therefore this can be considered as
an additional criterion for choosing the set of the random variables to be used in the reliability
analysis In Figure 710(a) and (b) the difference between the two indices is presented for the
anchor and the sheet pile in (a) and the soil in (b) as it was deduced from the first step of the SA
Regarding the sheet pile wall the most interactive variables are the shear stiffness and the unit
weight of the clay layer the friction angle of the sand and the cohesion shear stiffness interface
strength and unit weight of the Dikeold material Concerning the anchorrsquos limit state the
cohesion of the Dikeold material seems to have the highest interaction with the rest of the
variables Last but not least regarding the soil performance the stiffness of all the soil layers
and the friction angle of the clay and the sand layer have a higher interaction effect amongst the
other parameters The importance of the aforementioned variables can be enhanced
considering both total indices and the difference that they display between their first and total
indices
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 96
(a)
(b)
Figure 710 Interaction effect amongst the parameters in (a) anchor and sheet pile and in (b) soil This effect is expressed as the difference between the total order indices T and the first order indices S as they were deduced from the first step of the SA for both anchorsheet pile and soil respectively
Considering the discussion above some preliminary conclusions can be drawn regarding the
influence of the several variables to the different components of the system
1 both anchor and sheet pile wallrsquos limit states are sensitive mostly towards the shear
stiffness of the soil
2 sheet pile wall stress level is also affected by the unit soil weight more than the anchor
does
3 soil body is mostly influenced by the strength properties and the unit soil weight
4 the interaction effect among the soil properties is more profound for the soil and the
sheet pile limit states rather than the anchorrsquos
It must be mentioned that this SA is a global one that gives an overview of the impact of the soil
parameters in terms of the system behaviour However near the design point a local sensitivity
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 97
can possibly give a better estimation of the variablesrsquo impact Near the design point the
structural elements as well as the soil body usually reaches plasticity that implies a non-linear
behaviour Such a behaviour is unlikely to be initiated during a SA that usually samples around
the mean values This can be counteracted by conducting more iterations so as to have a larger
amount of available samplings but this can make the SA to be quite time consuming and thus to
lose its benefit to give a fast preliminary estimation of the most influencing soil layers and
properties
Besides the purpose of a preliminary global SA is to reduce the number of the stochastic
parameters to a reasonable amount so as the initial reliability analysis to be more efficient and
computationally accessible The parameters that were eventually considered as stochastic in
each LSF are summarized in the table below and they were chosen according to both the SA and
the engineering judgement In the last column the stochastic variables used also for the
systemrsquos analysis are presented
Table 72 Stochastic variables for each LSF
Soil
Failure Anchor Failure
Sheet Pile failure
System failure
Sa
nd
γunsat radic
c
φ radic radic radic radic
G radic radic radic radic
v
radic
Rinter radic
Cla
y
γunsat radic
radic radic
c radic radic radic radic
φ radic radic radic radic
G radic radic radic radic
v
radic
Rinter radic radic radic
Dik
e_n
ew
γunsat radic
c radic
φ radic
G
v
Rinter
Dik
e_o
ld
γunsat
radic radic
c radic radic radic radic
φ radic radic radic radic
G
radic radic radic
v
radic
Rinter radic radic radic
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 98
74 Soil Shear Failure As far as the soil failure is concerned the limit state function that was implemented is given
below (see also sections 52 and 53) It was formulated as a single value depending on the
success of computation in Plaxis
(71)
The reliability method that was chosen in that case is DS due to the formulation of the limit state
function and the convergence issues of FORM There were used 12 input stochastic soil
parameters as it is indicated in Table 72 (see also input file of Soil Failure in Appendix D) The
variables were chosen based on the SA that was presented in the previous section For the soil
failure Poisson ratio was initially included in the analysis but it was noticed that values close to
04 during Plaxis calculations were leading to unrealistic behaviour of the dike More precisely
it has been observed that Poisson ratio equal to 04 can lead to swelling effects on the dike
which consequently lead to a raise of the dike crest instead of collapsing as it would be expected
in a soil failure analysis Therefore Poisson ratio was excluded from the reliability analysis of
the soil failure
The results of the reliability analysis according to DS are presented below In particular 300
iterations were carried out and 3 different failure directions were found Specifically in Table
73 the probability of failure the reliability index the number of iterations and the duration of
the analysis are firstly shown In the sequence the points from each failing direction with the
lowest reliability index are shown
Table 73 Reliability results of DS for the soil failure
DS Pf 1310-8 β 55 Number of LSF calls 1840 Number of iterations (or directions)
300
Elapsed time (hr) 34
Failure points (in real space) Parameter 1st direction 2nd direction 3rd direction Dikenew_phi [deg] 2277 2543 3021 Dikenew_c [kPa] 529 557 444 Dikenew_gammaunsat [kNm3]
1748 1848 1973
Dikenew_gammasat [kNm3]
1948 1984 1997
Sand_phi [deg] 2782 2039 3899 Sand_G [kPa] 808437 715473 715785 Sand_gammaunsat [kNm3]
2183 2122 2097
Sand_gammasat [kNm3] 2352 2145 2283 Clay_G [kPa] 56934 90063 80134 Clay_c [kPa] 704 1593 794
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 99
Clay_phi [deg] 1039 1629 1077 Clay_gammaunsat [kNm3]
1772 1413 1612
Clay_gammasat [kNm3] 1833 1433 1719 Dikeold_c [kPa] 1621 2978 915 Dikeold_phi [deg] 2474 3149 3281
According to the Shortest distance method the failure point from the 2nd direction turned out to
be the design point the importance factors of which are depicted in Figure 714 (first pie chart
at the left) However before concluding for the final design point an assessment of the above
failing directions was made
In particular the combination of the variables from each direction was later implemented in
Plaxis in order to illustrate the failure This is also necessary for verifying and assessing the
failure that Plaxis is giving and inspecting the error in the logfile in order to check if it is a ldquorealrdquo
(ie error=Soil collapses) or a ldquofakerdquo failure (ie error=Load advancement failed or Not enough
load steps) In Figure 711 the ldquorealrdquo failure is defined
Moreover a demonstration of the displacements is also essential in order to certify the failure
shape that normally has the dike crest settled down A different output of the displacements
formation (ie dike crest goes up) would probably imply a ldquofakerdquo failure However it should be
mentioned that a ldquofakerdquo failure in Mohr Coulomb model can be a ldquorealrdquo failure in another model
that can more realistically simulate the deformation patterns of the structure In general even
though Mohr Coulomb can be successfully applied for detecting failure it is not very suitable for
the determination of the displacements and especially for soft soils such that the dike under
investigation is constructed by It should be also mentioned at that point that in this case study
the system is subjected to unloading conditions24 (ie decrease of effective stresses) in soft soil
layers after phase 4 which cannot be modelled successfully by the Mohr-Coulomb model In that
case other constitutive models such as the Soft Soil Model and the Hardening Soil model are
presumably more reliable for simulating such situations For further information for this model
a reference is made to Plaxis (2015) Therefore someone should be always critical and verify
the failure points based on hisher engineering judgement
Figure 711 Prerequisites for real failure in Plaxis
24 From phase 4 to 5 there is an increase in water pressures and thus due to constant total stresses there is a decrease of the effective stresses This can be characterised as an unloading situation
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 100
In order to make it more understandable the different failure points were illustrated in Plaxis
and their deformed mesh and total displacements are presented in Figure 712 (a)-(c) for each
failure direction As one can observe in the first two directions the dike crest settles down
whereas in the 3rd one the dike body seems to swell and move to the right (the blue lines that
surrounds the dike soil layers indicate the initial dike geometry) However even if the 2nd
direction gave a normal pattern of displacements the error message was ldquoNot enough load
stepsrdquo The combination of the variables of the 3rd direction was then tested again with the
Hardening Soil model (HS) in order to illustrate the deformations with a more suitable model
and it was proved that the deformations seem to be realistic as it is depicted in Figure 713 With
Mohr Coulomb only the 1st direction seems to comply with the diagram in Figure 711 and that
would properly be considered as the real design point
(a) Deformed mesh
(b) Total displacements |u| in m (shadings)
(c) Total displacements direction (arrows)
1st direction 2nd direction 3rd direction
Figure 712 Illustration of the design point in soil body failure
(a) Deformed mesh (b) Total displacements |u| in m (shadings)
(c) Total displacements direction (arrows)
Figure 713 Displacements pattern of the 3rd failure direction according to HS
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 101
However taking into account that the other directions showed characteristics of failing behavior
they should also be accounted for the determination of the importance factors Excluding them
from the analysis would not be wise as the error message or the deformation schematization can
be a limitation of FEM or Mohr Coulomb model respectively In any case the points that lead to
failure shall be treated critically in order to decide if they should be included or not In this case
the points from the three directions seem to be realistic as both strength and stiffness
parameters are low enough in order to lead to a potential failure Therefore they were finally
considered for the analysis and the importance factors were averaged over all the points as it is
depicted in Figure 714 (last graph at the right)
Figure 714 Importance factors α2 for soil failure
Another option is also to average over the 10 range of close to the lowest reliability index that
was calculated with the shortest distance method (ie Average 10 method) in order to come
up with the α2 values However such methodology can lead to averaging over points of the same
direction that are not very different with each other Therefore the rest of the failure points are
automatically omitted from the consideration This can be also verified from Figure 714 where
the pie chart of the Shortest Distance and the Average 10 are quite similar whereas the Average
all chart introduces the importance of other soil variables as well
From a general perspective the soil properties of the clay layer seems to be determinant
whereas also sand and Dikeold material play an important role to the soil failure More precisely
according to the Average all method the unit weight and the friction angle of the clay layer
turned out to be the most influencing while also the cohesion of the Dikeold material and
friction angle of the sand layer contribute to the failure domain From a qualitative point of view
the weakening and consequently the settlement of the Dikeold material pushes the subsoil
creating an additional surcharge for the underlying clay and sand layer which act like the
foundation soil the incapability then of the clay layer to withstand the overlying load due to the
low unit weight in combination with the low friction angle of the clay as well as the low stiffness
and strength of the sand layer can lead to the creation of an inner slip surface and thus failure
Since the clay layer can be considered as a foundation soil the importance of the friction angle
and the unit weight can be also explained from the Terzaghirsquos bearing capacity theory (Terzaghi
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 102
1943) where these two factors are crucial for the determination of the ultimate bearing capacity
of the soil However these can be better verified by conducting large scale experiments in dikes
and inspecting the sensitivity of the dikersquos stability towards the soil properties
75 Anchor Failure The reliability of the support system is estimated based on the anchor and the waling probability
of failure However the load on the waling is proportional to the anchor force as it appeared in Eq
57 Therefore after the estimation of the anchor reliability by calculating the probability of
exceedance of a certain admissible anchor force the waling can be designed deterministically
Besides the failure probability of the waling must be lower than the anchorrsquos (given that no
uncertainties of the strength and the structural properties are considered) For the anchor failure
the LSF was formulated as follows
(72)
where [kN] is the anchor force is the yield stress and the cross
section area The reliability method to be utilized in this analysis is FORM and the selection of the
random variables to be used was made according to the sensitivity analysis results and the
engineering judgment and they are listed in Table 72 In particular the variables from the second
step of the sensitivity analysis were considered which will also help at a later stage to compare
FORM sensitivity method and FAST In Table 74 the results of the reliability analysis with FORM
are listed Additionally the design point together with the importance factors of each variable is
presented It should be mentioned at that point that the sensitivity analysis helped to define a
starting point closer to the design point and thus reduce the computational time required for the
analysis More precisely the analysis lasted approximately 16 hours whereas by starting from
mean values from which by default FORM starts the iterations it would take almost a day
Table 74 Reliability results for the anchor failure with FORM
FORM Pf 28410-6 β 45 Number of LSF calls 119 Maximum number of iterations
100
Elapsed time (hr) 16
Design point Parameter value α-value Clay_phi [deg] 1814 -0059 Clay_nu [-] 033 0086 Clay_c [kPa] 1487 -0089 Clay_G [kPa] 41739 0911 Clay_Rinter [-] 051 -0021 Sand_Rinter [-] 069 -0051 Sand_nu [-] 030 0007 Sand_phi [deg] 3259 -0174 Sand_G [kPa] 1130770 0048 Dikeold_phi [deg] 2842 -0020 Dikeold_nu [-] 036 -0045
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 103
Dikeold_c [kPa] 1873 -0045 Dikeold_G [kPa] 243183 -0326 Dikeold_Rinter [-] 053 -0072
Figure 715 Importance factors α2 in anchor failure from FORM analysis
The influence coefficients in Figure 715 indicate that this limit state is governed by the shear
stiffness of the clay layer G Furthermore the shear stiffness of the Dikeold material seems to
contribute significantly to the anchor yield stress exceedance It can be concluded that the
problem is still in the elastic domain as far as the soil behaviour is concerned In case of
predominately plastic behaviour the strength properties of the soil (φ and c) become more
important In Figure 716 a demonstration of the design point as it was acquired by the FORM
analysis is presented The figures indicate that presumably due to the clay layer withdrawal in the
passive side as it is depicted in Figure 716(b) with the red shadings sheet pile moves to the right
and subsequently anchor is tensioned enough to reach its yield stress25
(a) Deformed mesh (b) Total displacements |u| in m (shadings)
(c) Total displacements direction (arrows)
Figure 716 Design point illustration from the anchor failure
25 A better representation of the soil-anchor interaction could be possibly obtained if the anchor was modelled as a node-to-node element and the anchorage body as an embedded element
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 104
The calculations were repeated with the 6 most influencing parameters both with FORM and DS
in order to investigate the effect of reducing the number of random variables and in order to
validate the results of FORM analysis In Table 75 the reliability results of respectively FORM and
DS are presented while in Figure 717 the importance factors of both methods are juxtaposed The
design point and the importance factors of DS were estimated based on the Average 10 method
(see section 71)
Table 75 Reliability results of FORM and DS with reduced variables in anchor failure
FORM Pf 8510-6 DS Pf 1510-5 β 43 β 42 Number of LSF calls
28 Number of LSF calls
776
Maximum number of iterations
100 Number of iterations
100
Elapsed time (hr) 034 (starting point close to the design point)
Elapsed time (hr)
115
Design point Design point Parameter value α-value value α-value Clay_nu [-] 033 0119 033 0242 Clay_c [kPa] 1404 -0026 1535 -0212 Clay_G [kPa] 43236 0928 44548 0807 Sand_phi [deg] 3163 -0110 2954 0203 Dikeold_G [kPa] 239465 -0330 251723 -0366 Dikeold_Rinter [-] 053 -0064 060 -0263
Figure 717 Importance factors α2 with FORM and DS analysis respectively
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 105
As it can be noticed comparing FORM results from Table 74 and 75 the reliability index changes
from β=45 with 14 parameters to β=43 with the 6 most influencing parameters This verifies the
importance of the most influencing parameters that it was found in the analysis with the 14
parameters The small discrepancy though between the two reliability indexes might be
attributed to the interaction effect of the soil variables on the output performance
The estimation of failure probability with DS is almost the same with FORM As far as the
importance factors are concerned DS also came up with the conclusion that the shear stiffness of
the clay and the Dikeold material are the most crucial soil properties for the anchor stress level
However according to DS the stiffness of clay ν and the interface strength Rinter between the
Dikeold material and the sheet pile wall seem to also contribute to the limit state of the anchor
The reason why the strength of the clay layer under the dike and the Dikeold material (the part of
the dike that comes in contact with the anchor and the sheet pile wall) are the most important
variables can be explained by the principle of soil arching According to this principle the soil
columns on both sides of the rigid sheet pile wall are more compressive than the soil columns on
the top of the wall (ie overlying dike body) because of the higher stiffness of the wall when
compared with soils As such soil columns on both sides tend to settle more than the soils on top
of the rigid wall and this differential settlement causes a downward shear force acting along the
sides of soil columns on top of the wall As such the vertical load on the wall becomes larger than
the sole weight of soil columns on its top and the anchor that actually receives the most of this
vertical load reaches eventually its yielding stress
Last but not least it should be mentioned that the response surface which was constructed based
on the SA worked efficiently in most of the Plaxis errors cases in a sense that it gave reasonable
results of the anchor stress level
76 Sheet pile wall failure For the reliability analysis of the sheet pile wall the strength property ie the yield stress was
considered constant and equal to 355 kPa Therefore the LSF that was applied for this case was
formulated as follows
[
] (73)
where [kNm] and [kN] are the maximum bending moment and the axial force
that are exerted on the wall respectively is the elastic section modulus ( =1245 cm3m)
and is the cross sectional area ( =1201 cm2m for a AZ12 profile)
In sections 643 and 72 the maximum bending moments of the sheet pile wall were calculated
with the design values of the soil properties and the mean values respectively It has been
observed that the bending moments were not so significant and determinant in comparison with
the anchor force and the safety factor of the system that were also considered (in section 643) in
order to determine the structural properties
In a later stage a SA was carried out for the sheet pile wall that has been further discussed in
section 73 Figure 718 shows the stress level of the wall that was developed after 3150 iterations
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 106
by considering as random variables the ones that are defined in Table 72 As someone can
observe the stress level is generally quite low and even an order lower than the yield stress (ie
355105 Pa)
That small magnitude of bending moments can be probably attributed to the counteraction of the
active earth pressures from the passive ones that due to the homogeneity of the soil in both sides
of the wall are quite similar
Figure 718 Stress level [in Pa] of the sheet pile wall after 3150 iterations of the SA
The reliability analysis has been carried out with FORM method and considering as random
variables those listed in Table 72 However due to the low level of stresses that are developed
and the simultaneously high yield stress the method converged after almost 35 days of
calculations As it can be noticed in Table 76 the analysis came up with an extremely low
probability of failure It should be mentioned at that point that the convergence of the analysis
was successful thanks to the RS that was fitted This means that a Plaxis error of soil collapse
preceded and it activated the RS which implies that the soil failure is more likely to occur before
the sheet pile failure
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 107
Table 76 Reliability results for the sheet pile wall failure with FORM
FORM Pf 7210-38 β 13 Number of LSF calls 2619 Maximum number of iterations
100
Elapsed time (hr) 84 (35 days)
Design point Parameter value α-value Clay_phi [deg] 1164 0261 Clay_c [kPa] 586 0336 Clay_G [kPa] 14275 0663 Clay_Rinter [-] 052 -0016 Clay_gammaunsat [kNm3]
1848 0000
Clay_gammasat [kNm3] 1948 0000 Sand_phi [deg] 3020 0000 Sand_G [kPa] 1193650 0000 Dikeold_phi [deg] 1360 0406 Dikeold_c [kPa] 592 0438 Dikeold_G [kPa] 168798 0000 Dikeold_Rinter [-] 053 -0024 Dikeold_gammaunsat [kNm3]
2275 -0150
Dikeold_gammasat [kNm3]
2375 0000
This can be also verified from the importance factors that are depicted in Figure 719 where apart
from the shear stiffness of the clay layer which seems to be the predominant soil property for the
sheet pile LSF the strength parameters (φ and c) of the clay and the Dikeold materials also play
an important role This implies that the soil behaviour has already passed into its plastic domain
where the governing properties are the strength parameters before the sheet pile exceeds its
yield stress
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 108
Figure 719 Importance factors α2 of the sheet pile stress level in the plastic domain of soil
Nevertheless in order to investigate the potential most influencing soil properties for the sheet
pile LSF in the elastic domain as well the available output data of the FORM analysis have been
assessed The highest observed value of the sheet pile stress before its failure was around
235000 kPa After tracing the results it seems that when FORM tries to reduce further the
stiffness and strength properties of the soil layers soil collapses before the sheet pile failure The
vectors of the random variable that led to a stress level on the sheet pile wall in the range of
230000-235000 kPa were chosen in order to evaluate the importance factors The method that
was followed in that case is the Shortest Distance and the results are shown in Figure 720
Figure 720 Importance factors α2 of the sheet pile stress level in the elastic domain of soil
As it can be noticed from Figure 720 the results in the elastic domain regarding the influence of
the stochastic parameters are comparable with anchorrsquos analysis The stiffness parameter of the
clay layer (69) is the most dominant variable for the system behaviour whereas the stiffness of
the sand (15) and the Dikeold (4) layers seems also to have a significant influence This is not
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 109
surprising as the nature of the loads that generate high stresses on the anchor as well as the sheet
pile is the same namely the horizontal earth and water forces
Concluding in this case study the sheet pile wall is not expected to show a decisive behaviour to
the systemrsquos reliability according to the reliability analysis and the deterministic analysis (see
section 643) It should be stressed the fact that the fitted RS in a lot of calculations seemed not to
perform suitably for the sheet pile and this can be attributed to the fact that the it was created
based on the SA which was far from the plastic domain In that case a better fit of a RS is
recommended However such an action is still under investigation that for future considerations
in the design concept will be contemplated as a valuable technique for complex structures
modelled in FEM whose reliability analysis with conventional methods is time consuming and the
coupling may lead to numerical problems In the next section the analysis of the systemrsquos
reliability is presented
77 System Failure The estimation of the systemrsquos reliability was based on the systemrsquos fault tree of Figure 53
according to which the general LSF was formulated as the minimum of the three LSF that were
presented above for the soil the sheet pile and the anchor and it is given below
[ ]
[(
)]
(74)
According to this LSF the failure that occurs first is counted as a system failure The analysis was
carried out with DS reliability method and the variables that were taken as stochastic are listed in
Table 72 As it was discussed in section 51 the systemrsquos probability of failure differs from the
probability of flooding for which a combination between the systems reliability and the water
level uncertainty shall be made
In Table 77 the reliability analysis results are presented as they were deduced from DS Similarly
to the soil failure in Figure 721 the importance factors according to the Shortest Distance the 10
Average and the Average all methods are illustrated
Table 77 Reliability analysis results from the system failure
DS Pf 2410-3 β 30 Number of LSF calls 1807 Number of iterations (or directions)
300
Elapsed time (hr) 39
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 110
Figure 721 Importance factors α2 of the system failure
The analysis took almost one and a half days and it came with indicative valuable though results
about the system behaviour According to the averaged importance factors in Figure 721
(deduced by the Average all method) we can notice features from both soil and anchor failure
For example the unit weight (14) the friction angle (7) and the cohesion (8) of the clay
layer as well as the friction angle (8) and the stiffness (7) of the sand layer that played an
essential role in the soil failure analysis Moreover the stiffness of the clay (11) and the Dikeold
(9) layers that appeared to be important for the anchor and sheet pile LSF They seem to be also
essential in the system analysis
However the other two methods (Shortest Distance and Average 10) came up with the cohesion
and the interface strength of the Dikeold material to be the most influencing soil properties for
the system For that reason an investigation was conducted regarding the output results of the
system analysis in order to obtain an insight into the different reasons of the system failure More
precisely the failure points were divided into those that come from Plaxis Error and those that
come from the yield stress exceedance of the structural elements It was then observed that many
failure points originate from Plaxis errors that are related to ldquoNot enough load stepsrdquo (Error codes
102 and 112) rather than ldquoSoil seems to collapserdquo (Error code 101)
Therefore in order to estimate the actual important soil properties the vectors of the random
variables that led to errors of ldquoSoil seems to collapserdquo and to anchorsheet pile (SP) failure were
studied separately The importance factors are depicted in Figure 722 which actually verifies the
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 111
importance of the variables that were also deduced from the Averaged all method in Figure 721
(which are basically the unit weight and the stiffness of the clay layer and the friction angle of the
sand layer) It should be mentioned that the failure of the structural elements contributes 65
whereas the soil failure contributes 35 on the total probability of failure The latter would be
possibly lower than 2410-3 that was initially found as several failure points are not attributed to
real failure but to Plaxis numerical errors
Figure 722 Influence factors α2 for the structural elements (left) and the soil LSF (right) respectively
The system analysis can lead to valuable conclusions for the system behaviour as a whole and
individually for the different elements given though that an inspection of the results is followed
Moreover such an analysis is taking into account correlations among the three sub-failure
mechanisms implicitly For example the different LSFs may be affected by the same soil variable
such as the shear stiffness of the clay layer for the anchor and the sheet pile LSF In that case the
most dominant failure mechanism will occur first and the rest are excluded26 However during
the system analysis someone is not aware of what failed or not and therefore heshe is not
capable of distinguishing the weaknesses of the system that need further improvement Besides
with this technique is not also easy to inspect the failure mechanisms that occur in order to get
some knowledge about the systemrsquos behaviour Therefore the reliability analysis of the
components is recommended to be carried out separately if time is available
78 Comparison between Global Sensitivity Analysis and Local Sensitivity
Analysis (FORM)
The aim of this section is a comparison between the importance factors that were deduced in
section 75 according to a FORM analysis and the results of the global SA that were presented in
section 73 for the anchor failure
The output of the global SA is actually the total indices that show the influence of the random
variables The importance factors of FORM and the total indices of global SA cannot be directly
26 This method is not capable though of accounting for the simultaneous occurrence of different failure mechanisms as the system was considered a serial system
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 112
compared as they are different values More precisely the total index of a random variable is the
ratio between the total variance and the portion of the total variance that stems from the
uncertainty of the specific variable and its interactions with the rest (see Eq A10 in Appendix A)
On the other hand importance factors of FORM express the ratio between the normalized value of
the variable in the design point and the reliability index (see Eq 72) Therefore the comparison
between these values is carried out qualitatively in terms of the variables that are distinguished
as the most influencing
In particular according to global SA the shear stiffness G of the clay layer and the cohesion c of
the Dikeold material seem to be the most determinant soil properties Moreover the strength
reduction parameters Rinter of the Dikeold and the clay layers and the friction angle φ of the
Dikeold and the sand layer appear to have a relatively notable contribution to the anchors limit
state
On the other hand FORM analysis has concluded to the same results as the global SA with the
difference that Dikeold material affects the anchorrsquos stress level more with its shear stiffness
rather than its cohesion In Table 78 the results of both FORM and global SA are presented
separately for each soil layer and property (in blue for the global SA and in black for the FORM
analysis)
Table 78 Important soil properties according to global SA and FORM analysis for the anchor failure (radicradic=very important radic=less important but still considerable)
FORM global SA Sand Clay Dikeold
ν
radicradic
ϕ radicradic
radic
c radic radic radicradic
G radicradicradicradic radicradic
Rinter
radic radicradic
The validation of the most important parameters was carried out with the repetition of the
analysis with DS as it was showed in section 75 The small difference of the probability of failure
as well as the same tendency of the important factors of both DS and FORM verified the reliability
of FORM analysis Therefore the similarity of the results deduced according to global SA with the
ones given by FORM indicates the fact that in this case study the global SA should be considered
as a trustable way of SA for obtaining a first overview of the dominant variables that affect the
limit state under investigation Furthermore it is capable of giving a general insight into the
outputrsquos (in this case anchor stress level) behavior and the soil layers that can potentially stand as
predominant during the reliability analysis For example in order to obtain a first sense of the
reliability and the magnitude of the expected probability of failure someone can inspect the
output samples of the LSF deduced from the global SA In Figure 723 the samples of the anchor
stress level are presented as they were deduced after 3150 iterations during the global SA
(computation time =75 hr = 3 days) The fact that after 3150 iterations and variables
combinations the anchor did not reach its ultimate stress level (ie 450 kPa) implies that the
probability of failure might be small This can also help for qualitative decision making concerning
a systemrsquos behavior without conducting a reliability analysis Moreover global SA proved the high
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 113
interaction (large difference between first and total order indices) among the parameters that can
affect the reliability analysis and can help to decide the amount of the random variables
Figure 723 Anchor stress level during global SA (step 1)
However as someone can observe from Table 78 there are also some discrepancies concerning
the dominance of some variables This can be attributed to the fact that the influence of all
parameters in specific points such as the design point cannot be identified by the global SA
Moreover it is possible that the samples used in the SA could not reach the field where the
system has a plastic behavior and where presumably other variables may also affect the limit
state
As general conclusions for global SA some possibilities and limitations are listed below
Possibilities
bull Indicates the level of interaction effect on the output variance
bull Indicate the amount of the random variables that should be considered in the reliability
analysis
bull Gives a general insight into the output behavior
bull Valuable for future use in the construction of response surfaces
Limitations
bull Cannot rely on global SA for the evaluation of the most important parameters near the
design point
bull Difficult to reach plastic zone (large amount of samples are probably needed)
79 Discussion
In this chapter the reliability analysis results were presented for the soil the anchor the sheet
pile and eventually the system failure taking soil properties of the several soil layers as stochastic
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 114
Initially a SA was carried out for each system component in order to identify the major soil
properties and reduce the number of the stochastic parameters The reliability analysis was
performed either with FORM or DS and in both cases the design point and the importance factors
were estimated Eventually the list of subquestions that is proposed in section 11 was answered
The way they were treated and the corresponding conclusions are elaborated below
Which reliability methods are computationally tractable in combination with FEM
Before starting analyzing the reliability of the system it was essential to evaluate and
assess the performance of the available reliability methods in order to be aware of their
performance in terms of computational time and flexibility in different kind of LSFs After
a literature review and the implementation of some of them in simple LSF problems a
general overview has been obtained for the most applied reliability methods concerning
the previous knowledge required for the system behavior in order to provide the right
input parameters its accuracy and the computational effort Eventually FORM and DS
were chosen to be employed for this case study considering the no preliminary knowledge
of the LSF is required their reliable accuracy and the limited computational effort that
they have in comparison with other methods
The anchorrsquos and sheet pilersquos reliability has been evaluated with FORM method whereas
soil and systemrsquos reliability have been assessed by DS Carrying out a FORM analysis and
steering the algorithmrsquos parameters in order to converge proved to be more difficult than
it was expected However an efficient performance was finally achieved and the failure
probability of the anchor failure was successfully estimated together with the importance
factors (or else the α values) The analysis was repeated with the most important variables
and they were validated with DS The FORM analysis took approximately one and half
hour to converge (with 14 random variables) after providing a starting point that it was
estimated to be close to the design point whereas for the validation it took almost half an
hour On the other hand in case of the sheet pile wall FORM analysis took almost 4 days
The computational time also depends on the LSF to be approximated the number of
variable and the complexity of the problem that each time is analyzed As far as the soil
analysis is concerned DS took approximately one and a half day to converge for 300
iterations and 14 random variables
Which limit states are relevant for the envisaged application of retaining walls in dikes and
how can they be formulated using FEM analysis outcomes
The relevant LSFs of a dike with an anchored sheet pile wall were investigated For each
element ie the sheet pile the anchor and the soil body a separate LSF has been set up
The focus of this research was on the ULS that in case of the structural elements it was
represented by the ultimate yield stress of the anchor and the sheet pile respectively as
they were expressed in Eq 72 and 73 The determination of the soil failure has been also
elaborated and the available alternatives that can be considered in that case were
elaborated After a short evaluation the Plaxis definition of soil failure was selected that it
was formulated as it is shown in Eq 71 Finally the systemrsquos LSF was formed as a
combination of the aforementioned limit states More precisely the system was
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 115
considered as a serial and thus system failure occurs when one of the elements fails first
(see Eq 74)
How robust (convergence) are the tractable methods
The robustness of the different reliability methods can be controlled by the convergence
criteria that are defined by the user and depend on the required accuracy In case of FORM
the convergence criteria consist of a set of errors that has mostly to do with the
approximation of the reliability index and the limit state threshold whereas in DS the
convergence is controlled through the coefficient of variation (CoV)
In order to reassure the robustness of FORM the method has been validated with DS The
validation showed that the FORM method complies with the result of DS with a small
difference though in the design point and the importance factors FORM analysis seemed
to converge sufficiently and faster than DS However the choice of the optimization
algorithm plays a key role to the methodrsquos efficiency In this research the convergence
errors were manipulated for increasing the efficiency of the method More precisely it has
been observed that by relaxing the relative error and the constraint error (the explanation
of these errors and an evaluation of the different optimization algorithms are available in
Appendix A2) FORM convergence can be accelerated
The reliability analysis of the soil body and the system was performed with DS Similarly
to FORM analysis in order to improve the performance of the searching algorithm some
of the methodrsquos parameters should be stipulated for improving its efficiency and
reliability Therefore the maximum step size of the algorithms was tripled and it was
proved that for the same time period double iterations were feasible to be carried out
This does not mean that the larger the step size the better it is The choice of the step size
must be made wisely depending on the reliability index that it is expected and
respectively it is allowed to change for achieving a better performance In this case study
it was proved that after 300 iterations the CoV of DS tends to be high and it can therefore
still be improved by carrying out more iterations In Figure 724 an example of a DS
convergence graph concerning the system analysis is shown The middle red line shows
the probability estimation whereas the green lines show the boundaries At that example
the CoV was 040
DS seems to perform successfully both in soil failure with one LSF and in system failure
where 3 different LSF were included In such cases performing an approximation method
such as FORM would not be wise as the output from each iteration does not provide
information for going to the next that a FORM analysis normally needs In contrast it just
gives an indication of failure or not This means that only a random sampling method can
handle it and converge successfully as it was proved
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 116
Figure 724 Directional Sampling convergence graph (at confidence level 095) of the system LSF
What is the contribution of different uncertainties in the failure mechanisms of the system
According to the global SA and the reliability analysis of FORM and DS the friction angle
the cohesion and the shear stiffness properties appeared to have an influence on the
different LSFs This is not surprising as they all have large CoV of 10 20 and 25
respectively Thus their uncertainty generally showed an impact on the systemrsquos
behaviour It is also essential to mention that the global SA has shown that the uncertainty
of the interface strength Rinter (CoV=20) has also a notable contribution to the structural
elements LSFs However the FORM analysis which is more representative close to the
design point came up with a less important impact of that variable
The reliability analysis showed that as far as the structural elements (anchor and sheet
pile) are concerned the shear stiffness of the clay and the Dikeold material is determinant
and especially for the sheet pile the strength parameters of the soft soils seem also to play
a crucial role In contrast with the anchor where the soil body seems still to behave in an
elastic manner during the sheet pile failure soil appears to reach plasticity in more
locations where the strength properties (ϕ and c) are also influential The most dominant
failure mode appeared to be the anchor failure whereas the soil failure is always preceded
the sheet pile failure The dominance of the anchor failure can be explained from the fact
that the most influent parameter in this LSF is the stiffness that is also the variable with
the highest uncertainty (CoV=25) Therefore the plastic limit of the anchor is more
likely to be reached
Regarding the soil body reliability the analysis detected 3 different failure directions in
which the clay and the sand layer below the dike seem to be the most influent More
specifically the friction angle φ the cohesion c the unit weight γ of the clay layer the
friction angle of the sand layer and the cohesion of the Dikeold material turned out to play
an important role The decrease of the clay layerrsquos unit weight lead to the incapability of
the layer to withstand the overlying load of the dike body and at the same time the
weakening of the sand layer contribute to the creation of an inner sliding surface Even if
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 117
the uncertainty of the unit weight is small (CoV=5) it is still important for the stability
of the system as the maximum deformations appear in that layer (see Figure 712) Last
but not least the illustration of the several failure points in Plaxis allows someone to
inspect the failure mode and realize if it is a real failure of the structure or if it is a
weakness of the modelling
Finally during the system analysis multiple failure directions were detected whereas the
pie charts of the importance factors include soil properties that affect both soil and
structural elements as it was expected However due to Plaxis numerical errors several
failure points were found to be not realistic and they should be excluded in order to come
up with a more reliable failure probability which will be presumably lower The most
suitable method for estimating the importance factors though seems to be the Average all
method which is not distorted from the Plaxis numerical errors However it is
recommended that a further research be done on how realistic are the failure points
coming from Plaxis warnings and to verify if they comply with Figure 711 failure
definition
In Table 79 a synopsis of the most important soil properties for the different LSFs is
presented It is also apparent the common contribution that some of them have in
different LSFs In the last two columns of the table the results of the influencing soil
variables for the system LSF are shown with the Average 10 and Average all method
respectively For the reasons that were explained in the previous paragraph the Average
all method is considered to be the most suitable in that case and as it can be noticed from
the table it gives the most representative picture of the governing soil variables of each
LSF
Table 79 Important soil properties for the different LSFs (radicradic=very important radic=less important but still considerable the most important parameters for each LSF is painted in orange)
Soil material
Soil property
Anchor Sheet pile
Soil body
System Average
10
System Average
all
Clay
G radicradic radicradic
radic radicradic
φ
radic radicradic
c
radic radic radic
Rinter γ
radicradic
radicradic
Sand
G
radic radic
φ
radicradic radicradic
c
Rinter γ
Dikeold
G radicradic
radic
φ
radic
c
radicradic radic radicradic radic
Rinter radic
radicradic radic
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 118
γ
radic radic
Rinter was taken from the validation of FORM with DS Figure 717 (right graph)
taken from the sensitivity of the sheet pile in the soil elastic domain Figure 720
Can response surface techniques help to increase the efficiency and robustness of the
reliability model
In this research RS techniques were used only as a source for providing a reasonable
value to the FORM analysis in order to continue running in case of Plaxis errors These
response surfaces were constructed based on the SA results for both the anchor and the
sheet pile wall Apparently the domain of the SA is limited and most of the times it is
incapable of approaching the design point However RS worked efficiently in many cases
where Plaxis failed to give an output result such as in case of the sheet pile LSF Moreover
the RS could be used individually for the estimation of the failure probability (instead of
Plaxis calculations) but a better fit is advised to be firstly achieved that would be
representative in a wider plastic domain of the structural elements Therefore the RS
technique can be used to enhance the efficiency of the reliability analysis of a complex
system in the sense that it can be considered as solution for FEM errors What is more RS
techniques are a quite promising method for the reliability analysis of a system as it can
also be used alone without coupling with FEM and eliminate the computational effort
However this requires more investigation of such a technique
How can the current design approach be improved
In section 22 the current design concept of the case study that is under investigation in
this research is described At that section the design values of the moments and the forces
on the structural elements are presented as well as the required total safety factor FEM
=18 is given as it was initially calculated by Breedeveld (2011) It has been observed that
a significant overestimation is made in terms of the design values that might later lead to
heavy and cost inefficient structures For that purpose alterations was made on the
original case study in terms of the structural elements and some soil properties and a new
deterministic design was made in order to come up with a less heavy structure This
would also render it possible to investigate the reliability of the system rather than taking
zero probabilities or non-convergence Moreover it would finally help to see if there are
any potentials for optimizing the design by applying a full probabilistic analysis
According to the soil reliability analysis the probability of failure was found 1310-8 and
the reliability index β=55 As far as the anchor failure is concerned the probability of
failure found to be 28410-6 with a reliability index β=45 according to the FORM analysis
The sheet pile failure seem not be significant as it will more likely occur after the soil
failure Eventually the system analysis was performed the probability of failure was
found to be 2410-3 and the reliability index β=30 which is high enough to consider the
system safe
The systemrsquos reliability index that came out of this analysis cannot be directly compared
with the required one (ie β =40) because they stem from different probabilities The
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 119
former refers to the probability of failure due to the global instability given a design water
level whereas the latter refers to the probability of flooding that incorporates also the
water level uncertainty It must be mentioned that in reality there is no sheet pile wall
inside the dike section in the dike ring 43 but this section is under investigation for
reinforcement purposes as it does not comply with the safety standards after a recent
inspection that has been carried out Moreover the original case study taken from
Breedeveld (2011) was subjected to multiple alterations for the needs of the current
research Therefore a comparison of the reliability index with the required one from the
regulations would not be meaningful to make However what it can be compared in that
case is the order of magnitude of the reliability index After the analysis a β of 30 was
estimated for the probability of failure whereas for the overall probability of flooding a β
of 40 is required whose magnitude is the comparable with the former This can firstly be
considered a sort of the probability ldquovalidationrdquo in the sense that the analysis gave
realistic results and secondly it can imply that the structure that has been investigated in
this research is likely to be safe towards macro-instability as its coupling with the water
level uncertainty would certainly give a reliability index above 40 Furthermore a
reliability index of 30 indicates that there is room for optimizing the design concept of
such a system if someone considers also the fact that the soil strength of the subsoil was
further reduced for the needs of the current study
Due to the limitations of the modeling and the alterations of the original case study these
reliability indices should not be taken into account as an indicative number for the safety
of this dike section but rather as an order of magnitude in order to realize what the
potentials for improving the design procedure are This research can also contribute to
the way of designing not only the system but also each structural element as the reliability
of each can be calculated in advance According to the methodology developed in this
thesis a comparison can later on be made for a real case situation between the results of a
full reliability analysis and the current design procedure In that sense a tangible
improvement can be suggested for the design concepts and even to introduce a new way
for the future design of dikes with retaining walls
It should be stressed that the conclusions are based on the results of this case study with its
geometry and set of material parameters and the variation coefficients Changes in the geometry
the material parameters or the statistical properties could lead to different results The
generalization of these conclusions should thus be treated carefully
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 120
MSc Thesis A Rippi 121
8 Conclusions and Recommendations
81 Conclusions The advances in structural reliability analysis have made a large variety of methods available The
applicability and efficiency of these methods depends on the problem that is analyzed and on the
models that are used One the other hand the necessity for optimizing the design procedure of
several structures is becoming more and more intense A solution to this demand can be the
reliability analysis of a system for which rare information about its application on real-world
problems is available This work can contribute to making probabilistic analysis more accessible
for dikes reinforced with a sheet pile wall and at the same time to analyzing such a complex
system in terms of the factors that mainly influence the system behavior In particular the main
research question of this research was
How can the probability of failure due to global instability of a dike with a sheet pile wall modeled
by a Finite Element Model be analyzed
For that purpose in this research a reliability analysis was conducted by making use of a soil-
structure model that was simulated in FEM The output of the FEM calculations was utilized as an
input in the probabilistic model that was used in this thesis in order to assess the reliability of a
dike with an anchored sheet pile wall In that case the load was basically the earth and water
pressure whose impact was investigated on the anchor (LSF 1) the sheet pile (LSF 2) and the soil
body (LSF 3) respectively as well as on the system as a whole
It was demonstrated that the calculation effort of this kind of analysis does not necessarily have to
be extremely high Especially when previous knowledge about the system behavior and the
reliability models used is available This can be achieved by carrying out a sensitivity analysis In
this thesis a global SA was conducted that was proved capable of giving a representative
estimation of the most influencing variables and give a general overview of the systemrsquos
performance However close to the design point local SA such as FORM becomes more suitable
As far as the reliability of the structural elements is concerned their limit state was evaluated in
terms of the exceedance of their yielding stress The reliability assessment was carried out with
FORM from which the probability of failure and the influence factors were concluded More
precisely the anchor failure mode seemed to be the most determinant one with the shear
stiffness of the soft soils to be the most important soil properties The sheet pile failure is quite
rare and it is unlikely to occur before the soil failure since the soil body has already entered its
plastic domain
The soil and the system failure were challenging to assess as there are multiple criteria that
someone can choose In this thesis regarding the soil failure the Plaxis definition of soil collapse
was used and it was evaluated with DS In that case the unit weight and the strength parameters
of the soft soil layers seemed to be influencing The detected failure directions were investigated
Conclusions amp Recommendations
MSc Thesis A Rippi 122
in terms of their validity More precisely the failure points were illustrated in Plaxis in order to
visualize the failure and see if it is real failure or not (real if dike crest settles) That procedure
revealed the possibilities and limitations of FEM modeling The advantage is that a potential
failure can be detected and its visualization is possible On the other hand one should be aware of
the limitations related with the constitutive model that is used In this study the Mohr Coulomb
method was applied that although it can efficiently identify possible failure mechanisms the
representation of the deformations is not so realistic
The failure of the system was considered as a serial system of LSF 1 2 and 3 and it was also
assessed with DS The importance factors stemmed from the reliability analysis indicated soil
features that were important in all the LSFs However it should be stressed that the frequency of
the Plaxis numerical errors was such that quite a few failure points seem to be unrealistic In this
thesis the identification of those errors and their partially exclusion from the analysis was
achieved The latter was attained by the construction of the so-called Response Surfaces (RS)
based on a preliminary sensitivity analysis Such a technique found to be valuable in order to
overcome such a problem however a study for fitting more suitable RS in each problem is
recommended Nevertheless the further elimination of those numerical errors is strongly
recommended to be investigated in future research given the results from the current study
Retrieving the importance factors from a DS analysis was also a challenging part In this study
additional methods were developed in order to evaluate the importance factors and define the
design point after a DS analysis It revealed that alternative possible approaches of the
importance factors (Average 10 and Average all methods) can presumably provide a better
understanding of the systemrsquos behaviour Such methods seemed to be valuable for excluding the
ldquonoiserdquo of the possible fake failure points that originated from FEM numerical errors However
due to the inherent randomness of DS in terms of the selection of the directions the design point
shall be inspected and compared with the other failure points that were found Then someone can
contemplate if it is necessary to proceed with an averaging of the importance factors in order to
get a clearer view the variables influence Of course such a procedure takes time that is not always
available but according to the requirements of each research the most preferable assumptions
should be made
Last but not least the reliability analysis came up with relatively low probabilities of failure which
implies that there is still room for optimizing the design concept of a dike with retaining wall
More precisely the original case study of Breedeveld (2011) has been altered by reducing the
strength properties of the subsoil and installing a less heavy sheet pile wall after following a less
strict deterministic design It was then observed that even then the reliability analysis came up
with a respectively high reliability index It should be stressed that the failure probability of the
system should be also coupled with the uncertainty of the water level in order to obtain an overall
picture of the total probability of failure Nevertheless this research proves the potentials to
optimize the design of such a system which presumably will lead to a less expensive construction
However additional failure mechanisms such as piping internal erosion etc should be also
investigated and the design properties of the system (such as the length and cross section area of
the sheet pile wall) shall be based on them
Conclusions amp Recommendations
MSc Thesis A Rippi 123
82 Recommendations The following topics are recommended for further research
Additional research is required on FEM modeling both for the structural elements and the
soil body More in particular in this research it has been proved that Mohr Coulomb
model cannot realistically estimate the deformations in a soft soil In that case there are
other more advanced models such Hardening Soil model and Soft Soil model that are
capable of simulating the soil displacements and the stress-strain relationship closer to
the actual ones However these methods require more parameters to be specified and
their coupling with reliability packages might not be so robust and functional
As far as the structural elements modeling is concerned there are two major topics that
future research should focus on Firstly the anchor rod can be simulated in Plaxis as a
node-to-node element and at the bottom an embedded beam can be installed in order to
represent the anchorage body This would provide a better reaction with the surrounding
soil body and it could probably simulate the anchor-soil interaction more realistically
Another alternative would be to simulate it as a plate element in order to be able to test
the moments that are also developed on the anchor separately Secondly it is
recommended that the elements be investigated close to their plastic area and a reliability
analysis to be carried out with plastic elements instead In this research an attempt was
made for such an analysis but it appeared not to be functional in cooperation with FORM
analysis
Another issue that is recommended to be investigated is the inspection and the validation
of the failure that Plaxis is giving During the soil failure analysis different types of failure
were observed from which some might not be realistic but due to the modeling limitations
Therefore an attempt should be made in order to further eliminate the effect of Plaxis
numerical errors on the reliability analysis that can lead to fake failure modes A check of
the failure points can be also conducted by using a more suitable constitutive model for
the deformations rather than Mohr Coulomb However this is not always feasible for each
individual case Therefore the illustration of the failure mechanisms in large scale
experiments can give a better insight into the systemrsquos behavior and valuable knowledge
can be acquired of the expected failure modes In this way also 3D effects can be
investigated see what is their effect to the dike instability and at what extend they shall be
considered
Apart from the soil body the stresses and the displacements of the structural elements
shall be also verified It is thus recommended that field tests and measurements take place
in embedded elements like sheet piles and anchor in order to validate Plaxis results and
learn what is the most suitable manner for the their simulation (fixed-end-anchor node-
to-node element plate etc)
Concerning the reliability analysis of the specific case study a research on the potential
relevant LSFs is recommended for the different elements Specifically in this thesis the
LSF were chosen in order to represent the ULS of the element There is the possibility that
the SLS is used and certain acceptable deformations can be settled as a threshold for the
Conclusions amp Recommendations
MSc Thesis A Rippi 124
limit states However the choice of this displacement limit is not straightforward and a
qualitative research shall also be conducted in order to identify the impacts of different
values Besides it is not only the value that should be selected but also the location on the
structure that is going to be checked
It is also advised that the potential correlations among the soil parameters to be included
In Wolters (2012) and Bach (2014) correlation matrices are introduced for specific soil
properties that can be included in a future research and reveal what is their effect on the
reliability analysis The influence of correlations depends on what kind of variables are
correlated and what is the influence of those variables on the system behavior However a
preliminary analysis with independent variables is advisable to be carried out at the
beginning so as for someone to be able to analyze the effect of the variables individually
and then evaluate the correlation effect more wisely
It would be also advisable a comparison to be made between the design of a real case
according to a full probabilistic analysis and partial safety factors In that way a direct
insight into the advantages and disadvantages of each method can be gained and
fundamental improvements of the design concept can be made for dikes with retaining
walls
At this moment the most reliable way of conducting a reliability analysis is with sampling
methods like MC or DS However such methods are usually unattractive due to their large
computational time Thus the tolerance of sampling methods accuracy is lowered in order
also to reduce the number of iterations or approximating methods are used like FORM
analysis However even if approximating methods are used they are usually preferred to
be validated with sampling methods in order for their results to be acceptable A real
validation of the failure probability would mean that a structure should be experimentally
tested in different positions and under the same conditions in order to see if the
probability of failure that is calculated in each case is the same Therefore a real
validation of the failure probability cannot yet be initiated but the computational effort of
the sampling methods can be eliminated with the development of technology and the
improvement of the computer science
MSc Thesis A Rippi 125
References
Abdo T amp Rackwitz R (1990) A new beta-point algorithm for large time-invariant and time
variant reliability problems Proceedings of the 3rd IFIP WG 75 Conference Berkeley pp1-12
California USA
Ang AHS and Tang W (1984) Probability Concepts in Engineering Planning and Design Volume
IBasic Principles John Wiley and Sons New York USA
Baecher GB and Christian JT (2003) Reliability and Statistics in Geotechnical Engineering
Chichester West Sussex John Wiley amp Sons Ltd
Behaviour of Soils and Rocks GEO-ENGINEERING - CIE4361 (Vol 2015) Retrieved from
httpsblackboardtudelftnlbbcswebdavpid-2386735-dt-content-rid-
7926659_2courses32281-14150221-possibilities26limitationspdf
Breedeveld J (2011) Technisch Rapport Analyse Macrostabiliteit Dijken met de Eindige Elementen
Methode Bijlage Voorbeeldcase Deltares Delft the Netherlands
Brinkgreve RBJ (2015) Personal discussion TU Delft the Netherlands
Brinkgreve RBJ and Bakker HL (1991) Non-linear finite element analysis of safety factors In
G Beer JR Booker and J P Carter (Eds) Proceedings of the Seventh International Conference on
computer methods and advances in Geomechanics Rotterdam A A Balkema Brookfield
Cukier RI Fortuin M Shuler KE Petschek AG and Schaibly JH (1973) A study of the
sensitivity of coupled reaction systems to uncertainties in rate coefficients I Theory California
University USA
Cundall PA (2002) A discontinuous future for numerical modeling in soil and rock Published in
Proceedings of the third international conference Discrete Element Methods Numerical Modeling
of Discontinua Edited by Cook and Jensen Santa Fe pp 3-4
CUR (2005) Handbook Quay Walls Gouda the Netherlands
CUR-publication 190 (1997) Probabilities in civil engineering Part 1 Probabilistic design in theory
Stichting CUR Gouda
CUR 166 ndash Damwandconstructies (2005) Technical Recommendation 4e herziene uitgave 2005
The Netherlands
Deltaprogramma (2014) Synthesedocument Veiligheid Achtergronddocument B1 Ministrie van
Infrastructuur en Milieu
EN 1990 2002 Basis of Structural Design European Committee for Standardization
References
MSc Thesis A Rippi 126
Farrokh N (2007) Tools and Strategies for Dealing with Uncertainty in Geotechnics In DV
Griffiths and GA Fenton (Ed) Probabilistic methods in Geotechnical Engineering (pp 71-95) New
York NY Springer Vienna
Fenton GA and Vanmarcke EH (1990) Simulation of random fields via local average
subdivision J Eng Mech ASCE 116(8)1733ndash1749
Flood Defences HYDRAULIC ENGINEERING ndash CIE5314 (Vol 2015) Retrieved from
httpsblackboardtudelftnlbbcswebdavpid-2125084-dt-content-rid-
7621002_2courses29759-
131404CIE5314_FloodDefences_05_failure_mechanisms_stability_Schweckendiekpdf
Gemeentewerken-Rotterdam (2003) Alle triaxiaal proeven rotterdam Technical report Gemeentewerken Rotterdam Rotterdam the Netherlands
Ghanem R and Spanos P (1991) Stochastic Finite Elements ndashA Spectral Approach
Springer New York
Gonzaacutelez A Cuadrado A Josa A and Olivella S (2013) Safety assessment of limit equilibrium
methods for the design of cantilever and anchored sheet pile walls Escola de Camins Barcelona
Griffiths DV and Fenton GA (Eds) (2007) Probabilistic Methods in Geotechnical Engineering Lecture notes monographs edited works and proceedings in the field of Mechanics Engineering Computer Science and Applied Mathematics Italy Hasofer AM and Lind NC (1974) An exact and invariant first-order reliability format Journal of
Engineering Mechanics Division ASCE 100(1) 111-121
Hoek E Read J Karzulovic A and Chen ZY (2000) Rock slopes in civil and mining engineering
Published in Proceedings of the International Conference on Geotechnical and Geological
Engineering GeoEng2000 19-24 November Melbourne
Homma T and Saltelli A (1996) Importance measures in in global sensitivity analysis of
nonlinear models Reliab Eng Syst Saf 52(1) pp1-17
Hydra-Ring 12 (2015) Probabilistic toolbox for the WTI2017 Technical Reference Manual Version
12 Deltares Delft
Jaky J (1944) ldquoThe coefficient of earth pressure at rest In Hungarian (A nyugalmi nyomas tenyezoje)rdquo J Soc Hung Eng Arch (Magyar Mernok es Epitesz-Egylet Kozlonye) 355ndash358
Johansson E amp Sandeman E (2014) Modelling of Deep Excavation in Soft Clay A comparison of different calculation methods to in-situ measurements Chalmers University of Technology Gothenburg Sweden
Joint Committee on Structural Safety (1981) General principles on reliability for structural design
International Association for Bridge and Structural Engineering
References
MSc Thesis A Rippi 127
Larsen H Lubking P and Breedeveld J (2013) Ontwerp stabiliteitsschermen (type II) in primaire
waterkeringen (groene versie) Deltares Delft
Melchers R E (2002) Safety and risk in structural engineering Prog Struct Engng Mater
4 193ndash202
Moumlllman A amp Vermeer P (2008) Reliability analysis of a dike failure case study Elbe river
University of Stuttgart Germany
Nelson RB (1999) Lecture Notes in Statistics An Introduction to Copulas Springer New York
OpenTURNS 15 (2015a) Reference Guide OpenTURNSrsquo methods for Step C uncertainty
propagation pp155-161
OpenTURNS 15 (2015b) Use Cases Guide for the Textual User Interface pp190-192
Pandya N Marsdea E Floquet P and Gabas N (2008) Sensitivity analysis of a model for
atmospheric dispersion of toxic gases In B Braunschweig and X Joulia (Ed) 18th European
Symposium on Computer Aided Process Engineering-ESCAPE18 (pp1143-1148) Elsevier BV
Phoon K (Ed) (2008) Reliability-based Design in Geotechnical Engineering Computations and Applications Taylor amp Francis Britain
Plaxis 2D (2015a) Tutorial manual Delft the Netherlands
Plaxis 2D (2015b) Reference Manual The Netherlands
Plaxis (2015) Material Models Manual The Netherlands
Powell MJD (1994) A direct search optimization method that models the objective and
constraint functions by linear interpolation In S Gomez amp JP Hennart (Ed) Advances in
Optimization and Numerical Analysis (pp51-67) Dordrecht Kluwer Academic
Rankine W (1857) On the stability of loose earth Philosophical Transactions of the Royal Society
of London Vol 147
Saltelli A (2004) What is Sensitivity Analysis In A Saltelli K Chan and EM Scott (Ed)
Sensitivity Analysis (pp 3-13) John Wiley amp Sons Publication
Saltelli A S Tarantola F Campolongo and M Ratto (2004) Sensitivity Analysis in Practice
A Guide to Assessing Scientific Models John Wiley amp Sons Ltd
Schittkowski K (1985) NLPQL A Fortran subroutine solving constrained nonlinear
programming problems Annals of Operations Research 5 485-500
Schweckendiek T (2006) Structural Reliability applied to Deep Excavations Coupling Reliability
Methods with Finite Elements TU Delft Delft
References
MSc Thesis A Rippi 128
Schweckendiek T Vrouwenvelder ACWM Calle EOF Jongejan RB and Kanning W (2013)
Partial Factors for Flood Defenses in the Netherlands Modern Geotechnical Codes of Practice ndash
Development and Calibration Fenton GA et al (eds) Special Geotechnical Publication Taylor amp
Francis
Seed RB Bea R G Abdelmalak R I Athanasopoulos A G Boutwell G P Bray J D Briaud J-L Cheung C Cobos-Roa D Cohen-Waeber J Collins B D Ehrensing L Farber D Hanemann M Harder L F Inkabi K S Kammerer A M Karadeniz D Kayen RE Moss R E S Nicks J Nimmala S Pestana J M Porter J Rhee K Riemer M F Roberts K Rogers J D Storesund R Govindasamy A V Vera-Grunauer X Wartman J E Watkins C M Wenk Jr E and Yim SC (2006) Investigation of the Performance of the New Orleans Flood Protection Systems in Hurricane Katrina on August 29 2005 Volume I Main Text and Executive Summary New Orleans USA
Steenbergen RDJM (2013) Lecture notes Probabilistic Design Delft the Netherlands TU Delft
TAW (Technische Adviescommissie voor de Waterkeringen) (2004)Technisch Rapport Waterspanningen bij dijken The Netherlands
Terzaghi K (1942) Theoretical Soil Mechanics New York Wiley ISBN 978-0-471-85305-3 Teunissen JAM (2009) SBW Analyse macrostabiliteit van dijken met Eindige Elementen
Modellen Achtergronden bij Activiteit EEM 04a Opstellen stappenplan Deltares the Netherlands
Terzaghi K (1943) Theoretical Soil Mechanics John Wiley and Sons New York
Turaacutengi T amp Rabitz H (2004) Local Methods In A Saltelli K Chan and EM Scott (Ed)
Sensitivity Analysis (pp 81-99) John Wiley amp Sons Publication
Vergouwe R Huting RJM and van der Scheer P (2014) Veiligheid Nederland in Kaart 2 Overstromingsrisico dijkring 43 Betuwe Tieler- en Culemborgerwaarden The Netherlands
Vrijling JK Schweckendiek T and Kanning W (2011) Safety Standards of Flood Defenses Keynote at ISGSR 2011 (International Symposium on Geotechnical Safety and Risk) Munich Vogt Schuppener Straub amp Braumlu (eds)
Vrijling JK and Gelder PHAJM van (2002) Probabilistic Design in Hydraulic Engineering Lecture notes TUDelft Faculty of Civil Engineering and Geosciences Delft Vrouwenvelder ACWM Steenbergen HMGM and Slijkhuis KAH (1999) Theoriehandleiding PC-Ring Delft the Netherlands TNO
Waarts PH (2000) Structural reliability using Finite Element Analysis An appraisal of DARS
(Directional Adaptive Response surface Sampling) Delft University of Technology Delft the
Netherlands
Wyllie DC amp May CW (2004) Rock Slope Engineering (4th edition) Spon Press London
MSc Thesis A Rippi 129
Appendix A OpenTURNS features
In this Appendix an overview of the most relevant of OT is given Firstly in section 112 an
introduction of the applicable methods in SA is made while in section 52 the results from such an
analysis are presented In Appendix A1 a more detailed elaboration is made of how the FAST
sensitivity method that was applied in this project works
In Appendix A2 an explanation of the different optimization algorithms available for FORM
analysis is given The theory behind these algorithms and their convergence criteria are discussed
while finally an evaluation of their performance after applying them in case of the anchor failure
is presented This evaluation has been conducted in order to decide about the most relatively
reliable algorithm that should be adopted for the reliability analysis
Finally in Appendix A3 the types of probability distributions that have been used in this study
for the random variables and which are available in OT are displayed
A1 Fourier Amplitude Sensitivity Test (FAST) In this section the extended FAST sensitivity method is discussed as it is also described in the
Reference Guide of OpenTURNS 15 (2015) FAST is based on the Fourier decomposition of the
variance of the model response ( ) the latter being represented by its Fourier expansion
is an input random vector of independent components Its key idea is to
recast this representation as a function of a scalar parameter by defining
exploring the support of the input random vector
For each input the same procedure is realized in three steps
1) Sampling
Deterministic space-filling paths with random starting points are defined ie each input Xi is
transformed as follows
( ( )) (A1)
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 130
Figure A1 Search curves in the input space as they are defined from the transformation (the blue
dot indicates the direction of the current search path)
where is the number of input variables N is the length of the discretization of the s-space with
s varying in (-π π) by step 2πN is a random phase-shift chosen uniformly in [0 2π]
is a set of integer frequencies assigned to each input Xi The frequency
associated with the input of interest is set to the maximum admissible frequency satisfying the
Nyquist criterion (which ensures to avoid aliasing effects)
(A2)
with M the interference factor usually equal to 4 or higher It corresponds to the truncation level
of the Fourier series ie the number of harmonics that are retained in the decomposition realised
in the third step of the procedure And the maximum frequency of the complementary set of
frequencies is
(A3)
with the index ldquo-irdquo which meaning ldquoall but irdquo
2) Simulations
Output is computed such as
Then is expanded onto a Fourier series
sum [ ] (A4)
where and are Fourier coefficients defined as follows
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 131
int
(A5)
int
(A6)
These coefficients are estimated thanks to the following discrete formulation
sum
(A7)
sum
(A8)
3) Estimations by frequency analysis
The quantitative measure of sensitivity is represented by Sensitivity Indices The first-order
sensitivity index of an input pi is the measure of the main (direct) effect of pi on the output
variance (where i ne j) the second-order sensitivity indices measures the interaction effect of
pi and pj on the output variance Other higher-order indices are defined in the same manner The
total sensitivity index is the sum of all sensitivity indices involving factor pi (Homma amp Saltelli
1996) For example the total sensitivity index of factor 1 for a model with 3 input factors is
given as
The first-order (Si) and total ( ) indices can be interpreted as following (Pandya et al 2008)
high pi is an influent parameter
and both small pi is not an influent parameter (neither alone nor in interaction with
other parameters)
and nearly the same no interaction of pi with the other parameters
and very different high interaction of pi with the other parameters
Total indices are especially suited to apportion the model output variation to the input factors in a
comprehensive manner The FAST method calculates the first-order and the total sensitivity
indices whereas Sobolrsquo method in addition to these also provides all higher-order sensitivity
indices to determine quantitatively the interaction between parameters However the
computational cost and calculation time of Sobolrsquos method tends to be higher than that of the
FAST method
Combining equations A1-A8 the first-order indices are estimated as follows
sum (
)
sum (
)
(A9)
where is the total variance the portion of D arising from the uncertainty of the ith input and N
the size of the sample using to compute the Fourier series
Subsequently the total-order indices are estimated as follows
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 132
sum (
)
sum (
)
(A10)
where is the part of the variance due to all inputs except the ith input
A2 Optimization Algorithms in FORM
Principles of optimization algorithms
The possible optimization algorithms in OT are
Abdo-Rackwitz (ARF)
Cobyla
Sequential Quadratic Programming (SQP)
Below a description of the main principles of the optimization methods mentioned above is
quoted ARF and SQP methods are elaborated according to Abdo and Rackwitz (1990) while
Cobylarsquos principles according to Powell (1994) Finally an evaluation of the performance of the
different methods follows after applying them in case of the anchor failure
The general optimization problem (objective function) can be written in the classical form
subject to inequality constraints (constraint function) (A11)
The Lagrangian function of the general problem is defined by
sum (A12)
where are the so-called Lagrangian multipliers The SQP method (developed by Schittkowski
(1985)) replaces the original problem by a sequence of quadratic programming problems which
are exactly solvable and which approximate the original one This is done by approximating the
Lagrangian function by its second order Taylor expansion in an initial point u0
(A13)
where
sum
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 133
sum
sum
in which is the gradient operator and represents the Hessian matrix27 of the function f in
the point u0 The optimality conditions for any iteration point k of the sequence of quadratic
expansions are
sum [ sum
] (A14)
(A15)
The exact calculation of the Hessian matrix is generally too expensive and cannot be efficiently
implemented for a general case Therefore the gradient information obtained in each point
during iteration is used to build up an approximation of this matrix using one of the known
update formulas
The new iteration point is defined by
(A16)
where is the step length and is a direction in which a line search is performed The process
stops when the optimality conditions of the original problem are satisfied
The most time consuming part in this algorithm is the updating of the Hessian matrix and the
solution of the system of equations A fair approximation of the Hessian of non-quadratic
functions is also obtained with about n updates of the matrix This means that the approximation
used in the few (say ten) iterations to reach convergence cannot be very good when the problem
has large number of variables The rounding errors during the updating process in large problems
can make the approximate Hessian to become singular Close to singularity the search direction
can be significantly distorted In this case the algorithm has to restart the iteration with a unit
Hessian matrix in the point where singularity occurred
The posterior algorithm ARF developed by Abdo and Rackwitz (1990) overcomes this problem
by obtaining a constant approximation of the true Hessian matrix Only the contribution of the
objective function to the Hessian is considered The numerical decomposition of the matrix
contains the scalar product of the gradients of the constraints as elements in each iteration
27 Hessian is a square matrix of second-order partial derivatives of a scalar-valued function or scalar field It describes the local curvature of a function of many variables
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 134
Cobyla algorithm has been developed by Powell (1994) and it constructs linear polynomial
approximation to the objective and constraint functions by interpolation at the vertices of
simplices (a simplex in n dimensions is the convex hull of n+1 points n being the number of
variables) It generates the next vector of variables from function values at the vertices
of a nondegenerate simplex in In this case there are unique linear
functions and say that interpolate f and at the vertices
and the optimization problem A15 by the linear programming problem
( ) (A18)
( )
The iterative use of expression A18 puts this method in the class of ldquosequential linear
programming algorithmsrdquo It is also essential to notice that in this method the gradients are
derived by interpolation at the vertices of simplices and not analytically as it happens in SQP and
ARF
Additionally Cobyla algorithm makes use of a trust region bound ρ so that the trust region
condition on the new vector of variables is
(A19)
Such a region gives the user some control over the steps that are taken automatically and which
respond satisfactorily to the fact that there may be no finite solution to the linear programming
problem The algorithm also employs a merit function of the form
( ) ( ) [ ( ) ] (A20)
in order to compare the goodness of two different vectors of variables Here is a parameter that
is adjusted automatically depending on the how close to optimizing the objective function the
analysis is and the subscript ldquo+rdquo means that the expression in square brackets is replaced by zero
if and only if its value is negative so we have ( ) ( ) whenever is feasible Parameters
and are changing ( only reduces) automatically according to the improvement of the
optimization problem that sequential vectors might cause Therefore if the change in the merit
function does not happen to improve the optimization then these values are changing Especially
regarding the parameter user gives the initial and the final values of it namely and (it
is recommended that be a reasonable change to make the variables for a coarse exploration
of the calculation while should be approximately the required distance from the final vector
of variables to the solution of the optimization problem)
Convergence criteria
The convergence of the aforementioned algorithms is controlled by the evaluation of the
following errors expressed in the standard space (for the relationship between the real space (x-
space) and the standard space (u-space) see section 422) (OpenTURNS 15 (b) 2015 pp 190-
192)
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 135
The absolute error which is the distance between two successive iterates
(A21)
The constraint error which is the absolute value of the limit state function minus the
threshold
(A22)
The relative error which is the relative distance between two successive iterates (with
regards to the second iterate)
(A23)
The residual error which is the orthogonality error (lack of orthogonality between the
vector linking the center and the iterate and the limit state function)
( ) (A24)
The algorithm converges if all the final error values are less than the maximum value specified by
the user The convergence can be also controlled by altering the maximum number of iterations
which should be higher in case of stricter required errors What is more there is the possibility to
manipulate the starting point that the algorithm uses for its first iteration Usually the default
starting point in a FORM analysis consists of the mean values of the random variables However if
there is already some knowledge about the potential design point it can be set as the starting
point of the algorithm in order to accelerate the convergence and save a lot of time
The aforementioned errors the maximum number of iterations and the starting point were
manipulated during the analysis in order to understand the functionality of the different
algorithms and to improve their efficiency
Evaluation of the algorithms performance
An evaluation of the above optimization algorithms has been made in order to make a decision of
the most suitable for the current project In principle a literature overview regarding the
performance of the algorithms is presented Then the 3 algorithms were tested considering the
anchor failure limit state function (see Eq 55) with a stress limit of 450000 Pa and their results
are discussed below
According to Abdo and Rackwitz (1990) the SQP method seems to have better convergence
behavior for highly curved constraint functions which is especially true when very expensive
structural state functions are involved such as finite element structural analyses However the
mentioned algorithm fails to reach convergence at a problem dimension (amount of random
variables) of around n=50 Therefore for increased dimensionality SQP is less efficient in terms
of storage and CPU time compared with other methods due to singularity of the updated Hessians
matrices In OT a warning is introduced since an analysis has started that a default
implementation of Hessian is being used and that the computation can be severely wrong
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 136
Abdo and Rackwitz (1990) also proved that ARF algorithm can handle problems of up to 2000
variables while the storage requirements and the CPU time are much smaller than with SQP
Therefore optimization problems of high dimensionality can be evaluated
In OT a warning appears at the beginning of a FORM calculation saying that a default
implementation of the gradient is being used and that the computation can be wrong A similar
message also appears for the SQP algorithm saying that a default implementation of the Hessian
matrix is used and that the computationrsquos results might be unreliable Moreover in both
algorithms it is advisable to check the values of the random variables that were used during the
analysis in order to ensure that they fluctuate within acceptable boundaries (Figure A4 is an
example of a variable that surpassed the minimum boundary at some stage)
Firstly the Cobyla algorithm was applied on the case study in combination with a FORM analysis
Cobyla does not require a gradient evaluation of the limit state function and does not use a default
implementation of the Hessians matrix as ARF and SQP does while the computational time is
quite small comparable to the aforementioned methods
The Cobyla algorithm was tested with (i) 3 (ii) 10 and (iii) 19 parameters respectively in order to
check its efficiency28 As it has been mentioned in the previous section the convergence criteria to
be steered are the errors (absolute relative residual and constraint) the maximum number of
iterations and the starting point At first the default errors were used with a value of 10-3 but it
was realised that such an order of errors lead to non-convergence of the algorithm Therefore the
algorithm was then checked after relaxing the maximum errors to values of 01 05 10 50 and
100 Finally the combination of the errors that led to convergence was 01 05 01 02 for the
maximum absolute relative residual and constraint error respectively After the analysis the
number of LSF calls in case (i) (ii) and (iii) was 25 even if the maximum one was set to 100 and
the probability of failure was around 035 The probability of failure was quite high and that is
why an inspection of the output file was carried out In Figures A2 and A3 the values of the
anchor limit state function are presented in the right graph during the iterations It was then
observed that the threshold of 450 kPa was never surpassed during the iterations
28 All of the algorithms were first checked with very simple LSF (linear and quadratic) and their results were verified by MC and DS
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 137
Figure A2 Friction angle and cohesion progress during a FORM analysis with Cobyla and with 3 random variables The figure at the right illustrates the limit state function evolution
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 138
Figure A3 Friction angle and cohesion progress during a FORM analysis with Cobyla and with 19 input parameters The last figure at the right illustrates the limit state function evolution
An additional action was then taken by relaxing the maximum number of iterations from 100 to
200 but no improvement was noticed (the LSF calls were again 25) It was also given extreme
values to the errors and especially to the absolute constraint and relative error in order to check
the response of the algorithm but that did not lead to better results Furthermore the
parameter (see theory part for explanation) was manipulated but even then no further
improvement was noticed in terms of the results In Figures A2 and A3 the performance of the
algorithm during the analysis can be seen through the values of the variables that were assigned
during the iterations
As someone can notice from the graphs above it is obvious that Cobyla does not perform very
well in high dimensionality problems as the variables do not seem to vary a lot More specifically
in Figure A3 with the 19 parameters it can be noticed that the random variable does not change
at all apart from two peaks at the beginning and at the end of the analysis respectively The
incapability of the algorithm to change the parameters in order to reach the limit of the anchor
stress ie 450 kPa (while here a maximum of approximately 267 kPa is reached) leads to a poor
approximation of both the design point and the probability of failure In Figure A2 with only 3
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 139
parameters the variability of the parameter is higher during the analysis which leads to a better
approach of the limit (around 370 kPa which is larger than with the 19 parmaters) but even then
a failure probability of 035 was not a realistic value On the other hand Cobyla is easy to use only
for small number of variables (ie less than 9) with linear LSF otherwise the linear
approximations can be highly inefficient Apparently the limit state function of the anchor stress
level as a function of the soil parameters is much more complex in order to be approximated by
this algorithm
Therefore the other available optimization algorithms ie ARF and SQP were examined as well
Because of the similarity of the two algorithms only the ARF performance is presented below and
a link with SQP is made when it is needed
The ARF algorithm was tested initially with 19 parameters The settings for the errors that were
used were 01 for the absolute relative residual and constraint error respectively and the
maximum number of iterations was set to 100
Figure A4 Friction angle and cohesion progress during a FORM analysis with ARF algorithm and with 19 input parameters The figure at the right illustrates the limit state function evolution while the red spots indicate the points where the 450 kPa threshold was surpassed
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 140
The algorithm worked efficiently which means that the algorithm managed to converge towards
the threshold of 450 kPa of the limit state function as it is depicted in Figure A4 (see red dots) It
was also noticed that the random variables fluctuated a lot during the iteration steps even if the
amount was significant (19 parameters) The capability of the algorithm to vary all the 19
variables evenly and in a wide range makes it possible to find the real design point and thus to
come up with a more reliable failure probability The resulted probability of failure was 3810-9
and the reliability index was ϐ=58
However the analysis took almost 4 days to finish which is a long duration and the number of LSF
calls counted at 4151 Moreover even if the results were promising after tracing all the 19
parameters in order to inspect the algorithmrsquos performance it was found out that some variables
took unacceptable values29 during the analysis As an example in Figure A5 Poissonrsquos ratio
received negative values (see red circle) at the beginning of the iterations There were also other
such examples in the same analysis such as the friction angle which took sky-high values and the
interface strength which took negative values as well as values above 1 which is unacceptable It
should be mentioned that the design point of the algorithm was completely reasonable but even
then such a performance can prevent the algorithm from converging to the right design point
quickly and increase the computational effort
Figure A5 Poisson ratio progress during the FORM analysis with ARF algorithm
Taking into account the above implications there are two main concerns at that stage (a) the
high computational time and (b) the unreasonable values that the variables took Firstly in order
to reduce the computational time the maximum errors were relaxed More precisely it was
noticed that after relaxing the maximum relative error a lot which is more related to the
29 In some cases a value of ldquoinfrdquo was given to the variables that led to termination of the analysis without results
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 141
relaxation factor (in this case an extreme value of 1000030 was tried) and increasing the absolute
error to 05 the number of LSF calls was reduced to 724 and the analysis was finished after
almost 10 hours In Figure A6 the values that the cohesion and the friction angle took during the
iterative procedure while the red spots in the right figure indicate the points that are included in
the failure space Figure A4 shows the progress of friction angle after relaxing the errors
Figure A6 Progress of the friction angle during the FORM analysis with ARF algorithm with errors
05 10000 01 01 for the absolute relative residual and constraint error respectively
As a general conclusion after the test of different combination of maximum errors by relaxing the
constraint absolute and relative error too much successive iterations of a certain variablersquos value
seem to have a wide difference which can detain the algorithm from converging Therefore the
wisest in this case study is to keep the maximum errors below or around 10 except for the
relative and the constraint error that can relax more than 1 (depending on the case) in order to
reduce the computational time to the minimum possible
To prevent having unacceptable values of the parameters and in order to increase the efficiency of
the algorithm the random variables were reduced to 11 after a SA and some of them were
truncated in a certain range according to observations and experience until now However the
algorithm then started to give an error due to zero gradient of the objective function at a specific
point (similarly SQP gave an error due to zero hessian matrix at a specific point) while the upper
limit of the LSFs could not be reached As a countermeasure the errors were relaxed (05 10 05
10 for the absolute relative residual and constraint error respectively) in order for the algorithm
not to be trapped in a specific point while also the step of the algorithm was manipulated It
was finally proved that this kind of error stemmed from the fact that the structural elements were
considered as elastoplastic and thus a limit of the stress level was considered that misled the
algorithm The elements were eventually considered as elastic and the numerical problem was
solved
30 Practically such a value has the meaning that there is no limitation of the relative error during the optimization In the standard space (where these errors are expressed) a value also greater than 10 can be also quite high However here extreme values were given due to the time limit to test multiple values
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 142
Before figuring out that the elastoplastic behaviour was giving the problem Cobyla as a gradient
free algorithm was tested again as a potential solution but with changing the starting point close
to the design one that was found with ARF It is worthwhile to take a look into the performance of
the algorithm with 19 random variables in Figure A7 In that way we boost Cobyla to converge
and it eventually performed quite differently than in Figure A3 and without giving convergence
problems The computational time was estimated to be 9 hours and the probability of failure was
very small in the order of 10-120 However the validation of the failure probability with DS
showed that the result of Cobyla was far from correct and thus ARF was preferred for continuing
with the rest of the analysis For educational reasons a FORM analysis with ARF was again
performed after changing the starting point and it was found that the computational time was
reduced to 4 hours Therefore it is still worthy to be aware of the fact that by changing the
starting point the performance of the algorithm can improve a lot However that requires a prior
knowledge of the potential design point that usually does not exist
Figure A 7 Cobyla performance after changing the starting point closer to the failure point
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 143
It must be also mentioned that by reducing the number of the variables to 11 as it was mentioned
before the algorithms could not converge to the errors that has been set and the computational
time increased compared with the case of 19 variables Taking into account the high interaction
among the variables that was proved in section 73 through the SA it can be possible that the
combination of the soil variables could lead to failure rather than the variables individually
Therefore an analysis was tried again with the complete set of the parameters and it was realised
that the algorithm did converge to the limit value
A3 Distribution Types
Uniform Distribution
If the probability of a result is between a lower limit a and an upper limit b and if every result
between these limits is equally probable a uniform distribution should be used This distribution
is defined by the two limit parameters and the probability density function is written
(A25)
and the CDF is
(A26)
Figure A8 Probability density function of a uniform distribution with a=-2 and b=2
The mean and the variance of the distribution are also defined by the limit parameters
(A27)
(A28)
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 144
Normal Distribution
One of the most commonly used distribution types is the normal distribution or Gauss-
distribution and it is defined by its first two central moments the mean micro and the standard
deviation σ The PDF is given by
radic
(A29)
and its CDF is
int
radic
(A30)
Figure A9 Probability density function of a normal distribution with different mean and standard
deviation values (notice that the curve with micro=0 and σ2=1 belongs to the standard normal distribution)
The normal distribution with mean and standard deviation is called the standard
normal distribution In the most reliability methods such as FORM SORM DS etc the analysis is
carried out in the standard normal space which actually means that the stochastic parameters are
transformed into standard normal variables
Lognormal Distribution
If a variable X has a lognormal distribution the variable Y=ln(X) has a normal distribution The
probability density function is defined as
radic (
) (A31)
where and the mean and the standard deviation of the normally distributed random
variable Y Sometimes it is useful to work with the Yrsquos parameters rather than Xrsquos These
parameters can be expressed as follows
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 145
radic
(A32)
(A33)
where and are the mean and the standard deviation of Y In Figure A10 an example is given
of a log-normally distributed variable with different values of mean and standard deviation
Figure A10 Probability density function of a log-normal distribution with different mean and standard deviation values
Truncated Normal Distribution
The truncated normal distribution is the probability distribution of a normally distributed
random variable whose value is either bounded below or above or both
Suppose has a normal distribution and lies within the interval
Then conditional on has a truncated normal distribution whose probability
density function is
(
) (
)
(A34)
and otherwise and is the probability density function and the CDF of the standard
normal distribution similar to Eq (1) and (2) If
and
then the CDF is
defined as
(A35)
Then the mean value and the variance of the conditional will be respectively
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 146
(A36)
[
(
) ] (A37)
In the next figure the PDF histograms of a variable with a mean of 0 and a standard deviation 1
were plotted for two cases (a) variable is normally distributed (b) variable is truncated-normally
distributed with =-2 and b=2
Figure A 11 Truncated normal distribution in comparison with normal distribution
MSc Thesis A Rippi 147
Appendix B Plaxis 2D (2015) features
In this Appendix the basic features of Plaxis 2D that were treated in this thesis are discussed First
of all an introduction over the Mohr Coulomb model is made as the soil failure criteria stem from
that model Then the initial stress generation in Plaxis is discussed and how that was applied in
the current master thesis while also the φ-c reduction technique is described which was used in
the SA of the soil failure Last but not least a more detailed explanation of the interface strength
Rinter is given which represents the modeling of the soil-structure interaction in Plaxis and it has
been considered as a random variable in the current project
B1 Mohr Coulomb failure criterion Soil tends to behave in a highly non-linear way under load This non-linear stress-strain behaviour
can be modelled at several levels of sophistication There are different constitutive models
available for simulating the behaviour of the soil whose parameters can change depending on the
level of sophistication that is required in each case
As someone can notice in Table B1 different models can be applied in different cases of soil
behaviour Hardening Soil model is becoming more and more interesting in soil structures as it
has been proved to approach the plastic behaviour of the soil and the non-linear stress-strain
relationship in a reliable and promising level Moreover UBCSAND31 and the Hypoplastic model
are more sophisticated models capable of representing several aspects of soil behaviour
however the amount of the model parameters is large and for that reason they are recommended
only in case that enough soil data are available for determining these parameters
Table B 1 Features of material models (source Behaviour of Soils and Rocks 2015)
ModelFeature Elasticity Failure Hardening Softening Small strain stiffness
Cyclic loading
Liquefaction Creep
Mohr-Coulomb x x
Duncan-Chang x x
Hardening Soil x x x
HSsmall x x x
x
Modified Cam-Clay
x x x x
Soft Soil x x x
Soft Soil Creep x x x
x
UBCSAND x x x
x x
Hypoplastic (x) x x x
x (x)
31 UBC are the initials for University of British Columbia in Vancouver in Canada
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 148
In this master thesis the Mohr-Coulomb model is going to be applied which is simple but an
efficient model to describe the soilrsquos failures condition Below a description of that model is made
The Mohr-Coulomb model is a linear elastic perfectly plastic model (see Figure B1) The linear
elastic part is based on Hookersquos law while the perfectly plastic part is based on the Mohr-Coulomb
failure criterion formulated in a non-associated plasticity framework
The main principle of elastoplasticity is that total strains and strain rates consist of elastic and
plastic strain components
(B1)
(B2)
in which the elastic strain rate component abides by the Hookersquos law and thus it can be expressed
as follows
(B3)
where is the elastic stress-strain matrix and the effective stress rate The plastic strain
component is given by
(B4)
where is a scalar defining the magnitude of the plastic strains called plastic multiplier and
is
a vector describing their direction with to be the so-called plastic potential function
Figure B1 Basic principle of elastoplasticity
The occurrence of (elasto)plasticity can be indicated by the yield function which is
illustrated in Figure B2 in principle stress space and it is derived according to the Mohr-Coulomb
failure criterion expressed in the next formula
(B5)
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 149
Considering the sign of this yield function the different soil states can be recognised as follows
if flt0 (inside yield contour) pure elastic behaviour
if f=0 and dflt0 unloading from aplastic state (=elastic behaviour)
if f=0 and df=0 elastoplastic behaviour
if fgt0 (outside yield contour) impossible stress state
Figure B2 Mohr-Coulomb yield surface in principle stress space
Finally in Mohr-Coulomb model the calculation of the plastic strains is achieved by the so-called
non-associated flow rule where with the dilatancy angle
This makes a difference between the friction and the dilatancy angle which prevents the model
from overestimating potential uplift
Concluding this model uses 5 input parameters
Stiffness parameters E (Youngrsquos modulus) and v (Poissonrsquos ratio) as they are included in
the elastic stress-strain matrix according to Hookersquos law
Strength parameters c (cohesion) and (friction angle)
Dilative behaviour ψ (dilatancy angle)
The author is aware of the possibilities and limitations of this model In particular the model is
capable of
a first order approach of the soil behaviour in general
describing quite well the (drained) failure behaviour
It should be mentioned at that point that the Youngrsquos modulus E cannot be steered via the
Python interface (as it appears to be a readonly parameter) and thus the shear stiffness G was
considered instead that is defined as follows
[kPa]
f=0 (yield contour)
flt0
fgt0
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 150
On the other hand the model has also limitations that are referring to
the linear elastic behaviour until failure (no strain- or stress path-dependent stiffness
behaviour)
overestimation of the shear strength in case of an undrained behaviour
However this model is considered reliable for detecting failure and giving realistic stresses
magnitudes in the soil and structural elements
Moreover in this project a non-associative behaviour was assumed by considering a zero
dilatancy angle However it has been proved by Teunissen (2009) that in case of deep failure
mechanisms in a dike structure with soil materials of a friction angle above 30deg problems arise in
the numerical analysis in finite element calculations Particularly the strength in the sliding
surface decreases due to rotation and the ground stresses are relieving around it by releasing
energy which might lead to calculation implications These phenomena are not treated in this
thesis but it is advisable to check and review the results after such an analysis
B2 φ-c Reduction Technique In this section the ldquoSafety Calculationrdquo is applied as it is described in Plaxis (Plaxis 2D (a) 2015)
In the Safety Calculation the φ-c reduction technique (Brinkgreve amp Bakker 1991) is adopted
where the shear strength parameters tanφ and c of the soil as well as the tensile strength are
successively reduced until failure of the structure occurs The dilatancy angle ψ is in principle not
affected by the phic reduction procedure However the dilatancy angle can never be larger than
the friction angle When the friction angle φ has reduced so much that it becomes equal to the
(given) dilatancy angle any further reduction of the friction angle will lead to the same reduction
of the dilatancy angle The strength of interfaces if used is reduced in the same way
The total multiplier is used to define the value of the soil strength parameters at a given
stage in the analysis
(B6)
where the strength parameters with the subscript ldquoinputrdquo refer to the properties entered in the
material sets and the parameters with the subscript ldquoreducedrdquo refer to the reduced values used in
the analysis is set to 10 at the beginning of a calculation to set all material strengths to
their input values
The incremental multiplier is used to specify the increment of the strength reduction of the
first calculation step (this increment is by default 01 but user can change it) The strength
parameters are successively reduced automatically until all required steps have been performed
(by default these steps are set to 100 but this can be subjected to changes depending on the
situation) It must always be checked whether the final step has resulted in a fully developed
failure mechanism In other words if has reached an almost constant value while the
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 151
deformation is continuing that indicates an equilibrium state In that case the factor of safety is
given by
(B7)
B3 Initial Stress Generation The first construction phase in Plaxis was the initial stress generation before the dike has been
constructed In that phase for horizontal soil surfaces with any soil layers parallel to this surface
and horizontal phreatic level32 a K0-procedure can be applied where K0 is the ratio between
horizontal and vertical stresses
(B8)
In practice the value of for a normally consolidated soil is often assumed to be related to the
friction angle by Jakyrsquos empirical expression (Jaky 1944)
Using very low or very high K0-values in the K0 procedure may lead to stresses that violate the
Mohr-Coulomb failure criterion In this case Plaxis automatically reduces the lateral stresses such
that the failure condition is obeyed Hence these stresses points are in a plastic state and are thus
indicated as plastic points Although the corrected stress state obeys the failure condition it may
result in a stress field which is not in equilibrium
B4 Interface Strength In FEM for a proper modelling of soil-structure interaction interface elements can be added to
plates in order to simulate the thin zone of intensely shearing material at the contact between the
plate and the soil
The strength properties of the interface are linked to the strength properties of the adjacent soil
layer and they are reduced by a factor of as follows
(B9)
(B10)
(B11)
where index ldquoirdquo refers to the interfacersquos parameters whereas index ldquosoilrdquo refers to the adjacent soil
parameters The behaviour of the interfaces is defined according to the Coulomb criterion which
distinguishes between elastic behaviour where small displacements can occur within the
interface and plastic interface behaviour when permanent slip may occur
32 For non-horizontal surfaces a Gravity loading is recommended
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 152
When the interface is elastic then both slipping (relative movement parallel to the interface) and
gapping (relative displacements perpendicular to the interface) can occur The magnitudes of the
interface displacements are
(B12)
(B13)
where is the shear modulus of the interface is the 1-D compression modulus of the
interface is the virtual thickness of the interface (generated automatically from Plaxis) is
the elastic interface normal stiffness and is the elastic interface shear stiffness
The shear and compression moduli are related as follows
(B14)
(B15)
Taking into account the above relations the displacements can be highly influenced by the elastic
parameters and thus the strength reduction factor can have a considerable impact on the
systemrsquos behaviour In the Reference manual of Plaxis 2D (2015) a value of 23 is generally
recommended while a value of 1 would not normally be used It should be stated that in the
framework of this thesis there are no available measurements in order to define the sensitivity to
this parameter and hence it is considered uncertain variable following a truncated normal
distribution
Appendix C ndash NEN 6740 ndash Table 1
MSc Thesis A Rippi 153
Appendix C NEN 6740 - Table 1
MSc Thesis A Rippi 154
MSc Thesis A Rippi 155
Appendix D Input Files for the Reliability Analysis
Below an example of the input files that were used for the anchor the sheet pile the soil body and
the system failure is presented in Figures D1 D2 D3 and D4 respectively In particular
regarding the input file of the anchor failure the meaning and the scope of each line is briefly
explained in Table D1
Figure D 1 Anchor analysis input file (with FORM method)
APPENIDX D ndash Input Files for the Reliability Analysis
MSc Thesis A Rippi 156
Table D1 Explanation of input components
A Reliability method and its parameters eg FORM with a maximum number of iterations 100 a maximum absolute error of 05 a relative error of 1000 a residual error of 05 and a constraint error of 05
B Number of random variables
C Soil material as it is specified in Plaxis eg AKleiVenigOnder is the clay layer under the dike
D Soil parameter eg Rinter is the interface strength of the clay layer
E Probability distribution and its parameters eg Truncated normal distribution with mean 066 standard deviation 013 and bounds from 00 to 10
F Copula type
G I Limit state function I is the threshold of the function
H Response surface function that is activated in case of Plaxis errors
Figure D 2 Sheet pile wall analysis input file (with FORM method)
APPENIDX D ndash Input Files for the Reliability Analysis
MSc Thesis A Rippi 157
Figure D 3 Soil analysis input file (with DS method)
APPENIDX D ndash Input Files for the Reliability Analysis
MSc Thesis A Rippi 158
Figure D 4 System analysis input file (with DS method)
MSc Thesis A Rippi 159
Appendix E Characteristic and mean values
The characteristic value of a parameter implies that there is 5 probability that the real value is
higher (load) or lower (resistance) than the characteristic In Figure 22 the probability density
function of a standard normal distributed parameter (mean value equals to zero and standard
deviation 1) is depicted In case of such a distribution the characteristic value is 164σ higher than
the mean value
Figure E1 Characteristic value illustration for a normal distribution (source Wolters 2012)
In this study full probabilistic calculations are taking place in order to estimate the corresponding
structural reliability According to the probabilistic design concept the mean values and the
standard deviations are used instead of characteristic values In that case characteristic values
and shall be converted to their mean values and as follows
where
(E1)
(E2)
and are the values for load or resistance respectively to which a probability of (non-
)exceedance of 95 corresponds (non-exceedance for the load and exceedance for the resistance)
(for a standard normal distributed parameter this is 1645) The definition of the rest parameters
is given in section 423
MSc Thesis A Rippi 160
MSc Thesis A Rippi 161
Appendix F Reliability methods
In this appendix an overview of different reliability methods is given More precisely in F1 an
explanation is given concerning the generation of random samples that is applied in MC as well as
in other sampling methods In Appendix F2 more advanced sampling methods than crude MC are
presented that can be used as accelerating reliability methods in the sense that the computational
time can be reduced Finally in Appendix F3 some basic principles of the FOSM are elaborated
that can help to better understand the concept of FORM as well
F1 Generation of random samples in Monte Carlo
The non-exceedance probability of an arbitrary random variable is uniformly distributed between
zero and one regardless of the distribution of the variable In formula (CUR ndashpublication 190
1997)
(F1)
is the uniformly distributed variable between zero and one and is the non-exceedance
probability Thus for the variable X
(F2)
in which is the inverse of the PDF of X Using this formula a random number X can be
generated from an arbitrary distribution by drawing a number of from the uniform
distribution between zero and one
More or less the same way base variables of a statistical vector can be drawn from a known joint
probability distribution function However the joint PDF must then be formulated as the product
of the conditional probability distributions of the base variables of the vector In formula this is
( )
By taking m realizations of the uniform probability distribution between zero and one a value can
be determined for every
(F3)
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 162
If the variables are statistically independent this can be simplified to
(F4)
By inserting the values for the LSF one can check whether the obtained vector ( ) is
located in the safe area
F2 Other Sampling Methods
Quasi-Monte Carlo Simulation
As an alternative to pseudo-random number generation (used in Crude MC) the quasi-random
number sequences are introduced from where the method Quasi-MC took its name Quasi-MC is a
technique which approximates the probability of failure using low discrepancy sequences33
x1hellipxN instead of randomly generated sequences
If the domain of failure is the goal is to estimate the following
probability
int ( )
(F5)
The main principle of this method is the integral of the function on [ ] can be
approximated by using some low discrepancy sequence x1hellipxN as follows
int ( )
sum
(F6)
The low discrepancy sequence is generated in Δ according to Lebesgue measure and then may be
transformed to any measure micro thanks to the inverse CDF technique in order to approximate the
integral
int ( )
sum
(F7)
In Figure 4 a comparison is made between quasi-random sample and pseudo-random sample in
order to illustrate the difference in sampling for the two methods
33 The discrepancy of a sequence is low if the number of points in the sequence falling into an arbitrary set B is close to proportional to the measure of B as would happen on average (but not for particular samples) in the case of a uniform distribution
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 163
Figure F1 Example of Quasi random Sample (left) and Pseudo Random Sample (right) (source
Waarts 2000)
This method a priori provides an asymptotically higher convergence rate than traditional MC but
no general rule can guarantee a better efficiency of the Quasi-MC sampling than the classical MC
sampling The advantages of Quasi-MC sampling tend to disappear with the increase of the
number of dimensions of the input variables It is recommended to use the Quasi-MC technique
with high sampling sizes or with very low dimensionality (in OT this method is valid only if the
input random variables are independent) (OpenTURNS 15 2015)
Importance Simulation
The main idea of the Importance Sampling method is to replace the initial probability distribution
of the input variables by a more efficient one Efficient means that more events will be counted
in the failure domain and thus reduce the variance of the estimator of the probability of
exceeding a threshold In other words importance sampling starts with the observation that if we
are going to sample randomly we should distribute the points to put the most of them in the
region that contains information (ie the failure region) and to waste as few as possible
In this method a sampling density is used instead of the actual probability density function
For a MC simulation the probability of failure is calculated by
sum ( ( )) ( )
( )
(F8)
where ( ( )) ( )
( ( )) ( )
The problem is that prior knowledge is needed of the failure area which in most cases is not
available In Figure 5 examples of the different sampling methods are shown
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 164
(a) (b)
(c)
Figure F2 Comparison of Importance Sampling (a) MC (b) and Numerical integration (c) (source
Latin Hypercube Simulation
Latin Hypercube Sampling is a sampling method enabling to better cover the domain of variations
of the input variables thanks to a stratified sampling strategy (in OT this method is applicable in
the case of independent input variables) Stratified sampling tries to place the sampling points so
that more of them will be found in regions where the variance of the limit state function g(x) is
largest In Figure F3 a comparison is made between the LHS and MC in terms of their sampling
technique The sampling procedure according to Baecher and Christian (2003) is based on the
next steps
1 Divide the range of each variable into several intervals of equal probability
2 If there are D variables to be sampled each in k regions then for each variable the k
sampling points are placed in independent random order
3 K samplings are made by choosing the first randomized value for each variable the second
randomized value for each variable and so on until k sets of randomized variables have
been chosen
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 165
Figure F 3 Comparison between Latin Hypercube sampling and MC (source OpenTurns 15 2015a)
F3 First Order Second Moment (FOSM) Method The First-Order Second-Moment (FOSM) approach (Ang amp tang 1984) provides analytical
approximation for the mean and the standard deviation of a parameter of interest as a function of
the mean and standard deviations of the various input factors and their correlations
Consider Z to be a function of random variables x1 x2 hellip xn that is
(F9)
In general case are correlated with covariance matrix [C] ie [C]=[σ][R][σ] where [σ]
is a diagonal matrix of standard deviations and [R] is the (positive-definite and symmetric)
correlation matrix with diagonal elements Rii=1 and non-diagonal elements Rij=ρij (ρij is the
correlation coefficient between variables i and j) In scalar notation Cij= σiσjRij (Farrokh 2007)
The first order approximation to the mean variance and standard deviation of the Z function is
based on the first terms of a Taylor series expansion of Z The following first-order estimates of
mean and variance are obtained
[ ]
(F10)
[ ] (F11)
Where the vector b denotes evaluated at the mean values of ie
(F12)
If there is no correlation among the variables Eq 223 can be written as
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 166
sum
(F13)
Then the reliability index can be calculated in 6 steps (Baecher amp Christian 2003)
1 Identify all variables that affect the mechanism that is researched
2 Determine the best estimate (usually the mean value) of each variable E[ ] and use these
to calculate the best estimate of the function E[Z]
3 Estimate the uncertainty in each variable and in particular its variance Var[ ]
4 Perform sensitivity analyses by calculating the partial derivatives of Z with respect to each
of the uncertain variables or by approximating each derivative by the divided difference
5 Use the equation of Var[Z] to obtain the variance of the function Z
6 Calculate the reliability index
[ ]
(F14)
MSc Thesis A Rippi 167
MSc Thesis A Rippi iv
Abstract
Some dike sections in the Netherlands failed to comply with the safety standards and one of the
most promising countermeasures is the construction of retaining walls inside the dike The
Dutch design codes for dikes with retaining walls rely on Finite Element Analysis (FEM) in
combination with partial safety factors However this can lead to conservative designs For this
reason in this research a reliability analysis is carried out with FEM calculations aiming to
demonstrate the feasibility of reliability analysis for such a soil-structure interaction problem
The case study concerns a (river) dike with an anchored sheet pile wall modelled in PLAXIS The
sensitivity and reliability analyses were enabled by coupling the uncertainty software package
OpenTURNS and PLAXIS through a Python interface The most relevant (ultimate) limit states
concern the anchor the sheet pile wall and global instability (soil body failure) The case was
used to investigate the applicability of the First Order Reliability Method (FORM) and
Directional Sampling (DS) to analysing these limit states Finally also the system reliability was
evaluated using sampling-based methods (DS)
Due to the considerable number of random variables before starting the reliability analysis a
sensitivity analysis was conducted for each limit state This indicated the most important soil
layers to be accounted for and the variables to be considered as stochastic The sensitivity
analysis and later on the reliability analysis were based on analytical formulations of the limit
state functions The anchor and the sheet pile limit states were formulated in terms of their
yield stress for global instability loss of equilibrium in the FEM analysis was used to define
failure Moreover the systemrsquos reliability was evaluated by taking into account all the three limit
states that were mentioned previously
The goal is to implement the coupling between FEM and reliability methods in order to analyse
the components of such a system (ie anchor sheet pile wall and dikersquos soil body) estimate the
probability of failure and identify the most important soil properties that affect the behaviour of
each component and the system as a whole The results of this research can be used to assess
and optimize the current design procedure for dikes with retaining walls
MSc Thesis A Rippi v
MSc Thesis A Rippi vi
Table of Contents
Preface ii
Abstract iv
List of abbreviations x
1 Introduction 1
11 Project objective and main research questions 2
12 Research approach and outline 3
2 System description and current design concept 7
21 System description and forces configuration 7
22 Current design concept 12
23 Safety standards 17
3 Literature study 21
31 Background 21
311 Finite Element Modeling 21
312 Uncertainties and Sensitivity analysis 24
32 Previous Studies 29
33 Overview 34
4 Structural Reliability Analysis 35
41 Basics of Reliability Analysis 35
42 Overview of Reliability Analysis Methods 36
421 Level III Methods 36
422 Level II Methods 40
423 Level I Methods (semi-probabilistic) 42
424 Response Surface Techniques (RS) 44
43 Coupling Reliability Analysis with FEM 46
431 The functionality and possibilities of OT 46
432 Coupling OpenTURNS-Plaxis 47
44 Overview 49
5 Failure Mechanisms and Limit State Functions 53
51 Introduction to the system analysis and the limit states 53
52 Limit State Functions 55
521 Serviceability Limit State 55
522 ULS for Structural Members 57
MSc Thesis A Rippi vii
523 ULS for Soil Failure 60
53 Overview 66
6 Case Study-Dike with an anchored sheet pile wall 69
61 Case Description 69
62 Soil Parameters 70
63 Finite Element Model 73
64 Deterministic Analysis 74
641 Calculation Scheme and Design Values 75
642 Construction Stages 77
643 Determination of the structural elementsrsquo characteristics 78
65 Overview 81
7 Reliability analysis results with stochastic soil properties 85
71 Method description 85
72 Mean values calculations 88
73 Sensitivity Analysis Results 91
74 Soil Shear Failure 98
75 Anchor Failure 102
76 Sheet pile wall failure 105
8 Conclusions and Recommendations 121
81 Conclusions 121
82 Recommendations 123
References 125
Appendix A 129
OpenTURNS features 129
A1 Fourier Amplitude Sensitivity Test (FAST) 129
A2 Optimization Algorithms in FORM 132
Principles of optimization algorithms 132
Convergence criteria 134
Evaluation of the algorithms performance 135
A3 Distribution Types 143
Uniform Distribution 143
Normal Distribution 144
Lognormal Distribution 144
Truncated Normal Distribution 145
Appendix B 147
MSc Thesis A Rippi viii
Plaxis 2D (2015) features 147
B1 Mohr Coulomb failure criterion 147
B2 φ-c Reduction Technique 150
B3 Initial Stress Generation 151
B4 Interface Strength 151
Appendix C NEN 6740 - Table 1 153
Appendix D 155
Input Files for the Reliability Analysis 155
Appendix E 159
Characteristic and mean values 159
Appendix F 161
Reliability methods 161
F1 Generation of random samples in Monte Carlo 161
F2 Other Sampling Methods 162
F3 First Order Second Moment (FOSM) Method 165
MSc Thesis A Rippi ix
MSc Thesis A Rippi x
List of abbreviations
CDF Cumulative Distribution Function
CoV Coefficient of Variation
CUR Civieltechnisch Centrum Uitvoering Research en Regelgeving
DS Directional Sampling
FAST Fourier Amplitude Sensitivity Analysis
FEA Finite Element Analysis
FEM Finite Element Model
FERM Finite Element Reliability Method
FORM First Order Reliability Method
FOSM First Order Second Moment
LEM Limit Equilibrium Method
LRFD Load and Resistance Factor Design
LSF Limit State Function
MC Monte Carlo
OT OpenTURNS
PDF Probability Distribution Function
RFEM Random Finite Element Method
RS Response Surface
SA Sensitivity Analysis
SLS Serviceability Limit State
SORM Second Order Reliability Method
ULS Ultimate Limit State
VNK Veiligheid Nederland in Kaart
MSc Thesis A Rippi xi
The roots of education are bitter but the fruit is sweet
Aristotle
MSc Thesis A Rippi 1
1 Introduction
In the Netherlands according to the Flood Protection Program (Hoogwater
Beschermingsprogramma1) and the Room for the River program (Ruimte voor de Rivier2)
alternative structural techniques for the reinforcement of existing dikes or for future dike
constructions additional to the conventional ones have been introduced and tend to be
attractive The heightening of the dike crest and the construction of a stability berm are some of
the most common current actions for dike strengthening Additionally filter layers geotextiles
and drainage systems can be applied in advance in order to prevent several failure mechanisms
Recently cantilever or anchored sheet pile walls and diaphragm walls tend to be used as an
alternative procedure for embankment reinforcing (see Figure 11) Such an alternative can be
chosen for an improvement of the macro stability of the dikersquos inner slope and additionally for
saving space in the land area that needs to be habited
Figure 11 Design options for dike reinforcement (source Flood Defences 2015)
For that purpose engineers need a concrete and unequivocal design methodology for such
combined structures In the meantime Deltares in cooperation with the Water Board of
Rivierenland is working on a design guideline which will be applicable to design the so-called
Type II stability fences3 A draft of such a guideline is elaborated in the report of Larsen et al
(2013) In this report suggestions and recommendations are outlined for the design of such
structures with Finite Element Modeling (FEM) and partial factors Remarks are also made for
the advantages and the limitations of such models and how they should be handled so that
reasonable and trustworthy results can be retrieved from the analysis
1 httpwwwhoogwaterbeschermingsprogrammanldefaultaspx 2 httpalbertawatercomhow-is-water-governedwhat-is-room-for-the-river 3 Type II structures are structures which in combination with a soil construction fulfil water retaining functionalities
Introduction
MSc Thesis A Rippi 1
The concept behind the recommended design criteria (Larsen et al 2013) is the definition and
the evaluation of several partial factors with the view to determine overall safety factors These
factors are further discussed in section 22 However validation of these factors shall be carried
out before being used in any case otherwise either the safety of the structure is jeopardized or
the structure will be overdesigned and cost inefficient
For improving the design criteria researchers for a long time focused on enhancing structural
models (beams shells etc) and constitutive laws (elasticity plasticity damage theories etc)
With the development of computer science a great amount of work has been devoted to
numerically evaluate approximated solutions of the boundary value problems describing the
mechanical system FEM is probably nowadays the most applied approach for the solution of
these problems
However the increasing level of detail of the constitutive models and the constant enhancement
of the computational tools do not solve the problem of identification of the model parameters
and the inherent physical and modelling uncertainties Moreover in most civil engineering
applications the intrinsic randomness of materials (soil rock concrete etc) or loads (water
elevation wind earthquake motion etc) is such that deterministic models are using average or
later on characteristic values of the properties at best lead to rough representations of the
reality
As a counteraction a semi-probabilistic methodology has been developed that was based on the
application of characteristic and design values by using partial factors Current design codes
such as Eurocode provide target reliabilities for different types of structures and structural
elements according to the potential consequences of failure However these partial factors are
not always equally suitable and efficient for all types of structural applications since they have
been calibrated under specific conditions Besides the consequences of failure of flood defences
such as dikes can be comparable to the investments in increasing the reliability of such systems
are For that reason it would be advisable for these systems to define target reliabilities based
on a risk assessment (ie tailor-made solution) rather than using the standard partial factors
coming from general geotechnical design codes which may be either too low or too high for a
given flood defence system
One step of such a risk assessment is accounting for randomness and spatial variability of the
mechanical properties of materials is one of the tasks of stochastic or probabilistic mechanics
which has developed fast in the last decade In this master thesis project the uncertainty of soil
properties is going to be treated in terms of its contribution to failure For that purpose
probabilistic methods are going to be implemented for a dike with an anchored sheet pile wall
(see Figure 12) simulated in a FEM software that is specialized in soil mechanics The successful
implementation and in future research the verification of such methods can be considered as
the most preferable and cost efficient way to design structures with high safety requirements
and not only for the validation of the partial factors Of course such a procedure tends to be
time consuming However the gradual improvement of the current probabilistic methods in
combination with the state-of-the-art computer capabilities as well as the scientific knowledge
gained in terms of different systems behaviour and failure modes can introduce a more
optimized way of designing structures with considerable investments
Introduction
MSc Thesis A Rippi 2
Figure 12 Reinforced dike section with an anchored sheet pile wall
11 Project objective and main research questions At the beginning of this research it has been observed that the design procedure being followed
for dikes with retaining walls modeled in FEM can lead to an overestimation of the critical loads
and thus to a potential cost inefficient final structure An example of the magnitude of the design
values in such a case study is given in the next chapter where also a more detailed explanation
of the current design concept is described
The objective of this study is to implement a full probabilistic analysis for evaluating the
reliability of a dike reinforced with an anchored sheet pile wall Such a research analysis can
subsequently come up with valuable recommendations for the improvement of the present
design approach Soilrsquos and structural elementsrsquo behavior should be taken into account both
separately and as a system Some of the reliability methods can deal with system reliability
problems such as Monte Carlo (MC) and Directional Sampling (DS) whereas for others like
FORM and SORM additional methods should be applied that use reliability information for each
individual limit state function to obtain the systemrsquos reliability
In principle the probability of failure of different limit states is to be computed individually
while in the sequence the system reliability is going to be estimated As far as the soil failure
mechanisms are concerned this research will focus mainly on the global stability of the dike
slope while for the sheet pile wall and the anchorrsquos failure the exceedance of the steel yield
stress is going to be considered
The scope of this thesis is not only to test different reliability models robustnessrsquo in conjunction
with FEM simulations but also to get a better insight into the specific system behavior (ie of a
dike with a retaining wall) analyze each component separately and investigate its response
under certain load conditions Furthermore the minimization of the computational effort and
time could also be carried out meaningfully under the constraint of sufficient accuracy The
accuracy should be such that the probability of failure is acceptable for the ultimate limit state
(ULS) in a normal design process and according to the current safety standards Finally the
Sheet pile
wall
Anchor
Dike section
Introduction
MSc Thesis A Rippi 3
robustness of the coupling between the reliability model and the FEM will be tested via their
capability of adapting to new input parameters without encountering convergence errors
during execution
The main research question of this master thesis project is thus formed as follows
How can the probability of failure of a dike with a sheet pile wall due to global instability modeled
by a Finite Element Model be analyzed
Essential questions regarding the soil models and failure criteria as well as the reliability
methods and the systems behavior are generated which are listed below
Subquestion 1 Which reliability methods are computationally tractable in
combination with FEM
Subquestion 2 How robust (convergence) are the tractable methods
Subquestion 3 Which limit states are relevant for the envisaged application of
retaining walls in dikes and how can they be formulated using FEM
analysis outcomes
Subquestion 4 What is the contribution of different uncertainties in the failure
mechanisms of the system
Subquestion 5 Can response surface techniques help to increase the efficiency and
robustness of the reliability model
Subquestion 6 How can the current design approach for dikes with sheet piles be
improved
In the next chapters the above research question and the related subquestions are going to be
answered by following the methodology that is described in the next section
12 Research approach and outline In this section an overview is given of how the aforementioned objective and sub-questions are
approached The thesis is mainly divided into 8 chapters and 6 appendices In Figure 13 the
thesis outline is illustrated and an indication of which subquestions (Sq1-6) are treated to
which chapter is given In Chapter 1 an introduction into this research content is made and a
first illustration of the system under investigation is presented together with the main research
questions
Chapter 2 has been devoted to the description of the system (ie a dike with an anchored sheet
pile wall) and to the case study that has been adopted for being analyzed in terms of its
reliability The forces configuration is also illustrated and the results of the case study according
to the current design concept are presented and evaluated Last but not least the new
recommended safety standards are included and the current required reliability of the specific
dike section is presented
Introduction
MSc Thesis A Rippi 4
In Chapter 3 a literature study related to FEM uncertainties and sensitivity analysis is
presented Moreover previous studies that are associated with the objective of this research are
mentioned whereas some of their results were also taken into account for proceeding with this
research
The main scope of this thesis is the implementation of different reliability methods on a specific
case study with the view to investigate and analyze its behavior Some of these methods are
continuously mentioned through the test and thus in Chapter 4 an overview is given of the
most well-known reliability methods Eventually an evaluation of these methods is made based
on literature and preliminary testing with simple case studies and the procedure of their
coupling with FEM is discussed The evaluation helped to qualitatively answer subquestion 1
while also an introduction of how special reliability methods such Response Surfaces were used
in this thesis that partially covers subquestion 5 In this chapter an introduction into the Limit
State Functions (LSF) concept is made and how they are considered for the reliability analysis of
a system An LSF actually represents a failure mode that can be detected in a structural or soil
element and it is expressed as a function of several variables In Chapter 5 the failure
mechanisms and the corresponding LSF that are related to the specific case study are identified
and formulated which answers subquestion 3
In Chapter 6 the case study whose reliability is to be evaluated is presented as it was modeled
in FEM The boundary conditions are specified and the soil and structural properties are
indicated and illustrated Moreover a deterministic design is also taking place in order to
roughly estimate the structural elementsrsquo properties This will help to make a qualitative
comparison between the current design procedure and the design according to a fully
probabilistic approach by referring thus to subquestion 6
Eventually the results of the aforementioned analysis are presented in Chapter 7 The results
mainly include the estimated probability of failure of each of the system components as well as
of the system as a whole by considering the soil properties as stochastic Moreover an
interpretation of the failure points is made and an assessment of the level of impact of the
random variables on the systemrsquos behavior is carried out At that chapter subquestions 2 4 and
5 are mainly treated
Last but not least in Chapter 8 the general conclusions are presented together with some
valuable recommendations for future consideration and research Furthermore a reflection on
the methodology and how the different research questions were approached is made
Introduction
MSc Thesis A Rippi 5
Figure 13 Thesis outline
Introduction
MSc Thesis A Rippi 6
MSc Thesis A Rippi 7
2 System description and current design concept
In this section the system to be analyzed is described more in detail and the individual
components are identified Moreover the forcesrsquo configuration is explained and illustrated as
they would have been calculated with the conventional way in parallel with a qualitative
evaluation of the potential expected deformations Finally the current design concept and the
safety standards that are referred to such a system are introduced
21 System description and forces configuration The system to be investigated in this thesis consists of a dike section reinforced with a one-layer
anchored sheet pile wall The system has been simulated by FEM and later on it was coupled
with a reliability package for carrying out a reliability analysis In the figure below an
illustration of the system is depicted and the several elements of an anchored sheet pile wall are
showed
Figure 21 System layout and different components of the anchored sheet pile wall
In particular an anchored sheet pile wall consists of the sheet piles which are anchored in the
soil via a tie rod The rod is grouted most of the times in a sandy layer with a cemented grout
body and it both alleviates the sheet piles from the axial forces exerted by the upper structure
and keeps the wall stable in case of excessive developed moments due to the lateral earth
pressure Last but not least a waling system is applied mainly downstream of the sheet piles in
order to transfer the loads from the piles to the anchors in such a manner so as to avoid
excessive local stresses on the intersection between the sheet piles and the tie rod
Cross-Section
Top View
System description and current design concept
MSc Thesis A Rippi 8
The main load configuration acting on a dike section as well as a sheet pile wall is depicted in
Figures 22-24 together with the possible displacements As far as the soil body stability is
concerned most of the slope stability analysis computer programs are based on the limit
equilibrium concept according which a soil mass tends to slide down under the influence of
gravity The transitional or rotational movement is considered on an assumed or known slip
surface below the soil There an equilibrium should be achieved among the driving and the
resisting forces In that case the driving moments consist of the soil weight the water pressures
and the loads around the center of the slip surface such as a possible vertical load on the top of
the dike crest (ie traffic load) The magnitude of the water pressures is controlled by the water
elevation on the river side
In Figure 22(a) the external and internal forces acting on a slice of the slip surface are
illustrated The driving forces are the soil weight W the water pressure U and any additional
load that can contribute to the rotation of the slip surface which in turn activate the lateral
active earth pressure The resisting forces consist of the lateral passive earth pressure
(Rankine 1857) and the shear strength (Terzaghi 1943) Their expressions are written as
follows
[kN] (21)
[kN] (22)
[kPa] (23)
where is the active lateral earth pressure coefficient and the passive lateral earth
pressure [kNm3] is the unit weight of the corresponding soil layer and H [m] is the thickness
of the soil layer that the lateral earth pressure comes from (in case of multiple soil layers there
should be a distinction among the different forces) and [kPa] is the effective cohesion of the
soil layer The safety factor that is written in the Figure 22(a) as FS is defined by the ratio of the
shear strength (excluding the pore water pressure) divided by the weight of the earth body
(including the pore water pressure)
The loss of equilibrium between the driving and the resisting moments lead to the rotation and
instability of a slip surface as it is illustrated in Figure 22(b) A slide surface can take various
shapes according to the Limit Equilibrium Method (LEM) that is followed According to Bishop
method this surface tends to be circular in Uplift-Van method a horizontally compressed zone
can be also considered whereas in Spencer method the shape of the slide body can be arbitrary
The slip surface can be located either in the landside or in the river side of the dike depending
on the load the soil characteristics and the design of the dike (ie inner or outer berm
reinforcing revetment on the outer slope etc) Most of the times the landside area of the dike is
jeopardized due to different failure mechanisms that are described in section 523 The
instability of the outer slope is mainly relevant in case of sea dikes and a potential damage can
be usually reconstructed until the next flooding
System description and current design concept
MSc Thesis A Rippi 9
(a)
(b) Figure 22 (a) External and internal forces acting on a slip surface of a dike section and (b) possible deformation pattern
As a counteraction for the slope instability a sheet pile wall can be installed inside the dike The
special thing about these structures is that they make possible a greater freedom in form and
functionality than a traditional dike design This structure derives its strength from the
materials used such as steel which are able to withstand higher pressures than clay for instance
The general stability is due to friction and wedging in the bottom
After the reinforcement of the inner side of the dike with a retaining wall the strength capacity
of the wall is also important for the global stability of the system The forces to be taken into
account for the sheet pile wall stability are the active and the passive earth pressures (effective
pressures) and the water pressure as they are illustrated in Figure 23(a) In that figure the
System description and current design concept
MSc Thesis A Rippi 10
forces were simplified in a singular triangular shape as in a homogeneous soil body in order to
explain and depict the overall picture of the acting forces However the stresses distribution
over depth can be more complex depending on the variety of the soil layers that are present In
Figure 23(a) it can be noticed that a sheet pile wall which is installed close to the inner berm
might not have significant instability issues as there is the passive side that contributes to
resisting forces However in the passive side the soil can differ and be weaker than this of the
active side Therefore the passive force in that case might not be very supportive and thus the
wall shall be designed cautiously
Additionally the anchor resistance shall be reassured for the sake of the stability of both the
sheet pile wall and the overlying soil body In Figure 23(a) the forces acting on the anchorage
are displayed The distributed load qs comes from the earth pressure of the upper soil layer (it
might not be uniformly distributed as it is depicted in the figure but more in a trapezoidally type
of load) whereas the tension force Ft4 stems both from the earth pressure and the displacement
of the sheet pile wall The bending and the tension capacity of the anchor are determinant for its
stability In Figure 23(b) a possible displacement pattern of the retaining wall and the
anchorage is illustrated
Furthermore corrosion is an additional weakening impact on the wall that depends on the
water level and the pore water pressures near the structure The thickness of the wall and the
material properties are playing a key role to the resistance towards corrosion This will not be a
subject of this thesis however it should be taken into account in case of design purposes
(a)
4 where O is the average circumference of the pile P [kPa] is the shaft friction at depth z and L is the length of the pile
System description and current design concept
MSc Thesis A Rippi 11
(b) Figure 23 (a) Mobilizing forces acting on the sheet pile wall and the anchorage and (b) possible deformation pattern
It should be mentioned that as far as the pore pressures are concerned in Figure 23(a) only the
hydrostatic pore pressures are depicted Generally the pore pressures (or active pore pressures)
are the sum of the steady state pore pressures and the excess pore pressures
[kNm2] (24)
Steady state pore pressures can be hydrostatic (based on a horizontal phreatic line) or non-
hydrostatic (based on steady state groundwater flow) Excess pore pressures are based on
loading of undrained soils In FEM these various pore pressures are taken into account
automatically In Figure 24 an example of a potential distribution of the active pressures
nearby the sheet pile wall is depicted as it was deduced from Plaxis Therefore as it can be
noticed from the figure the distribution can indeed sometimes approximated as triangular
System description and current design concept
MSc Thesis A Rippi 12
Figure 24 Example distribution of the active pore pressures acting nearby the sheet pile wall
The knowledge of the type of forces that are exerted on both the structural components and the
soil body as well as of the most expected form of displacements is valuable in order for
someone to be able to evaluate the results deduced from FEM and detect potential modelling
errors Therefore this validation was necessary before starting with the reliability analysis
In the next sections the current design approach of the system described above is elaborated
Additionally the Dutch recommended safety standards are presented for this type of structures
which show the need of carrying out a reliability analysis For that purpose a case study was
adopted that it is presented first as it was designed according to the current regulations while
finally a comparison is made between the original and the new case study as it was found to be
according to the reliability analysis
22 Current design concept In the Netherlands dike reinforcement is becoming a major issue concerning the design of flood
defenses Special reinforcing structures such as cofferdams anchored sheet piles and diaphragm
walls are used for strengthening the dike Especially in the context of the Flood Protection
Program and the Room for the River program many dike reinforcement projects have been
suggested
In the report of Larsen et al (2013) the design procedure of a dike reinforced with sheet pile
wall is prescribed using FEM The safety philosophy that is followed in this technical report
actually composes the current procedure for designing dikes with sheet pile walls and it is
linked with the usage of partial safety factors Particularly the required overall safety factor
FEM that should be compared with the one from FEM calculations is determined as follows
SFEM
b d m n
(25)
where
System description and current design concept
MSc Thesis A Rippi 13
Partial safety factor indicating the uncertainties of the soil composition and
the water pressures (also called schematization factor)
Partial safety factor which is related to the calculation model and the way the
calculations have been carried out (also called model factor)
Partial safety factor which is related to the material parameters (also called
material factor)
Partial safety factor associated with damage caused during the soil tests (also
called loss factor)
Safety factor of load
Each of these partial factors shall be at first calibrated by conducting a full probabilistic analysis
of the system under consideration In most of the cases these partial factors have been already
calibrated on previous similar projects and then they are reused for any similar case In
Schweckendiek et al (2013) a new approach of using partial factors for flood defences is
proposed whose application is still under consideration At the moment the aforementioned
partial factors for dike reinforced with sheet pile walls modelled in FEM are estimated
according to mostly engineering judgement and political negotiations
In sequence the FEM design procedure that is applied for sheet pile walls inside soil structures
is based on a so-called ldquosafety calculationrdquo (for information for this type of calculation see
Appendix B2) In this calculation an artificial reduction to the soil strength parameters (friction
angle and cohesion) is applied in order to evaluate the overall safety factor and confirm if this is
lower or higher than the required one that was estimated according to Eq 25 Furthermore the
resulting moments and forces developed on the structural elements during the specific
calculation are used for their design
This is a quite helpful and indicative calculation type available in Plaxis that can give a sense of
the structurersquos safety factor and the possible failure mechanisms that can occur under the
predefined load configuration However such a method of calculating the developed stresses
can be misleading in terms of the moments and forces acting on the structural elements and the
total deformations of the system
In this thesis the reliability of a dike with an anchored sheet pile wall is going to be evaluated
and in particular the case study of Breedeveld (2011) is to be used for that purpose In Figure
26 the location of the existing dike section is presented while in Figure 25 the structure as it
was modelled in FEM is illustrated It is essential to mention that the retaining wall does not
exist in reality but it was placed so as to implement and demonstrate the current design
regulations in the report of Breedeveld (2011) The anchorrsquos angle has been taken equal to 30deg
from the vertical which is generally a steep anchor inclination That is expected to reduce the
part of the horizontal force that is undertaken by the anchor and thus the most of it exerts on
the sheet pile wall The angle is usually taken 30deg-45deg from the horizontal axis (CUR 2005) but
this is of course depended on the load conditions and the construction requirements for each
System description and current design concept
MSc Thesis A Rippi 14
case The influence of the anchor orientation and the construction methods that are generally
followed for an anchored sheet pile wall are not considered in this thesis The case study was
taken as it was modelled in the aforementioned project and the reader should be aware that this
serves the proceedings of a reliability analysis and not of a perfectly constructed retaining wall
Figure 25 Original dike section in Varik-Heesselt as it is presented in Breedeveld (2011)
In particular based on this case study and trying to understand the design criteria a
comparison was carried out between the ldquosafety calculationrdquo and a normal ldquoplastic calculationrdquo
(ie elastoplastic drained or undrained analysis) In both cases the boundary conditions the
external loads and the input parameters are the same The difference between them is the
calculation procedure followed in order to come up with final stresses which in the case of the
ldquoplastic calculationrdquo does not include a reduction of the soil strength properties In Figures 27
and 28 the results from both calculations are presented In Figure 27 the total deformations
are exhibited while in Figure 28 the moments and the axial forces on the sheet pile are shown
Figure 26 Location of the Varik-Heeselt dike section (section number TG-078) in the Dike ring 43
BetuweTieler-en Culemborgerwaarden west (source Breedeveld 2011)
System description and current design concept
MSc Thesis A Rippi 15
As it is realised from this figure the discrepancy between the two results is quite high while as
far as the deformations are concerned no realistic conclusions can be drawn from a ldquosafety
calculationrdquo Moreover taking as design moments the moments that are deduced from the
ldquosafety calculationrdquo will lead to a much heavier and more expensive sheet pile cross section than
the ldquoplastic calculationrdquo
It is also essential to mention that the required safety factor was estimated up to FEM =18
according to the partial-factor procedure that is defined in Eq 25 Such a safety factor as
prerequisite for coming up with the design values of the moments and forces of the sheet pile
and the procedure of the ldquosafety calculationrdquo in FEM can lead to a conservative and financially
inaccessible design
(a) ldquoPlastic calculationrdquo Deformed mesh and maximum deformations (umax= 008 m)
(b) ldquoSafety calculationrdquo Deformed mesh and maximum deformations (umax= 9041103 m)
Figure 27 Deformed mesh and maximum deformations in the end of (a) plastic calculation and
(b) safety calculation
30deg
System description and current design concept
MSc Thesis A Rippi 16
(a) ldquoSafety
calculationrdquo
Mmax = 9392 kNmm Nmax= -4402 kNm
(b) ldquoPlastic
calculationrdquo
Mmax = 6833 kNmm Nmax= -1372 kNm
Figure 28 Bending moments and axial forces exerted on the sheet pile during the (a) safety
calculation and (b) plastic calculation
On the other hand the aforementioned design procedure guaranties a strong structure capable
of probably undertaking more than the expected loads and thus ensuring the safety against
flooding of the landside However the knowledge on fully probabilistic methods that has been
obtained as well as the development of the technology can introduce a reliability analysis of the
system as a preferable way for the future design concept Such a procedure should be
accompanied also by field test that would be able to validate FEM results as well as in-situ
measurements and inspections of the soil properties for calibrating their statistical
characteristics
Due to the conservative results that the aforementioned procedure came up with and because of
the simplifications that this case study was later subjected to in order to be used in this thesis a
new design was carried out in Chapter 6 and the properties of the structural elements were
redefined
The inclination towards probabilistic methods has not only been created due to the incomplete
design regulations for the system under consideration but also the current proposed safety
standards which are discussed in the section below recommends a risk based safety assessment
of the primary flood defenses which in principle implies the evaluation of the failure probability
In the next section an overview of the new recommended safety standards related to primary
flood defenses is given and the required reliability for the stability of the aforementioned case
study is elaborated
System description and current design concept
MSc Thesis A Rippi 17
23 Safety standards An important factor for the evaluation of a structurersquos reliability is the safety standards that
have been set and according which the acceptance or not of the failure probability is made After
the safety assessment5 of 2011 it has been declared that 33 (1225 km) of the primary flood
defenses (3767 km) in the Netherlands does not comply with the safety standards For this
purpose a new session of investigations has commenced in order to assess the reliability of the
existing dikes and the potential amendment of the safety standards for future constructions
Taking into account this latest information and within the framework of the Flood Risk in the
Netherlands 2 project6 (VNK2) a new Delta program was founded whose one of the main
decisions was the ldquoDelta Safety Decisionrdquo This decision contains a proposal for new safety
standards of the primary dikes as they are depicted in Figure 29 which are stricter than the
previous ones and formulated in terms of failure probability
According to this project a new policy regarding the safety assessment of the flood defenses has
been settled and proposed for the future design criteria which is based on the evaluation of the
acceptable flood risk (probability of failure (x) consequences) of each dike ring rather than the
probability of exceedance in order to achieve a level of protection that is in balance with the
societal value (Cost Benefit Analysis and life loss)
VNK has been already using reliability evaluation techniques in order to calibrate partial safety
factors Especially in geotechnical engineering the high inherent uncertainty of the soil
properties renders the evaluation of the structural reliability essential for the safety assessment
of the structure Subsequently this implies the investigation of the response of the different
reliability methods applied in a real case study of a dike The techniques that are used to
evaluate the probability of failure are discussed thoroughly in Chapter 4 and in Appendix F In
case of a complicated structure such as a dike with a retaining wall both the reliability of the
dike body and the reliability of the wall shall be evaluated and eventually a common standard
for the reliability of the system shall be defined
5 A safety assessment is an evaluation of the function of the flood protection structure and it is being conducted every approximately 6 years after the reassessment of the existing hydraulic boundary conditions 6 in Dutch Veiligheid Nederland in Kaart 2 or VNK2 It has been initiated by the Ministry of Infrastructure and the Environment (IampM) the Association of Regional Water Authorities (UvW) and the Association of Provincial Authorities (IPO)
System description and current design concept
MSc Thesis A Rippi 18
Figure 29 Maximum admissible flooding probability for primary defences according to Delta
programme 2015 (Deltaprogramma 2014)
As far as the retaining walls are concerned CUR (2005) design procedure distinguishes the
following three safety classes for retaining walls with corresponding reliability indexes
Table 21 Safety classes and corresponding reliability indices (CUR 2005)
Class I Relatively simple constructions no personal safety risks and relatively minor
damage in the case of overall failure ϐ=25
Class II Considerable damage in the case of overall failure minor personal safety
risks ϐ=34
Class III Major damage in the case of overall failure andor considerable personal
safety risks ϐ=42
System description and current design concept
MSc Thesis A Rippi 19
Additionally as far as a dike section is concerned the required overall reliability index shall be
concluded according to the new norms that are shown in Figure 29 However the reliability
index of the different failure mechanisms of a dike shall be estimated separately
According to Vergouwe et al (2014) that is part of the VNK2 report the major failure
mechanisms in dike ring 43 where the case study under investigation is located (see Figure 26)
is uplift and piping and macro-instability of the inner slope Especially in Figure 210 the
estimated percentage of the contribution of each failure mechanism to the overall probability of
flooding7 is indicated In Vergouwe et al (2014) the overall flooding probability of the specific
dike section that the case study concerns was evaluated to 17800 Therefore the macro-
instability on which this thesis is concentrated with a contribution percentage of 226 has a
probability of about 310-5 (
) and thus a target reliability index β of
approximately 4 This reliability level can be considered as a benchmark for evaluating the one
that will be entailed in the end of this thesis
Figure 210 Percentage contribution of the different failure mechanisms to the probability of flooding (source Vergouwe et al 2014)
It should be mentioned though that the aforementioned safety standards concerning the dike
ring 43 were defined according to the probability of exceedance of a certain water level that the
dike has to retain Therefore this complies with the old norms and not with those appeared in
Figure 29 However for large engineered systems such as flood defence systems with large
potential consequences and substantial investments it is worthwhile to assign target reliability
levels based on a risk assessment of the area surrounded by the specific dike ring There are
three widely used types of criteria for evaluating the risks related to floods and major industrial
hazards (Vrijling et al 2011)
Economic criteria
Individual Risk Criteria
Societal Risk criteria
7 The dependencies among the failure mechanisms were not taken into account Therefore this probability stated in Vergouwe et al (2014) is not exactly equal to the probability of flooding
System description and current design concept
MSc Thesis A Rippi 20
In order to check then if a certain system abides by the target reliability that is required the
failure probability of each failure mechanism shall be calculated based on a LSF The
aggregation of the failure probabilities of all the modes gives the failure probability of the
system which is supposed to be compared with the required in order to reassure the reliability
of the structure Therefore this thesis is dealing with the calculation of this failure probability of
a certain system modelled in FEM analysis and understanding of the systemrsquos behaviour and
finally the investigation of the robustness of coupling FEM with reliability methods for
evaluating the reliability of a structure In the next chapter an overview over the research has
been done until now concerning coupling and reliability methods applications is presented and
a discussion over the most important is made
MSc Thesis A Rippi 21
3 Literature study
In this chapter an overview of some principles related to FEM and the uncertainties in
geotechnical engineering is presented More precisely an introduction into FEM concept and a
discussion over the different types of FEM is made Moreover the uncertainties in geotechnical
engineering are stressed and the general framework according to which they are handled is
described Finally previous studies that are associated with the application of FEM on soil
structures and the reliability analysis are discussed
31 Background
311 Finite Element Modeling
FEM is a numerical method whose essence is to convert a problem described by partial
differential equations over space and time into one by dividing the space-time continuum into a
set of discrete elements and assuming that the unknowns vary over each element FEM solution
process is as follows
1 Divide structure into pieces (elements with nodes) (discretizationmeshing)
2 Connect the elements at the nodes to form an approximate system of equations for the
whole structure (forming element matrices)
3 Solve the system of equations involving unknown quantities at the nodes (eg
displacements)
4 Calculate desired quantities (ie strains and stresses) at selected elements
The properties of each element are set the same as the material properties that have been
defined by the user Then the Deterministic Finite Element Method (the finite element method
that was formulated with deterministic variables) can be used in conjunction with means and
standard deviations of the input variables to obtain reliability estimates
In the simple coupling of FEM with reliability analysis each parameter that is considered as
stochastic is given a particular probability density function estimated either by field tests or by
engineering judgement This approach is also called ldquosingle random variablerdquo and it assumes
that the spatial correlation length is infinite In other words the soil is considered to be
homogeneous and the stochastic property assigned to the soil is taken at random from a certain
probability distribution
Nevertheless a more realistic model should take into account the spatial correlation within
smaller regions where the stochastic property is allowed to vary For that purpose the Random
Finite Element Method (RFEM) was introduced by Fenton and Vanmarcke (1990) in which the
random variables are correlated to one another using auto-correlation functions
The most sophisticated method of FEM is the Stochastic Finite Element Method introduced by
Ghanem and Spanos (1991) which accounts for the uncertainties in the input parameters
Literature study
MSc Thesis A Rippi 22
implicitly within the finite element calculation This aims at representing the complete response
probability distribution function (PDF) in an intrinsic way Two steps are basically applied for
that purpose
Discretization of the problem and definition of the random vector with an unknown
joint PDF and
Expansion of the response in a particular basis of random vectors with a finite variance
called the polynomial chaos
In this thesis the Deterministic Finite Element Method is to be used coupled with reliability
analysis According to Waarts (2000) coupling FEM with reliability methods (FERM) would lead
to the following advantages
In comparison to standard finite element analysis it gives direct insight into the
structural reliability and decisive parameters
Structures designed using FERM will either be safer orand more economically built in
comparison to structures designed using safety factors and classical constitutive models
FERM is expected to be useful for the determination of LSF that cannot be explicitly formulated
and that differ in each case such as soil limit state in different structural schematizations
Moreover it is likely to be valuable in areas where little knowledge exists on the systems
reliability of structures with multiple components (ie a dike with a sheet pile wall)
FEM in practice
The conventional method for stability analysis in a soil body is represented by LEM although
FEM is increasingly used by designersresearchers The latter has been proved to be quite
realistic for the progressive behaviour (ie stress-strain development in different construction
phases) of a soil system under the effect of stress redistribution in comparison with classical
models Especially in their master thesis Johansson amp Sandeman (2014) compared the
deformations and the forces measured at a deep excavation supported by anchored sheet pile
wall in a railway tunnel located in Gothenburg in Sweden with 1D finite element software the
2D finite element software (Plaxis) and hand calculations They proved that Plaxis produces
reliable results for horizontal deformations in the sheet pile wall and anchor forces when
compared to in-situ measurements
Moreover in Gonzaacutelez et al (2013) it was proved that the simplification of reality done by the
classical methods such as Blumrsquos Engelrsquos Kreyrsquos methods etc (for further information for these
methods a reference is made to Gonzaacutelez et al (2013)) allows us to generally understand the
behaviour of the system wall-soil Nonetheless the results that came out of this analysis were
found to be quite conservative whereas FEM managed to give a more realistic interpretation of
the wallrsquos movement
Seed et al (2006) investigated the performance of the New Orleans Flood Protection Systems
during hurricane Katrina in August 2005 by simulating the different levee sections with FEM As
an example in Figure 31 pictures of the failure of the Lower Ninth ward IHNC East Bank in
New Orleans after the hurricane Katrina are presented FEM was proved to be an efficient tool
to realistically interpret the shape and the triggers of the failure mechanism
Literature study
MSc Thesis A Rippi 23
Figure 31 Failure of the Lower Ninth ward IHNC East Bank in New Orleans modelled in FEM with Plaxis (source Seed et al 2006) At the bottom right the shear strain contours are depicted that are developed through the embankment and at the bottom left the deformed mesh is presented illustrating the failure mode of the levee
In recent implementation of design concepts and technical recommendations such as CUR 166
ndash Damwandconstructies (2005) which refers to the design of retaining walls it can be observed
that guidelines for Finite Element Analysis (FEA) have been set parallel to the existing analytical
or empirical calculation methods Therefore FEA tends to become more and more accepted as
an alternative for Serviceability Limit State (SLS) as well as for ULS design
According to Wyllie and Mah (2004) LEM give an estimation of the factor of safety with no
information on deformation of the slope However in numerical analysis the failure surface can
evolve during the calculation in a way that is representative of the natural evolution of the
physical failure plane in the slope In that way a better insight into the evolution of failure
mechanisms can be gained
Cundall (2002) compared the characteristics of numerical solutions and LEM in solving the
factor of safety of slopes and concluded that continuum mechanics-based numerical methods ie
FEM have the following advantages
No pre-defined slip surface is needed
The slip surface can be of any shape (in contrast with Bishop and Kranz stability)
Multiple failure surfaces are possible
No static assumptions are needed
Literature study
MSc Thesis A Rippi 24
Structures (such as footings embankments etc) and structural elements (such as
beams cables sheet piles etc) and interface can be included without concern about
compatibility
It is also important to recognize that LEM only identifies the onset of failure whereas FEM
includes the effect of stress redistribution and progressive failure after its initiation Numerical
models can also be used to determine the factor of safety of a slope in which a number of failure
mechanisms can exist simultaneously or where the mechanism of failure may change as
progressive failure occurs (Hoek et al 2000)
On the other hand the accuracy of FEM is dependent on usersrsquo settings such as the refinement
of the mesh generation the loading steps etc Moreover even if FEM can be functional and easy
to use it is essential that the user has a deep knowledge of the subject under investigation (soil
and structural mechanics) in order to be sceptical and critical with the FEM results able to
interpret the output behaviour of the structure and changecorrect everything that seems
peculiar Moreover the usage of FEM implies also some basic knowledge on numerical methods
and the general background of the FEM so as to solve possible numerical errors
An ideal way of modelling in FEM would be a prior calibration of the program according to data
related with the structurersquos properties and performance (ie stress generation) However this
means that enough field measurements should be carried out and under specific conditions in
order to be considered as a base for FEM calibration Unfortunately such field tests are not
always available and if so their reliability is on doubt For example measurements related with
the developed moments or displacements of a retaining wall inside a dike section are rare to be
found but even if there are some available they cannot represent the situation near the failure
domain where most of the engineers are worried about
312 Uncertainties and Sensitivity analysis
Uncertainty in geotechnical engineering can be categorized into aleatoric epistemic and
decision model uncertainty as they are depicted in Figure 32 (Baecher amp Christian (2003))
Aleatoric uncertainty is associated with the natural variability which is defined as the intrinsic
randomness of natural processes Such variability can be expressed by changes of the soil
properties over time at a certain location (temporal variability) or over space at a single time
(spatial variability) Epistemic uncertainty can be divided into the site characterization model
and parameters uncertainty and it is attributed to lack of information about events or lack of
understanding the physical laws that prohibits the ability of creating a realistic model Finally
the decision uncertainty describes the difficulty of being aware of social objectives defining
social values such as discount rates and predicting the planning horizon
Literature study
MSc Thesis A Rippi 25
Figure 32 Categories of uncertainty in geotechnical engineering (modified from Baecher amp
Christian (2003))
Below the general steps of an uncertainty study are described and an introduction into the
sensitivity analysis concept is made
Global methodology of an uncertainty study
A first step of an uncertainty study8 can be described as ldquothe definition of the problemrdquo Initially
the variables of interest (or else the output variables) of which the uncertainty is to be
quantified shall be specified In sequence given several input variables for which the user may
have data andor expertengineering judgment a model denoted usually by a mathematical
function should be introduced that enables the computation of the set variable of interest
After the general context has been staged we should choose the criteria with which the
uncertainty can be evaluated The most complete measure of uncertainty when dealing with a
random vector is the probability distribution In order to assess the value of this distribution
function the following criteria can be followed
Probability of exceeding a threshold the aim is to assess the probability that the variable
of interest exceeds a threshold important for the goals at stake
Quantiles the aim is to assess the threshold that a variable of interest may exceed with a
probability equal to a given value
Central dispersion the aim is to obtain an ldquoorder of magnituderdquo of the uncertainty by
specifying the average value and the variance of a variable of interest
8 In the engineering world this is also called reliability analysis However in this section the general uncertainty study is described In particular reliability describes the ability of a system or component to function under stated conditions for a specified time period
Literature study
MSc Thesis A Rippi 26
The next step is to define a model to represent and quantify the uncertainties of the input
variables One shall investigate each variable as a singularity and come up with the most
suitable probability density function (mostly depending on available data) Besides it is
essential to assess also the potential correlations among the variables that can be included in a
joint probability density function which is discussed later in this section
Eventually the uncertainties of the input variables shall be translated in terms of uncertainty on
the variables of interest This procedure is called uncertainty propagation and can be carried
out via several reliability methods (approximation methods or sampling methods) that are
extensively elaborated in Chapter 4 In Figure 33 a schematization of the different steps during
an uncertainty study is given
Figure 33 Methodology of uncertainty study
Last but not least a better understanding of uncertaintiesrsquo influence can be achieved by
analyzing the contribution of the different uncertainty sources to the uncertainty of the
variables of interest via a sensitivity analysis Such an analysis aims at identifying important
parameters for the system response besides it gives a better insight into the model used to
quantify the uncertainties Before conducting an uncertainty analysis it is advisable to filter out
parameters of less significance in order to reduce the modeling computational effort Below
different methods for sensitivity analysis are described
Sensitivity analysis
Sensitivity analysis (SA) is the study of how the variation in the output of a model can be
apportioned quantitatively and qualitatively to the variation in the model parameters (Saltelli
et al 2004) Saltelli (2004) proposes one possible way of grouping these methods into three
classes (i) screening methods (ii) global SA methods and (iii) local SA methods
i Screening methods
Screening is a particular instance of a sampling-based method The objective here is to identify
which input variables are contributing significantly to the output uncertainty in high-
dimensionality models rather than exactly quantifying sensitivity (ie in terms of variance)
Screening tends to have a relatively low computational cost when compared to other
Literature study
MSc Thesis A Rippi 27
approaches and can be used in a preliminary analysis to weed out less influential variables
before applying a more informative analysis to the remaining set One of the most commonly
used screening methods is the elementary effect method
ii Global SA methods
Global SA techniques incorporate the whole range of variation and the probability density
function of the input parameters to calculate their influence on the output Many global
sensitivity analysis techniques are now available such as
Variance-based methods
o Sobolrsquos method9
o Fourier Amplitude Sensitivity Test (FAST)10
o Analysis of Covariance (ANCOVA)
Regression-based methods11
Both FAST and Sobolrsquos method rely on the assumption of parameter independence while
ANCOVA can also handle correlated input parameters The main principles of FAST method are
presented in Appendix A1 For more information about the other methods the reader can
betake himself to the related references
iii Local SA methods
Local SA methods provide the slope of the calculated model output in the parameter space at a
given set of values (Turanyi and Rabitz 2004) This basically means that local methods involve
taking the partial derivative of the output Y with respect to an input factor Xi |
|
where the
subscript indicates that the derivative is taken at some point in the space of the input
Examples for these are (Schweckendiek 2006)
the α-values in a FORM calculation (further discussion in section 422)
Local Probabilistic Sensitivity Measure
In reliability analysis the local sensitivities will be more important than the global ones in the
sense that the influence of all variables in specific points such as the design point cannot be
identified by the global methods However local SA can only inspect one point at a time and the
sensitivity index of a specific parameter is dependent on the central values of the other
parameters
In this thesis the influence of the different parameters on the response of the system is
evaluated according to local sensitivity indices However before the main part of the reliability
analysis starts it was necessary to filter out the less influencing variables in order to reduce the
number of the major variables and make the analysis more efficient and computationally
affordable For this purpose the global sensitivity method FAST was used whose results are
presented in Chapter 7
9 created by Sobolrsquo (1993) 10 introduced by Cukier et al (1973) 11 introduced by Helton et al (2005)
Literature study
MSc Thesis A Rippi 28
Probabilities as a measure of uncertainties
Most engineers today use the concept of probabilities as the standard way to talk about
uncertainty One way to quantify a probability of a random variable is the calculation of the
cumulative probability function (CDF) of non-exceedance which can be obtained by
the probability distribution function For a random vector this reads
( ) (11)
From this the joint probability density function of this random vector can be determined as
( )
( )
(12)
This function is depicted in Figure 34 by means of contour levels The joint probability density
function has the shape of a hill as it is showed in the 3D figure and the inclination of the ellipses
reveals a correlation between the two variables and
Figure 34 Joint probability density function in 2D and 3D
The functions that join or couple multivariate distribution functions to their one-dimensional
marginal distribution functions are called copulas These are going to be used in order to define
the joint density probability distribution from which the random variables are taking their
values However no detailed explanation is made in this report regarding copulas and its
characteristics as it is automatically implemented in OT whereas the user has to define only the
type of copula that heshe considers the most appropriate (a reference is made to Nelson (1999)
for further information) According to the type of copula that is selected the order of the values
of the variables can be chosen within an iterative reliability method In this thesis an
independent copula was used considering that the input variables are independent the ones
from the others due to the time limit
Literature study
MSc Thesis A Rippi 29
32 Previous Studies In this section a summary of the most relevant researches on the topic of FEM applied in soil
structures and coupling with reliability methods is given
To begin with an attempt to estimate the reliability of a structure modeled in FEM has been
made by Waarts (2000) by introducing an optimized reliability method in terms of
computational effort and efficiency
In particular in Waarts (2000) two adopted reliability methods are introduced both making
use of a response surface (a detailed explanation of the reliability methods and the response
surface techniques is made in Chapter 4 and Appendix F) These adaptive response surfaces are
used in combination with FORM and DS respectively The accuracy and the effectiveness of
these methods are investigated on the basis of artificial LSFs and a comparison is made with the
existing standard reliability methods The most efficient combinations of response surface
techniques and reliability methods were with FORM (FORM-ARS) and DS (DARS) Comparing
these two methods DARS predominated over FORM-ARS as it can cope with a much wider
range of limit state functions
In a later stage the above best performing reliability method (ie DARS) is further investigated
in terms of its efficiency on the basis of complex structures reliability In Figure 35 a couple of
case studies used to verify the performance of DARS are given
Figure 35 Left Two-story two-bay frame structure Right overview of a sheet pile installation
(source Waarts 2000)
The conclusions of this research showed that DARS serves its purpose and that the benefit from
using it increases with increasing number of random variables In Figure 36 the Limit State
Function Evaluations (LSFE) that are carried out as a function of the number of variables are
presented
Literature study
MSc Thesis A Rippi 30
Figure 36 Number of LSFE as a function of the number of variables n The upper line depicts the
required number of LSFE N=160 (it is supposed that on average 3 LSFE per sample are required
and consequently the number of LSFE equals 3x160n=480n) The lower line shows the
performance of DARS (source Waarts 2000)
More recently a methodology for assessing the reliability of geotechnical structures and in
particular deep excavations has been developed by Schweckendiek (2006) More precisely this
study deals with the application of different reliability methods in combination with FEM which
carries out the LSFE The reliability methods are applied initially on simple examples in order to
be calibrated and eventually the most efficient methods are used for the reliability assessment
of a deep excavation with a sheet pile wall The several failure mechanisms are treated in detail
concerning the structural elements of the excavation as well as the soil medium
Finally the failure probability of the system is estimated according to a fault tree
schematization of a deep excavation with one layer anchored sheet pile wall Uncertainties in
the soil properties the phreatic levels and also the corrosion of the sheet pile wall were taken
into account In Figure 37 a picture of the case study is shown together with a FEA illustration
and the importance factors of different soil properties in terms of the probability of failure after
a FORM analysis
Literature study
MSc Thesis A Rippi 31
Figure 37 Finite Element simulation and contribution of several structural and soil properties in
the failure mechanism of an anchored retaining wall in layered soil (source Schweckendiek 2006)
For a similar case study as in Schweckendiek (2006) a cantilever and an anchored sheet pile
wall the safety of the structure based on LEM is evaluated in a research conducted by Gonzaacutelez
et al (2013) Moreover in order to compare the efficiency of the classical methods with FEM
the finite element software Plaxis has been used This research shows the limitation of the
classical models and the efficiency of FEM as far as an anchored sheet pile wall in concerned
More precisely LEM results found to be conservative as they do not take into account the
confinement around the free length of the anchor rod that increases the passive pressure After
comparing numerical results with the classical methods in anchored walls the failure
mechanism given by the FEM suggests an intermediate failure state where the sheet pile wall
describes a translation movement on the deep zone and at the same time a rotation movement
around the anchor point that is closer to the reality
Furthermore Moumlllmann and Vermeer (2008) conducted a reliability analysis of a dike failure A
case study at river Elbe in Saxony Germany has been used and the failure probabilities of
different dike cross sections involving different failure modes were compared More precisely
overflowwave overtopping upliftpiping slope stability and damage of the revetment on the
waterside were taken into account For each failure mode the particular failure probability was
determined using the software package PC-Ring (Vrouwenvelder et al 1999) while the
reliability method used for the assessment of the failure probability was FORM That was
coupled with a slope stability software where Bishoprsquos approach was performed
However this method is limited to circular slip surfaces and prescribed pore pressure
distributions within the dike In order to overcome this limitation FEM was used for further
analysis In combination with FEM an adaptive response surface technique was used in order to
represent the limit state function of each failure mode In Figure 38 the simulated structure
Literature study
MSc Thesis A Rippi 32
together with the response surface of the slope instability failure mode is showed Note that this
structure is approached with a single soil layer (same soil within and under the dike) and that
no structural element is implemented inside the dike
Figure 38 Dike model for the probabilistic FEA of dike stability and the corresponding response
surface of the limit state function for 3 stochastic input parameters (source Moumlllmann and
Vermeer 2008)
The failure probability of each of the aforementioned failure mechanisms were finally calculated
and compared with those recommended in each case while the performance of the coupling
between FEM and response surface method was assessed
Additionally a more recent investigation in terms of dikes with sheet piles employing FEA was
carried out by Breedeveld (2011) The main objective of this study was to display the
implementation of partial factors for design purposes using FEM and especially the software
Plaxis in simulating dikes with a sheet pile wall The dike was firstly simulated as a single
structure without reinforcement while in sequence the occurring stresses and pore pressures
were calculated with an existing sheet pile wall The results of the distribution of the effective
stresses within and below the dike are illustrated in Figure 39 His case study has been
introduced in Chapter 2 while in Chapter 6 a more detailed explanation of how it was modelled
in FEM is given This case study will be used as an example case in this thesis
(a)
Literature study
MSc Thesis A Rippi 33
(b)
Figure 39 Effective stress distribution before (a) and after (b) the installation of the sheet pile wall (source Breedeveld 2011)
Last but not least Bach (2014) has been occupied with the reliability analysis of bored piles and
the case study that he examined is presented in Figure 310 The major objectives of this
research were (1) Propose models to calibrate resistance factors for the Load and Resistance
Factor Design (LRFD) (see section 423 for further explanation of this method) approach and
find a suitable model aiming to directly determine reliability of a bored pile considering some
types of defect that may occur in the bored pile (2) Select a quality control method and evaluate
its reliability when applied to bored piles
For that purpose he made a coupling calculation between the finite element software Plaxis
(version 90) and the numerical probabilistic toolbox Prob2B in order to design bored pile
foundations in light of the reliability-based design Two failure modes the geotechnical failure
mode and the structural failure mode were proposed in order to comprehensively assess the
reliability of an axially loaded pile The parameter uncertainty was considered through the use
of statistical parameters and probability distributions for material parameters in soil models
The soil parameters were treated as random variables The geometry parameters of pile were
used as deterministic quantities because a change in pile shape in the calculation process
requires establishing a new mesh which is now impossible with regard to the Plaxis software
Figure 310 Nine calculation cases considered with different pile toe levels (source Bach 2014)
Literature study
MSc Thesis A Rippi 34
For the reliability of the pile the ULS of the pile was investigated and finally a displacement
criterion of a percentage of the pile diameter used depending on the soil type and
characteristics that the structure is founded on
33 Overview In this chapter concepts that are related to FEM probabilities uncertainty and sensitivity
analysis were discussed The introduction of FEM applications and the principles of how to
handle uncertainties will help the reader to better understand and follow the process of this
thesis
Furthermore preliminary researches related to several aspects of this thesis are presented that
mainly concern
Reliability methods and their applications [Waarts 2000 Schweckendiek 2006 Bach
2014]
Coupling FEM with reliability methods [Waarts 2000 Schweckendiek 2006 Bach
2014]
(Reliability) Analysis of geotechnical structures and especially dikes reinforced with
retaining walls [Waarts 2000 Schweckendiek 2006 Moumlllmann and Vermeer 2008
Breedeveld 2011 Gonzaacutelez et al 2013 Bach 2014]
As someone might have noticed from the overview of the existing research on the topic both the
performance of different reliability methods and the simulation of a dike with a sheet pile wall
in FEM have been carried out These references were quite helpful in order to get an idea of
coupling FEM with reliability methods as well as FEM and reliability methods individually
Moreover a first dive into the system under investigation and the variables to be handled was
made However noone has analysed the reliability of such a complicated system that is more
and more often seemed to be applied in the Netherlands Thus it would be essential to evaluate
the implementation and the results of the reliability analysis of a dike with sheet piles simulated
in FEM The conclusions of such a research are going to contribute to design and optimization
concepts and hopefully to a better understanding of the system behaviour
MSc Thesis A Rippi 35
4 Structural Reliability Analysis
The aim of this chapter is to introduce the basic theory of a reliability analysis and to summarize
the principles of different reliability methods that are widely used for the uncertainty
evaluation of a system Particularly three of them (MC DS and FORM)12 were used in this thesis
and are described in this chapter Finally a brief evaluation of the selected reliability methods
and the way that they are applied on this thesis by coupling them with FEM are discussed
41 Basics of Reliability Analysis In the consideration of the reliability of an element the determination of the probability of
failure is the central issue The limit between failure and non-failure is defined as a limit state
and the reliability is the probability that this limit state is not exceeded The limit states are
interpreted through the so-called limit state functions (LSF) whose general form is
(41)
in which R is the strength or more general the resistance to failure and S is the load or that
which is conducive to failure (ldquosolicitationrdquo) The basic principle of structural design is that the
resistance needs to be higher than the load or in other words that the LSF is larger than zero
(Zgt0) The main objective of the design is to ensure that this performance criterion is valid
throughout the lifetime of a structure Nevertheless the majority of the quantities that both
resistance and load consists of are uncertain Therefore a probability of satisfying the preceding
criterion is estimated and expresses the reliability of the structure P(Zgt0) The probability of
failure is
(42)
Hence the probability of failure is complement to the reliability to the sense that
(43)
The LSF is plotted in Figure 41 indicating the failure space and the limit state function by Z=0
12 Besides RS techniques were used for proceeding with the analysis but most for solving numerical problems rather than evaluating the failure probability The RS techniques are still under investigation and need to be tested before being applied in real cases However it is highly possible that this would be a potential way of evaluating the failure probability in the future
Structural Reliability Analysis
MSc Thesis A Rippi 36
Figure 41 Limit Sate function
In this case the design concept is based on the evaluation of the so-called design point which is
the point in the failure space with the greatest probability density Generally this point is
located on the border between the safe and the unsafe areas
In the structural domain the Joint committee on structural safety (1981) proposed a level-
classification of the calculation methods This classification includes the following three levels
Level III these methods calculate the probability of failure by considering the probability
density functions of all strength and load variables The reliability of an element is linked
directly to the probability of failure
Level II this level comprises a number of methods for determining the probability of
failure and thus the reliability It entails a linearization of the LSF in a carefully selected
point These methods approximate the probability distribution of each variable by a
standard normal distribution
Level I at this level no failure probabilities are calculated The level I calculation is a
design method according to the standards which consider an element sufficiently reliable
if a certain margin is present between the representative values of the strength and the
loads This margin is created by taking the so-called partial safety factors into account in
the design
In the next section the different reliability methods to be used in this thesis are further
elaborated
42 Overview of Reliability Analysis Methods
421 Level III Methods
The foundation of the Level III failure probability calculation is the mathematical formulation of
the subset of the probability space which involves failure (Zlt0) Level III reliability methods
Structural Reliability Analysis
MSc Thesis A Rippi 37
(also known as fully probabilistic methods) compute the probability of failure based on the
exact probability density function and the exact limit state functions Therefore such methods
are considered to be the most accurate Well-known sampling methods are (Quasi-) MC
simulation Importance Simulation DS and the Latin Hybercube Simulation Below the former is
discussed as it represents the fundamental theory on which the sampling methods are based on
while DS is also further explained as it is continuously used in this thesis For the rest a brief
explanation is provided in Appendix F There are also other sampling methods such as the
Numerical and the Directional integration which will not be described in this study However
for further information about them a reference though is made to Waarts (2000) In Level III
methods errors can only occur by a too low number of simulations
Crude Monte Carlo Simulation
The MC method uses the possibility of drawing random numbers from a uniform probability
density function between zero and one If and are the
vectors for the resistance and the load respectively that consists of n variables then by taking
multiple realizations of the uniform probability distribution a value can be determined for
every and
By repeating this procedure a large number of times the failure probability can be estimated
(44)
where is the total number of simulations (nm draws from the uniform distribution in which
m is the number of base variables) and is the number of simulations for which Zlt0
The simulationrsquos relative error is
(45)
If the expected value of the relative error is zero the standard deviation is
radic
(46)
Based on the central limit theorem the error is normally distributed provided is sufficiently
large The probability that the relative error is smaller than the given value E is then
(47)
Thus for the reliability of the relative error is smaller than For the wanted k and
E the required number of simulations can be determined with
Structural Reliability Analysis
MSc Thesis A Rippi 38
(48)
If for instance a reliability of 95 is required for a maximum relative error E=01 the required
number of simulations amounts to
(49)
The number of simulations is therefore still dependent on the probability of failure
MC simulation is applicable whatever the complexity of the deterministic model is However the
number of samples required to get an accurate estimation of may be dissuasive especially
when the value of is small (10-k) Thus crude MC is not applicable when small values of are
sought andor when the CPU cost of each run of the model is non-negligible
In Appendix F and especially in sections F1 and F2 a further elaboration of the MC simulation
is presented Moreover the methods of Quasi-MC Importance sampling and Latin Hybercube
are described
Directional Sampling
The directional simulation method is an accelerated sampling method It implies a preliminary
iso-probabilistic transformation as for FORM and SORM methods however it remains based on
sampling and is thus not an approximation method Below the method as it is described in Open
TURNS 15 Reference Guide (2015 pp190-193) is presented
Each simulation of the directional simulation algorithm is made of three steps Using the
probability distribution of a random vector we seek to evaluate the following probability
( ) (410)
Here is a deterministic vector and ( ) the limit state function which enables the
definition of the event For the iteration these steps are the
following
Let A point is drawn randomly on S according to a uniform
distribution
In the direction starting from the origin and the passing through solutions of the
equation ( ) (ie limits of ) The set of values of that belong to is
deduced for these solutions it is a subset of
Then one calculates the probability By property of independent
standard variable
is a random variable distributed according to chi-square
distribution which makes the computation effortless
Structural Reliability Analysis
MSc Thesis A Rippi 39
Finally the estimate of the probability after N simulations is the following
sum
(411)
The main idea is that for each direction we go along the direction by step of a specified length
from the origin to the maximum distant point and we check if there is a sign change on each
segment so formed (see Figure 42) We go until the maximum distant point Then for all the
segments where a sign change is detected we research the root λ in the segment (there are
also other strategies available in OT however all of them follow the same principle of
searching) The following figure illustrates the principle of an iteration in dimension 2
Figure 42 DS of a 2-dimensional problem
Similar to MC method the outcome of the estimated probability of failure is a random variable
and the error in the estimate can be made as small as possible by taking a sufficient number of
samples For DS the standard deviation σ of the estimated failure probability can be quantified
as follows (Melchers 2002)
radic
sum
(412)
From this equation relative errors and the confidence intervals can be estimated As can be
seen the error in the estimated probability of failure will decrease with increasing number of
sampled directions Equation (412) can be used to determine the number of sampled directions
that is required for a reliable estimate of the failure probability The convergence criterion in
this method is usually the coefficient of variation (CoV) which is defined as
Structural Reliability Analysis
MSc Thesis A Rippi 40
422 Level II Methods
Level II methods (known also as fully probabilistic methods with approximations) can take all
the probabilistic properties of the random variables into account but they include
approximations of the limit state function and therefore their use and outcomes should be
inspected and evaluated in order to be considered as reliable However experience in that
methods has shown that the computational effort is profoundly decreasing in comparison with
Level III methods and that the application of them can provide important parameters such as
the reliability index (β) and the influence factors (α) Some of the most known Level II methods
are First Order Second Moment (FOSM) Method First and Second Order Reliability Methods
(FORM and SORM) and the Point Estimate method (PEM) In the next paragraphs FORM is
elaborated as it will be mainly used in this thesis while in Appendix F3 the principles of FOSM
are discussed
First Order Reliability Method (FORM)
The principles of FORM are based on the First-Order Second-Moment (FOSM) approach which
first introduced the reliability index concept This method is presented in Appendix F3 in detail
However the reliability index estimated using the FOSM approach is not ldquoinvariantrdquo which
means that the index value depends on the limit state function format (Farrokh 2007) Hasofer
and Lind (1974) proposed an invariant definition of the reliability index that leads to a
geometric interpretation known as first-order reliability method (FORM)13 The first step of this
method is the transformation of the random variables to equivalent standard normally
distributed variables and the whole procedure is carried out in u-space (or else standard space)
For variables that are normally distributed this step is as follows
(413)
For other types of distributions there are procedures available for carrying out this
transformation such as Generalised Nataf and Rosenblatt transformations In this study these
transformations are not discussed however for further explanation a reference is made to the
Reference Guide of OpenTURNS 15 (2015)
In the sequence the limit state function Z is expressed in terms of
The second step is the approximation of the function with the first two terms of the Taylor-
polynomial The approximation reads (CUR 1997)
( ) ( ) sum
(414)
where is the point where the linearization takes place This approximation of Z is linear and
according to the central limit theorem it is normally distributed The expected value of the LSF
can be approximated by the expected value of the linearized function
13 VNK2 uses FORM in view of the calculation time required In case of inaccuracies VNK2 switches to more time-consuming techniques such as DS whereby a large number of draws are made from the probability density functions of the various stochastic variables (uncertain quantities)
Structural Reliability Analysis
MSc Thesis A Rippi 41
( ) sum
(415)
While the standard deviation (sum of contributions of each variable to the variance of Z) is
defined as
radicsum
( )
(416)
Using the definition of FOSM for the reliability index [ ]
where [ ] is the mean value of
the limit state function and substituting Eq 415 and Eq 416 in it the reliability index can be
approximated However linearization in different points can lead to different values of the
reliability index Hasofer and Lind definition of the reliability index overcomes this limitation
and renders it equal to the minimum distance from the origin to the design point
(417)
where radic
Looking for the design point is basically an optimization problem Many analytical and
numerical approaches can be used for that purpose A relatively straightforward method to do
this is by firstly assuming that the design point is the mean value (the starting point can also be
another point according to an engineering judgement in order to accelerate the optimization
procedure) The obtained β-value is used to determine a new point in which the LSF is
linearized In this case the importance factors αi are calculated as
( )
radicsum (
( ) )
(418)
where is the design point and is the number of variables expresses the
contribution of the variance of each variable to the total variance of Z in the design point Figure
43 illustrates the linear approximation of the limit state function and the aforementioned
parameters for a two dimensional problem The probability that Zlt0 can be determined using
the standard normal CDF
(
) (419)
Structural Reliability Analysis
MSc Thesis A Rippi 42
The new calculation point is determined by
(420)
After some iterations the chosen optimization algorithm finds the final design point and the
new reliability index In Appendix A2 an overview of the different optimization algorithms
available in OT is presented together with an evaluation of their performance in the specific
case study
Figure 43 Two-dimensional illustration of u-space LSF and design point
423 Level I Methods (semi-probabilistic)
At the beginning of the probabilistic concepts incorporation in the field of structural
engineering the most notable development was the implementation of LFRD At the past single
factors of safety on the ratio of total resistance to total load were used in order to ensure that
the stresses developed from an applied load on a structure were lower than the allowable ones
LRFD replaces those factors by introducing a set of partial safety factors on the individual
components of load and resistance These partial factors can be selected such as they account
for uncertainties associated with the resistance and the load respectively
The current design philosophy in CUR 166 and Eurocode is characterized by the use of
characteristic values of the parameters (in Appendix E a further explanation of the
characteristic values is given together with the way that they are converted into the mean
values for being used in the reliability analysis later on) The values of the partial safety factors
are additional factors to the characteristic values Therefore the characteristic values are
multiplied with the aforementioned safety factors ( ) and the following criterion has
to be satisfied for a reliable structural performance (see Figure 44)
(421)
Structural Reliability Analysis
MSc Thesis A Rippi 43
The main disadvantage of LRFD is the fact that there is a set of load and resistance factors that
need to be calibrated in order to cover the specific cases that are likely to occur This is because
of the lack of sufficient statistical data that are necessary for this calibration procedure
Consequently it cannot reassure that all the designs to which the set of factors is applied result
in the reliability level that was aimed for As a result in majority of the cases a conservative
calibration that gives the required reliability is chosen which can lead to ldquoover-designrdquo
structures
Figure 44 Design with partial factors for the load and resistance parameters (source
Schweckendiek 2006)
The evaluation of an elementrsquos reliability starts in principle with calculating the probability of
failure and subsequently the reliability for the given strength and load In practise the problem
is often that the strength is unknown but it has to be determined for a given reliability The
determination of the required reliability can be estimated with the help of Level II and III
methods by iteratively adjusting the strength in the calculation until a sufficiently small
probability of failure is found
In the design domain regulations and guidelines follow the standard that the characteristic
value14 of the strength is divided by a factor and that the characteristic value of the load is
multiplied by a factor as it is described in Eq 421
The link between Level I and the estimation of the failure probability has been achieved through
Level II methods The design point that results from a Level II method calculation is the point
with the greatest joint probability density of the strength and the load and it is therefore
possible that when failure occurs the strength and load values will be close to that point The
design values can be specified as
(422)
(423)
14 The characteristic value of a parameter is its main representative value and its value can be specified as a mean value an upper or lower value or a nominal value according to EN 1990 definition (2002)
Structural Reliability Analysis
MSc Thesis A Rippi 44
where are the coefficient of variation for the resistance and the solicitation respectively
15 and are the importance factor and β is the reliability factor These can be also expressed
in terms of characteristic values as follows
(424)
By substituting function 423 with 421 and 422 the partial safety factors are defined as
(425)
(426)
where 16 and are the values for load or resistance respectively to which a probability of
(non-)exceedance of 95 corresponds (non-exceedance for the load and exceedance for the
resistance) (for a standard normal distributed parameter this is 1645)
424 Response Surface Techniques (RS)
In case that the models presented above tend to be time consuming for the limit state function
evaluation or convergence problems of the optimization algorithms occur a better methodology
is recommended known as response surface The RS is mainly used when (unknown) response is
only available from experiments or complex FEM computations (for example large highly-non-
linear FEM models) Indeed once a RS has been built up the various standard methods may be
applied at a negligible cost An analytical limit state function replaces the real response function
The main idea is that the response consisting of a complex function of input variables is
approximated by a simple function of the input variables
A list of possible response surfaces techniques is given below
Linear and quadratic Taylor expansions
Polynomial RS based on least squares
Kriging method
Polynomial chaos expansion
A detailed explanation of the aforementioned methods can be found in OpenTURNS 15 (2015a)
The first method is associated with the approximation of the model response around a specific
set of input parameters while the rest seek a global approximation of the model over its whole
domain of definition The most sophisticated one is the polynomial chaos expansion but one the
15 α values are negative for solicitation (ie RV for which increasing their values lead to adversely effects on the reliability) and positive for resistance (ie RV for which increasing their values implies a beneficial effect on the reliability) 16 is negative and can be negative or positive
Structural Reliability Analysis
MSc Thesis A Rippi 45
other hand its complexity does not make it attractive In this research a polynomial RS is
applied based on least squares
In combination with FEA the standard procedure is as follows
1 Select the most important random variables on the basis of engineering basis
2 A SA is carried out in combination with FEM
3 Reduce the stochastic variables if needed according to the SA
4 A RS is constructed through the response data
5 A reliability calculation is carried out using the RS instead of the real response
A polynomial RS is generally constructed by fitting a quadratic function to the sampling points
whose general expression is
sum
sum
sum sum
(427)
The type of the reliability method to be used is of little importance since the time consuming
LSFE (using FEM) are replaced by analytical expressions However of main importance is now
the accuracy of the RS compared to the real response There are several parameters in order to
evaluate the goodness of fit of the RS to the real model among which is the well-known R-
squared
In Figure 45 an example of a quadratic response surface is showed that is fitted on the data
response indicated with the black circles This was a 6-dimensional problem that means 6
different variables were included in the system (non-visualized dimensions are kept constant at
their mean value) In this thesis the response of the system is to be acquired after the coupling
of FEM with a reliability method and that would give the ldquoblackrdquo circles that are shown in Figure
45 The response surface technique applies a curve fitting on the real response of the system
(ldquoblackrdquo circles) whose analytical formula can later be used for the reliability analysis of similar
type of systems with negligible computation time
Structural Reliability Analysis
MSc Thesis A Rippi 46
Figure 45 Quadratic response surface with 28 terms (ie ) for 6 dimensions ( ie x
=[ x1 hellip x6])
43 Coupling Reliability Analysis with FEM The limit state function evaluations (LSFE) will be carried out with the software Plaxis 2D 2015
which is a special two-dimensional finite element software used to perform deformations and
stability analysis for various types of geotechnical applications Moreover considering the case
study under investigation in this thesis Plaxis in contrast with other FEM such as Abacus
Comsol DIANA etc offers several techniques to realistically simulate structural elements such
as sheet pile walls and anchors and their interaction with soil while the variety of the
constitutive models for the soil body that are available and the ability to include the history of
the construction phases can lead to a better analysis of the systemrsquos behavior in terms of the
stress level and the deformations It is essential at that point to mention that using FEM for this
purpose means that the limit state formulation is implicit and can only be solved numerically
The reliability analysis is carried out through an uncertainty package In this section firstly a
description of the reliability package and its possibilities are given Finally an explanation of the
coupling procedure between the reliability tool and FEM is given together with the calculation
process that was followed for the parameters manipulation
431 The functionality and possibilities of OT
The reliability analysis tool to be utilized in this research is OpenTURNS (OT) which is a
scientific library usable as a Python module dedicated to the treatment of uncertainties and it is
still under development during the work of this thesis Several reliability packages are available
such as OpenEarth (applied in Matlab) Prob2b (applied in Probox software of TNO) and
Probabilistic Toolkit (software of Deltares) Plaxis 2015 features Python connection possibility
and thus coupling was decided to be carried out with an uncertainty package in Python and OT
is the most developed one Moreover until now only OT is an open source package and can be
Structural Reliability Analysis
MSc Thesis A Rippi 47
used easily from anyone as there is a wide community of experienced people supporting with
their knowledge while many related manual reports are already available for starting learning
A list of the current available reliability methods in OT is given in Figure 46 In this figure apart
from the standard methods the alternative method of the Adaptive Response surfaces is
introduced that is mainly used when (unknown) response is only available from complex FE
computations
Figure 46 Reliability methods available in OpenTURNS
In this thesis the methods to be basically applied are the FORM and the DS analysis while MC is
also used mainly for confirming the application of FORM Additionally the RS technique has
been mainly implemented for enhancing the performance of FORM and DS analysis
OT handles 47 types of distributions amongst which the ones used in this research are Normal
Lognormal Uniform and Truncated Normal distribution Regarding the joint distributions that
are available to be selected for random vectors 11 types of copulas are existing in OT amongst
which the most known ones are the Independent the Gumbel and the Normal copula In this
thesis the independent copula has been used during the reliability analysis which means that
the variables were assumed to be independent among each other However it is strongly
advised a further elaboration and research considering correlation matrices for specific soil
parameters
432 Coupling OpenTURNS-Plaxis
The finite element simulation will be carried out in Plaxis 2015 Plaxis is a finite element
software for plane strain and axi-symmetric modelling of soil and rock behaviour Moreover it
supports a fully automatic mesh generation allowing for a virtually infinite number of 6-node
and 15-node elements
Structural Reliability Analysis
MSc Thesis A Rippi 48
The coupling of reliability analysis and FEA requires an interface for the communication
between each other When a reliability tool is coupled with another software program the
reliability program carries out the whole reliability analysis and it uses the other program only
for the evaluation of the limit state function More precisely OT should be able to read and
amend Plaxis output values for important variables such as material parameters pore pressures
generation and stresses development and corresponding deformations inside the dike
Respectively Plaxis has to be also capable of obtaining the new values that has been set by OT
for the variables that are treated as stochastic during an iterative process according to the
reliability assessment In Figure 47 an illustration of the coupling methodology and its function
is shown
Figure 47 Coupling scheme OpenTURNS-Plaxis
In principle an input file is firstly required where the user set the preferable reliability method
to be used the stochastic input parameters and their probability distributions the joint
probability distribution and the corresponding correlation matrix and finally the limit state
function is formed depending on the situation In Appendix D an example of the input files that
were used for soil sheet pile wall and anchor analysis is shown However such input files
should be interpreted so as to be readable by both Plaxis and OT Therefore an input interpreter
was created which is actually a python script that helps OT to start up the reliability analysis
according to the assigned method variables distributions and LSF As it was mentioned before
the evaluation of the limit state function is conducted by Plaxis For that purpose the input
interpreter should be also able to send the next set of input parameters to Plaxis However an
additional means of connecting the interpreter with Plaxis is also needed This is can be
Structural Reliability Analysis
MSc Thesis A Rippi 49
achieved via a Plaxis interface which actually call Plaxis to make the calculations while it also
transfers the required value of the limit state function to input interpreter and this in turn to OT
Eventually the probability of failure is obtained as a model result However it is also essential
that Plaxis simulation procedure converges to the desirable criteria and under the physical
boundary conditions that have been determined Likewise the convergence criteria of the
reliability methods shall be manipulated so as the optimization algorithms to able to converge
efficiently (see Appendix A2 for further explanation)
44 Overview In section 42 a summary of the principles of the main reliability methods was made From
Level III methods MC and DS were presented whereas from Level II the basics of a FORM
analysis were introduced Besides the concept of the Level I method was explained Even
though this method was not used in this research for the evaluation of the probability of failure
it was applied in section 643 for the deterministic analysis of the system under investigation
and the preliminary calculations of the dimensions In Appendix F more reliability methods are
introduced and they are further described In the sequence their implementation in conjunction
with Plaxis was discussed through the coupling of FEM with OT For more information
concerning OT special features (ie SA optimization algorithms and probability distributions) a
reference is made to Appendices A and E
Before any of these reliability methods was applied on the case study of this thesis their
performance and their compatibility with Plaxis were tested with simple examples At first the
probability of failure was evaluated with FORM SORM Importance Sampling (IS) crude MC and
DS for simple artificial linear and non-linear limit state functions
and
respectively where R B and F are the random variables distributed normally
or lognormally The crude MC technique has been applied in order to validate the failure
probability in case the various methods give different results Secondly a flexible circular
footing over a one-layer soil material was used from Plaxis tutorial in order to test the coupling
between Plaxis and the reliability method and reassure the functionality of the interface In this
pilot example the objective was to determine the probability that the settlements of the footing
exceed a certain threshold of settlements and thus The respective
components that were taken into account for the failure mechanism was the cohesion (c) the
friction angle (φ) and the specific weight (γ) of the soil
From a qualitative evaluation of the reliability methods that has been done and according to the
experience gained from the aforementioned applications of some methods valuable conclusions
were drawn in terms of their efficiency More precisely regarding the sampling methods it does
not required previous knowledge of the failure domain apart from the IS method in which a
starting point inside the failure domain shall be provided Moreover the calculation effort
depends on the magnitude of the failure probability and the required accuracy whereas for the
IS it depends also on the selection of the sampling region In Waarts (2000) it has been proved
that DS tends to be more efficient than MC for low dimensional problems as it is indicated in
Figure 48 (number of random variables nlt100)
Structural Reliability Analysis
MSc Thesis A Rippi 50
Figure 48 Required number of samples for MC and DS as a function of the random variables
(source Waarts 2000)
As far as the approximation reliability methods (FORM and SORM) are concerned it is not
necessary to be aware of the failure region in advance Furthermore it has been observed that
the required iterations and the calculation time is quite lower in comparison with the sampling
methods However the accuracy of the method is highly dependent on the shape of the LSF
In Table 41 a summary of the most applied reliability methods evaluation is presented
According to this evaluation it was decided that the methods of FORM and DS as well as the
combination of them with RS techniques are going to be tested and evaluated in terms of their
efficiency and robustness These methods were chosen from both Level II and Level III methods
due to their expected reduced computational time the non-requirement of previous knowledge
about the LSF and their satisfactory performance in similar case studies In Chapter 5 the
different failure mechanisms and the relevant LSF that are going to be utilized in this thesis are
presented whereas in Chapter 7 the results deduced from the reliability analysis are elaborated
In Chapter 7 there is also an indication of which reliability method was used for each systemrsquos
component the selection of which was based on the aforementioned evaluation
Table 41Evaluation of reliability methods
Method Previous knowledge Accuracy Calculation Effort
FORM not required full accuracy for
Gaussian variables and linear LSF
depends on LSF linearity and the number of random
variables
SORM not required exact up to 2nd order
LSF error dependent on the shape of LSF
depends on LSF linearity and the number of random
variables
Structural Reliability Analysis
MSc Thesis A Rippi 51
Crude MC not required can be controlled by convergence criteria
depends on the magnitude of failure probability and the
required accuracy
IS required can be controlled by convergence criteria
depends on the choices made for the sampling region
DS not required can be controlled by convergence criteria
for low dimensional problems (nlt100) DS is more efficient
than MC
Especially RS were created based on a SA sampling in order basically to increase the efficiency
of FROM and DS analysis in case of FEM numerical errors Precisely an analytical formula or a
response surface was needed that can approximately represent the performance of the structure
under investigation and that is called in case of Plaxis errors This formula was essential in
order to keep the reliability analysis running by providing a response value for the LSF when
Plaxis calculation is unsuccessful and thus incapable of returning a result It should be
mentioned that in case a RS is implemented for the reliability analysis Figure 47 can be
reduced to a graph consisting of only the input file OT and the input interpreter as no Plaxis
calculations would be necessary any more The LSFE will then be performed through the
response surface that is provided by the user Due to this limited use of the RS method it will
not be further treated in this thesis
Structural Reliability Analysis
MSc Thesis A Rippi 52
MSc Thesis A Rippi 53
5 Failure Mechanisms and Limit State Functions
The reliability analysis of a structure requires the definition of the different failure modes that
are relevant to the corresponding structural elements In this chapter the possible failure
mechanisms of the systemrsquos elements expressed in LSF are identified and their role in the
system reliability is explained Based on these LSF the reliability of the system components was
evaluated These LSFs focus on the ULS whose exceedance implies failure of the corresponding
component
51 Introduction to the system analysis and the limit states For system reliability analysis all the possible failure mechanisms have to be identified and
summarized in a fault tree In Figures 51 and 52 the fault trees of a dike section without
structural elements and a retaining structure with sheet piles are depicted respectively In this
project the combination of these two separate structures ie a reinforced dike section with an
anchored sheet pile wall is to be simulated and studied in terms of its failure modes
As it is described later in this chapter failure can be expressed in different ways depending on
the structure and stakeholders demands and safety standards This research is mainly focused
on the ULS of failure and Mohr Coulomb was chosen as the soil failure criterion due to its
efficiency on detecting failure Such a failure criterion is actually defining the stress strain
relationship and the gradual weakening of the soil (ie the gradual reduction of the strength
parameters such as the friction angle and the cohesion due to the stresses development until
soil collapses) under a stress condition In Plaxis new version there is the ability of gradually
reducing structural properties such the strength of steel elements However the results
retrieved of such a safety analysis have not been yet investigated and for that reason such
calculations are not included in this project
As far as the dike section is concerned the failure mechanism to be investigated in this project is
the macro instability of the slope rather than the other mechanisms depicted in Figure 51 On
the other hand because of Plaxis limitations to distinguish among the different structural
elements failure mechanisms each element was considered as a singular case for investigating
its failure mode In the next section a distinction is also made between the ULS and the SLS
For a reinforced dike with retaining walls there are basically four classes of structural elements
(see Figures 12 and 21)
∙ Retaining wall (ie sheet piles)
∙ Anchors
∙ Walings
∙ Soil structure (ie dike section)
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 54
For the reliability analysis of each of the elements the probability of failure is to be determined
In the next sections of this chapter the respective LSF that are considered for each class are
further elaborated according to the ULS criterion while in the end the general combined fault
tree is to be schematized
Figure 51 Fault tree of a dike section
Figure 52 Fault tree for retaining structures using anchored sheet piles (source Schweckendiek 2006)
The system failure in this research is considered as a serial system of the anchor sheet pile
wall and soil body LSFs The fault tree of the system failure is presented in Figure 53 together
with the LSF of each component as they are formulated according to sections 522 and 523
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 55
Figure 53 Fault tree of the system failure
It should be mentioned at that point that the probability of failure of the system differs from the
overall probability of flooding which takes into account any potential failure mechanism of the
system combined with the uncertainties of the water elevation More precisely as far as the dike
safety is concerned a probability of failure (where ) under a certain
water level is estimated from the fault tree of Figure 51 which in sequence is multiplied
with the occurrence probability of the corresponding water level in order to estimate the
overall probability of failure for the specific water elevation Then that product is accumulated
over a required range of water levels that can jeopardise the overall stability of the structure in
order to calculate the overall probability of failure or else the probability of flooding This
probability is then compared to the one established from the safety standards so as to reassure
the safety or not of the structure can be estimated as follows
int
sum (51)
In this thesis the probability is to be estimated where is a specific water level
and especially the design water level as it is considered to be the most challenging part of the
procedure described above gives the probability of occurrence of the water level (ie it is
the PDF of ) After setting up the steps and implementing them successfully for the estimation
of the calculation of the overall probability of failure is just a repetition of the same
procedure for more water levels It should be mentioned that in this thesis failure consists only
of the macro-instability failure mode whereas the rest are excluded for the time being A
simplified alternative for the overcoming the integral of Eq51 is to repeat the calculations for
certain water levels and then sum their products instead of integrating the full range of the
them
52 Limit State Functions
521 Serviceability Limit State
The SLS is evaluated in design calculations in order to avoid excessive deformations of a
structure that could lead to the loss of its service requirements and its functionality In some
cases the SLS criteria are used to prevent damage to adjacent structures in the surroundings
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 56
For example concerning the design of a dike section attention should be paid to deformations
caused to buildings that are located in the inland part of the dike in case of an extreme event
In Figure 54 an example of a deformed dike is depicted In this figure some of the potential
locations that excessive deformations might be experienced are illustrated such as
∙ the vertical settlement of top of dike (arrow 1)
∙ the horizontal displacement of the top of the sheet pile (it might also be the middle part
of the sheet pile that ends up with the largest deformation according to the loading
conditions and the specific soil structure different deformations might occur) (arrow 2)
∙ Uplift upstream of the dike (arrow 3) special attention shall be paid to this type of
deformation not only due to the direct effect on inland structures but also as a sign of
developing piping mechanism
(a)
(b)
Figure 54 Figure (a) shows the initial undeformed structure while (b) illustrates the deformed mesh (not to scale) (the plotted length of the anchor does not have any physical meaning) In (b) the possible locations and directions (black arrows) of potentially excessive deformations
According to the current Dutch regulations (Larsen et al 2013) the SLS criteria for designing
reinforced dikes with sheet pile walls require that
∙ the settlement on the top of dike does not exceed 10 cm and
∙ the horizontal displacements of the sheet pile wall does not overcome 10 cm
In this research the SLS criteria and especially the dike settlements were used roughly for the
preliminary determination of the structures characteristics that are discussed in section 643
Dike Anchor
Sheet pile wall
1
2 3
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 57
More precisely displacements at the top of the dike were also taken into account as
complementary to the safety factor in order to determine the required length of the sheet pile
wall As far as the reliability analysis in concerned the LSF were formulated based on the ULS of
the different systemrsquos components that are presented in the next section
522 ULS for Structural Members
In the design process one is most interested in the ULS of a failure mechanism This state
describes the situation wherein the acting extreme loads are just balanced by the strength of the
construction If that limit state is exceeded the construction will lose its functionality and thus
collapse or fail In general most attention is paid to the behaviour of the structure after
completion However during construction there are also periods in which the construction may
fail The different phases of construction are listed in section 642 In the present section the
analytical LSF of the structural elements are given as they are going to be used in the reliability
analysis
Sheet pile wall
The most relevant failure mode for the sheet pile wall is the exceedance of the yield strength
which corresponds to the ultimate steel strength The response of the structure is mainly due to
bending moments and the axial forces (shear forces are considered to be negligible) Where an
axial force is present allowance should be made for its effect on the moment resistance
Figure 55 The water and earth pressures as well as the anchor forces lead to bending moments and axial forces on the sheet pile wall
In Figure 55 an example of the axial forces and the bending moments that can be developed in
the sheet pile wall with one anchor layer are illustrated after an extreme water level loading
Accordingly the maximum stresses on the sheet pile wall are composed of a bending moment
and a normal force component17
[
] (52)
where [kNm] and [kN] are the bending moment and the axial normal force
respectively that depend on the depth level where they are calculated over the sheet pile length
[m3] is the elastic section modulus and [m2] the cross-sectional area of the sheet pile
wall
17 the vertical anchor force component is reducing by its interaction with the soil over depth
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 58
Bending moment and axial force can be variable over the depth and that is why they are
expressed as a function of z-depth FEM has the advantage to take into account second order
effects ie a stiffer structure will experience higher bending moments than a more flexible one
Taking into account the above the limit state function can be formed as the difference between
the maximum developed stress and the yield stress
[
] (53)
where and can be characterized as the load variables while and can be considered
as the resistance variables and are assumed to be constant over depth
Concerning the permissible displacements for the ULS an upper limit for the maximum
horizontal displacements of the top of the sheet pile has been set to 150L where L is the
vertical length of the sheet pile wall (Larsen et al 2013) This requirement was mainly used in
section 643 where the structural properties were determined after the deterministic analysis
Anchors
Anchors are loaded by their reaction to the horizontal loads on the retaining walls The failure of
the anchor element is actually represented by the failure of the steel members of the anchor
(tubes bars cables etc) that are loaded by traction forces
As it is illustrated in Figure 56 the anchor experiences an axial tensile N that is nearly constant
over its length
Figure 56 Axial loading of anchor inside a dike
The elastic behaviour of an anchor involves only a relationship between axial force N and
displacement (elongation) u of the form
[ ] (54)
where EA [kN] is the anchor stiffness consisting of the steel Youngrsquos modulus E [kNm2] and the
anchor cross section A [m2] and L [m] in the length of the anchor
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 59
Similarly to the sheet pile wall the limit state function of the anchor involves the certain yield or
ultimate strength of the steel members and the maximum stress that the anchor experiences
during its loading Consequently the LSF is as following
(55)
where [kN] is the calculated anchor force and [m2] is the cross sectional area of the
anchor (both of them considered to be constant over the depth) It is essential to mention that
the anchor is also subjected to bending moments due to soil settlements (that are implicitly
illustrated via the uniformly distributed load q over the tie rod) that should be taken into
account in order to investigate the displacements of the tie rod itself However in this thesis
only the axial forces on the anchor are considered without taking into account the individual
deformations and its reaction with the surrounding soil
Walings
The waling is the element that transfers the loads from the retaining wall to the anchors (see
Figure 21) The loading of the walings can be schematized as a continuous beam on several
supports as it is depicted in Figure 57
Figure 57 Loading of walings
If Nα [kNm] is the calculated anchor force per m then the developed bending in the support Ms
and in the opening Mo can be approximated as follows
[ ] (56)
(57)
where La is the mutual anchor distance Considering as the design moment of the waling to be
the limit state function can be formulated as follows
Waling Sheet pile Anchor
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 60
(58)
For a conservative design of the waling the limit state function will give the same or lower
failure probability than the anchor itself Therefore it will actually not be necessary to carry out
this analysis The anchor failure will be the determinant mechanism (Schweckendiek 2006)
523 ULS for Soil Failure
In this thesis emphasis is given on the dike global instability which actually consists of several
failure modes Figure 58 indicates the possible failure mechanisms of flood defences The
combination if these can lead to the overall instability of the dike and thus to soil body failure
However Plaxis is not capable of accounting for all of them More precisely Plaxis assumes the
soil to be a continuous body and thus it can model movements in the scale of soil bodies (using a
relative fine mesh in relation to the moving soil body) but not on a soil particle level Plaxis can
simulate the groundwater flow in a soil body and from this someone can deduce input for a
piping analysis (see failure mechanism G in Figure 58) (using other softwaremethods) but as
it has been mentioned above it cannot determine movement of soil particles due to
groundwater flow Moreover Plaxis cannot deal with the flow and waves occurring in ldquoopen
waterrdquo ie water outside the soil in a canal a lake or sea for instance
Figure 58 Possible failure mechanisms of flood defences (source Flood Defences 2015)
Therefore Plaxis can handle the failure modes (C) (D) (E) (J) (K) and (L) of those illustrated in
Figure 57 In this thesis loads due to extreme temperature conditions or ships are not
considered while the settlements are taken into account implicitly though the mechanisms (C)
(D) and (E) These mechanisms are thus investigated in this thesis and they mainly refer to the
macro-instability of the inner and outer slope of the dike or the overall dike body (ie horizontal
sliding (mechanism D))
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 61
For these failure modes soil failure is defined as shear failure In Figure 59 the most relevant
patterns of the macro-instability regarding the shear strength of the soil are illustrated
Applying FEA the most critical failure mode is determined automatically However it is not
always straightforward what the trigger mechanism of the failure was and that is why more
investigation and FEA tests are needed in order to obtain a clear view
(a) Outer slope failure (b) ldquoActive siderdquo failure
(c) ldquoPassive siderdquo failure (d) Overall failure
(e) ldquoKranzrdquo stability
Figure 59 Macro-instability failure modes in a dike with an anchored sheet pile wall
In Figure 59 failure modes (a)-(d) follows Mohr-Coulomb failure criterion In principle the
total stress state inside a dike section consists of the effective stresses and the pore pressures
according to Terzaghi principle
[
] (59)
Mohr-Coulomb failure criterion basically considers as limit for the failure to occur the
maximum shear stress that the soil can withstand This is determined from the friction
angle the cohesion of the soil and the current stress state The drained shear strength
(similarly in for the undrained shear strength) in the soil according to Mohr-Coulomb is defined
as follows
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 62
[
] (510)
where is the effective friction angle In Appendix B1 a detailed explanation of this failure
criterion is attached
Taking into account the above mentioned failure mechanism (a) is mainly triggered by the
increase of the pore pressures in the outer slope which subsequently causes a decrease in the
effective stresses ( ) which at the same time leads to a reduction of the shear strength
(considering Mohr-Coulomb failure criterion) When the shear stresses in the soil exceed the
shear strength a slip plane forms and a soil wedge collapses
Failure modes (b) and (c) are primarily determined from the active and passive effective
stresses respectively As far as the ldquoactive siderdquo failure is concerned displacements of the sheet
pile wall towards the downstream side of the dike lead to development of active stresses
where is the active pressure coefficient During the ldquoactiverdquo failure the
retaining soil is allowed to relax which leads to a decrease of horizontal stresses and
simultaneously to a decrease in shear strength Similarly to failure mode (a) a soil wedge
collapses The result is an increased earth pressure on the wall for which it is not designed
However this type of failure in dikes is not such determinant for the soil and the sheet pile wall
failure because there is always the resistance of the passive side of the dike that keeps the active
side stable enough Such a mechanism is thus more relevant for deep excavations with retaining
walls
In mechanism (c) the passive soil resistance is exceeded by the horizontal loads In this case the
wall moves inland and a wedge of soil compresses The shear strength is larger due to the
deformation of the wall and the horizontal stresses increase stresses where is
the passive pressure coefficient ( ) This failure mechanism usually occurs due to an
underestimation of the sheet pile length or due to the presence of a weak soil layer in the
passive side
Failure mode (d) is actually a combination of (b) and (c) possibly accompanied also with the
development of excess pore pressures as it was discussed in failure mechanism (a)
Moreover too short anchors could cause a soil block to collapse as it is showed in Figure 59(e)
(ldquoKranzrdquo stability)
As it has been already mentioned for mechanism (d) failure modes can be correlated and
combined with each other in order to lead to a final failure state Therefore the common failure
probability would be smaller than the sum of the singular probabilities In this thesis the total
probability of failure of the soil body due to global instability is to be determined This is due to
the fact that the application of FEM is capable of simulating the combination of the failure
modes depicted in Figure 58 and thus the overall failure (related to (C) (D) and (E) modes of
Figure 58) can be evaluated However with the classical engineering approach for the stresses
calculation each one of the failure modes depicted in Figure 59 should be separately evaluated
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 63
Below a description of the available methods to formulate the LSF of the soil failure is given
After the evaluation of these alternatives a selection was made regarding the most suitable LSF
for the current case study
Limit State functions for soil failure
Soil mechanical failure can be analysed in several ways (Schweckendiek 2006)
1 Excessive Deformations
2 φ-c Reduction
3 Relative Shear Resistance
4 Plaxis definition of soil collapse
The possibilities and limitations of the aforementioned methods are briefly discussed below
1 Excessive Deformations
Similarly to the SLS a limit state function can be formulated by deformations that are
unacceptable As it is illustrated in Figure 54(b) there are some points (the top of the dike the
sheet pile top and the inland soil level behind the dike) the displacements of which shall be
limited to the minimum possible during the structurersquos lifetime in order for the structure to
meet the required service standards Therefore the calculated deformations at locations 1 2
and 3 could give a clear indication of failure given the maximum acceptable deformation
In case that there is a limit value for each location then the limit state function can be formed as
follows
[ ] (511)
with this criterion designer can control the developed deformation on the system and improve
the structure so as to limit the displacements to the required level and subsequently increase
the reliability
On the other hand such an approach might exhibit several problems during the reliability
analysis that are listed below
Many failure mechanisms suddenly happen and as a result no significant displacements
are observed before moving very close to failure In sequence this may cause problems
for the iterative procedures of some reliability methods such as FORM and DS
The determination of suitable maximum admissible deformations is not
straightforward They have to be large enough to serve as failure criterion and
simultaneously they shall not be larger than the values that can be calculated within
the limits of equilibrium in FEM-calculations This requires previous knowledge on the
analysed system as well as on the feasibility of FEM-calculations on the specific subject
Last but not least the location of the undesirable displacements in a dike section is also
under investigation as there can be multiple vulnerable spots on the dike that should
be kept under a certain range of deformations in order not to cause a collapse and that
they are not always known in advance
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 64
2 φ-c Reduction
An available type of calculation in Plaxis is the so-called ldquoSafety Calculationrdquo with which global
safety factors can be computed In Appendix B2 a more detailed elaboration of this method is
exhibited This feature can also be used in reliability analysis by formulating the LSF as follows
(512)
where is the Multiplier Safety Factor obtained by the φ-c reduction The main concept is
that if the safety factor is smaller than 1 then it is considered as failure This method can provide
us with the probability of failure for a general soil body collapse However even if this method
seems to be quite simple there are some issues during its implementation in a reliability
analysis
For complex limit state functions there can be convergence problems for some
reliability methods with iterative procedures (ie FORM) This can be explained from the
fact that in a φ-c reduction soil strength properties follows a certain path as it is
described in Appendix B2 independently of the values for the random variables that
have been set from the reliability method
The safety factor is a general safety factor regarding the failure of the system
Therefore sometimes it is not straightforward what the ldquotriggerrdquo factors are that led to a
certain type of failure
A safety calculation in Plaxis is time-consuming and the outcome can be unstable (Plaxis
calculationsrsquo convergence depends on the number of calculation steps)
Plaxis cannot handle safety factors below 1 Therefore a new limit of the safety factor
shall be introduced in this limit state function
3 Relative Shear Resistance
In this method the basic idea is to define soil failure according to the failure criterion that Plaxis
is set to use for the analysis In this research the Mohr Coulomb failure criterion is applied in
combination with a linear ndashelastic perfectly-plastic stress-strain relationship (that is the so-
called Mohr-Coulomb model) For this model the initiation of plasticity is considered as failure
Therefore the maximum shear resistance is defined just before plastic yielding occur for any
given stress state Of course the occurrence of plasticity does not directly indicate the failure of
the soil structure however this model can give a first estimate for the stress state and the
deformations
According to this method the relative shear resistance is defined as the ratio between the
mobilized shear stress and the maximum shear resistance In Figure 510 the Mohr-Coulomb
model is illustrated in a Cartesian stress-plane In the graph the principle stresses σ1 and σ3 are
indicated More precisely the principle stresses are defined as follows
radic
(513)
(514)
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 65
radic
(515)
Figure 510 Mohrs stress circle and Mohr-Coulomb failure criterion
Hence the mobilized shear stress is
radic
(516)
And the maximum shear resistance which is the distance of the mean stress point to the yield
surface is defined as
(517)
Then the relative shear resistance is a measure for load-resistance ratio that can be determined
in any integration point in the soil continuum
(518)
However this criterion requires prior knowledge of the possible relevant failure mechanisms
and their potential locations on the soil structure This is necessary in this method in order to
choose a suitable cluster of integration points where the average value of the relative shear
strength is to be determined Therefore the limit state function to be considered is not
straightforward and it needs problem investigation in advance
4 Plaxis definition of soil collapse
In Plaxis the construction stages are analysed by performing a Load advancement ultimate level
procedure which is controlled by a total multiplier
(519)
where is the load that is applied by Plaxis is the load at the beginning of the
calculation phase (ie the load that has been reached at the end of the previous calculation
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 66
phase) and is the defined load configuration (ie water level soil weight vertical load
on the top of the dike etc) Thus the main idea is that Plaxis applies an incremental load until it
reaches the defined one When has reached an ultimate level (which by default is 1)
the current phase is finished However if a staged construction calculation has not properly
finished the multiplier fails to converge to the desired ultimate level at the end of the
phase and a warning of ldquoSoil collapsesrdquo appears in the Log info box
In other words a collapse load has been reached In this case the total specified load has not
been applied In physical terms this means that the current value of the stiffness parameter CSP
is less than 0015 (Plaxis 2D 2015b) It is a measure for the amount of plasticity that occurs
during the calculation When the solution is fully elastic CSP is equal to unity whereas at failure
it approaches zero Therefore soil reaches its upper limit of plasticity and it collapses which can
be visualized as a settlement of the dike body Such a warning in the analysis is thus assumed to
be a possible failure situation
53 Overview In this chapter an overview of the relevant possible LSF of the structural elements and the dike
body respectively has been given The different LSFs are formulated and discussed whereas the
failure of each component is going to be evaluated in the ULS of each component
The system failure in this research is considered as the combination of the anchor sheet pile
wall and soil body LSF Particularly for the anchor and the sheet pile failure their yield stress
was chosen as a limitation for their failure by adopting the next LSFs
[(
)] sheet pile wall LSF (520)
anchor LSF (521)
As far as the soil body is concerned in this research method 4 is to be used in order to identify
soil failure The limit state function is then formed as a single value in case of a successful and an
unsuccessful computation Therefore the limit state function was chosen to be equal to 1
(actually the value of is retrieved from Plaxis that in case of a successful calculation is
1) in case of a successful Plaxis analysis and -1 in case that soil collapses in any phase (error in
Plaxis Log file ldquoSoil Collapsesrdquo) This formula can be written as follows
(522)
This accounts for the global instability of the dike body in contrast with the other methods that
are related to a certain dike area (methods 1 and 3) or they are cannot handle overall failure
due to numerical issues (method 2) More precisely methods 1 and 3 were excluded due to the
prior knowledge that is required in method 3 while in method 1 a definition of the maximum
admissible deformations shall be first introduced However attention shall be paid to the type
of the warning that Plaxis gives because it might also be that the maximum number of load steps
was not sufficient In that case the phase must be recalculated using a larger number of steps
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 67
Someone could argue that such a failure can be considered as a system failure rather than a soil
failure alone Partially this is true as the incapability of the sheet pile wall or the anchor to keep
the upstream part of the dike stable can lead to non-convergence and thus according to Plaxis
definition to failure This incapability can be translated into various scenarios of failures that
stem from either the structural elements or the soil body weakness and that are illustrated in
Figure 511 This picture shows the different scenarios that can take place due to macro
instability that was discussed in 523 and the failure of the retaining wallrsquos elements
Figure 511 Combined failure scenarios (initial and developed mechanisms) (source Larsen et al 2013)
However this does not mean that the sheet pile wall or the anchor fail because of exceeding
their yield stress but rather due to possible unacceptable deformations This is also an
advantage of the inherent residual strength of the structural elements that allows a large
deformation before they fail However the residual strength of the elements was not taken into
account in this thesis as they were considered as elastic Moreover the message for the ldquosoil
collapsesrdquo can appear in a phase prior to the installation of the wall which means that this type
of failure is exclusively attributed to the deficiency of the soil strength Therefore it is not
always straightforward what is failing first and lead to the system failure Consequently by
Failure Mechanisms and Limit State Functions
MSc Thesis A Rippi 68
considering such an analysis representative of a system failure it is likely to underestimate the
overall probability of failure whereas if it is considered as a soil failure only it might
overestimate the total probability of failure since it is later combined with the rest LSF of the
anchor and the sheet pile wall In this research it will be considered only for the soil failure that
brings the results to the safe side even if it does not lead to the best optimization In this thesis a
serial system of the anchor the sheet pile and the soil body LSF was considered the fault tree of
which is depicted in Figure 53
Another way to evaluate the reliability of the different elements could be the reach of a
threshold displacement according to the SLS criteria Such a threshold could be also considered
for the system as a whole rather than the individual elements However in that case someone
should be cautious regarding the choice of the value for the maximum deformations as well as
the location of their occurrence
Concluding in this section the possible failure mechanisms of a flood defence (Figure 58) were
shown Then it was clarified which of these mechanisms can be modelled in Plaxis and they
were further explained and schematized in Figure 58 with the presence of the retaining wall as
they would have to be handled in case of hand calculations Plaxis can incorporate these
mechanisms automatically and thus no distinguish among the different patterns of these
specific modes is necessary Last but not least in Figure 511 the possible failure modes for the
system of the dike and the retaining wall that can be simulated in Plaxis are depicted In this
thesis the plasticity of the structural elements is not considered and thus possible plastic hinges
in the anchor and the sheet pile are not taken into account Therefore scenario 3 is excluded
In Chapter 7 the results after the reliability analysis of the aforementioned LSFs are given It
was essential though before proceeding with the reliability analysis to get a better
understanding into the system under investigation For this purpose in the next chapter a
description of the case study as it has been simulated in Plaxis is given while also the different
structural properties are defined for the following reliability analysis The soil variables to be
considered as stochastic are presented and a first insight into of the systemrsquos behaviour is
obtained through the mean values calculation
MSc Thesis A Rippi 69
6 Case Study-Dike with an anchored sheet pile wall
In this chapter a description of the case study is presented together with the relevant soil and
structural parameters In sequence a deterministic analysis follows in order to specify the basic
structural characteristics and obtain a first sense of the stresses magnitude developed on the
structure This was carried out based on the recommendations given in the CUR 166 (2005)
The aim of this deterministic analysis is to redefine the structural properties in order to avoid
having an overdesigned structure understand the current design procedure and see if there are
any possibilities of improvement by applying the proposed reliability analysis
61 Case Description The case study concerns a dike section with an anchored sheet pile wall which was initially used
to showcase the current methodology of designing dikes with sheet piles using partial factors
For more information a reference is made to Breedeveld (2011)
In this research this case study has been modified and simplified for the needs of the thesis and
it is illustrated in Figure 61 while in Table 61 the soil materials are identified Details for the
properties of the soil layers are given in section 62 We assume the structure to be
homogeneous to the third dimension and therefore a plane-strain model is applied The dike
soil materials (layers 3 and 4 in Figure 61) were adopted from the case study used in
Breedeveld (2011) that was presented in Chapter 2 Especially in Figure 25 the original case
study as it was modeled in Plaxis is shown
Table 61 Soil materials that corresponds to Figure 61
1 Aquifer Sand layer 2 Aquitard Clay layer 3 Dike new material Clay layer 4 Dike old material Clay layer
The geometrical properties the groundwater level and the design water level were taken
identical to the original case study as well as the soil properties of the dike materials The soil
layers under the dike were modified for the need of the thesis and were modeled using random
average quantities for which the statistics were chosen arbitrarily but in a realistic range
according to the NEN6740- Table 1 (see Appendix C) included in the Dutch code for
geotechnical structures Moreover the vertical load was removed in order to examine the
impact of the water elevation on the dike stability
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 70
Figure 61 Case study Dike with anchored sheet pile wall Dimensions and Soil materials
62 Soil Parameters The soil parameters for the dike material were obtained from the report of Breedeveld (2011)
where for some parameters the characteristic values were available whereas for some others
the mean values were given For the soil layers below the dike Table 1 of NEN6740 was used as
it was mentioned before which gives characteristic values For those parameters whose value
was a characteristic 18 one a transformation was made according to their probability
distributions in order to acquire the mean values that are required for the reliability analysis
Only the properties required for the use of the Mohr-Coulomb model are presented
The parameters of each soil layer are presented in Tables 63(a) and (b) and the front number of
each layer indicates the corresponding soil layer as showed in Figure 61
The distribution types and the coefficients of variation for each parameter were chosen
according to the knowledge that has been obtained until now about the physically possible
ranges of such parameters and the recommendations that have been given in several researches
up to now (Baecheramp Christian 2003 Schweckendiek 2006 GriffithsampFenton 2007 Phoon
2008 Wolters 2012 Bach 2014) The distributions and the CoV of all parameters are displayed
in Table 62 They were considered to be the same for all soil layers Particularly for the
saturated and unsaturated volumetric weight γsat and γunsat respectively a relationship was
established in order to derive the one from the other In general γunsat varies between the real
dry weight and γsat depending each time on the degree of saturation An estimated maximum
difference between γunsat and γsat is approximately 4 kNm3 (Brinkgreve 2015) In this thesis γsat
was expressed as the summation of γunsat and a variable with uniform distribution in the range
of [0 2]19 Initially a deterministic relationship was defined between the two variables but later
on it was proved the variable was important for the soil failure and thus it was decided to
consider it as random
18 for the definition of the characteristic value see Appendix E 19 The range of [0 4] was not chosen in order to avoid extreme conditions
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 71
Table 62 Soil parameter distributions
Soil parameter Symbol DistributionRelation COV Unit
Unsaturated
Volumetric weight γunsat Normal (microσ) 5 [kNm3]
Saturated Volumetric
weight γsat γsat= γunsat+U(02) 5 [kNm3]
Cohesion c Lognormal (microσ0) 20 [kPa]
Friction angle φ Truncated normal (microσ045) 10 [ ˚]
Youngrsquos Modulus E Lognormal (microσ0) 25 [kPa]
Poissonrsquos ratio ν Truncated normal (microσ005) 10 [-]
Interface strength Rinter Truncated normal (microσ0099) 20 [-]
Table 63 (a) Soil statistical properties under the dike and (b) soil statistical properties in the dike
(a) 1 SAND (very silty)
Soil property Symbol Characteristic
value
Mean
value STD Unit
Saturated
Volumetric
weight
γsat 20 22 11 [kNm3]
Unsaturated
Volumetric
weight
γunsat 19 21 103 [kNm3]
Cohesion c 0 0 0 [kPa]
Friction angle φ 30 36 36 [ ˚]
Dilatancy angle ψ 0 0 0 [ ˚]
Youngrsquos Modulus E 20000 30769 7692 [kPa]
Poissonrsquos ratio ν 025 03 003 [-]
Interface strength Rinter 044 066 013 [-]
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 72
2 CLAY (clean medium)
Soil property Symbol Characteristic
value
Mean
value STD Unit
Saturated
Volumetric
weight
γsat 17 185 093 [kNm3]
Unsaturated
Volumetric
weight
γunsat 17 185 093 [kNm3]
Cohesion c 10 14 282 [kPa]
Friction angle φ 175 21 21 [ ˚]
Dilatancy angle ψ 0 0 0 [ ˚]
Youngrsquos Modulus E 2000 3077 769 [kPa]
Poissonrsquos ratio ν 029 035 0035 [-]
Interface strength Rinter 034 05 01 [-]
(b) 3 DIKE NEW (very sandy clay)
Soil property Symbol Characteristic
value
Mean
value STD Unit
Saturated
Volumetric
weight
γsat 17 185 093 [kNm3]
Unsaturated
Volumetric
weight
γunsat 17 185 093 [kNm3]
Cohesion c 4 564 113 [kPa]
Friction angle φ 29 347 347 [ ˚]
Dilatancy angle ψ 0 0 0 [ ˚]
Youngrsquos Modulus E 1625 2500 625 [kPa]
Poissonrsquos ratio ν 029 035 0035 [-]
Interface strength Rinter 034 05 01 [-]
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 73
4 DIKE OLD (little sandy medium clay)
Soil property Symbol Characteristic
value
Mean
value STD Unit
Saturated
Volumetric
weight
γsat 195 212 106 [kNm3]
Unsaturated
Volumetric
weight
γunsat 19 207 103 [kNm3]
Cohesion c 13 183 367 [kPa]
Friction angle φ 28 335 335 [ ˚]
Dilatancy angle ψ 0 0 0 [ ˚]
Youngrsquos Modulus E 2925 4500 1125 [kPa]
Poissonrsquos ratio ν 029 035 0035 [-]
Interface strength Rinter 034 05 01 [-]
63 Finite Element Model The structure has been modelled with the finite element software Plaxis 2D 2015 In Figure 62
the model as it was created and the automatically generated mesh are presented The ldquoboxrdquo
around the dike section was used to refine the mesh in this area The mesh in the interfaces
between the sheet pile and the soil was also refined for a better representation of the potential
high stresses that can develop in that area
Figure 62 FEM model and generated mesh for case study
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 74
The sheet pile has been modelled with elastic20 plate elements and the free anchor length with
an elastic fixed-end-anchor element The grout body of the anchor was not explicitly modelled
but it is implied in the end of the anchor as its final point is considered to be fixed in the soil A
fixed-end-anchor was chosen because even if it is a simplification of an anchorage system it can
still give reliable results in terms of the anchor axial force which is actually included in the limit
state function However the interaction of the soil and the grout body in that case cannot be
simulated The interaction and the modelling of a grouted body in the soil are still under
investigation due to several difficulties in modelling the real conditions
For all the steel members a Youngrsquos modulus of E=210 GPa was used while the specific
structural parameters are to be specified after the deterministic analysis that is described in the
next section
64 Deterministic Analysis The design concept of dikes with sheet piles modelled in FEM is based on the combination of
several partial factors in order to evaluate the overall required safety factor Additionally the
design recommendations of CUR 166 (2005) are considered for using the design or the
characteristic values of the structural parameters and the loads The design procedure (see
section 22) that has been initially followed in this structure is reported in Breedeveld (2011)
and the following structural elements have been defined accordingly
Cross sectional area of the sheet pile (per meter)
Length of the sheet pile wall
Steel quality of the piles
Cross sectional area of the anchor (per meter)
Free length of the anchor (without the grounded body)
Steel quality of the anchorrsquos reinforcement
Profile of waling
The required overall safety factor was calculated up to 18 according to the current design
procedure described in section 22 according which the above mentioned structural parameters
were defined Moreover apart from the overall safety factor in Breedeveld (2011) additional
partial factors are applied on the deduced forces and bending moments of the anchor and the
sheet pile wall as they are defined in CUR 166 (2005) In particular in Table 64 the different
partial factors that are applied on the several structural developed actions are given
Table 64 Partial factors for the different structural actions according to CUR 166 (2005)
Structural action Partial factor γ
Sheet pile normal force Fn 115
Sheet pile bending moment M 115
Anchor normal force FA 125
Waling bending moment Mw 110
20 The elements were considered as elastic in order to avoid numerical problems in FORM analysis For a better explanation a reference is made to the evaluation of the optimization algorithms in Appendix A2
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 75
However as it is mentioned in section 62 alterations have been made in the soil layers below
the dike while also the vertical load has been removed On the top of that in section 22 it was
shown that the large applied partial factors led to an overestimation of the design values for the
moments of the sheet pile Therefore a redesign of the sheet pile and the anchor is carried out
in this section in order to determine the new required characteristics of the structural elements
according to the new loading situation with less strict partial factors In particular in this thesis
the required overall safety factor was considered 12 according to CUR 166 (2005) calculation
scheme that is described in the next section The reason for using a lower safety factor was to
avoid an overestimation of the design forces and moments of the sheet pile wall that would
render a reliability analysis meaningless as the structure would be quite safe Such an analysis is
also important in order to show the discrepancy between the two different design procedures
and indicate the optimal one
However the reliability analysis was carried out for the mean values of the soil parameters and
for a safety factor of 10 in order to account only for soil properties uncertainties A safety factor
of 10 was not directly used as the design procedure followed in section 643 accounted only for
the macro-instability of the system and thus other failure mechanisms such as piping
overtopping infiltration etc were not taken into account In that way we prevent coming up
with a retaining wall that would be volatile under other crucial failure mechanisms
In section 641 the calculation scheme to be used in this thesis for the deterministic analysis is
further described In the sequence in section 642 the construction phases that were followed
in Plaxis simulation are listed and illustrated and finally in section 643 the results of the
preliminary design are displayed
641 Calculation Scheme and Design Values
In this research only the design philosophy suggested in CUR 166 is applied and the desired
safety factor of the structure is considered to be 12 for structures classified in ldquosafety class IIIrdquo
in order to avoid an overestimation of the structurersquos dimensions as it happened in the original
case study At this point it should be also clarified that the redesign was conducted considering
only the overall stability of the structure and no other failure mechanisms such as piping
internal erosion or overtopping
The prescribed calculation schemes in CUR 166 (2005) for designing structures with retaining
walls are basically the following
Calculation scheme A Calculations with design values The calculations are executed
using the design values for the soil parameters retaining height water levels and
stiffness of the structure It uses two different soil stiffnesses When using a high soil
stiffness the anchor force is generally relatively high The advantage of this scheme is
that it requires relative little effort The disadvantage is that the deformations may be
overestimated due to the use of design values in every stage However a designer is not
interested in deformations when performing a ULS calculation
Calculation scheme B Calculations with characteristic values The calculations are
executed using characteristic values of the soil strength (crsquo and φrsquo) Design values are
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 76
used for the retaining height water levels external loads and stiffness of the soil Here
again two different soil stiffnesses are used In the end of the governing stage a phi-c
reduction should be carried out As the characteristic values for the soil parameters are
used the deformations are most probably smaller than in case of calculation scheme A
The safety margin lies in the fact that phi-c reduction is imposed to obtain a safety factor
of 115 (class II) or 12 (class III)
For the soil stiffness parameters two options are given in CUR 166 a low and a high value
(61)
where is the (low and the high) design value the (low and the high) characteristic value
and is the mean value of the stiffness parameter In Woltersrsquos master thesis (2012) both
have been applied and the appeared to be governing for the bending moment in the wall
and the anchor force
In this research the Calculation scheme B was used and the as the soil stiffness
parameter It should be mentioned that regarding the soil parameters the characteristic values
were available with an exemption for the stiffness for which the mean values were given The
groundwater level was manually schematized according to TAW (2004) because of lack of data
while the maximum water level in the river side was taken equal to the design water level
according the Dutch regulations for the specific dike region
In Tables 65 and 66 the partial factors according to CUR 166 and the corresponding design
values for the different soil properties are presented As someone can notice in these tables
only the strength parameters (phi and c) and the stiffness parameter (E) are exhibited For the
rest parameters the design value is identical to the characteristic value (see Tables 63(a) and
(b)) as a partial factor of 1 is used
Table 65 Partial factors and design values for the soil layers under the dike
1 SAND (very silty)
Soil property Symbol Partial factor γk Design value Unit
Cohesion c 11 0 [kPa]
Friction angle φ 12 2569 [ ˚]
Youngrsquos Modulus E 13 1538462 [kPa]
2 CLAY (medium)
Soil property Symbol Partial factor γk Design value Unit
Cohesion c 11 91 [kPa]
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 77
Friction angle φ 12 1472 [ ˚]
Youngrsquos Modulus E 13 153846 [kPa]
Table 66 Partial factors and design values for the soil materials in the dike
3 DIKE NEW (medium clay)
Soil property Symbol Partial factor γk Design value Unit
Cohesion c 11 364 [kPa]
Friction angle φ 12 2479 [ ˚]
Youngrsquos Modulus E 13 1250 [kPa]
4 DIKE OLD (stiff clay)
Soil property Symbol Partial factor γk Design value Unit
Cohesion c 11 1182 [kPa]
Friction angle φ 12 239 [ ˚]
Youngrsquos Modulus E 13 2250 [kPa]
642 Construction Stages
The construction and the gradual loading of the dike ware modelled as follows (see Figure 63)
1 K0-procedure for the generation of the initial stresses under horizontal groundwater
level
2 Dike self-weight under horizontal groundwater level
3 Rise groundwater level to the phreatic water elevation
4 Interpolation21 of the clay layer under the dike and activation of the sheet pile wall and
the anchor
5 Apply extreme water level conditions on the dike structure
6 (φ-c reduction for the determination of the safety factor)
21 Due to different permeability between the clay and the sand layer an approximation of the developed pore pressures is carried out through interpolation between the two layers
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 78
1
2
3
4
5
Figure 63 Construction and loading stages
For the soil modelling the Mohr-Coulomb model was used with a non-associative flow rule
(φneψ) For phases 1-4 the calculations were conducted under drained conditions whereas in
phases 5 and 6 the undrained behaviour of clay layers was considered
643 Determination of the structural elementsrsquo characteristics
According to the calculation scheme that was described above the design characteristics of the
structural elements were obtained from phase 6 In particular the target overall safety factor is
12 with which initially the length of the sheet pile wall was calculated
+370 NAP
+700 NAP
+1050 NAP
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 79
Firstly a length of 10 m and a profile of AZ 12 with a steel quality of 240 kNm2 yield stress
were used while for the anchor a cross-section of 628 mm2 and S355 were chosen for the first
calculation However even if the safety factor reached 12 it seemed to be quite unstable during
the calculation steps while the developed anchor stress was above the ultimate yield stress
What is more the deformations resulted in that case were more than 30 cm Therefore in order
to stabilize the safety factor and to decrease the displacements the length of the sheet pile was
increased 1 m (so 11 m in total) while also a better steel quality for the sheet pile was chosen
S355 Furthermore the anchor cross section was increased to 933 mm2 with a steel quality of
MW450 Last but not least the waling was designed based on the anchor force
In the following paragraphs the design procedure of each element is described providing also
the results of moments and forces that were retrieved from Plaxis
Sheet Pile wall
As it was partially mentioned in the previous paragraph the length of the sheet piles is
considered to be sufficient when the safety factor of the system22 reaches at least 12 For the
reasons discussed above the sheet pile length was decided to be 11 m and with a profile of AZ
12 S355
With this configuration the safety factor reached is 13 as it is depicted at the right graph of
Figure 64 Someone can notice that the safety factor from 117 in the situation without a sheet
pile structure increased to 13 with the structure
Figure 64 Safety factor deduced from the phi-c reduction calculation without the anchored sheet pile wall at the left (safety factor=117) and with the wall at the right (safety factor=13) The steps on the x-axis represent the calculation steps have been carried out in order to reach an equilibrium after incrementally reducing the strength parameters ie phi and c
The design moment is to be taken for a safety factor of 12 that is reached in step 34 where it
was found to be 1519 kNmm and the corresponding axial force 9133 kNm Therefore
considering Eq 52 for checking the exceedance of the ultimate yield stress and applying the
prescribed partial factor of 115 for both the moment and the axial force we get
22 Plaxis always gives the safety factor of the system which is under investigation
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 80
In Figure 65 the maximum bending moments with a safety factor 10 and 12 are respectively
showed From the values being shown in that figure someone can easily distinguish the
difference between the two moments It must be also mentioned in that point that the required
safety factor defined in Chapter 2 was estimated at 18 (and not 12 that is considered here)
which gives excessive design moments (almost 7 times larger) For that reason in this research
a lower safety factor was assumed avoiding thus the design of a quite conservative structure
Figure 65 Bending moments on the sheet pile for a safety factor safety factor=10 at the left and
safety factor=12 at the right
The explanation behind this discrepancy is that in case of the phi-c reduction procedure the
artificially decreasing strength of the soil and on the other hand the non-reduced strength and
stiffness23 parameters of the structural elements tends to lead in an increased arching
phenomenon that makes the structure attract the most of the load
In this research the reliability of the structure will be evaluated for a safety factor of 1 or else
during phase 5 and by considering the mean values of the soil parameters As someone can
notice from Figure 65 the magnitude of the maximum bending moment for a safety factor =10
is quite low which implies a possible minor contribution to the overall failure
At this point it is essential to mention that Mohr-Coulomb model is capable of simulating the
elastic behaviour and the state at failure Therefore in case that plasticity of the soil or the
structural elements plays an important role a different constitutive model is recommended to
be applied
Anchor
For a safety factor 12 the anchor force was found to be 10387 kNm Considering a mutual
anchor distance of 3 m the total axial force that is undertaken by the rod multiplied also with
the prescribed design partial factor is
23 ldquoStrengthrdquo and ldquostiffnessrdquo are usually confused terms Strength is a measure of the maximum load that can be placed on a material before it permanently deforms or breaks and it is represented by the yield stress σy in metals and by φ and c in soil Stiffness is a measure of the amount of deflection that a load causes in a material and it is represented by the Youngrsquos modulus E and the Poisson ratio ν both in soil and metals
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 81
Therefore choosing an anchor with a cross sectional area of A = 933 mm2 and a steel quality of
MW450 ( ) the design stress level is
In comparison with the sheet piles the anchor does develop large stresses as it was concluded
from the results in the deterministic calculations This is probably attributed to the high
stiffness of the anchor that apparently attracts the load and thus the anchor develops large axial
force
Waling
The type of the waling system is defined according to the anchor force as it was described in
section 522 Therefore by choosing a waling type of 2UPE 200 the developed bending
moment can be estimated as follows
where is the design bending moment of the walling system according to the chosen profile
In Table 67 the characteristics of each structural element that was described above are
displayed in detail
65 Overview In this chapter an introduction of the case study as it has been used in this research is made The
soil materials and their properties are summarized and their mean and design values are
defined The probability distributions of the soil parameters to be considered as stochastic were
proposed according to the literature study and the engineering judgement Moreover the mesh
generation and the modelling features are discussed
Last but not least due the observed overestimation of the retaining wallrsquos design moment and
axial force according to the current design regulation a new deterministic analysis was carried
out in order to redefine the structural properties and come up with a less conservative structure
This makes the reliability analysis of the system meaningful in a sense that an extremely safe
structure would give a failure probability of zero Besides the scope of a reliability analysis is
also the optimization of the construction that corresponds to a financially attractive and
simultaneously safe structure This can be made by conducting a risk analysis whose first step
would be the analysis that is carried out in this research Therefore the aim is to start with the
marginal required structural properties according to the safety standards In a risk analysis that
should conducted for several dimensions of the structure until the most optimized one is
identified However this is out of the scope of this thesis
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 82
In Table 67 the structural properties of the sheet pile wall the anchor and the waling are
summarized as they were found in the deterministic analysis
Table 67 Design parameters of structural elements
SHEET PILES
Property Symbol Value Unit
Profile - AZ 12 [-]
Steel quality - S355 [-]
Length L 11 [m]
Moment of inertia I 21430 [cm4m]
Elastic section modulus Wel 1245 [cm3m]
Mass w 943 [kgm]
Sheet pile thickness d 85 [mm]
Cross sectional area A 1201 [cm2m]
ANCHOR
Property Symbol Value Unit
Steel quality - MW450 [-]
Free length Lafree 104 [m]
angle φ 30 [deg]
Cross sectional area A 933 [mm2]
Mutual anchor distance s 3 [m]
WALING
Property Symbol Value Unit
Profile - 2UPE200 [-]
Steel quality - S355 [-]
Elastic section modulus Wel 191 [cm3m]
In the original case study the sheet pile wall is 162 m long and it has a profile of AZ 28 (S240)
while the anchor is an MW450 with a cross-section of 1134 mm2 As it is noticed both sheet
pilesrsquo and anchorrsquos properties have been reduced after the new deterministic analysis However
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 83
it should be mentioned that in the original case there was a vertical load on the top of the dike
that was displaced while the clay layers under the dike was merged into one in the new case
study These changes might lead to less heavy structure However the difference among the
design properties between the new and the original case study is mainly due to the required
safety factor that was assumed In the new case study the safety factor was set to 12 whereas in
the original cases study a safety factor of 18 was considered as it was calculated according to
the partial safety factors
In the next chapter the results after the reliability analysis of the just designed case study are
presented
Case Study ndash Dike with an Anchored Sheet Pile Wall
MSc Thesis A Rippi 84
MSc Thesis A Rippi 85
7 Reliability analysis results with stochastic soil properties
In this chapter the probability of failure of the different structural elements that were discussed
in Chapter 4 will be estimated by varying the soil properties of the several soil layers under and
inside the dike At the beginning a better insight into the system behavior is gained by
inspecting the deterministic Plaxis calculations for the mean values of the soil variables After
that a global SA is carried out in order to evaluate the most important soil parameters and
inspect their impact on the different structural elements in combination with each other and
individually Finally the results regarding the probability of failure of the anchor the sheet pile
wall and the soil are presented as they deduced by the reliability analysis (concerning the ULS of
the elements) Finally the systemrsquos reliability is evaluated The methodology that was followed
during the probabilistic analysis is described below and it is mainly dedicated to the DS
rationale that was used in order to obtain the influence factors and to evaluate the results of
Plaxis calculations
71 Method description For the calculation of the failure probability of the systemrsquos different components FORM and DS
were used for the retaining wall and the soil body respectively From these methods the
probability of failure and the important factors are going to be deduced for each component (ie
anchor sheet pile wall and soil body) However before starting with the reliability analysis it is
essential to carry out a sensitivity analysis in order to obtain a first impression of what are the
most influencing variables and thus determine the input random parameters for the reliability
analysis In the figure below the steps for the reliability analysis as they were followed in this
thesis are presented
Figure 71 Steps of the reliability analysis
Below the methodology of both the sensitivity and the reliability analysis are discussed and
explained in detail
Sensitivity analysis
In this thesis the FAST method was utilized for conducting the sensitivity analysis which is
further elaborated in Appendix A1 Here an explanation is given for the quantitative measure of
sensitivity which is represented by Sensitivity Indices The first-order sensitivity index of an
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 86
input variable pi is the measure of the main (direct) effect of pi on the output variance
(where i ne j) the second-order sensitivity indices measures the interaction effect of pi and pj on
the output variance Other higher-order indices are defined in the same manner The total
sensitivity index is the sum of all sensitivity indices involving factor pi (Homma amp Saltelli
1996) For example the total sensitivity index of factor 1 for a model with 3 input factors is
given as
The first-order (Si) and total ( ) indices can be interpreted as following (Pandya et al 2008)
high pi is an influent parameter
and both small pi is not an influent parameter (neither alone nor in interaction
with other parameters)
and nearly the same no interaction of pi with the other parameters
and very different high interaction of pi with the other parameters
Reliability analysis
As far as the FORM analysis is concerned section 422 gives sufficient information about the
methodology that is followed in order to obtain the failure probability and the influence factors
(ie α-values) The main concept and principles of DS are briefly explained in section 421
whereas here a more precise description of the implementation of DS and the how the α2-
values can be deduced is given
In particular DS is a method in which ldquodirectionsrdquo are sampled The concept of this method is
schematically depicted in Figure 72(a) For each sampled direction it is evaluated whereas
along this line the LSF Z equals to zero or not This procedure is repeated for a number of
directions and this method is applied in the standard normal space (u-space see also section
422) using equations to transform the standard normal variables u1 hellip un to their
corresponding ldquorealrdquo variables (in the x-space) X1 hellipXn and vice versa
(a) (b)
Figure 72 (a) Schematic view of the DS method in the u-space The numbers 0 1 2 3 4 show successive LSF evaluations along one direction (b) Schematic view of the line search algorithm β is the distance from the origin in the u-space along one direction The numbers 0 1 2 3 4 correspond to the ones shown in figure (a) (source Hydra-Ring 12 2015)
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 87
The search procedure for the location on the line where Z=0 is often referred to as the ldquoline
search methodrdquo and it is depicted in Figure 72 (a) and (b) In both figures the numbers 0-4 refer
to successive Z-function evaluations along one direction Figure 72(a) shows 3 evaluated
directions in the u-space whereas Figure 72(b) shows the (fictional) value of the Z-function
along one direction (where ||u|| is the distance from the origin along the line) The maximum
length of the direction line as well as the step size along the direction (for example the distance
between 0 and 1 along one line) can be steered in order to increase the efficiency of the method
and presumably decrease the computational time
In the sequence a method was developed in this thesis in order to estimate the α2-values of the
random variables out of a DS analysis as OT does not have an available method so far For that
purpose a transformation of the output samples to u-space was firstly carried out Then the
distance to the origin of all samples (in u-space) that are located on a direction where failure
(Z=0) was detected is calculated as follows
radicsum
i=1hellip (71)
where n is the number of random variables and uij is the ith ldquofailurerdquo sample of the jth random
variable is equivalent with the reliability index of this direction βi The α-values for each
random variable can then be calculated as (for more information about these values see
sections 422- Eq 417 and 423)
(72)
After that three different methods were created in order to evaluate the influence of the
random variables on the response of the LSF ie the α2-values (square of Eq 72)
Shortest distance ldquoβminrdquo
Average 10
Average all
In the Shortest distance method the sample with the smallest distance to the origin is
considered to be the design point and thus the corresponding influence factors are estimated
according to that point The smallest distance is also equivalent to the reliability index βmin This
method is well-known for estimating the design point out of MC techniques (Vrijling amp Van
Gelder 2002) However for a soil failure modeled in FEM this might not be always indicative of
the real influence that the soil properties can have on the system failure Therefore the last two
techniques were used as additional in order to get a better insight into the dominant soil
properties In Figure 73 a general scheme of the methodology that was followed for obtaining
the α2-values is depicted where also the different techniques are presented
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 88
Figure 73 DS methodology for calculating α2-values
The second method Average 10 takes into account the α2-values of the samples in the failure
domain whose distance to the origin lies within the 10 higher than βmin and averages them
Therefore it is firstly required the calculation of βmin as it was described for the Shortest distance
method Such a technique helps to identify the other directions that are close to the one which
gives the shortest distance to the origin and thus inspect if the important variables that were
indicated according to the first method are influential for other failure mechanisms as well
Last but not least the third method Average all averages the α2-values of all the samples
located in the failure domain This shows the overall contribution of the random variables to the
response of the LSF under investigation Concerning the soil failure LSF that is described in
section 73 such a method can be more suitable for distinguishing the soil properties that are
involved in the majority of the failure mechanisms rather than only in the one that the first
method indicates
72 Mean values calculations The results of the deterministic analysis that are presented in the previous chapter included a
degree of conservatism due to the load and material factors (partial safety factors) In order to
overcome this limitation structural reliability calculations were carried out and are discussed in
the next sections by considering relevant soil parameters as random However before starting
the probabilistic calculations by coupling Plaxis with OT that cannot be fully interpreted and
verified to a certain extent it is essential to first carry out some deterministic calculation Such
calculations can help to get a better insight into the behavior of the system and show
qualitatively what kind of response to wait later in the analyses These calculations were taken
place based on the mean values of the soil properties
Deformations
The analysis of this case study focuses on the ULS of the structure For the stresses and
deformation calculations the Mohr-Coulomb model is used that is generally accepted for the
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 89
simulation of the soil behavior until failure and for detecting the failure modes (a more detailed
explanation of this model is provided in Appendix B1) However this model is not so suitable
for analyzing the deformations of soft soils or the settlements around the wall
The deformation pattern that would more likely occur due to primarily the soil body failure can
be illustrated by conducting a ϕ-c reduction calculation In Figure 74 (a)-(c) the displacement
pattern after the ϕ-c reduction of the mean values of the friction angle and the cohesion
respectively As it can be seen a slip surface in the inner side of the dike has been created that
rotates towards the inland
(a) deformed mesh
(b) shadings
(c) arrows
Figure 74 Deformations after the ϕ-c reduction (mean values)
Stresses on the structural elements and the soil body
The anchor and the sheet pile wall are installed in phase 4 In phase 5 the anchor normal force
increases and especially for the mean values it amounts to Na = 7983kNm
As far as the sheet pile wall is concerned the bending moments and the normal forces that are
developed on the wall are indicated in Figure 75 as they were deduced from phase 5 Similarly
to the anchor the maximum stresses occur in phase 5 and are mainly located in the area of the
wall which is surrounded by the clay layer (brownish soil layer in Figure 74(a)) As someone
can observe the bending moments are not significant taking into account that the maximum one
observed reaches 2132 kNm
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 90
Figure 75 Bending moments and normal forces for the mean values calculation
For the interpretation of the stress field of the soil body the effective principle stresses and the
relative shear stress τrel (see Eq 517) were illustrated in Figure 76 (a) and (b) respectively In
Figure 76 (a) the crosses indicate the direction and the magnitude of the principle stresses σ1rsquo
and σ3rsquo and from which an active soil behavior (
) can be noticed on the river side of
the retaining wall whereas a passive behavior (
) is dominant in the inland side
(a) Effective principle stresses
(b) Relative
shear stress τrel
Figure 76 Effective principle stresses and relative shear stress in mean values calculation
In Figure 76 (b) the distribution of the relative shear strength shows the potential developing
plastic area in case of failure that expands from the river side of the wall and around it It can
also betoken the possible shape of formation of a slip surface that starts from the outer side of
the dike and goes around the wall
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 91
73 Sensitivity Analysis Results In section 312 different types of SA are presented and that can be applied through OT In this
research due to the time limit and the amount of parameters FAST method was used for
carrying out the SA as it allows the calculation of the first order sensitivity indices as well as the
total order indices and the computational time required is less than the other available methods
In this section the results from the conducted SA are presented separately for the anchor the
sheet pile and the soil performance Regarding the sheet pile wall and the anchor the sensitivity
of the total developed stresses towards the soil properties was evaluated according to Eq 53
and 55 respectively The influence of the soil parameters on the soil behaviour was evaluated
based on the calculation of the overall safety factor after a safety calculation (ϕ-c reduction
method) The total order indices are considered for distinguishing the most important
parameters while their difference with the first order indices Si indicates the level of interaction
effect amongst the parameters on the output variance
As far as the anchor and the sheet pile wall are concerned the SA was carried out in two steps
(1) SA with all the soil parameters (listed in Table 62) and (2) SA with the most important
parameters found in step 1 In Figure 77(a) and (b) the sensitivity indices from step (1) are
presented as they were found for the anchor and the sheet pile respectively Step 1 mainly
provides an overview of the contribution of the variablesrsquo variance to the total variance of the
limit state response and it helps to identify the most influencing parameters Step 2 was
basically carried out in order to build the response surface based on the SA output that was
later utilized in the FORM analysis as a complementary function in case of Plaxis errors At that
point it should be mentioned that the Dikenew material was not considered neither for the SA
nor for the reliability analysis of the structural elements as its influence was assumed to be
insignificant relative to the other soil layers However Dikenew was considered in case of the
soil body investigation
From Figure 77(a) it can be concluded that as far as the anchor is concerned the shear stiffness
G of the clay layer seems to be the most crucial soil property In particular the anchorrsquos stress
level seems also to be sensitive towards the Poisson ratio of the clay layer ν the friction angle φ
of the sand layer and the strength parameters φ c of the Dikeold material Moreover the
interface strength Rinter between the Dikeold material and the sheet pile wall seems to display
an additional notable influence as well The unit soil weight γ does not appear to affect the
anchor performance significantly
As far as the sheet pile wall is concerned the shear stiffness of the clay layer and the Dikeold
material seem to be dominant according to Figure 77(b) From these two soil layers it appears
that also the strength properties (φ c and Rinter) as well as the unit soil weight γ have a relative
impact on the stress level of the wall Furthermore the sand layer contributes via especially its
friction angle and shear stiffness
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 92
(a)
(b)
Figure 77 Total and First order sensitivity indices for the (a) anchor and the (b) sheet pile in step 1
As it was mentioned before step 2 was used for building the response surfaces for the anchor
and the sheet pile limit states The input random variables in this step are the most influencing
ones that were found in step 1 In Table 71 the soil properties that were considered as random
during step 2 are presented for both anchor and sheet pile It is essential to mention that the RS
was constructed by fitting a quadratic polynomial function to the sampling points of the SA
whose general expression is given in Eq 425 In Figure 78(a) and (b) an illustration of these
response surfaces is made However the figures depict the response of the anchor and the sheet
pile stress level over the fluctuation of only two variables amongst the 14 that are presented in
Table 71
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 93
Table 71 Soil properties considered for the RS construction of the anchor and the sheet pile respectively
Soil layer Anchor Sheet pile
Clay Sand
Dikeold
The linear regression of all the parameters was based on least squares and the R2 parameter
was found 0989 for the sheet pile and 0996 for the anchor Therefore the regression was quite
representative for the domain of the samples over which it was adjusted However there were
many residuals (ie difference between the observed and the predicted value) in the order of
103 and 104 which means that a possible expansion of this RS to an extended domain of the
input variables may lead to an underestimation of the limit state response
(a)
(b)
Figure 78 Polynomial response surface adjusted for the (a) anchor and (b) sheet pile stress level For the anchor the friction angle of the sand layer and the shear stiffness of the clay layer were accounted for this illustration whereas for the sheet pile wall the shear stiffness of the sand and the cohesion of the clay layer were used
The SA for the soil performance was also conducted in two steps However in that case there
was no need of a response surface construction since the LSF for the reliability analysis (Eq
522) takes the value of -1 in case of Plaxis error due to soil collapse Eq 522 though takes only
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 94
two different values -1 and 1 for soil collapse or not respectively which cannot be used for
carrying a SA For that reason the φ-c reduction method was used and the sensitivity of the
value of the safety factor (Msf) was investigated instead The second step of the SA was carried
out in order to divide the soil properties into two different groups of variables and thus
investigate the sensitivity of the soil in a more efficient way rather than including all the
parameters at one SA It should be mentioned that in the soil analysis soil properties of the
Dikenew material were added In Figure 79(a) and (b) the total and the first order indices are
depicted for step 1 and 2 respectively Step 1 includes all the stiffness and strength parameters
of all the soil layers whereas in step 2 the most important variables that were found in step 1
together with the unit soil weight γ are included Subsequently step 1 detects the most
influencing stiffness and strength properties and in step 2 the impact of γ is investigated
As it is observed in Figure 79(a) and (b) the safety factor appeared to be sensitive towards the
sandrsquos friction angle while also the strength properties (c and φ) and the shear stiffness (G) of
the clay layer play an important role Moreover the cohesion of the Dikeold material and the
unit soil weight of the clay sand and Dikenew layer seem to be influencing as well As it has
been mentioned above for this SA the LSF was different than the one considered in the
reliability analysis Therefore even if the SA at that point helps to evaluate in general the most
dominant soil properties the engineering judgement should also be included for the selection of
random variables that are going to be utilized in the reliability analysis
(a)
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 95
(b)
Figure 79 Sensitivity indices for the soil in step 1 and 2
It is also advisable that the difference between the total and the first order sensitivity indices be
investigated in order to have an insight into the most interactive variables The magnitude of the
interaction effect of a variable is crucial for the system behaviour because a variable might not
be important as a singularity but it is possible its combination with another variable to have a
considerable effect on the limit state under investigation Therefore this can be considered as
an additional criterion for choosing the set of the random variables to be used in the reliability
analysis In Figure 710(a) and (b) the difference between the two indices is presented for the
anchor and the sheet pile in (a) and the soil in (b) as it was deduced from the first step of the SA
Regarding the sheet pile wall the most interactive variables are the shear stiffness and the unit
weight of the clay layer the friction angle of the sand and the cohesion shear stiffness interface
strength and unit weight of the Dikeold material Concerning the anchorrsquos limit state the
cohesion of the Dikeold material seems to have the highest interaction with the rest of the
variables Last but not least regarding the soil performance the stiffness of all the soil layers
and the friction angle of the clay and the sand layer have a higher interaction effect amongst the
other parameters The importance of the aforementioned variables can be enhanced
considering both total indices and the difference that they display between their first and total
indices
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 96
(a)
(b)
Figure 710 Interaction effect amongst the parameters in (a) anchor and sheet pile and in (b) soil This effect is expressed as the difference between the total order indices T and the first order indices S as they were deduced from the first step of the SA for both anchorsheet pile and soil respectively
Considering the discussion above some preliminary conclusions can be drawn regarding the
influence of the several variables to the different components of the system
1 both anchor and sheet pile wallrsquos limit states are sensitive mostly towards the shear
stiffness of the soil
2 sheet pile wall stress level is also affected by the unit soil weight more than the anchor
does
3 soil body is mostly influenced by the strength properties and the unit soil weight
4 the interaction effect among the soil properties is more profound for the soil and the
sheet pile limit states rather than the anchorrsquos
It must be mentioned that this SA is a global one that gives an overview of the impact of the soil
parameters in terms of the system behaviour However near the design point a local sensitivity
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 97
can possibly give a better estimation of the variablesrsquo impact Near the design point the
structural elements as well as the soil body usually reaches plasticity that implies a non-linear
behaviour Such a behaviour is unlikely to be initiated during a SA that usually samples around
the mean values This can be counteracted by conducting more iterations so as to have a larger
amount of available samplings but this can make the SA to be quite time consuming and thus to
lose its benefit to give a fast preliminary estimation of the most influencing soil layers and
properties
Besides the purpose of a preliminary global SA is to reduce the number of the stochastic
parameters to a reasonable amount so as the initial reliability analysis to be more efficient and
computationally accessible The parameters that were eventually considered as stochastic in
each LSF are summarized in the table below and they were chosen according to both the SA and
the engineering judgement In the last column the stochastic variables used also for the
systemrsquos analysis are presented
Table 72 Stochastic variables for each LSF
Soil
Failure Anchor Failure
Sheet Pile failure
System failure
Sa
nd
γunsat radic
c
φ radic radic radic radic
G radic radic radic radic
v
radic
Rinter radic
Cla
y
γunsat radic
radic radic
c radic radic radic radic
φ radic radic radic radic
G radic radic radic radic
v
radic
Rinter radic radic radic
Dik
e_n
ew
γunsat radic
c radic
φ radic
G
v
Rinter
Dik
e_o
ld
γunsat
radic radic
c radic radic radic radic
φ radic radic radic radic
G
radic radic radic
v
radic
Rinter radic radic radic
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 98
74 Soil Shear Failure As far as the soil failure is concerned the limit state function that was implemented is given
below (see also sections 52 and 53) It was formulated as a single value depending on the
success of computation in Plaxis
(71)
The reliability method that was chosen in that case is DS due to the formulation of the limit state
function and the convergence issues of FORM There were used 12 input stochastic soil
parameters as it is indicated in Table 72 (see also input file of Soil Failure in Appendix D) The
variables were chosen based on the SA that was presented in the previous section For the soil
failure Poisson ratio was initially included in the analysis but it was noticed that values close to
04 during Plaxis calculations were leading to unrealistic behaviour of the dike More precisely
it has been observed that Poisson ratio equal to 04 can lead to swelling effects on the dike
which consequently lead to a raise of the dike crest instead of collapsing as it would be expected
in a soil failure analysis Therefore Poisson ratio was excluded from the reliability analysis of
the soil failure
The results of the reliability analysis according to DS are presented below In particular 300
iterations were carried out and 3 different failure directions were found Specifically in Table
73 the probability of failure the reliability index the number of iterations and the duration of
the analysis are firstly shown In the sequence the points from each failing direction with the
lowest reliability index are shown
Table 73 Reliability results of DS for the soil failure
DS Pf 1310-8 β 55 Number of LSF calls 1840 Number of iterations (or directions)
300
Elapsed time (hr) 34
Failure points (in real space) Parameter 1st direction 2nd direction 3rd direction Dikenew_phi [deg] 2277 2543 3021 Dikenew_c [kPa] 529 557 444 Dikenew_gammaunsat [kNm3]
1748 1848 1973
Dikenew_gammasat [kNm3]
1948 1984 1997
Sand_phi [deg] 2782 2039 3899 Sand_G [kPa] 808437 715473 715785 Sand_gammaunsat [kNm3]
2183 2122 2097
Sand_gammasat [kNm3] 2352 2145 2283 Clay_G [kPa] 56934 90063 80134 Clay_c [kPa] 704 1593 794
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 99
Clay_phi [deg] 1039 1629 1077 Clay_gammaunsat [kNm3]
1772 1413 1612
Clay_gammasat [kNm3] 1833 1433 1719 Dikeold_c [kPa] 1621 2978 915 Dikeold_phi [deg] 2474 3149 3281
According to the Shortest distance method the failure point from the 2nd direction turned out to
be the design point the importance factors of which are depicted in Figure 714 (first pie chart
at the left) However before concluding for the final design point an assessment of the above
failing directions was made
In particular the combination of the variables from each direction was later implemented in
Plaxis in order to illustrate the failure This is also necessary for verifying and assessing the
failure that Plaxis is giving and inspecting the error in the logfile in order to check if it is a ldquorealrdquo
(ie error=Soil collapses) or a ldquofakerdquo failure (ie error=Load advancement failed or Not enough
load steps) In Figure 711 the ldquorealrdquo failure is defined
Moreover a demonstration of the displacements is also essential in order to certify the failure
shape that normally has the dike crest settled down A different output of the displacements
formation (ie dike crest goes up) would probably imply a ldquofakerdquo failure However it should be
mentioned that a ldquofakerdquo failure in Mohr Coulomb model can be a ldquorealrdquo failure in another model
that can more realistically simulate the deformation patterns of the structure In general even
though Mohr Coulomb can be successfully applied for detecting failure it is not very suitable for
the determination of the displacements and especially for soft soils such that the dike under
investigation is constructed by It should be also mentioned at that point that in this case study
the system is subjected to unloading conditions24 (ie decrease of effective stresses) in soft soil
layers after phase 4 which cannot be modelled successfully by the Mohr-Coulomb model In that
case other constitutive models such as the Soft Soil Model and the Hardening Soil model are
presumably more reliable for simulating such situations For further information for this model
a reference is made to Plaxis (2015) Therefore someone should be always critical and verify
the failure points based on hisher engineering judgement
Figure 711 Prerequisites for real failure in Plaxis
24 From phase 4 to 5 there is an increase in water pressures and thus due to constant total stresses there is a decrease of the effective stresses This can be characterised as an unloading situation
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 100
In order to make it more understandable the different failure points were illustrated in Plaxis
and their deformed mesh and total displacements are presented in Figure 712 (a)-(c) for each
failure direction As one can observe in the first two directions the dike crest settles down
whereas in the 3rd one the dike body seems to swell and move to the right (the blue lines that
surrounds the dike soil layers indicate the initial dike geometry) However even if the 2nd
direction gave a normal pattern of displacements the error message was ldquoNot enough load
stepsrdquo The combination of the variables of the 3rd direction was then tested again with the
Hardening Soil model (HS) in order to illustrate the deformations with a more suitable model
and it was proved that the deformations seem to be realistic as it is depicted in Figure 713 With
Mohr Coulomb only the 1st direction seems to comply with the diagram in Figure 711 and that
would properly be considered as the real design point
(a) Deformed mesh
(b) Total displacements |u| in m (shadings)
(c) Total displacements direction (arrows)
1st direction 2nd direction 3rd direction
Figure 712 Illustration of the design point in soil body failure
(a) Deformed mesh (b) Total displacements |u| in m (shadings)
(c) Total displacements direction (arrows)
Figure 713 Displacements pattern of the 3rd failure direction according to HS
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 101
However taking into account that the other directions showed characteristics of failing behavior
they should also be accounted for the determination of the importance factors Excluding them
from the analysis would not be wise as the error message or the deformation schematization can
be a limitation of FEM or Mohr Coulomb model respectively In any case the points that lead to
failure shall be treated critically in order to decide if they should be included or not In this case
the points from the three directions seem to be realistic as both strength and stiffness
parameters are low enough in order to lead to a potential failure Therefore they were finally
considered for the analysis and the importance factors were averaged over all the points as it is
depicted in Figure 714 (last graph at the right)
Figure 714 Importance factors α2 for soil failure
Another option is also to average over the 10 range of close to the lowest reliability index that
was calculated with the shortest distance method (ie Average 10 method) in order to come
up with the α2 values However such methodology can lead to averaging over points of the same
direction that are not very different with each other Therefore the rest of the failure points are
automatically omitted from the consideration This can be also verified from Figure 714 where
the pie chart of the Shortest Distance and the Average 10 are quite similar whereas the Average
all chart introduces the importance of other soil variables as well
From a general perspective the soil properties of the clay layer seems to be determinant
whereas also sand and Dikeold material play an important role to the soil failure More precisely
according to the Average all method the unit weight and the friction angle of the clay layer
turned out to be the most influencing while also the cohesion of the Dikeold material and
friction angle of the sand layer contribute to the failure domain From a qualitative point of view
the weakening and consequently the settlement of the Dikeold material pushes the subsoil
creating an additional surcharge for the underlying clay and sand layer which act like the
foundation soil the incapability then of the clay layer to withstand the overlying load due to the
low unit weight in combination with the low friction angle of the clay as well as the low stiffness
and strength of the sand layer can lead to the creation of an inner slip surface and thus failure
Since the clay layer can be considered as a foundation soil the importance of the friction angle
and the unit weight can be also explained from the Terzaghirsquos bearing capacity theory (Terzaghi
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 102
1943) where these two factors are crucial for the determination of the ultimate bearing capacity
of the soil However these can be better verified by conducting large scale experiments in dikes
and inspecting the sensitivity of the dikersquos stability towards the soil properties
75 Anchor Failure The reliability of the support system is estimated based on the anchor and the waling probability
of failure However the load on the waling is proportional to the anchor force as it appeared in Eq
57 Therefore after the estimation of the anchor reliability by calculating the probability of
exceedance of a certain admissible anchor force the waling can be designed deterministically
Besides the failure probability of the waling must be lower than the anchorrsquos (given that no
uncertainties of the strength and the structural properties are considered) For the anchor failure
the LSF was formulated as follows
(72)
where [kN] is the anchor force is the yield stress and the cross
section area The reliability method to be utilized in this analysis is FORM and the selection of the
random variables to be used was made according to the sensitivity analysis results and the
engineering judgment and they are listed in Table 72 In particular the variables from the second
step of the sensitivity analysis were considered which will also help at a later stage to compare
FORM sensitivity method and FAST In Table 74 the results of the reliability analysis with FORM
are listed Additionally the design point together with the importance factors of each variable is
presented It should be mentioned at that point that the sensitivity analysis helped to define a
starting point closer to the design point and thus reduce the computational time required for the
analysis More precisely the analysis lasted approximately 16 hours whereas by starting from
mean values from which by default FORM starts the iterations it would take almost a day
Table 74 Reliability results for the anchor failure with FORM
FORM Pf 28410-6 β 45 Number of LSF calls 119 Maximum number of iterations
100
Elapsed time (hr) 16
Design point Parameter value α-value Clay_phi [deg] 1814 -0059 Clay_nu [-] 033 0086 Clay_c [kPa] 1487 -0089 Clay_G [kPa] 41739 0911 Clay_Rinter [-] 051 -0021 Sand_Rinter [-] 069 -0051 Sand_nu [-] 030 0007 Sand_phi [deg] 3259 -0174 Sand_G [kPa] 1130770 0048 Dikeold_phi [deg] 2842 -0020 Dikeold_nu [-] 036 -0045
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 103
Dikeold_c [kPa] 1873 -0045 Dikeold_G [kPa] 243183 -0326 Dikeold_Rinter [-] 053 -0072
Figure 715 Importance factors α2 in anchor failure from FORM analysis
The influence coefficients in Figure 715 indicate that this limit state is governed by the shear
stiffness of the clay layer G Furthermore the shear stiffness of the Dikeold material seems to
contribute significantly to the anchor yield stress exceedance It can be concluded that the
problem is still in the elastic domain as far as the soil behaviour is concerned In case of
predominately plastic behaviour the strength properties of the soil (φ and c) become more
important In Figure 716 a demonstration of the design point as it was acquired by the FORM
analysis is presented The figures indicate that presumably due to the clay layer withdrawal in the
passive side as it is depicted in Figure 716(b) with the red shadings sheet pile moves to the right
and subsequently anchor is tensioned enough to reach its yield stress25
(a) Deformed mesh (b) Total displacements |u| in m (shadings)
(c) Total displacements direction (arrows)
Figure 716 Design point illustration from the anchor failure
25 A better representation of the soil-anchor interaction could be possibly obtained if the anchor was modelled as a node-to-node element and the anchorage body as an embedded element
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 104
The calculations were repeated with the 6 most influencing parameters both with FORM and DS
in order to investigate the effect of reducing the number of random variables and in order to
validate the results of FORM analysis In Table 75 the reliability results of respectively FORM and
DS are presented while in Figure 717 the importance factors of both methods are juxtaposed The
design point and the importance factors of DS were estimated based on the Average 10 method
(see section 71)
Table 75 Reliability results of FORM and DS with reduced variables in anchor failure
FORM Pf 8510-6 DS Pf 1510-5 β 43 β 42 Number of LSF calls
28 Number of LSF calls
776
Maximum number of iterations
100 Number of iterations
100
Elapsed time (hr) 034 (starting point close to the design point)
Elapsed time (hr)
115
Design point Design point Parameter value α-value value α-value Clay_nu [-] 033 0119 033 0242 Clay_c [kPa] 1404 -0026 1535 -0212 Clay_G [kPa] 43236 0928 44548 0807 Sand_phi [deg] 3163 -0110 2954 0203 Dikeold_G [kPa] 239465 -0330 251723 -0366 Dikeold_Rinter [-] 053 -0064 060 -0263
Figure 717 Importance factors α2 with FORM and DS analysis respectively
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 105
As it can be noticed comparing FORM results from Table 74 and 75 the reliability index changes
from β=45 with 14 parameters to β=43 with the 6 most influencing parameters This verifies the
importance of the most influencing parameters that it was found in the analysis with the 14
parameters The small discrepancy though between the two reliability indexes might be
attributed to the interaction effect of the soil variables on the output performance
The estimation of failure probability with DS is almost the same with FORM As far as the
importance factors are concerned DS also came up with the conclusion that the shear stiffness of
the clay and the Dikeold material are the most crucial soil properties for the anchor stress level
However according to DS the stiffness of clay ν and the interface strength Rinter between the
Dikeold material and the sheet pile wall seem to also contribute to the limit state of the anchor
The reason why the strength of the clay layer under the dike and the Dikeold material (the part of
the dike that comes in contact with the anchor and the sheet pile wall) are the most important
variables can be explained by the principle of soil arching According to this principle the soil
columns on both sides of the rigid sheet pile wall are more compressive than the soil columns on
the top of the wall (ie overlying dike body) because of the higher stiffness of the wall when
compared with soils As such soil columns on both sides tend to settle more than the soils on top
of the rigid wall and this differential settlement causes a downward shear force acting along the
sides of soil columns on top of the wall As such the vertical load on the wall becomes larger than
the sole weight of soil columns on its top and the anchor that actually receives the most of this
vertical load reaches eventually its yielding stress
Last but not least it should be mentioned that the response surface which was constructed based
on the SA worked efficiently in most of the Plaxis errors cases in a sense that it gave reasonable
results of the anchor stress level
76 Sheet pile wall failure For the reliability analysis of the sheet pile wall the strength property ie the yield stress was
considered constant and equal to 355 kPa Therefore the LSF that was applied for this case was
formulated as follows
[
] (73)
where [kNm] and [kN] are the maximum bending moment and the axial force
that are exerted on the wall respectively is the elastic section modulus ( =1245 cm3m)
and is the cross sectional area ( =1201 cm2m for a AZ12 profile)
In sections 643 and 72 the maximum bending moments of the sheet pile wall were calculated
with the design values of the soil properties and the mean values respectively It has been
observed that the bending moments were not so significant and determinant in comparison with
the anchor force and the safety factor of the system that were also considered (in section 643) in
order to determine the structural properties
In a later stage a SA was carried out for the sheet pile wall that has been further discussed in
section 73 Figure 718 shows the stress level of the wall that was developed after 3150 iterations
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 106
by considering as random variables the ones that are defined in Table 72 As someone can
observe the stress level is generally quite low and even an order lower than the yield stress (ie
355105 Pa)
That small magnitude of bending moments can be probably attributed to the counteraction of the
active earth pressures from the passive ones that due to the homogeneity of the soil in both sides
of the wall are quite similar
Figure 718 Stress level [in Pa] of the sheet pile wall after 3150 iterations of the SA
The reliability analysis has been carried out with FORM method and considering as random
variables those listed in Table 72 However due to the low level of stresses that are developed
and the simultaneously high yield stress the method converged after almost 35 days of
calculations As it can be noticed in Table 76 the analysis came up with an extremely low
probability of failure It should be mentioned at that point that the convergence of the analysis
was successful thanks to the RS that was fitted This means that a Plaxis error of soil collapse
preceded and it activated the RS which implies that the soil failure is more likely to occur before
the sheet pile failure
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 107
Table 76 Reliability results for the sheet pile wall failure with FORM
FORM Pf 7210-38 β 13 Number of LSF calls 2619 Maximum number of iterations
100
Elapsed time (hr) 84 (35 days)
Design point Parameter value α-value Clay_phi [deg] 1164 0261 Clay_c [kPa] 586 0336 Clay_G [kPa] 14275 0663 Clay_Rinter [-] 052 -0016 Clay_gammaunsat [kNm3]
1848 0000
Clay_gammasat [kNm3] 1948 0000 Sand_phi [deg] 3020 0000 Sand_G [kPa] 1193650 0000 Dikeold_phi [deg] 1360 0406 Dikeold_c [kPa] 592 0438 Dikeold_G [kPa] 168798 0000 Dikeold_Rinter [-] 053 -0024 Dikeold_gammaunsat [kNm3]
2275 -0150
Dikeold_gammasat [kNm3]
2375 0000
This can be also verified from the importance factors that are depicted in Figure 719 where apart
from the shear stiffness of the clay layer which seems to be the predominant soil property for the
sheet pile LSF the strength parameters (φ and c) of the clay and the Dikeold materials also play
an important role This implies that the soil behaviour has already passed into its plastic domain
where the governing properties are the strength parameters before the sheet pile exceeds its
yield stress
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 108
Figure 719 Importance factors α2 of the sheet pile stress level in the plastic domain of soil
Nevertheless in order to investigate the potential most influencing soil properties for the sheet
pile LSF in the elastic domain as well the available output data of the FORM analysis have been
assessed The highest observed value of the sheet pile stress before its failure was around
235000 kPa After tracing the results it seems that when FORM tries to reduce further the
stiffness and strength properties of the soil layers soil collapses before the sheet pile failure The
vectors of the random variable that led to a stress level on the sheet pile wall in the range of
230000-235000 kPa were chosen in order to evaluate the importance factors The method that
was followed in that case is the Shortest Distance and the results are shown in Figure 720
Figure 720 Importance factors α2 of the sheet pile stress level in the elastic domain of soil
As it can be noticed from Figure 720 the results in the elastic domain regarding the influence of
the stochastic parameters are comparable with anchorrsquos analysis The stiffness parameter of the
clay layer (69) is the most dominant variable for the system behaviour whereas the stiffness of
the sand (15) and the Dikeold (4) layers seems also to have a significant influence This is not
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 109
surprising as the nature of the loads that generate high stresses on the anchor as well as the sheet
pile is the same namely the horizontal earth and water forces
Concluding in this case study the sheet pile wall is not expected to show a decisive behaviour to
the systemrsquos reliability according to the reliability analysis and the deterministic analysis (see
section 643) It should be stressed the fact that the fitted RS in a lot of calculations seemed not to
perform suitably for the sheet pile and this can be attributed to the fact that the it was created
based on the SA which was far from the plastic domain In that case a better fit of a RS is
recommended However such an action is still under investigation that for future considerations
in the design concept will be contemplated as a valuable technique for complex structures
modelled in FEM whose reliability analysis with conventional methods is time consuming and the
coupling may lead to numerical problems In the next section the analysis of the systemrsquos
reliability is presented
77 System Failure The estimation of the systemrsquos reliability was based on the systemrsquos fault tree of Figure 53
according to which the general LSF was formulated as the minimum of the three LSF that were
presented above for the soil the sheet pile and the anchor and it is given below
[ ]
[(
)]
(74)
According to this LSF the failure that occurs first is counted as a system failure The analysis was
carried out with DS reliability method and the variables that were taken as stochastic are listed in
Table 72 As it was discussed in section 51 the systemrsquos probability of failure differs from the
probability of flooding for which a combination between the systems reliability and the water
level uncertainty shall be made
In Table 77 the reliability analysis results are presented as they were deduced from DS Similarly
to the soil failure in Figure 721 the importance factors according to the Shortest Distance the 10
Average and the Average all methods are illustrated
Table 77 Reliability analysis results from the system failure
DS Pf 2410-3 β 30 Number of LSF calls 1807 Number of iterations (or directions)
300
Elapsed time (hr) 39
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 110
Figure 721 Importance factors α2 of the system failure
The analysis took almost one and a half days and it came with indicative valuable though results
about the system behaviour According to the averaged importance factors in Figure 721
(deduced by the Average all method) we can notice features from both soil and anchor failure
For example the unit weight (14) the friction angle (7) and the cohesion (8) of the clay
layer as well as the friction angle (8) and the stiffness (7) of the sand layer that played an
essential role in the soil failure analysis Moreover the stiffness of the clay (11) and the Dikeold
(9) layers that appeared to be important for the anchor and sheet pile LSF They seem to be also
essential in the system analysis
However the other two methods (Shortest Distance and Average 10) came up with the cohesion
and the interface strength of the Dikeold material to be the most influencing soil properties for
the system For that reason an investigation was conducted regarding the output results of the
system analysis in order to obtain an insight into the different reasons of the system failure More
precisely the failure points were divided into those that come from Plaxis Error and those that
come from the yield stress exceedance of the structural elements It was then observed that many
failure points originate from Plaxis errors that are related to ldquoNot enough load stepsrdquo (Error codes
102 and 112) rather than ldquoSoil seems to collapserdquo (Error code 101)
Therefore in order to estimate the actual important soil properties the vectors of the random
variables that led to errors of ldquoSoil seems to collapserdquo and to anchorsheet pile (SP) failure were
studied separately The importance factors are depicted in Figure 722 which actually verifies the
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 111
importance of the variables that were also deduced from the Averaged all method in Figure 721
(which are basically the unit weight and the stiffness of the clay layer and the friction angle of the
sand layer) It should be mentioned that the failure of the structural elements contributes 65
whereas the soil failure contributes 35 on the total probability of failure The latter would be
possibly lower than 2410-3 that was initially found as several failure points are not attributed to
real failure but to Plaxis numerical errors
Figure 722 Influence factors α2 for the structural elements (left) and the soil LSF (right) respectively
The system analysis can lead to valuable conclusions for the system behaviour as a whole and
individually for the different elements given though that an inspection of the results is followed
Moreover such an analysis is taking into account correlations among the three sub-failure
mechanisms implicitly For example the different LSFs may be affected by the same soil variable
such as the shear stiffness of the clay layer for the anchor and the sheet pile LSF In that case the
most dominant failure mechanism will occur first and the rest are excluded26 However during
the system analysis someone is not aware of what failed or not and therefore heshe is not
capable of distinguishing the weaknesses of the system that need further improvement Besides
with this technique is not also easy to inspect the failure mechanisms that occur in order to get
some knowledge about the systemrsquos behaviour Therefore the reliability analysis of the
components is recommended to be carried out separately if time is available
78 Comparison between Global Sensitivity Analysis and Local Sensitivity
Analysis (FORM)
The aim of this section is a comparison between the importance factors that were deduced in
section 75 according to a FORM analysis and the results of the global SA that were presented in
section 73 for the anchor failure
The output of the global SA is actually the total indices that show the influence of the random
variables The importance factors of FORM and the total indices of global SA cannot be directly
26 This method is not capable though of accounting for the simultaneous occurrence of different failure mechanisms as the system was considered a serial system
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 112
compared as they are different values More precisely the total index of a random variable is the
ratio between the total variance and the portion of the total variance that stems from the
uncertainty of the specific variable and its interactions with the rest (see Eq A10 in Appendix A)
On the other hand importance factors of FORM express the ratio between the normalized value of
the variable in the design point and the reliability index (see Eq 72) Therefore the comparison
between these values is carried out qualitatively in terms of the variables that are distinguished
as the most influencing
In particular according to global SA the shear stiffness G of the clay layer and the cohesion c of
the Dikeold material seem to be the most determinant soil properties Moreover the strength
reduction parameters Rinter of the Dikeold and the clay layers and the friction angle φ of the
Dikeold and the sand layer appear to have a relatively notable contribution to the anchors limit
state
On the other hand FORM analysis has concluded to the same results as the global SA with the
difference that Dikeold material affects the anchorrsquos stress level more with its shear stiffness
rather than its cohesion In Table 78 the results of both FORM and global SA are presented
separately for each soil layer and property (in blue for the global SA and in black for the FORM
analysis)
Table 78 Important soil properties according to global SA and FORM analysis for the anchor failure (radicradic=very important radic=less important but still considerable)
FORM global SA Sand Clay Dikeold
ν
radicradic
ϕ radicradic
radic
c radic radic radicradic
G radicradicradicradic radicradic
Rinter
radic radicradic
The validation of the most important parameters was carried out with the repetition of the
analysis with DS as it was showed in section 75 The small difference of the probability of failure
as well as the same tendency of the important factors of both DS and FORM verified the reliability
of FORM analysis Therefore the similarity of the results deduced according to global SA with the
ones given by FORM indicates the fact that in this case study the global SA should be considered
as a trustable way of SA for obtaining a first overview of the dominant variables that affect the
limit state under investigation Furthermore it is capable of giving a general insight into the
outputrsquos (in this case anchor stress level) behavior and the soil layers that can potentially stand as
predominant during the reliability analysis For example in order to obtain a first sense of the
reliability and the magnitude of the expected probability of failure someone can inspect the
output samples of the LSF deduced from the global SA In Figure 723 the samples of the anchor
stress level are presented as they were deduced after 3150 iterations during the global SA
(computation time =75 hr = 3 days) The fact that after 3150 iterations and variables
combinations the anchor did not reach its ultimate stress level (ie 450 kPa) implies that the
probability of failure might be small This can also help for qualitative decision making concerning
a systemrsquos behavior without conducting a reliability analysis Moreover global SA proved the high
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 113
interaction (large difference between first and total order indices) among the parameters that can
affect the reliability analysis and can help to decide the amount of the random variables
Figure 723 Anchor stress level during global SA (step 1)
However as someone can observe from Table 78 there are also some discrepancies concerning
the dominance of some variables This can be attributed to the fact that the influence of all
parameters in specific points such as the design point cannot be identified by the global SA
Moreover it is possible that the samples used in the SA could not reach the field where the
system has a plastic behavior and where presumably other variables may also affect the limit
state
As general conclusions for global SA some possibilities and limitations are listed below
Possibilities
bull Indicates the level of interaction effect on the output variance
bull Indicate the amount of the random variables that should be considered in the reliability
analysis
bull Gives a general insight into the output behavior
bull Valuable for future use in the construction of response surfaces
Limitations
bull Cannot rely on global SA for the evaluation of the most important parameters near the
design point
bull Difficult to reach plastic zone (large amount of samples are probably needed)
79 Discussion
In this chapter the reliability analysis results were presented for the soil the anchor the sheet
pile and eventually the system failure taking soil properties of the several soil layers as stochastic
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 114
Initially a SA was carried out for each system component in order to identify the major soil
properties and reduce the number of the stochastic parameters The reliability analysis was
performed either with FORM or DS and in both cases the design point and the importance factors
were estimated Eventually the list of subquestions that is proposed in section 11 was answered
The way they were treated and the corresponding conclusions are elaborated below
Which reliability methods are computationally tractable in combination with FEM
Before starting analyzing the reliability of the system it was essential to evaluate and
assess the performance of the available reliability methods in order to be aware of their
performance in terms of computational time and flexibility in different kind of LSFs After
a literature review and the implementation of some of them in simple LSF problems a
general overview has been obtained for the most applied reliability methods concerning
the previous knowledge required for the system behavior in order to provide the right
input parameters its accuracy and the computational effort Eventually FORM and DS
were chosen to be employed for this case study considering the no preliminary knowledge
of the LSF is required their reliable accuracy and the limited computational effort that
they have in comparison with other methods
The anchorrsquos and sheet pilersquos reliability has been evaluated with FORM method whereas
soil and systemrsquos reliability have been assessed by DS Carrying out a FORM analysis and
steering the algorithmrsquos parameters in order to converge proved to be more difficult than
it was expected However an efficient performance was finally achieved and the failure
probability of the anchor failure was successfully estimated together with the importance
factors (or else the α values) The analysis was repeated with the most important variables
and they were validated with DS The FORM analysis took approximately one and half
hour to converge (with 14 random variables) after providing a starting point that it was
estimated to be close to the design point whereas for the validation it took almost half an
hour On the other hand in case of the sheet pile wall FORM analysis took almost 4 days
The computational time also depends on the LSF to be approximated the number of
variable and the complexity of the problem that each time is analyzed As far as the soil
analysis is concerned DS took approximately one and a half day to converge for 300
iterations and 14 random variables
Which limit states are relevant for the envisaged application of retaining walls in dikes and
how can they be formulated using FEM analysis outcomes
The relevant LSFs of a dike with an anchored sheet pile wall were investigated For each
element ie the sheet pile the anchor and the soil body a separate LSF has been set up
The focus of this research was on the ULS that in case of the structural elements it was
represented by the ultimate yield stress of the anchor and the sheet pile respectively as
they were expressed in Eq 72 and 73 The determination of the soil failure has been also
elaborated and the available alternatives that can be considered in that case were
elaborated After a short evaluation the Plaxis definition of soil failure was selected that it
was formulated as it is shown in Eq 71 Finally the systemrsquos LSF was formed as a
combination of the aforementioned limit states More precisely the system was
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 115
considered as a serial and thus system failure occurs when one of the elements fails first
(see Eq 74)
How robust (convergence) are the tractable methods
The robustness of the different reliability methods can be controlled by the convergence
criteria that are defined by the user and depend on the required accuracy In case of FORM
the convergence criteria consist of a set of errors that has mostly to do with the
approximation of the reliability index and the limit state threshold whereas in DS the
convergence is controlled through the coefficient of variation (CoV)
In order to reassure the robustness of FORM the method has been validated with DS The
validation showed that the FORM method complies with the result of DS with a small
difference though in the design point and the importance factors FORM analysis seemed
to converge sufficiently and faster than DS However the choice of the optimization
algorithm plays a key role to the methodrsquos efficiency In this research the convergence
errors were manipulated for increasing the efficiency of the method More precisely it has
been observed that by relaxing the relative error and the constraint error (the explanation
of these errors and an evaluation of the different optimization algorithms are available in
Appendix A2) FORM convergence can be accelerated
The reliability analysis of the soil body and the system was performed with DS Similarly
to FORM analysis in order to improve the performance of the searching algorithm some
of the methodrsquos parameters should be stipulated for improving its efficiency and
reliability Therefore the maximum step size of the algorithms was tripled and it was
proved that for the same time period double iterations were feasible to be carried out
This does not mean that the larger the step size the better it is The choice of the step size
must be made wisely depending on the reliability index that it is expected and
respectively it is allowed to change for achieving a better performance In this case study
it was proved that after 300 iterations the CoV of DS tends to be high and it can therefore
still be improved by carrying out more iterations In Figure 724 an example of a DS
convergence graph concerning the system analysis is shown The middle red line shows
the probability estimation whereas the green lines show the boundaries At that example
the CoV was 040
DS seems to perform successfully both in soil failure with one LSF and in system failure
where 3 different LSF were included In such cases performing an approximation method
such as FORM would not be wise as the output from each iteration does not provide
information for going to the next that a FORM analysis normally needs In contrast it just
gives an indication of failure or not This means that only a random sampling method can
handle it and converge successfully as it was proved
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 116
Figure 724 Directional Sampling convergence graph (at confidence level 095) of the system LSF
What is the contribution of different uncertainties in the failure mechanisms of the system
According to the global SA and the reliability analysis of FORM and DS the friction angle
the cohesion and the shear stiffness properties appeared to have an influence on the
different LSFs This is not surprising as they all have large CoV of 10 20 and 25
respectively Thus their uncertainty generally showed an impact on the systemrsquos
behaviour It is also essential to mention that the global SA has shown that the uncertainty
of the interface strength Rinter (CoV=20) has also a notable contribution to the structural
elements LSFs However the FORM analysis which is more representative close to the
design point came up with a less important impact of that variable
The reliability analysis showed that as far as the structural elements (anchor and sheet
pile) are concerned the shear stiffness of the clay and the Dikeold material is determinant
and especially for the sheet pile the strength parameters of the soft soils seem also to play
a crucial role In contrast with the anchor where the soil body seems still to behave in an
elastic manner during the sheet pile failure soil appears to reach plasticity in more
locations where the strength properties (ϕ and c) are also influential The most dominant
failure mode appeared to be the anchor failure whereas the soil failure is always preceded
the sheet pile failure The dominance of the anchor failure can be explained from the fact
that the most influent parameter in this LSF is the stiffness that is also the variable with
the highest uncertainty (CoV=25) Therefore the plastic limit of the anchor is more
likely to be reached
Regarding the soil body reliability the analysis detected 3 different failure directions in
which the clay and the sand layer below the dike seem to be the most influent More
specifically the friction angle φ the cohesion c the unit weight γ of the clay layer the
friction angle of the sand layer and the cohesion of the Dikeold material turned out to play
an important role The decrease of the clay layerrsquos unit weight lead to the incapability of
the layer to withstand the overlying load of the dike body and at the same time the
weakening of the sand layer contribute to the creation of an inner sliding surface Even if
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 117
the uncertainty of the unit weight is small (CoV=5) it is still important for the stability
of the system as the maximum deformations appear in that layer (see Figure 712) Last
but not least the illustration of the several failure points in Plaxis allows someone to
inspect the failure mode and realize if it is a real failure of the structure or if it is a
weakness of the modelling
Finally during the system analysis multiple failure directions were detected whereas the
pie charts of the importance factors include soil properties that affect both soil and
structural elements as it was expected However due to Plaxis numerical errors several
failure points were found to be not realistic and they should be excluded in order to come
up with a more reliable failure probability which will be presumably lower The most
suitable method for estimating the importance factors though seems to be the Average all
method which is not distorted from the Plaxis numerical errors However it is
recommended that a further research be done on how realistic are the failure points
coming from Plaxis warnings and to verify if they comply with Figure 711 failure
definition
In Table 79 a synopsis of the most important soil properties for the different LSFs is
presented It is also apparent the common contribution that some of them have in
different LSFs In the last two columns of the table the results of the influencing soil
variables for the system LSF are shown with the Average 10 and Average all method
respectively For the reasons that were explained in the previous paragraph the Average
all method is considered to be the most suitable in that case and as it can be noticed from
the table it gives the most representative picture of the governing soil variables of each
LSF
Table 79 Important soil properties for the different LSFs (radicradic=very important radic=less important but still considerable the most important parameters for each LSF is painted in orange)
Soil material
Soil property
Anchor Sheet pile
Soil body
System Average
10
System Average
all
Clay
G radicradic radicradic
radic radicradic
φ
radic radicradic
c
radic radic radic
Rinter γ
radicradic
radicradic
Sand
G
radic radic
φ
radicradic radicradic
c
Rinter γ
Dikeold
G radicradic
radic
φ
radic
c
radicradic radic radicradic radic
Rinter radic
radicradic radic
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 118
γ
radic radic
Rinter was taken from the validation of FORM with DS Figure 717 (right graph)
taken from the sensitivity of the sheet pile in the soil elastic domain Figure 720
Can response surface techniques help to increase the efficiency and robustness of the
reliability model
In this research RS techniques were used only as a source for providing a reasonable
value to the FORM analysis in order to continue running in case of Plaxis errors These
response surfaces were constructed based on the SA results for both the anchor and the
sheet pile wall Apparently the domain of the SA is limited and most of the times it is
incapable of approaching the design point However RS worked efficiently in many cases
where Plaxis failed to give an output result such as in case of the sheet pile LSF Moreover
the RS could be used individually for the estimation of the failure probability (instead of
Plaxis calculations) but a better fit is advised to be firstly achieved that would be
representative in a wider plastic domain of the structural elements Therefore the RS
technique can be used to enhance the efficiency of the reliability analysis of a complex
system in the sense that it can be considered as solution for FEM errors What is more RS
techniques are a quite promising method for the reliability analysis of a system as it can
also be used alone without coupling with FEM and eliminate the computational effort
However this requires more investigation of such a technique
How can the current design approach be improved
In section 22 the current design concept of the case study that is under investigation in
this research is described At that section the design values of the moments and the forces
on the structural elements are presented as well as the required total safety factor FEM
=18 is given as it was initially calculated by Breedeveld (2011) It has been observed that
a significant overestimation is made in terms of the design values that might later lead to
heavy and cost inefficient structures For that purpose alterations was made on the
original case study in terms of the structural elements and some soil properties and a new
deterministic design was made in order to come up with a less heavy structure This
would also render it possible to investigate the reliability of the system rather than taking
zero probabilities or non-convergence Moreover it would finally help to see if there are
any potentials for optimizing the design by applying a full probabilistic analysis
According to the soil reliability analysis the probability of failure was found 1310-8 and
the reliability index β=55 As far as the anchor failure is concerned the probability of
failure found to be 28410-6 with a reliability index β=45 according to the FORM analysis
The sheet pile failure seem not be significant as it will more likely occur after the soil
failure Eventually the system analysis was performed the probability of failure was
found to be 2410-3 and the reliability index β=30 which is high enough to consider the
system safe
The systemrsquos reliability index that came out of this analysis cannot be directly compared
with the required one (ie β =40) because they stem from different probabilities The
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 119
former refers to the probability of failure due to the global instability given a design water
level whereas the latter refers to the probability of flooding that incorporates also the
water level uncertainty It must be mentioned that in reality there is no sheet pile wall
inside the dike section in the dike ring 43 but this section is under investigation for
reinforcement purposes as it does not comply with the safety standards after a recent
inspection that has been carried out Moreover the original case study taken from
Breedeveld (2011) was subjected to multiple alterations for the needs of the current
research Therefore a comparison of the reliability index with the required one from the
regulations would not be meaningful to make However what it can be compared in that
case is the order of magnitude of the reliability index After the analysis a β of 30 was
estimated for the probability of failure whereas for the overall probability of flooding a β
of 40 is required whose magnitude is the comparable with the former This can firstly be
considered a sort of the probability ldquovalidationrdquo in the sense that the analysis gave
realistic results and secondly it can imply that the structure that has been investigated in
this research is likely to be safe towards macro-instability as its coupling with the water
level uncertainty would certainly give a reliability index above 40 Furthermore a
reliability index of 30 indicates that there is room for optimizing the design concept of
such a system if someone considers also the fact that the soil strength of the subsoil was
further reduced for the needs of the current study
Due to the limitations of the modeling and the alterations of the original case study these
reliability indices should not be taken into account as an indicative number for the safety
of this dike section but rather as an order of magnitude in order to realize what the
potentials for improving the design procedure are This research can also contribute to
the way of designing not only the system but also each structural element as the reliability
of each can be calculated in advance According to the methodology developed in this
thesis a comparison can later on be made for a real case situation between the results of a
full reliability analysis and the current design procedure In that sense a tangible
improvement can be suggested for the design concepts and even to introduce a new way
for the future design of dikes with retaining walls
It should be stressed that the conclusions are based on the results of this case study with its
geometry and set of material parameters and the variation coefficients Changes in the geometry
the material parameters or the statistical properties could lead to different results The
generalization of these conclusions should thus be treated carefully
Reliability analysis results with stochastic soil properties
MSc Thesis A Rippi 120
MSc Thesis A Rippi 121
8 Conclusions and Recommendations
81 Conclusions The advances in structural reliability analysis have made a large variety of methods available The
applicability and efficiency of these methods depends on the problem that is analyzed and on the
models that are used One the other hand the necessity for optimizing the design procedure of
several structures is becoming more and more intense A solution to this demand can be the
reliability analysis of a system for which rare information about its application on real-world
problems is available This work can contribute to making probabilistic analysis more accessible
for dikes reinforced with a sheet pile wall and at the same time to analyzing such a complex
system in terms of the factors that mainly influence the system behavior In particular the main
research question of this research was
How can the probability of failure due to global instability of a dike with a sheet pile wall modeled
by a Finite Element Model be analyzed
For that purpose in this research a reliability analysis was conducted by making use of a soil-
structure model that was simulated in FEM The output of the FEM calculations was utilized as an
input in the probabilistic model that was used in this thesis in order to assess the reliability of a
dike with an anchored sheet pile wall In that case the load was basically the earth and water
pressure whose impact was investigated on the anchor (LSF 1) the sheet pile (LSF 2) and the soil
body (LSF 3) respectively as well as on the system as a whole
It was demonstrated that the calculation effort of this kind of analysis does not necessarily have to
be extremely high Especially when previous knowledge about the system behavior and the
reliability models used is available This can be achieved by carrying out a sensitivity analysis In
this thesis a global SA was conducted that was proved capable of giving a representative
estimation of the most influencing variables and give a general overview of the systemrsquos
performance However close to the design point local SA such as FORM becomes more suitable
As far as the reliability of the structural elements is concerned their limit state was evaluated in
terms of the exceedance of their yielding stress The reliability assessment was carried out with
FORM from which the probability of failure and the influence factors were concluded More
precisely the anchor failure mode seemed to be the most determinant one with the shear
stiffness of the soft soils to be the most important soil properties The sheet pile failure is quite
rare and it is unlikely to occur before the soil failure since the soil body has already entered its
plastic domain
The soil and the system failure were challenging to assess as there are multiple criteria that
someone can choose In this thesis regarding the soil failure the Plaxis definition of soil collapse
was used and it was evaluated with DS In that case the unit weight and the strength parameters
of the soft soil layers seemed to be influencing The detected failure directions were investigated
Conclusions amp Recommendations
MSc Thesis A Rippi 122
in terms of their validity More precisely the failure points were illustrated in Plaxis in order to
visualize the failure and see if it is real failure or not (real if dike crest settles) That procedure
revealed the possibilities and limitations of FEM modeling The advantage is that a potential
failure can be detected and its visualization is possible On the other hand one should be aware of
the limitations related with the constitutive model that is used In this study the Mohr Coulomb
method was applied that although it can efficiently identify possible failure mechanisms the
representation of the deformations is not so realistic
The failure of the system was considered as a serial system of LSF 1 2 and 3 and it was also
assessed with DS The importance factors stemmed from the reliability analysis indicated soil
features that were important in all the LSFs However it should be stressed that the frequency of
the Plaxis numerical errors was such that quite a few failure points seem to be unrealistic In this
thesis the identification of those errors and their partially exclusion from the analysis was
achieved The latter was attained by the construction of the so-called Response Surfaces (RS)
based on a preliminary sensitivity analysis Such a technique found to be valuable in order to
overcome such a problem however a study for fitting more suitable RS in each problem is
recommended Nevertheless the further elimination of those numerical errors is strongly
recommended to be investigated in future research given the results from the current study
Retrieving the importance factors from a DS analysis was also a challenging part In this study
additional methods were developed in order to evaluate the importance factors and define the
design point after a DS analysis It revealed that alternative possible approaches of the
importance factors (Average 10 and Average all methods) can presumably provide a better
understanding of the systemrsquos behaviour Such methods seemed to be valuable for excluding the
ldquonoiserdquo of the possible fake failure points that originated from FEM numerical errors However
due to the inherent randomness of DS in terms of the selection of the directions the design point
shall be inspected and compared with the other failure points that were found Then someone can
contemplate if it is necessary to proceed with an averaging of the importance factors in order to
get a clearer view the variables influence Of course such a procedure takes time that is not always
available but according to the requirements of each research the most preferable assumptions
should be made
Last but not least the reliability analysis came up with relatively low probabilities of failure which
implies that there is still room for optimizing the design concept of a dike with retaining wall
More precisely the original case study of Breedeveld (2011) has been altered by reducing the
strength properties of the subsoil and installing a less heavy sheet pile wall after following a less
strict deterministic design It was then observed that even then the reliability analysis came up
with a respectively high reliability index It should be stressed that the failure probability of the
system should be also coupled with the uncertainty of the water level in order to obtain an overall
picture of the total probability of failure Nevertheless this research proves the potentials to
optimize the design of such a system which presumably will lead to a less expensive construction
However additional failure mechanisms such as piping internal erosion etc should be also
investigated and the design properties of the system (such as the length and cross section area of
the sheet pile wall) shall be based on them
Conclusions amp Recommendations
MSc Thesis A Rippi 123
82 Recommendations The following topics are recommended for further research
Additional research is required on FEM modeling both for the structural elements and the
soil body More in particular in this research it has been proved that Mohr Coulomb
model cannot realistically estimate the deformations in a soft soil In that case there are
other more advanced models such Hardening Soil model and Soft Soil model that are
capable of simulating the soil displacements and the stress-strain relationship closer to
the actual ones However these methods require more parameters to be specified and
their coupling with reliability packages might not be so robust and functional
As far as the structural elements modeling is concerned there are two major topics that
future research should focus on Firstly the anchor rod can be simulated in Plaxis as a
node-to-node element and at the bottom an embedded beam can be installed in order to
represent the anchorage body This would provide a better reaction with the surrounding
soil body and it could probably simulate the anchor-soil interaction more realistically
Another alternative would be to simulate it as a plate element in order to be able to test
the moments that are also developed on the anchor separately Secondly it is
recommended that the elements be investigated close to their plastic area and a reliability
analysis to be carried out with plastic elements instead In this research an attempt was
made for such an analysis but it appeared not to be functional in cooperation with FORM
analysis
Another issue that is recommended to be investigated is the inspection and the validation
of the failure that Plaxis is giving During the soil failure analysis different types of failure
were observed from which some might not be realistic but due to the modeling limitations
Therefore an attempt should be made in order to further eliminate the effect of Plaxis
numerical errors on the reliability analysis that can lead to fake failure modes A check of
the failure points can be also conducted by using a more suitable constitutive model for
the deformations rather than Mohr Coulomb However this is not always feasible for each
individual case Therefore the illustration of the failure mechanisms in large scale
experiments can give a better insight into the systemrsquos behavior and valuable knowledge
can be acquired of the expected failure modes In this way also 3D effects can be
investigated see what is their effect to the dike instability and at what extend they shall be
considered
Apart from the soil body the stresses and the displacements of the structural elements
shall be also verified It is thus recommended that field tests and measurements take place
in embedded elements like sheet piles and anchor in order to validate Plaxis results and
learn what is the most suitable manner for the their simulation (fixed-end-anchor node-
to-node element plate etc)
Concerning the reliability analysis of the specific case study a research on the potential
relevant LSFs is recommended for the different elements Specifically in this thesis the
LSF were chosen in order to represent the ULS of the element There is the possibility that
the SLS is used and certain acceptable deformations can be settled as a threshold for the
Conclusions amp Recommendations
MSc Thesis A Rippi 124
limit states However the choice of this displacement limit is not straightforward and a
qualitative research shall also be conducted in order to identify the impacts of different
values Besides it is not only the value that should be selected but also the location on the
structure that is going to be checked
It is also advised that the potential correlations among the soil parameters to be included
In Wolters (2012) and Bach (2014) correlation matrices are introduced for specific soil
properties that can be included in a future research and reveal what is their effect on the
reliability analysis The influence of correlations depends on what kind of variables are
correlated and what is the influence of those variables on the system behavior However a
preliminary analysis with independent variables is advisable to be carried out at the
beginning so as for someone to be able to analyze the effect of the variables individually
and then evaluate the correlation effect more wisely
It would be also advisable a comparison to be made between the design of a real case
according to a full probabilistic analysis and partial safety factors In that way a direct
insight into the advantages and disadvantages of each method can be gained and
fundamental improvements of the design concept can be made for dikes with retaining
walls
At this moment the most reliable way of conducting a reliability analysis is with sampling
methods like MC or DS However such methods are usually unattractive due to their large
computational time Thus the tolerance of sampling methods accuracy is lowered in order
also to reduce the number of iterations or approximating methods are used like FORM
analysis However even if approximating methods are used they are usually preferred to
be validated with sampling methods in order for their results to be acceptable A real
validation of the failure probability would mean that a structure should be experimentally
tested in different positions and under the same conditions in order to see if the
probability of failure that is calculated in each case is the same Therefore a real
validation of the failure probability cannot yet be initiated but the computational effort of
the sampling methods can be eliminated with the development of technology and the
improvement of the computer science
MSc Thesis A Rippi 125
References
Abdo T amp Rackwitz R (1990) A new beta-point algorithm for large time-invariant and time
variant reliability problems Proceedings of the 3rd IFIP WG 75 Conference Berkeley pp1-12
California USA
Ang AHS and Tang W (1984) Probability Concepts in Engineering Planning and Design Volume
IBasic Principles John Wiley and Sons New York USA
Baecher GB and Christian JT (2003) Reliability and Statistics in Geotechnical Engineering
Chichester West Sussex John Wiley amp Sons Ltd
Behaviour of Soils and Rocks GEO-ENGINEERING - CIE4361 (Vol 2015) Retrieved from
httpsblackboardtudelftnlbbcswebdavpid-2386735-dt-content-rid-
7926659_2courses32281-14150221-possibilities26limitationspdf
Breedeveld J (2011) Technisch Rapport Analyse Macrostabiliteit Dijken met de Eindige Elementen
Methode Bijlage Voorbeeldcase Deltares Delft the Netherlands
Brinkgreve RBJ (2015) Personal discussion TU Delft the Netherlands
Brinkgreve RBJ and Bakker HL (1991) Non-linear finite element analysis of safety factors In
G Beer JR Booker and J P Carter (Eds) Proceedings of the Seventh International Conference on
computer methods and advances in Geomechanics Rotterdam A A Balkema Brookfield
Cukier RI Fortuin M Shuler KE Petschek AG and Schaibly JH (1973) A study of the
sensitivity of coupled reaction systems to uncertainties in rate coefficients I Theory California
University USA
Cundall PA (2002) A discontinuous future for numerical modeling in soil and rock Published in
Proceedings of the third international conference Discrete Element Methods Numerical Modeling
of Discontinua Edited by Cook and Jensen Santa Fe pp 3-4
CUR (2005) Handbook Quay Walls Gouda the Netherlands
CUR-publication 190 (1997) Probabilities in civil engineering Part 1 Probabilistic design in theory
Stichting CUR Gouda
CUR 166 ndash Damwandconstructies (2005) Technical Recommendation 4e herziene uitgave 2005
The Netherlands
Deltaprogramma (2014) Synthesedocument Veiligheid Achtergronddocument B1 Ministrie van
Infrastructuur en Milieu
EN 1990 2002 Basis of Structural Design European Committee for Standardization
References
MSc Thesis A Rippi 126
Farrokh N (2007) Tools and Strategies for Dealing with Uncertainty in Geotechnics In DV
Griffiths and GA Fenton (Ed) Probabilistic methods in Geotechnical Engineering (pp 71-95) New
York NY Springer Vienna
Fenton GA and Vanmarcke EH (1990) Simulation of random fields via local average
subdivision J Eng Mech ASCE 116(8)1733ndash1749
Flood Defences HYDRAULIC ENGINEERING ndash CIE5314 (Vol 2015) Retrieved from
httpsblackboardtudelftnlbbcswebdavpid-2125084-dt-content-rid-
7621002_2courses29759-
131404CIE5314_FloodDefences_05_failure_mechanisms_stability_Schweckendiekpdf
Gemeentewerken-Rotterdam (2003) Alle triaxiaal proeven rotterdam Technical report Gemeentewerken Rotterdam Rotterdam the Netherlands
Ghanem R and Spanos P (1991) Stochastic Finite Elements ndashA Spectral Approach
Springer New York
Gonzaacutelez A Cuadrado A Josa A and Olivella S (2013) Safety assessment of limit equilibrium
methods for the design of cantilever and anchored sheet pile walls Escola de Camins Barcelona
Griffiths DV and Fenton GA (Eds) (2007) Probabilistic Methods in Geotechnical Engineering Lecture notes monographs edited works and proceedings in the field of Mechanics Engineering Computer Science and Applied Mathematics Italy Hasofer AM and Lind NC (1974) An exact and invariant first-order reliability format Journal of
Engineering Mechanics Division ASCE 100(1) 111-121
Hoek E Read J Karzulovic A and Chen ZY (2000) Rock slopes in civil and mining engineering
Published in Proceedings of the International Conference on Geotechnical and Geological
Engineering GeoEng2000 19-24 November Melbourne
Homma T and Saltelli A (1996) Importance measures in in global sensitivity analysis of
nonlinear models Reliab Eng Syst Saf 52(1) pp1-17
Hydra-Ring 12 (2015) Probabilistic toolbox for the WTI2017 Technical Reference Manual Version
12 Deltares Delft
Jaky J (1944) ldquoThe coefficient of earth pressure at rest In Hungarian (A nyugalmi nyomas tenyezoje)rdquo J Soc Hung Eng Arch (Magyar Mernok es Epitesz-Egylet Kozlonye) 355ndash358
Johansson E amp Sandeman E (2014) Modelling of Deep Excavation in Soft Clay A comparison of different calculation methods to in-situ measurements Chalmers University of Technology Gothenburg Sweden
Joint Committee on Structural Safety (1981) General principles on reliability for structural design
International Association for Bridge and Structural Engineering
References
MSc Thesis A Rippi 127
Larsen H Lubking P and Breedeveld J (2013) Ontwerp stabiliteitsschermen (type II) in primaire
waterkeringen (groene versie) Deltares Delft
Melchers R E (2002) Safety and risk in structural engineering Prog Struct Engng Mater
4 193ndash202
Moumlllman A amp Vermeer P (2008) Reliability analysis of a dike failure case study Elbe river
University of Stuttgart Germany
Nelson RB (1999) Lecture Notes in Statistics An Introduction to Copulas Springer New York
OpenTURNS 15 (2015a) Reference Guide OpenTURNSrsquo methods for Step C uncertainty
propagation pp155-161
OpenTURNS 15 (2015b) Use Cases Guide for the Textual User Interface pp190-192
Pandya N Marsdea E Floquet P and Gabas N (2008) Sensitivity analysis of a model for
atmospheric dispersion of toxic gases In B Braunschweig and X Joulia (Ed) 18th European
Symposium on Computer Aided Process Engineering-ESCAPE18 (pp1143-1148) Elsevier BV
Phoon K (Ed) (2008) Reliability-based Design in Geotechnical Engineering Computations and Applications Taylor amp Francis Britain
Plaxis 2D (2015a) Tutorial manual Delft the Netherlands
Plaxis 2D (2015b) Reference Manual The Netherlands
Plaxis (2015) Material Models Manual The Netherlands
Powell MJD (1994) A direct search optimization method that models the objective and
constraint functions by linear interpolation In S Gomez amp JP Hennart (Ed) Advances in
Optimization and Numerical Analysis (pp51-67) Dordrecht Kluwer Academic
Rankine W (1857) On the stability of loose earth Philosophical Transactions of the Royal Society
of London Vol 147
Saltelli A (2004) What is Sensitivity Analysis In A Saltelli K Chan and EM Scott (Ed)
Sensitivity Analysis (pp 3-13) John Wiley amp Sons Publication
Saltelli A S Tarantola F Campolongo and M Ratto (2004) Sensitivity Analysis in Practice
A Guide to Assessing Scientific Models John Wiley amp Sons Ltd
Schittkowski K (1985) NLPQL A Fortran subroutine solving constrained nonlinear
programming problems Annals of Operations Research 5 485-500
Schweckendiek T (2006) Structural Reliability applied to Deep Excavations Coupling Reliability
Methods with Finite Elements TU Delft Delft
References
MSc Thesis A Rippi 128
Schweckendiek T Vrouwenvelder ACWM Calle EOF Jongejan RB and Kanning W (2013)
Partial Factors for Flood Defenses in the Netherlands Modern Geotechnical Codes of Practice ndash
Development and Calibration Fenton GA et al (eds) Special Geotechnical Publication Taylor amp
Francis
Seed RB Bea R G Abdelmalak R I Athanasopoulos A G Boutwell G P Bray J D Briaud J-L Cheung C Cobos-Roa D Cohen-Waeber J Collins B D Ehrensing L Farber D Hanemann M Harder L F Inkabi K S Kammerer A M Karadeniz D Kayen RE Moss R E S Nicks J Nimmala S Pestana J M Porter J Rhee K Riemer M F Roberts K Rogers J D Storesund R Govindasamy A V Vera-Grunauer X Wartman J E Watkins C M Wenk Jr E and Yim SC (2006) Investigation of the Performance of the New Orleans Flood Protection Systems in Hurricane Katrina on August 29 2005 Volume I Main Text and Executive Summary New Orleans USA
Steenbergen RDJM (2013) Lecture notes Probabilistic Design Delft the Netherlands TU Delft
TAW (Technische Adviescommissie voor de Waterkeringen) (2004)Technisch Rapport Waterspanningen bij dijken The Netherlands
Terzaghi K (1942) Theoretical Soil Mechanics New York Wiley ISBN 978-0-471-85305-3 Teunissen JAM (2009) SBW Analyse macrostabiliteit van dijken met Eindige Elementen
Modellen Achtergronden bij Activiteit EEM 04a Opstellen stappenplan Deltares the Netherlands
Terzaghi K (1943) Theoretical Soil Mechanics John Wiley and Sons New York
Turaacutengi T amp Rabitz H (2004) Local Methods In A Saltelli K Chan and EM Scott (Ed)
Sensitivity Analysis (pp 81-99) John Wiley amp Sons Publication
Vergouwe R Huting RJM and van der Scheer P (2014) Veiligheid Nederland in Kaart 2 Overstromingsrisico dijkring 43 Betuwe Tieler- en Culemborgerwaarden The Netherlands
Vrijling JK Schweckendiek T and Kanning W (2011) Safety Standards of Flood Defenses Keynote at ISGSR 2011 (International Symposium on Geotechnical Safety and Risk) Munich Vogt Schuppener Straub amp Braumlu (eds)
Vrijling JK and Gelder PHAJM van (2002) Probabilistic Design in Hydraulic Engineering Lecture notes TUDelft Faculty of Civil Engineering and Geosciences Delft Vrouwenvelder ACWM Steenbergen HMGM and Slijkhuis KAH (1999) Theoriehandleiding PC-Ring Delft the Netherlands TNO
Waarts PH (2000) Structural reliability using Finite Element Analysis An appraisal of DARS
(Directional Adaptive Response surface Sampling) Delft University of Technology Delft the
Netherlands
Wyllie DC amp May CW (2004) Rock Slope Engineering (4th edition) Spon Press London
MSc Thesis A Rippi 129
Appendix A OpenTURNS features
In this Appendix an overview of the most relevant of OT is given Firstly in section 112 an
introduction of the applicable methods in SA is made while in section 52 the results from such an
analysis are presented In Appendix A1 a more detailed elaboration is made of how the FAST
sensitivity method that was applied in this project works
In Appendix A2 an explanation of the different optimization algorithms available for FORM
analysis is given The theory behind these algorithms and their convergence criteria are discussed
while finally an evaluation of their performance after applying them in case of the anchor failure
is presented This evaluation has been conducted in order to decide about the most relatively
reliable algorithm that should be adopted for the reliability analysis
Finally in Appendix A3 the types of probability distributions that have been used in this study
for the random variables and which are available in OT are displayed
A1 Fourier Amplitude Sensitivity Test (FAST) In this section the extended FAST sensitivity method is discussed as it is also described in the
Reference Guide of OpenTURNS 15 (2015) FAST is based on the Fourier decomposition of the
variance of the model response ( ) the latter being represented by its Fourier expansion
is an input random vector of independent components Its key idea is to
recast this representation as a function of a scalar parameter by defining
exploring the support of the input random vector
For each input the same procedure is realized in three steps
1) Sampling
Deterministic space-filling paths with random starting points are defined ie each input Xi is
transformed as follows
( ( )) (A1)
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 130
Figure A1 Search curves in the input space as they are defined from the transformation (the blue
dot indicates the direction of the current search path)
where is the number of input variables N is the length of the discretization of the s-space with
s varying in (-π π) by step 2πN is a random phase-shift chosen uniformly in [0 2π]
is a set of integer frequencies assigned to each input Xi The frequency
associated with the input of interest is set to the maximum admissible frequency satisfying the
Nyquist criterion (which ensures to avoid aliasing effects)
(A2)
with M the interference factor usually equal to 4 or higher It corresponds to the truncation level
of the Fourier series ie the number of harmonics that are retained in the decomposition realised
in the third step of the procedure And the maximum frequency of the complementary set of
frequencies is
(A3)
with the index ldquo-irdquo which meaning ldquoall but irdquo
2) Simulations
Output is computed such as
Then is expanded onto a Fourier series
sum [ ] (A4)
where and are Fourier coefficients defined as follows
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 131
int
(A5)
int
(A6)
These coefficients are estimated thanks to the following discrete formulation
sum
(A7)
sum
(A8)
3) Estimations by frequency analysis
The quantitative measure of sensitivity is represented by Sensitivity Indices The first-order
sensitivity index of an input pi is the measure of the main (direct) effect of pi on the output
variance (where i ne j) the second-order sensitivity indices measures the interaction effect of
pi and pj on the output variance Other higher-order indices are defined in the same manner The
total sensitivity index is the sum of all sensitivity indices involving factor pi (Homma amp Saltelli
1996) For example the total sensitivity index of factor 1 for a model with 3 input factors is
given as
The first-order (Si) and total ( ) indices can be interpreted as following (Pandya et al 2008)
high pi is an influent parameter
and both small pi is not an influent parameter (neither alone nor in interaction with
other parameters)
and nearly the same no interaction of pi with the other parameters
and very different high interaction of pi with the other parameters
Total indices are especially suited to apportion the model output variation to the input factors in a
comprehensive manner The FAST method calculates the first-order and the total sensitivity
indices whereas Sobolrsquo method in addition to these also provides all higher-order sensitivity
indices to determine quantitatively the interaction between parameters However the
computational cost and calculation time of Sobolrsquos method tends to be higher than that of the
FAST method
Combining equations A1-A8 the first-order indices are estimated as follows
sum (
)
sum (
)
(A9)
where is the total variance the portion of D arising from the uncertainty of the ith input and N
the size of the sample using to compute the Fourier series
Subsequently the total-order indices are estimated as follows
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 132
sum (
)
sum (
)
(A10)
where is the part of the variance due to all inputs except the ith input
A2 Optimization Algorithms in FORM
Principles of optimization algorithms
The possible optimization algorithms in OT are
Abdo-Rackwitz (ARF)
Cobyla
Sequential Quadratic Programming (SQP)
Below a description of the main principles of the optimization methods mentioned above is
quoted ARF and SQP methods are elaborated according to Abdo and Rackwitz (1990) while
Cobylarsquos principles according to Powell (1994) Finally an evaluation of the performance of the
different methods follows after applying them in case of the anchor failure
The general optimization problem (objective function) can be written in the classical form
subject to inequality constraints (constraint function) (A11)
The Lagrangian function of the general problem is defined by
sum (A12)
where are the so-called Lagrangian multipliers The SQP method (developed by Schittkowski
(1985)) replaces the original problem by a sequence of quadratic programming problems which
are exactly solvable and which approximate the original one This is done by approximating the
Lagrangian function by its second order Taylor expansion in an initial point u0
(A13)
where
sum
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 133
sum
sum
in which is the gradient operator and represents the Hessian matrix27 of the function f in
the point u0 The optimality conditions for any iteration point k of the sequence of quadratic
expansions are
sum [ sum
] (A14)
(A15)
The exact calculation of the Hessian matrix is generally too expensive and cannot be efficiently
implemented for a general case Therefore the gradient information obtained in each point
during iteration is used to build up an approximation of this matrix using one of the known
update formulas
The new iteration point is defined by
(A16)
where is the step length and is a direction in which a line search is performed The process
stops when the optimality conditions of the original problem are satisfied
The most time consuming part in this algorithm is the updating of the Hessian matrix and the
solution of the system of equations A fair approximation of the Hessian of non-quadratic
functions is also obtained with about n updates of the matrix This means that the approximation
used in the few (say ten) iterations to reach convergence cannot be very good when the problem
has large number of variables The rounding errors during the updating process in large problems
can make the approximate Hessian to become singular Close to singularity the search direction
can be significantly distorted In this case the algorithm has to restart the iteration with a unit
Hessian matrix in the point where singularity occurred
The posterior algorithm ARF developed by Abdo and Rackwitz (1990) overcomes this problem
by obtaining a constant approximation of the true Hessian matrix Only the contribution of the
objective function to the Hessian is considered The numerical decomposition of the matrix
contains the scalar product of the gradients of the constraints as elements in each iteration
27 Hessian is a square matrix of second-order partial derivatives of a scalar-valued function or scalar field It describes the local curvature of a function of many variables
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 134
Cobyla algorithm has been developed by Powell (1994) and it constructs linear polynomial
approximation to the objective and constraint functions by interpolation at the vertices of
simplices (a simplex in n dimensions is the convex hull of n+1 points n being the number of
variables) It generates the next vector of variables from function values at the vertices
of a nondegenerate simplex in In this case there are unique linear
functions and say that interpolate f and at the vertices
and the optimization problem A15 by the linear programming problem
( ) (A18)
( )
The iterative use of expression A18 puts this method in the class of ldquosequential linear
programming algorithmsrdquo It is also essential to notice that in this method the gradients are
derived by interpolation at the vertices of simplices and not analytically as it happens in SQP and
ARF
Additionally Cobyla algorithm makes use of a trust region bound ρ so that the trust region
condition on the new vector of variables is
(A19)
Such a region gives the user some control over the steps that are taken automatically and which
respond satisfactorily to the fact that there may be no finite solution to the linear programming
problem The algorithm also employs a merit function of the form
( ) ( ) [ ( ) ] (A20)
in order to compare the goodness of two different vectors of variables Here is a parameter that
is adjusted automatically depending on the how close to optimizing the objective function the
analysis is and the subscript ldquo+rdquo means that the expression in square brackets is replaced by zero
if and only if its value is negative so we have ( ) ( ) whenever is feasible Parameters
and are changing ( only reduces) automatically according to the improvement of the
optimization problem that sequential vectors might cause Therefore if the change in the merit
function does not happen to improve the optimization then these values are changing Especially
regarding the parameter user gives the initial and the final values of it namely and (it
is recommended that be a reasonable change to make the variables for a coarse exploration
of the calculation while should be approximately the required distance from the final vector
of variables to the solution of the optimization problem)
Convergence criteria
The convergence of the aforementioned algorithms is controlled by the evaluation of the
following errors expressed in the standard space (for the relationship between the real space (x-
space) and the standard space (u-space) see section 422) (OpenTURNS 15 (b) 2015 pp 190-
192)
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 135
The absolute error which is the distance between two successive iterates
(A21)
The constraint error which is the absolute value of the limit state function minus the
threshold
(A22)
The relative error which is the relative distance between two successive iterates (with
regards to the second iterate)
(A23)
The residual error which is the orthogonality error (lack of orthogonality between the
vector linking the center and the iterate and the limit state function)
( ) (A24)
The algorithm converges if all the final error values are less than the maximum value specified by
the user The convergence can be also controlled by altering the maximum number of iterations
which should be higher in case of stricter required errors What is more there is the possibility to
manipulate the starting point that the algorithm uses for its first iteration Usually the default
starting point in a FORM analysis consists of the mean values of the random variables However if
there is already some knowledge about the potential design point it can be set as the starting
point of the algorithm in order to accelerate the convergence and save a lot of time
The aforementioned errors the maximum number of iterations and the starting point were
manipulated during the analysis in order to understand the functionality of the different
algorithms and to improve their efficiency
Evaluation of the algorithms performance
An evaluation of the above optimization algorithms has been made in order to make a decision of
the most suitable for the current project In principle a literature overview regarding the
performance of the algorithms is presented Then the 3 algorithms were tested considering the
anchor failure limit state function (see Eq 55) with a stress limit of 450000 Pa and their results
are discussed below
According to Abdo and Rackwitz (1990) the SQP method seems to have better convergence
behavior for highly curved constraint functions which is especially true when very expensive
structural state functions are involved such as finite element structural analyses However the
mentioned algorithm fails to reach convergence at a problem dimension (amount of random
variables) of around n=50 Therefore for increased dimensionality SQP is less efficient in terms
of storage and CPU time compared with other methods due to singularity of the updated Hessians
matrices In OT a warning is introduced since an analysis has started that a default
implementation of Hessian is being used and that the computation can be severely wrong
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 136
Abdo and Rackwitz (1990) also proved that ARF algorithm can handle problems of up to 2000
variables while the storage requirements and the CPU time are much smaller than with SQP
Therefore optimization problems of high dimensionality can be evaluated
In OT a warning appears at the beginning of a FORM calculation saying that a default
implementation of the gradient is being used and that the computation can be wrong A similar
message also appears for the SQP algorithm saying that a default implementation of the Hessian
matrix is used and that the computationrsquos results might be unreliable Moreover in both
algorithms it is advisable to check the values of the random variables that were used during the
analysis in order to ensure that they fluctuate within acceptable boundaries (Figure A4 is an
example of a variable that surpassed the minimum boundary at some stage)
Firstly the Cobyla algorithm was applied on the case study in combination with a FORM analysis
Cobyla does not require a gradient evaluation of the limit state function and does not use a default
implementation of the Hessians matrix as ARF and SQP does while the computational time is
quite small comparable to the aforementioned methods
The Cobyla algorithm was tested with (i) 3 (ii) 10 and (iii) 19 parameters respectively in order to
check its efficiency28 As it has been mentioned in the previous section the convergence criteria to
be steered are the errors (absolute relative residual and constraint) the maximum number of
iterations and the starting point At first the default errors were used with a value of 10-3 but it
was realised that such an order of errors lead to non-convergence of the algorithm Therefore the
algorithm was then checked after relaxing the maximum errors to values of 01 05 10 50 and
100 Finally the combination of the errors that led to convergence was 01 05 01 02 for the
maximum absolute relative residual and constraint error respectively After the analysis the
number of LSF calls in case (i) (ii) and (iii) was 25 even if the maximum one was set to 100 and
the probability of failure was around 035 The probability of failure was quite high and that is
why an inspection of the output file was carried out In Figures A2 and A3 the values of the
anchor limit state function are presented in the right graph during the iterations It was then
observed that the threshold of 450 kPa was never surpassed during the iterations
28 All of the algorithms were first checked with very simple LSF (linear and quadratic) and their results were verified by MC and DS
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 137
Figure A2 Friction angle and cohesion progress during a FORM analysis with Cobyla and with 3 random variables The figure at the right illustrates the limit state function evolution
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 138
Figure A3 Friction angle and cohesion progress during a FORM analysis with Cobyla and with 19 input parameters The last figure at the right illustrates the limit state function evolution
An additional action was then taken by relaxing the maximum number of iterations from 100 to
200 but no improvement was noticed (the LSF calls were again 25) It was also given extreme
values to the errors and especially to the absolute constraint and relative error in order to check
the response of the algorithm but that did not lead to better results Furthermore the
parameter (see theory part for explanation) was manipulated but even then no further
improvement was noticed in terms of the results In Figures A2 and A3 the performance of the
algorithm during the analysis can be seen through the values of the variables that were assigned
during the iterations
As someone can notice from the graphs above it is obvious that Cobyla does not perform very
well in high dimensionality problems as the variables do not seem to vary a lot More specifically
in Figure A3 with the 19 parameters it can be noticed that the random variable does not change
at all apart from two peaks at the beginning and at the end of the analysis respectively The
incapability of the algorithm to change the parameters in order to reach the limit of the anchor
stress ie 450 kPa (while here a maximum of approximately 267 kPa is reached) leads to a poor
approximation of both the design point and the probability of failure In Figure A2 with only 3
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 139
parameters the variability of the parameter is higher during the analysis which leads to a better
approach of the limit (around 370 kPa which is larger than with the 19 parmaters) but even then
a failure probability of 035 was not a realistic value On the other hand Cobyla is easy to use only
for small number of variables (ie less than 9) with linear LSF otherwise the linear
approximations can be highly inefficient Apparently the limit state function of the anchor stress
level as a function of the soil parameters is much more complex in order to be approximated by
this algorithm
Therefore the other available optimization algorithms ie ARF and SQP were examined as well
Because of the similarity of the two algorithms only the ARF performance is presented below and
a link with SQP is made when it is needed
The ARF algorithm was tested initially with 19 parameters The settings for the errors that were
used were 01 for the absolute relative residual and constraint error respectively and the
maximum number of iterations was set to 100
Figure A4 Friction angle and cohesion progress during a FORM analysis with ARF algorithm and with 19 input parameters The figure at the right illustrates the limit state function evolution while the red spots indicate the points where the 450 kPa threshold was surpassed
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 140
The algorithm worked efficiently which means that the algorithm managed to converge towards
the threshold of 450 kPa of the limit state function as it is depicted in Figure A4 (see red dots) It
was also noticed that the random variables fluctuated a lot during the iteration steps even if the
amount was significant (19 parameters) The capability of the algorithm to vary all the 19
variables evenly and in a wide range makes it possible to find the real design point and thus to
come up with a more reliable failure probability The resulted probability of failure was 3810-9
and the reliability index was ϐ=58
However the analysis took almost 4 days to finish which is a long duration and the number of LSF
calls counted at 4151 Moreover even if the results were promising after tracing all the 19
parameters in order to inspect the algorithmrsquos performance it was found out that some variables
took unacceptable values29 during the analysis As an example in Figure A5 Poissonrsquos ratio
received negative values (see red circle) at the beginning of the iterations There were also other
such examples in the same analysis such as the friction angle which took sky-high values and the
interface strength which took negative values as well as values above 1 which is unacceptable It
should be mentioned that the design point of the algorithm was completely reasonable but even
then such a performance can prevent the algorithm from converging to the right design point
quickly and increase the computational effort
Figure A5 Poisson ratio progress during the FORM analysis with ARF algorithm
Taking into account the above implications there are two main concerns at that stage (a) the
high computational time and (b) the unreasonable values that the variables took Firstly in order
to reduce the computational time the maximum errors were relaxed More precisely it was
noticed that after relaxing the maximum relative error a lot which is more related to the
29 In some cases a value of ldquoinfrdquo was given to the variables that led to termination of the analysis without results
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 141
relaxation factor (in this case an extreme value of 1000030 was tried) and increasing the absolute
error to 05 the number of LSF calls was reduced to 724 and the analysis was finished after
almost 10 hours In Figure A6 the values that the cohesion and the friction angle took during the
iterative procedure while the red spots in the right figure indicate the points that are included in
the failure space Figure A4 shows the progress of friction angle after relaxing the errors
Figure A6 Progress of the friction angle during the FORM analysis with ARF algorithm with errors
05 10000 01 01 for the absolute relative residual and constraint error respectively
As a general conclusion after the test of different combination of maximum errors by relaxing the
constraint absolute and relative error too much successive iterations of a certain variablersquos value
seem to have a wide difference which can detain the algorithm from converging Therefore the
wisest in this case study is to keep the maximum errors below or around 10 except for the
relative and the constraint error that can relax more than 1 (depending on the case) in order to
reduce the computational time to the minimum possible
To prevent having unacceptable values of the parameters and in order to increase the efficiency of
the algorithm the random variables were reduced to 11 after a SA and some of them were
truncated in a certain range according to observations and experience until now However the
algorithm then started to give an error due to zero gradient of the objective function at a specific
point (similarly SQP gave an error due to zero hessian matrix at a specific point) while the upper
limit of the LSFs could not be reached As a countermeasure the errors were relaxed (05 10 05
10 for the absolute relative residual and constraint error respectively) in order for the algorithm
not to be trapped in a specific point while also the step of the algorithm was manipulated It
was finally proved that this kind of error stemmed from the fact that the structural elements were
considered as elastoplastic and thus a limit of the stress level was considered that misled the
algorithm The elements were eventually considered as elastic and the numerical problem was
solved
30 Practically such a value has the meaning that there is no limitation of the relative error during the optimization In the standard space (where these errors are expressed) a value also greater than 10 can be also quite high However here extreme values were given due to the time limit to test multiple values
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 142
Before figuring out that the elastoplastic behaviour was giving the problem Cobyla as a gradient
free algorithm was tested again as a potential solution but with changing the starting point close
to the design one that was found with ARF It is worthwhile to take a look into the performance of
the algorithm with 19 random variables in Figure A7 In that way we boost Cobyla to converge
and it eventually performed quite differently than in Figure A3 and without giving convergence
problems The computational time was estimated to be 9 hours and the probability of failure was
very small in the order of 10-120 However the validation of the failure probability with DS
showed that the result of Cobyla was far from correct and thus ARF was preferred for continuing
with the rest of the analysis For educational reasons a FORM analysis with ARF was again
performed after changing the starting point and it was found that the computational time was
reduced to 4 hours Therefore it is still worthy to be aware of the fact that by changing the
starting point the performance of the algorithm can improve a lot However that requires a prior
knowledge of the potential design point that usually does not exist
Figure A 7 Cobyla performance after changing the starting point closer to the failure point
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 143
It must be also mentioned that by reducing the number of the variables to 11 as it was mentioned
before the algorithms could not converge to the errors that has been set and the computational
time increased compared with the case of 19 variables Taking into account the high interaction
among the variables that was proved in section 73 through the SA it can be possible that the
combination of the soil variables could lead to failure rather than the variables individually
Therefore an analysis was tried again with the complete set of the parameters and it was realised
that the algorithm did converge to the limit value
A3 Distribution Types
Uniform Distribution
If the probability of a result is between a lower limit a and an upper limit b and if every result
between these limits is equally probable a uniform distribution should be used This distribution
is defined by the two limit parameters and the probability density function is written
(A25)
and the CDF is
(A26)
Figure A8 Probability density function of a uniform distribution with a=-2 and b=2
The mean and the variance of the distribution are also defined by the limit parameters
(A27)
(A28)
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 144
Normal Distribution
One of the most commonly used distribution types is the normal distribution or Gauss-
distribution and it is defined by its first two central moments the mean micro and the standard
deviation σ The PDF is given by
radic
(A29)
and its CDF is
int
radic
(A30)
Figure A9 Probability density function of a normal distribution with different mean and standard
deviation values (notice that the curve with micro=0 and σ2=1 belongs to the standard normal distribution)
The normal distribution with mean and standard deviation is called the standard
normal distribution In the most reliability methods such as FORM SORM DS etc the analysis is
carried out in the standard normal space which actually means that the stochastic parameters are
transformed into standard normal variables
Lognormal Distribution
If a variable X has a lognormal distribution the variable Y=ln(X) has a normal distribution The
probability density function is defined as
radic (
) (A31)
where and the mean and the standard deviation of the normally distributed random
variable Y Sometimes it is useful to work with the Yrsquos parameters rather than Xrsquos These
parameters can be expressed as follows
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 145
radic
(A32)
(A33)
where and are the mean and the standard deviation of Y In Figure A10 an example is given
of a log-normally distributed variable with different values of mean and standard deviation
Figure A10 Probability density function of a log-normal distribution with different mean and standard deviation values
Truncated Normal Distribution
The truncated normal distribution is the probability distribution of a normally distributed
random variable whose value is either bounded below or above or both
Suppose has a normal distribution and lies within the interval
Then conditional on has a truncated normal distribution whose probability
density function is
(
) (
)
(A34)
and otherwise and is the probability density function and the CDF of the standard
normal distribution similar to Eq (1) and (2) If
and
then the CDF is
defined as
(A35)
Then the mean value and the variance of the conditional will be respectively
Appendix A ndash OpenTURNS features
MSc Thesis A Rippi 146
(A36)
[
(
) ] (A37)
In the next figure the PDF histograms of a variable with a mean of 0 and a standard deviation 1
were plotted for two cases (a) variable is normally distributed (b) variable is truncated-normally
distributed with =-2 and b=2
Figure A 11 Truncated normal distribution in comparison with normal distribution
MSc Thesis A Rippi 147
Appendix B Plaxis 2D (2015) features
In this Appendix the basic features of Plaxis 2D that were treated in this thesis are discussed First
of all an introduction over the Mohr Coulomb model is made as the soil failure criteria stem from
that model Then the initial stress generation in Plaxis is discussed and how that was applied in
the current master thesis while also the φ-c reduction technique is described which was used in
the SA of the soil failure Last but not least a more detailed explanation of the interface strength
Rinter is given which represents the modeling of the soil-structure interaction in Plaxis and it has
been considered as a random variable in the current project
B1 Mohr Coulomb failure criterion Soil tends to behave in a highly non-linear way under load This non-linear stress-strain behaviour
can be modelled at several levels of sophistication There are different constitutive models
available for simulating the behaviour of the soil whose parameters can change depending on the
level of sophistication that is required in each case
As someone can notice in Table B1 different models can be applied in different cases of soil
behaviour Hardening Soil model is becoming more and more interesting in soil structures as it
has been proved to approach the plastic behaviour of the soil and the non-linear stress-strain
relationship in a reliable and promising level Moreover UBCSAND31 and the Hypoplastic model
are more sophisticated models capable of representing several aspects of soil behaviour
however the amount of the model parameters is large and for that reason they are recommended
only in case that enough soil data are available for determining these parameters
Table B 1 Features of material models (source Behaviour of Soils and Rocks 2015)
ModelFeature Elasticity Failure Hardening Softening Small strain stiffness
Cyclic loading
Liquefaction Creep
Mohr-Coulomb x x
Duncan-Chang x x
Hardening Soil x x x
HSsmall x x x
x
Modified Cam-Clay
x x x x
Soft Soil x x x
Soft Soil Creep x x x
x
UBCSAND x x x
x x
Hypoplastic (x) x x x
x (x)
31 UBC are the initials for University of British Columbia in Vancouver in Canada
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 148
In this master thesis the Mohr-Coulomb model is going to be applied which is simple but an
efficient model to describe the soilrsquos failures condition Below a description of that model is made
The Mohr-Coulomb model is a linear elastic perfectly plastic model (see Figure B1) The linear
elastic part is based on Hookersquos law while the perfectly plastic part is based on the Mohr-Coulomb
failure criterion formulated in a non-associated plasticity framework
The main principle of elastoplasticity is that total strains and strain rates consist of elastic and
plastic strain components
(B1)
(B2)
in which the elastic strain rate component abides by the Hookersquos law and thus it can be expressed
as follows
(B3)
where is the elastic stress-strain matrix and the effective stress rate The plastic strain
component is given by
(B4)
where is a scalar defining the magnitude of the plastic strains called plastic multiplier and
is
a vector describing their direction with to be the so-called plastic potential function
Figure B1 Basic principle of elastoplasticity
The occurrence of (elasto)plasticity can be indicated by the yield function which is
illustrated in Figure B2 in principle stress space and it is derived according to the Mohr-Coulomb
failure criterion expressed in the next formula
(B5)
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 149
Considering the sign of this yield function the different soil states can be recognised as follows
if flt0 (inside yield contour) pure elastic behaviour
if f=0 and dflt0 unloading from aplastic state (=elastic behaviour)
if f=0 and df=0 elastoplastic behaviour
if fgt0 (outside yield contour) impossible stress state
Figure B2 Mohr-Coulomb yield surface in principle stress space
Finally in Mohr-Coulomb model the calculation of the plastic strains is achieved by the so-called
non-associated flow rule where with the dilatancy angle
This makes a difference between the friction and the dilatancy angle which prevents the model
from overestimating potential uplift
Concluding this model uses 5 input parameters
Stiffness parameters E (Youngrsquos modulus) and v (Poissonrsquos ratio) as they are included in
the elastic stress-strain matrix according to Hookersquos law
Strength parameters c (cohesion) and (friction angle)
Dilative behaviour ψ (dilatancy angle)
The author is aware of the possibilities and limitations of this model In particular the model is
capable of
a first order approach of the soil behaviour in general
describing quite well the (drained) failure behaviour
It should be mentioned at that point that the Youngrsquos modulus E cannot be steered via the
Python interface (as it appears to be a readonly parameter) and thus the shear stiffness G was
considered instead that is defined as follows
[kPa]
f=0 (yield contour)
flt0
fgt0
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 150
On the other hand the model has also limitations that are referring to
the linear elastic behaviour until failure (no strain- or stress path-dependent stiffness
behaviour)
overestimation of the shear strength in case of an undrained behaviour
However this model is considered reliable for detecting failure and giving realistic stresses
magnitudes in the soil and structural elements
Moreover in this project a non-associative behaviour was assumed by considering a zero
dilatancy angle However it has been proved by Teunissen (2009) that in case of deep failure
mechanisms in a dike structure with soil materials of a friction angle above 30deg problems arise in
the numerical analysis in finite element calculations Particularly the strength in the sliding
surface decreases due to rotation and the ground stresses are relieving around it by releasing
energy which might lead to calculation implications These phenomena are not treated in this
thesis but it is advisable to check and review the results after such an analysis
B2 φ-c Reduction Technique In this section the ldquoSafety Calculationrdquo is applied as it is described in Plaxis (Plaxis 2D (a) 2015)
In the Safety Calculation the φ-c reduction technique (Brinkgreve amp Bakker 1991) is adopted
where the shear strength parameters tanφ and c of the soil as well as the tensile strength are
successively reduced until failure of the structure occurs The dilatancy angle ψ is in principle not
affected by the phic reduction procedure However the dilatancy angle can never be larger than
the friction angle When the friction angle φ has reduced so much that it becomes equal to the
(given) dilatancy angle any further reduction of the friction angle will lead to the same reduction
of the dilatancy angle The strength of interfaces if used is reduced in the same way
The total multiplier is used to define the value of the soil strength parameters at a given
stage in the analysis
(B6)
where the strength parameters with the subscript ldquoinputrdquo refer to the properties entered in the
material sets and the parameters with the subscript ldquoreducedrdquo refer to the reduced values used in
the analysis is set to 10 at the beginning of a calculation to set all material strengths to
their input values
The incremental multiplier is used to specify the increment of the strength reduction of the
first calculation step (this increment is by default 01 but user can change it) The strength
parameters are successively reduced automatically until all required steps have been performed
(by default these steps are set to 100 but this can be subjected to changes depending on the
situation) It must always be checked whether the final step has resulted in a fully developed
failure mechanism In other words if has reached an almost constant value while the
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 151
deformation is continuing that indicates an equilibrium state In that case the factor of safety is
given by
(B7)
B3 Initial Stress Generation The first construction phase in Plaxis was the initial stress generation before the dike has been
constructed In that phase for horizontal soil surfaces with any soil layers parallel to this surface
and horizontal phreatic level32 a K0-procedure can be applied where K0 is the ratio between
horizontal and vertical stresses
(B8)
In practice the value of for a normally consolidated soil is often assumed to be related to the
friction angle by Jakyrsquos empirical expression (Jaky 1944)
Using very low or very high K0-values in the K0 procedure may lead to stresses that violate the
Mohr-Coulomb failure criterion In this case Plaxis automatically reduces the lateral stresses such
that the failure condition is obeyed Hence these stresses points are in a plastic state and are thus
indicated as plastic points Although the corrected stress state obeys the failure condition it may
result in a stress field which is not in equilibrium
B4 Interface Strength In FEM for a proper modelling of soil-structure interaction interface elements can be added to
plates in order to simulate the thin zone of intensely shearing material at the contact between the
plate and the soil
The strength properties of the interface are linked to the strength properties of the adjacent soil
layer and they are reduced by a factor of as follows
(B9)
(B10)
(B11)
where index ldquoirdquo refers to the interfacersquos parameters whereas index ldquosoilrdquo refers to the adjacent soil
parameters The behaviour of the interfaces is defined according to the Coulomb criterion which
distinguishes between elastic behaviour where small displacements can occur within the
interface and plastic interface behaviour when permanent slip may occur
32 For non-horizontal surfaces a Gravity loading is recommended
Appendix B ndash Plaxis 2D (2015) features
MSc Thesis A Rippi 152
When the interface is elastic then both slipping (relative movement parallel to the interface) and
gapping (relative displacements perpendicular to the interface) can occur The magnitudes of the
interface displacements are
(B12)
(B13)
where is the shear modulus of the interface is the 1-D compression modulus of the
interface is the virtual thickness of the interface (generated automatically from Plaxis) is
the elastic interface normal stiffness and is the elastic interface shear stiffness
The shear and compression moduli are related as follows
(B14)
(B15)
Taking into account the above relations the displacements can be highly influenced by the elastic
parameters and thus the strength reduction factor can have a considerable impact on the
systemrsquos behaviour In the Reference manual of Plaxis 2D (2015) a value of 23 is generally
recommended while a value of 1 would not normally be used It should be stated that in the
framework of this thesis there are no available measurements in order to define the sensitivity to
this parameter and hence it is considered uncertain variable following a truncated normal
distribution
Appendix C ndash NEN 6740 ndash Table 1
MSc Thesis A Rippi 153
Appendix C NEN 6740 - Table 1
MSc Thesis A Rippi 154
MSc Thesis A Rippi 155
Appendix D Input Files for the Reliability Analysis
Below an example of the input files that were used for the anchor the sheet pile the soil body and
the system failure is presented in Figures D1 D2 D3 and D4 respectively In particular
regarding the input file of the anchor failure the meaning and the scope of each line is briefly
explained in Table D1
Figure D 1 Anchor analysis input file (with FORM method)
APPENIDX D ndash Input Files for the Reliability Analysis
MSc Thesis A Rippi 156
Table D1 Explanation of input components
A Reliability method and its parameters eg FORM with a maximum number of iterations 100 a maximum absolute error of 05 a relative error of 1000 a residual error of 05 and a constraint error of 05
B Number of random variables
C Soil material as it is specified in Plaxis eg AKleiVenigOnder is the clay layer under the dike
D Soil parameter eg Rinter is the interface strength of the clay layer
E Probability distribution and its parameters eg Truncated normal distribution with mean 066 standard deviation 013 and bounds from 00 to 10
F Copula type
G I Limit state function I is the threshold of the function
H Response surface function that is activated in case of Plaxis errors
Figure D 2 Sheet pile wall analysis input file (with FORM method)
APPENIDX D ndash Input Files for the Reliability Analysis
MSc Thesis A Rippi 157
Figure D 3 Soil analysis input file (with DS method)
APPENIDX D ndash Input Files for the Reliability Analysis
MSc Thesis A Rippi 158
Figure D 4 System analysis input file (with DS method)
MSc Thesis A Rippi 159
Appendix E Characteristic and mean values
The characteristic value of a parameter implies that there is 5 probability that the real value is
higher (load) or lower (resistance) than the characteristic In Figure 22 the probability density
function of a standard normal distributed parameter (mean value equals to zero and standard
deviation 1) is depicted In case of such a distribution the characteristic value is 164σ higher than
the mean value
Figure E1 Characteristic value illustration for a normal distribution (source Wolters 2012)
In this study full probabilistic calculations are taking place in order to estimate the corresponding
structural reliability According to the probabilistic design concept the mean values and the
standard deviations are used instead of characteristic values In that case characteristic values
and shall be converted to their mean values and as follows
where
(E1)
(E2)
and are the values for load or resistance respectively to which a probability of (non-
)exceedance of 95 corresponds (non-exceedance for the load and exceedance for the resistance)
(for a standard normal distributed parameter this is 1645) The definition of the rest parameters
is given in section 423
MSc Thesis A Rippi 160
MSc Thesis A Rippi 161
Appendix F Reliability methods
In this appendix an overview of different reliability methods is given More precisely in F1 an
explanation is given concerning the generation of random samples that is applied in MC as well as
in other sampling methods In Appendix F2 more advanced sampling methods than crude MC are
presented that can be used as accelerating reliability methods in the sense that the computational
time can be reduced Finally in Appendix F3 some basic principles of the FOSM are elaborated
that can help to better understand the concept of FORM as well
F1 Generation of random samples in Monte Carlo
The non-exceedance probability of an arbitrary random variable is uniformly distributed between
zero and one regardless of the distribution of the variable In formula (CUR ndashpublication 190
1997)
(F1)
is the uniformly distributed variable between zero and one and is the non-exceedance
probability Thus for the variable X
(F2)
in which is the inverse of the PDF of X Using this formula a random number X can be
generated from an arbitrary distribution by drawing a number of from the uniform
distribution between zero and one
More or less the same way base variables of a statistical vector can be drawn from a known joint
probability distribution function However the joint PDF must then be formulated as the product
of the conditional probability distributions of the base variables of the vector In formula this is
( )
By taking m realizations of the uniform probability distribution between zero and one a value can
be determined for every
(F3)
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 162
If the variables are statistically independent this can be simplified to
(F4)
By inserting the values for the LSF one can check whether the obtained vector ( ) is
located in the safe area
F2 Other Sampling Methods
Quasi-Monte Carlo Simulation
As an alternative to pseudo-random number generation (used in Crude MC) the quasi-random
number sequences are introduced from where the method Quasi-MC took its name Quasi-MC is a
technique which approximates the probability of failure using low discrepancy sequences33
x1hellipxN instead of randomly generated sequences
If the domain of failure is the goal is to estimate the following
probability
int ( )
(F5)
The main principle of this method is the integral of the function on [ ] can be
approximated by using some low discrepancy sequence x1hellipxN as follows
int ( )
sum
(F6)
The low discrepancy sequence is generated in Δ according to Lebesgue measure and then may be
transformed to any measure micro thanks to the inverse CDF technique in order to approximate the
integral
int ( )
sum
(F7)
In Figure 4 a comparison is made between quasi-random sample and pseudo-random sample in
order to illustrate the difference in sampling for the two methods
33 The discrepancy of a sequence is low if the number of points in the sequence falling into an arbitrary set B is close to proportional to the measure of B as would happen on average (but not for particular samples) in the case of a uniform distribution
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 163
Figure F1 Example of Quasi random Sample (left) and Pseudo Random Sample (right) (source
Waarts 2000)
This method a priori provides an asymptotically higher convergence rate than traditional MC but
no general rule can guarantee a better efficiency of the Quasi-MC sampling than the classical MC
sampling The advantages of Quasi-MC sampling tend to disappear with the increase of the
number of dimensions of the input variables It is recommended to use the Quasi-MC technique
with high sampling sizes or with very low dimensionality (in OT this method is valid only if the
input random variables are independent) (OpenTURNS 15 2015)
Importance Simulation
The main idea of the Importance Sampling method is to replace the initial probability distribution
of the input variables by a more efficient one Efficient means that more events will be counted
in the failure domain and thus reduce the variance of the estimator of the probability of
exceeding a threshold In other words importance sampling starts with the observation that if we
are going to sample randomly we should distribute the points to put the most of them in the
region that contains information (ie the failure region) and to waste as few as possible
In this method a sampling density is used instead of the actual probability density function
For a MC simulation the probability of failure is calculated by
sum ( ( )) ( )
( )
(F8)
where ( ( )) ( )
( ( )) ( )
The problem is that prior knowledge is needed of the failure area which in most cases is not
available In Figure 5 examples of the different sampling methods are shown
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 164
(a) (b)
(c)
Figure F2 Comparison of Importance Sampling (a) MC (b) and Numerical integration (c) (source
Latin Hypercube Simulation
Latin Hypercube Sampling is a sampling method enabling to better cover the domain of variations
of the input variables thanks to a stratified sampling strategy (in OT this method is applicable in
the case of independent input variables) Stratified sampling tries to place the sampling points so
that more of them will be found in regions where the variance of the limit state function g(x) is
largest In Figure F3 a comparison is made between the LHS and MC in terms of their sampling
technique The sampling procedure according to Baecher and Christian (2003) is based on the
next steps
1 Divide the range of each variable into several intervals of equal probability
2 If there are D variables to be sampled each in k regions then for each variable the k
sampling points are placed in independent random order
3 K samplings are made by choosing the first randomized value for each variable the second
randomized value for each variable and so on until k sets of randomized variables have
been chosen
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 165
Figure F 3 Comparison between Latin Hypercube sampling and MC (source OpenTurns 15 2015a)
F3 First Order Second Moment (FOSM) Method The First-Order Second-Moment (FOSM) approach (Ang amp tang 1984) provides analytical
approximation for the mean and the standard deviation of a parameter of interest as a function of
the mean and standard deviations of the various input factors and their correlations
Consider Z to be a function of random variables x1 x2 hellip xn that is
(F9)
In general case are correlated with covariance matrix [C] ie [C]=[σ][R][σ] where [σ]
is a diagonal matrix of standard deviations and [R] is the (positive-definite and symmetric)
correlation matrix with diagonal elements Rii=1 and non-diagonal elements Rij=ρij (ρij is the
correlation coefficient between variables i and j) In scalar notation Cij= σiσjRij (Farrokh 2007)
The first order approximation to the mean variance and standard deviation of the Z function is
based on the first terms of a Taylor series expansion of Z The following first-order estimates of
mean and variance are obtained
[ ]
(F10)
[ ] (F11)
Where the vector b denotes evaluated at the mean values of ie
(F12)
If there is no correlation among the variables Eq 223 can be written as
APPENDIX F ndash Reliability methods
MSc Thesis A Rippi 166
sum
(F13)
Then the reliability index can be calculated in 6 steps (Baecher amp Christian 2003)
1 Identify all variables that affect the mechanism that is researched
2 Determine the best estimate (usually the mean value) of each variable E[ ] and use these
to calculate the best estimate of the function E[Z]
3 Estimate the uncertainty in each variable and in particular its variance Var[ ]
4 Perform sensitivity analyses by calculating the partial derivatives of Z with respect to each
of the uncertain variables or by approximating each derivative by the divided difference
5 Use the equation of Var[Z] to obtain the variance of the function Z
6 Calculate the reliability index
[ ]
(F14)
MSc Thesis A Rippi 167