Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
SEVERAL NEW DESIGNS FOR PID, IMC,DECOUPLING AND FUZZY CONTROL
BY
YANG YONG-SHENG (B.ENG., M.ENG.)
DEPARTMENT OF ELECTRICAL AND
COMPUTER ENGINEERING
A THESIS SUBMITTED
FOR THE DEGREE OF PHILOSOPHY DOCTOR
NATIONAL UNIVERSITY OF SINGAPORE
2003
Acknowledgments
I would like to express my sincere appreciation to my advisor, Professor Wang,
Qing-Guo, for his excellent guidance and gracious encouragement through my
study. His uncompromising research attitude and stimulating advice helped me
in overcoming obstacles in my research. His wealth of knowledge and accurate
foresight benefited me in finding the new ideas. Without him, I would not able
to finish the work here. I am indebted to him for his care and advice not only in
my academic research but also in my daily life. I wish to extend special thanks to
Professor C. C. Hang for his constructive suggestions which benefit my research a
lot. It is also my great pleasure to thank Dr. Chen Ben Mei and Dr. Ge Shuzhi
Sam who have in one way or another give me their kind help.
Also I would like to express my thanks to Dr. Zheng Feng and Dr. Lin Chong,
Dr. Zhang, Yong, Dr. Zhang, Yu, and Dr. Bi, Qiang for their comments, advice,
and inspiration. Special gratitude goes to my friends and colleagues. I would like
to express my thanks to Dr. Yang, Xue-Ping, Mr. Huang Xiaogang, Mr. Huang,
Bin, Ms. He Ru, Mr. Guo Xin, Mr. Zhou Hanqin, Mr. Lu Xiang, Mr. Li Heng
and many others working in the Advanced Control Technology Lab. I enjoyed
very much the time spent with them. I also appreciate the National University of
Singapore for the research facilities and scholarship.
Finally, I wish to express my deepest gratitude to my wife Han Rui. Without
her love, patience, encouragement and sacrifice, I could not have accomplished this.
I also want to thank my parents and brothers for their love and support, It is not
possible to thank them adequately. Instead, I devote this thesis to them and hope
they will find joy in this humble achievement.
i
Contents
Acknowledgements i
List of Figures vii
List of Tables viii
Summary ix
Nomenclature xiii
1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Organization of the Thesis . . . . . . . . . . . . . . . . . . . . . . . 10
2 Three New Approaches to PID Controller Design 12
2.1 Preview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2 Robust PID Controller Design for Gain and Phase Margins . . . . . 13
2.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2.2 PID Controller Design Using LMI . . . . . . . . . . . . . . . 14
2.2.3 Tuning Guidelines . . . . . . . . . . . . . . . . . . . . . . . 19
2.2.4 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3 Quantitative Robust Stability Analysis and PID Controller Design . 25
2.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
ii
Contents iii
2.3.2 Review of Robust Control Theory . . . . . . . . . . . . . . . 26
2.3.3 Quantitative Robust Stability . . . . . . . . . . . . . . . . . 28
2.3.4 Second-Order Uncertain Model . . . . . . . . . . . . . . . . 30
2.3.5 Robust PID controller Design . . . . . . . . . . . . . . . . . 32
2.3.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.4 PI Controller Design for State Time-Delay Systems via ILMI . . . . 37
2.4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.4.2 Problem Description . . . . . . . . . . . . . . . . . . . . . . 38
2.4.3 Stabilizing Control . . . . . . . . . . . . . . . . . . . . . . . 39
2.4.4 Suboptimal H∞ Control . . . . . . . . . . . . . . . . . . . . 42
2.4.5 Control Design with PI Controllers . . . . . . . . . . . . . . 44
2.4.6 A Numerical Example . . . . . . . . . . . . . . . . . . . . . 45
2.4.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3 Advance in Robust IMC Design for Step Input and Smith Con-
troller Design for Unstable Processes 48
3.1 Robust IMC Design via Time Domain Approach . . . . . . . . . . . 49
3.1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.1.2 The Robust IMC Design . . . . . . . . . . . . . . . . . . . . 50
3.1.3 LMI Solution . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.1.4 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.1.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.2 Robust IMC Controller Design via Frequency Domain Approach . . 60
3.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.2.2 IMC Design Review and New Formulation . . . . . . . . . . 60
3.2.3 Controller Design with Fixed Poles . . . . . . . . . . . . . . 62
3.2.4 Controller Design with General Form . . . . . . . . . . . . . 66
3.2.5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.2.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
3.3 Modified Smith Predictor Control for Disturbance Rejection with
Unstable Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
Contents iv
3.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
3.3.2 The Proposed New Structure . . . . . . . . . . . . . . . . . 78
3.3.3 Internal Stability . . . . . . . . . . . . . . . . . . . . . . . . 79
3.3.4 Controller Design . . . . . . . . . . . . . . . . . . . . . . . . 80
3.3.5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
3.3.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4 Decoupling with Stability and Decoupling Control Design 87
4.1 Preview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.2 Decoupling Problem with Stability via Transfer Function Matrix
Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.2.2 Minimal C+-Decoupler . . . . . . . . . . . . . . . . . . . . . 89
4.2.3 Decoupling with Stability . . . . . . . . . . . . . . . . . . . 91
4.2.4 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.2.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.3 Decoupling Control Design via LMI Approaches . . . . . . . . . . . 99
4.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.3.2 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . 99
4.3.3 Controller Design via LMI . . . . . . . . . . . . . . . . . . . 101
4.3.4 Stability and Robustness Analysis . . . . . . . . . . . . . . . 103
4.3.5 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
4.3.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
5 Fuzzy Modelling and Control for F-16 Aircraft 111
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.1.1 F-16 Aircraft and Control . . . . . . . . . . . . . . . . . . . 111
5.1.2 Objective of the Design . . . . . . . . . . . . . . . . . . . . . 114
5.1.3 Organization of the Chapter . . . . . . . . . . . . . . . . . . 115
5.2 F-16 Aircraft Model . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.2.1 Need for Modelling . . . . . . . . . . . . . . . . . . . . . . . 115
Contents v
5.2.2 Modelling Method . . . . . . . . . . . . . . . . . . . . . . . 116
5.2.3 Workable Model . . . . . . . . . . . . . . . . . . . . . . . . . 118
5.3 TS Fuzzy Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . 120
5.3.1 The Technique . . . . . . . . . . . . . . . . . . . . . . . . . 120
5.3.2 TS Model of F-16 . . . . . . . . . . . . . . . . . . . . . . . . 123
5.3.3 Model Validation . . . . . . . . . . . . . . . . . . . . . . . . 127
5.4 Lyapunov Based Control . . . . . . . . . . . . . . . . . . . . . . . . 128
5.4.1 Stabilization . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
5.4.2 Tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
5.4.3 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
5.5 Gain Scheduling Control . . . . . . . . . . . . . . . . . . . . . . . . 140
5.5.1 Gain Scheduled Linear Quadratic Regulator Design . . . . . 140
5.5.2 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
5.6 Discussions and Conclusions . . . . . . . . . . . . . . . . . . . . . . 152
5.6.1 Comparative Studies . . . . . . . . . . . . . . . . . . . . . . 152
5.6.2 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
6 Conclusions 160
6.1 Main Findings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
6.2 Suggestions for Further Work . . . . . . . . . . . . . . . . . . . . . 162
Bibliography 165
Appendix A TS Fuzzy Basic Model 177
Appendix B TS Fuzzy Augmented Model I 181
Appendix C TS Fuzzy Augmented Model II 188
Appendix D Gain Scheduling Control for Tracking (φ, θ) 195
Appendix E Gain Scheduling Control for Tracking (α, β, φ) 198
Author’s Publications 201
List of Figures
2.1 Unity feedback system . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2 Step response of proposed method with Am = 3 and φm = 60 . . . . 22
2.3 Step response of proposed method with Am = 2 and φm = 45 . . . . 23
2.4 Step response with Am = 3 and φm = 60 . . . . . . . . . . . . . . . 24
2.5 Step response with Am = 2 and φm = 45 . . . . . . . . . . . . . . . 24
2.6 Controlled uncertain system . . . . . . . . . . . . . . . . . . . . . . 26
2.7 The plot of max|G(jω)| and minargG(jω) . . . . . . . . . . 31
2.8 The plot of max|G(jω)K(jω)| and minargG(jω)K(jω) . . . 37
2.9 Step response of the uncertain system . . . . . . . . . . . . . . . . . 37
3.1 Internal Model Control . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.2 Robust IMC design . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.3 Nominal step response for H2 optimal design . . . . . . . . . . . . . 58
3.4 Nominal step response for robust IMC design . . . . . . . . . . . . 58
3.5 Step response of Robust design for process with mismatch . . . . . 59
3.6 Step response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.7 Step response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.8 Proposed smith predictor control scheme . . . . . . . . . . . . . . . 78
3.9 Step responses for IPTD process . . . . . . . . . . . . . . . . . . . . 84
3.10 Step responses for unstable FOPTD process . . . . . . . . . . . . . 85
3.11 Step responses for unstable SOPTD process . . . . . . . . . . . . . 85
4.1 Step Tests of the Plant in Example 4.4 . . . . . . . . . . . . . . . . 105
4.2 Step Tests of the plant in Example 4.5 . . . . . . . . . . . . . . . . 107
vi
List of Figures vii
4.3 Robust Stability bound in Example 4.5 . . . . . . . . . . . . . . . . 107
4.4 Step response for perturbed process in Example 4.5 . . . . . . . . . 108
5.1 Definition of aircraft axes and angles. . . . . . . . . . . . . . . . . . 117
5.2 Fuzzy triangle membership functions . . . . . . . . . . . . . . . . . 126
5.3 Approximate error of TS-fuzzy and linear model (different α) . . . . 127
5.4 Approximate error of TS-fuzzy and linear model (different φ) . . . . 127
5.5 Lyapunov based stabilizing control . . . . . . . . . . . . . . . . . . 134
5.6 Linear stabilizing control . . . . . . . . . . . . . . . . . . . . . . . . 135
5.7 Lyapunov based stabilizing control with control signal constraints . 138
5.8 Lyapunov based tracking control . . . . . . . . . . . . . . . . . . . . 141
5.9 Linear tracking control . . . . . . . . . . . . . . . . . . . . . . . . . 142
5.10 The gain scheduled tracking control with φ θ . . . . . . . . . . . . . 146
5.11 The gain scheduled tracking control I with α β and φ . . . . . . . . 147
5.12 The gain scheduled tracking control II with α β and φ . . . . . . . 148
5.13 The gain scheduled tracking control III with α β and φ . . . . . . . 149
5.14 Lee’s Backstepping control . . . . . . . . . . . . . . . . . . . . . . . 150
5.15 Outputs of the two proposed methods . . . . . . . . . . . . . . . . . 153
5.16 Control signals of the two proposed methods . . . . . . . . . . . . . 154
5.17 Initial part of gain scheduling control . . . . . . . . . . . . . . . . . 155
5.18 Outputs of the two proposed methods . . . . . . . . . . . . . . . . . 156
5.19 Control signals of the two proposed methods . . . . . . . . . . . . . 157
5.20 Initial part of gain scheduling control . . . . . . . . . . . . . . . . . 158
List of Tables
2.1 SOF and PI controller and their performance . . . . . . . . . . . . . 46
2.2 P matrix in SOF and PI controller design . . . . . . . . . . . . . . . 47
4.1 Search for Gd,min in Example 4.1 . . . . . . . . . . . . . . . . . . . 92
4.2 Search for Gd,min in Example 4.1 . . . . . . . . . . . . . . . . . . . 93
4.3 Search for Gd,min in Example 4.2 . . . . . . . . . . . . . . . . . . . 96
4.4 Search for Gd,min in Example 4.3 . . . . . . . . . . . . . . . . . . . 97
5.1 The φ tracking specifications of the two control methods . . . . . . 152
5.2 The θ tracking specifications of the two control methods . . . . . . 159
viii
Summary
With the development of industrial competition, the performance requirements of
industrial processes become increasingly stringent. Moreover, it was known that
many controllers is sensitive to model uncertainty. To deal with this problem,
the framework for robustness analysis and design was developed in 1980s and
1990s. Recently, many researchers have developed various approaches for robust
control (Goodwin et al., 1999; Wang, 1999; Wang and Goodwin, 2000). Though the
framework for robust control is available, the method for robust design is usually
very complicated and the resultant controllers are generally of high order. The
implementation of such high order controllers in industrial application is usually
difficult. This thesis is devoted to the development of new control design techniques
for better performance or robustness with relatively simple controller or structure.
Proportional-Integral-Derivative (PID) controllers are the dominant choice in
process control and many researches have been reported in literature. In this thesis,
three schemes are developed to design new PID controllers. The first method is
designed for achieving optimal gain and phase margins for uncertain processes.
Gain and phase margins are typical control loop specifications associated with the
frequency response technique. In the proposed method, the objective is to develop
a scheme such that it can achieve desired gain and phase margins for the uncertain
system. The robust PID controller design problem is converted into a standard
convex optimization problem with linear matrix inequalities (LMI) constraints,
which may be solved effectively using the interior point method. A complete PID
tuning guideline is also presented. Simulation shows that the proposed method
gives good performance. The second proposed scheme is based on the extension of
ix
Summary x
the small gain theorem. The well-known small gain theorem was extensively used in
the analysis of the robust stability and performance robustness of uncertain system.
However, the small gain theorem only constraints the gain of the system, while its
phase may be arbitrary. Thus much conservativeness is introduced. In this thesis, a
new quantitative robust stability criterion is presented. In this criterion, both gain
and phase information is employed to reduce the conservativeness. Examples are
given to show the effectiveness of the proposed criterion. Based on the criterion,
a class of second-order plus dead time uncertain process is discussed and a robust
PID tuning scheme is proposed. Examples are provided to illustrate our analysis
and design. For the processes with state time delay, a new approach is proposed
to design PI controller with iterative LMI optimization. It shows that the problem
of PI controller design may be converted into that of static output feedback (SOF)
controller design after appropriate formulation. The difficulty of SOF synthesis is
that the problem inherently is a bilinear problem which is hard to be solved via an
optimization with LMI constraints. In the thesis, an iterative LMI optimization
method is developed to solve the problem.
For the Internal Model Control (IMC) system, two approaches are developed
to achieve good performance while maintaining the robustness. The first design is
in time domain. A new approach to IMC design is proposed, which aims at obtain-
ing optimal H2 performance under the robust stability constraints. Such a robust
optimal IMC design is formulated into a H2/H∞ multiobjective output-feedback
control problem and solved via a system of LMIs in time domain. The validity of
the approach is illustrated by two examples. The second method is based on the
design in frequency domain, an IMC controller design methodology is presented
to achieve the optimal performance with robust stability. The original problem is
nonlinear and thus difficult to solve. The upper bound and lower bound of the
optimal solution are formulated and converted into LMI or BMI optimization. It
is shown that the optimal solution can be approximated by the upper bound and
lower bound with any accuracy. Examples are given to demonstrate the effective-
ness of the proposed method. The advantage of time domain method is that the
Summary xi
optimization problem encountered is an LMI problem, which is easy to be solved.
However, the method cannot be used for processes with time delay and it intro-
duces some conservativeness in the problem formulation. For frequency domain
method, the global optimal solution may be found without conservativeness and it
can be used for processes with time delay. However, the BMI optimization must
be employed to find the solution. One shortcoming of IMC system is that the
presence of time delay forces the designer to choose lower controller gain to main-
tain stability. To our best knowledge, Smith predictor is the best way to control
the processes with time delay. A new modified Smith predictor control scheme
and its simple control design are proposed for unstable processes. The internal
stability of the proposed structure is analyzed. Simulation results show that the
proposed method yields significant performance improvement with load responses
over existing approaches.
In the decoupling design, we wish to find a systematic scheme to satisfy the
requirements of stability, decoupling, performance and robustness. Firstly, a sim-
ple necessary and sufficient condition for solvability of decoupling with internal
stability for unity output feedback for non-singular plants is proposed. Then, a
new method is proposed for the design of multi-variable IMC system aiming at
obtaining good loop performance and small loop couplings based on LMI opti-
mization. The decoupling design with performance constraint is formulated into
an optimization problem with LMI constraints, thus the problem can be solved
effectively using LMI toolbox. Robust stability is analyzed and simulations show
that good control effects can be achieved.
Takagi-Sugeno (TS) fuzzy modelling and control becomes an effective tool for
nonlinear complex processes. In this thesis, a framework for control of F-16 aircraft
with TS fuzzy systems is developed. First, based on the best-available nonlinear
dynamical model of F-16 aircraft in the open domain, the TS fuzzy model of F-16
aircraft is presented and validated with reasonable accuracy. Then, two control
strategies, namely, Lyapunov based control and gain scheduling control, are pro-
posed using the TS model. Each of them is applied to synthesize a F-16 flight
Summary xii
control system for both stabilizing control and attitude tracking control. Exten-
sive simulation is carried out and comprehensive comparative studies are made
with the normal linear control and among two approaches. It shows that the pro-
posed two control designs are feasible and both of them outperform the linear
control design significantly. In particular, the gain scheduling control has achieved
better performance, which is almost equivalent to the best nonlinear control of
high complexity.
The schemes and results presented in this thesis have both practical values and
theoretical contributions. The results of the simulation show that the proposed
methods are helpful in improving the performance or the robustness of industrial
control systems.
Nomenclature
Abbreviations
BMI Bilinear Matrix Inequalities
DPSUF Decoupling Problem with internal Stability
for Unity output Feedback system
FOPDT First Order Plus Dead Time
HJB Hamilton-Jacobin-Bellman
ILMI Iterative Linear Matrix Inequalities
IMC Internal Model Control
ISE Integral Squared Error
ISTE Integral Squared Time Error
LMI Linear Matrix Inequalities
PI Proportional-Integral
PID Proportional-Integral-Derivative
RHP Right Half Plane
SOF Static Output Feedback
SOPDT Second Order Plus Dead Time
SP Smith Predictor
TS Takagi-Sugeno
xiii
nomenclature xiv
Symbols
a(s) Pole Polynomial in Transfer Function
ai Coefficient of Pole Polynomial in Transfer Function
b(s) Zero Polynomial in Transfer Function
bi Coefficient of Zero Polynomial in Transfer Function
H(s) Closed-loop Transfer Function Without Uncertainty
hi(x) Weights in TS model
G(s) Transfer Function of Plant
G(s) Nominal Model of Plant
G−(s) Stable and Minimal Phase Part of G(s)
G−(s) Unstable and Non-minimal Phase Part of G(s)
Gd Decoupler for G
Gd,min minimal C+- decoupler for G
K(s) Feedback Controller
kp P parameter in PID Controller
ki I parameter in PID controller
kd D parameter in PID controller
L(s) Open-loop Transfer Function
L(s) Open-loop Transfer Function for Nominal Model
L Time delay in Transfer Function
la(s) Transfer Function of Additive Uncertainty Bound
lm(s) Transfer Function of Multiplicative Uncertainty Bound
Q(s) IMC Controller
S(s) Sensitivity Transfer Function
T (s) Complementary Sensitivity Transfer Function
Tr Closed-loop Transfer Function for Set-point Tracking
Td Closed-loop Transfer Function for Disturbance Response
Wi(s) or Vi(s) Weighting Function
nomenclature xv
δp(G) McMillan degree of a transfer function G at the RHP pole p
∆(s) Uncertainty Description Associated with Closed-loop System
¯∆(s) System Determinant in Mason’s Formula
ωg Gain Crossover Frequency
ωp Phase Crossover Frequency
Am Gain Margin
φm Phase Margin
Re Real Part
Im Imaginary Part
Notations
‖ · ‖2 L2 norm
‖ · ‖2 ∞ norm
trace(·) Trace of Matrix
arg(·) Phase Angle of Transfer Function
sup(·) Upper Limit
max(·) Maximal Value
min(·) Minimal Value
diag(·) Diagonal Matrix
Chapter 1
Introduction
1.1 Motivation
Today, the automatic controller can be found in many applications in our lives.
It may range from missile tracking in the military area to water and temperature
control in washing machines. Development of analysis and design of controller
has been a goal for control engineering for a long time. The classical frequency
domain methods were developed during the 1930s and 1940s, the renowned classi-
cal stability theory was proposed by Nyquist (1932) and many methods of system
analysis were found by Black (1977) and Bode (1964). In the fifties of last century,
many analytical design methods were developed, which made it possible to design
a controller for a given model to satisfy the transient performance specifications.
With the development of computer technology, controller design based on the opti-
mization method became the main current from the 1960s. Such methods had the
advantage that many different aspects of the design problem are considered. Both
the analytical and the optimization design are based on the exact model of the pro-
cesses. However, controllers are always designed based on incomplete information.
The accuracy of the model varies but it never perfect. Moreover, the behavior of
the plant may change with time. Thus, the controller should be designed on the
basis of mathematical models with consideration for uncertain description of the
process. Since the early of 1980s, the robust control theory has become a major
1
Chapter 1. Introduction 2
area in control research. For dealing with complex nonlinear processes and using
inaccurate information, the fuzzy control theory and neural network methods are
developed rapidly since the 1980s. Throughout the years, control theory has made
important contributions to our world. As the process industries continue to in-
crease, the performance and robustness requirements of control systems become
more important to ensure strong competitiveness. Thus, it is a strong need to look
for new approaches to increase the performance and guarantee the robustness of
the control systems. This thesis motivated to develop new control techniques for
better performance or robustness.
Among most unity feedback control structures, the majority of regulators used
in the industry are of Proportional-Integral-Derivative (PID) type and a large in-
dustrial plant may have hundreds of such regulators (Astrom et al., 1993). They
have to be tuned individually to match process dynamics for acceptable perfor-
mance. The tuning procedure, if done manually, is very tedious and time consum-
ing, especially for those slow dynamics loops, and the resultant system performance
will mainly depend on the experience and the process knowledge the engineers have.
It is recognized that in practice, many industrial control loops are poorly tuned.
Gain and phase margins are typical control loop specifications associated with
the frequency response technology. Many controller designs about tuning gain and
phase margins have been presented. Ogata (1990) gave solutions using a numerical
method and Franklin et al. (1994) solved the problem using a graphical approach.
Using some approximation. Ho et al. (1995) presented an analytical formulae to
design the PID controller for the first-order and second-order plus dead time pro-
cesses to meet gain and phase margin specifications. However, all these methods
did not consider the uncertainty of the processes. Thus, there is a need to design
a new PID tuning scheme to achieve desired gain and phase margin specifications
for a family of uncertain systems.
One of the significant work on robust stability and performance robustness
of uncertain systems was done by Zames (1981), where he derived a well-known
theorem called the small gain theorem. In recent years, robust stability analysis
Chapter 1. Introduction 3
of closed-loop system is heavily based on the small gain theorem. Haddad et al.
(2000) discussed the problem of fixed-structure robust controller synthesis using
the small gain theorem. Jiang and Marcels (1997) presented a recursive robust
control scheme using the nonlinear small gain theorem. The usefulness of the
small gain theorem and its variants in addressing a variety of feedback stabilization
problem was clearly established by Praly (1996). However, since small gain analysis
allows uncertainty with arbitrary phase in the frequency domain, it can be overly
conservative, Thus it is desirable to develop a criterion which can employ both
gain and phase information in order to reduce the conservativeness. Moreover, it
is useful to develop improved PID tuning methods for uncertain processes based
on the less conservative criterion.
For processes with state time delay, Niculescu (1998) proposed an approach to
design H∞ state feedback controller via LMI optimization. Later, Mahmoud and
Zribi (1999) developed a scheme for H∞ static output feedback (SOF) control. In
their method, under the strict-positive realness condition, the SOF control design
problem is simplified into a state feedback problem. Obviously, the method cannot
be used for general processes. The difficulty of SOF synthesis is that the problem
inherently is a bilinear problem which is hard to be formulated into an optimization
problem with LMI constraints. Since PID controller is most popular controller
used, a more effective method to cope with the PI/PID controller for general
processes with state time delay is needed.
It is well known that the PID controller is the most popular controller used
in process control. Although PID controller may achieve good performance for
many benign processes, it will lose its effectiveness in more complex environments.
Due to its simple yet effective framework for system design, Internal model con-
trol (IMC) scheme has been under intensive research and development in the last
decades. The main advantage of the IMC system is that if the model is perfect
the IMC system becomes an open-loop system and the stability analysis becomes
trivial. However, the model can never be perfect. Thus, it is of great practical
importance that the controller perform well when the model differs from the real
Chapter 1. Introduction 4
process. The IMC provides a simple yet effective structure for robust controller
design (Morari and Zafiriou, 1989), thus IMC structure is employed widely in the
robust control system design. Boulet et al. (2002) developed IMC robust tunable
control based on the performance robustness bounds of the system and knowledge
of the plant uncertainty. Chiu et al. (2000) developed IMC for transition control
with uncertain process. Litrico (2002) proposed robust IMC flow control for dam-
river open-channel systems, in which the robustness is estimated with the use of
a bound on multiplicative uncertainty taking into account the model errors, due
to the nonlinear dynamics of the system. However, most of the existing control
design with IMC structure is based on trial and error method. Moreover, the ex-
isting robust controller usually is not optimal for nominal performance. Clearly,
systematic IMC design methods which can achieve optimal nominal performance
and guarantee the robust stability are in demand.
It is well known that a Smith predictor controller, which is an effective dead-
time compensator (Smith, 1959), can be put into an equivalent IMC structure.
However, the original Smith predictor control scheme will be unstable when ap-
plied to an unstable process (Wang et al., 1999b). To overcome this obstacle,
many modifications to the original Smith scheme have been proposed. Astrom
et al. (1994) presented a modified Smith predictor for integrator plus dead-time
process and can achieve faster setpoint response and better load disturbance rejec-
tion with decoupling of the setpoint response from the load disturbance response.
Matausek and Micic (1996) considered the same problem with similar results but
their scheme is easier to tune. Majhi and Atherton (1999) proposed a modified
Smith predictor control scheme which has high performance particularly for unsta-
ble and integrating process. This method achieves optimal integral squared time
error criterion (ISTE) for setpoint response and employs an optimum stability ap-
proach with a proportional controller for an unstable process. Another paper of
Majhi and Atherton (2000) extends their work for better performance and easy
tuning procedure for first order plus dead time (FOPDT) and second order plus
dead time (SOPDT) processes. To our best knowledge, Majhi and Atherton (2000)
Chapter 1. Introduction 5
achieve best performance for setpoint response with unstable dead time processes
employing modified Smith predictor structure. In this thesis, a new scheme is
motivated to improve the performance of disturbance rejection for the unstable
processes.
As for most of the control systems are of multi-variable characteristics, multi-
variable control design and stability analysis is another important topic of interest
and this is addressed in this thesis. Among the multi-variable design, the problem
of decoupling linear time-invariant multi-variable systems received considerable at-
tention in the literature of system theory for a period of over two decades. Much of
this attention was directed toward decoupling by state feedback (Morgan, 1964).
The problem of block decoupling was investigated by Wood (1986). The neces-
sary and sufficient conditions of decoupling were developed by Wang and Davison
(1975). Decoupling through a combination of pre-compensation and output feed-
back was considered by Pernebo (1981) and Eldem (1996). In contrast to the
above, unity output feedback systems are more widely used in industry. But the
problem of decoupling in such a configuration while maintaining internal stability
of the decoupled system appears to be more difficult. The crucial assumption made
by Gundes (1990) is that the plant does not have unstable pole coinciding with
internal stability. Under this assumption, it has been shown that the problem is
solvable. The condition is, however, not necessary, and it can be relaxed.
For multi-variable systems, interactions usually exist between the control loops.
The goal of controller design to achieve satisfactory loop performance has hence
posed a great challenge in the area of control design. Although multi-variable con-
trollers are capable of providing explicit suppression of interactions, their designs
are usually more complex and their implementation inevitably more costly. More-
over, the real processes cannot be exactly decoupled because of the uncertainty
of the processes model used. Thus, exact decoupling is usually impossible in a
practical environment. Therefore, there is a need to propose a novel method for
general multi-variable processes to achieve good loop performance with relatively
small loop coupling. State space H∞ design has been well established since the
Chapter 1. Introduction 6
late 1980’s. Ball and Cohen (1987) and Doyle et al. (1989) are the two notable
ones among numerous relevant references. Extensive lists of references and de-
tailed accounts of various approaches are provided in Stoorvogel (1991) and Zhou
et al. (1996). One possible solution of the robust decoupling problem is to adopt a
loop-wise H∞ approach by designing each controller column such that a combined
loop and decoupling performance index is minimized.
For nonlinear system, the linear model is usually not complex enough to de-
scribe the dynamics. Since Takagi and Sugeno (1985) opened a new direction
of research by introducing the Takagi-Sugeno (TS) fuzzy model, there have been
several studies concerning the systematic design of stabilizing fuzzy controllers
(Tanaka and Sugeno, 1992; Tanaka et al., 1996b; Wang et al., 1996; Thathachar
and Viswanath, 1997). In the TS fuzzy model of Takagi and Sugeno (1985), the
overall system is described by several fuzzy IF-THEN rules, each of which rep-
resents a local linear state equation x = Aix + Biu. To derive the stabilizing
controller, the Lyapunov stabilizing theory and LMIs method may be employed.
In the thesis, the complete attitude motion of a rigid spacecraft (Shuster, 1963)
is considered. The TS fuzzy modelling is one of the ways to find a better model
for a complex process, say F-16 aircraft, when the linear model is not enough
to represent the dynamics of the process. Moreover, both stabilizing and attitude
tracking controllers need to be developed based on TS fuzzy model to achieve good
performance.
1.2 Contributions
In this thesis, new control system design issue along with stability, performance
and robustness are addressed for single variable linear processes, multi-variable
linear processes and nonlinear processes. In particular, the thesis has investigated
the following areas:
A. Robust PID Controller Design for Gain and Phase Margins
A new scheme for optimal PID controller is proposed to meet gain and phase
Chapter 1. Introduction 7
margins for a family of plants. The main contribution is that the uncertainty is
included in the procedure of the optimization. With the new idea, a new method to
design PID controller for uncertain processes is proposed. Using S-procedure and
Schur complement, the PID controller design problem is converted into a standard
convex optimization problem with LMI constraints, which can be solved effectively
using the interior point method. A complete PID tuning method is presented and
simulation examples are provided to show the effectiveness of the proposed method.
B. Quantitative Robust Stability Analysis and Design
A new method quantitative robust stability criterion is proposed and new PID
tuning scheme is developed based on the new criterion. The author begins with a
simple example which shows the conservativeness of the traditional robust stability
theorem, such as small gain theorem. From our analysis, the conservativeness
mainly comes from the unknown sources of uncertainty. In the small gain theorem,
only the gain information of the perturbation is considered, however, the phase
information is discarded. A new quantitative robust stability criterion is in turn
proposed employing both gain and phase information of the uncertain systems. An
example is employed to show that the conservativeness of the new stability criterion
is reduced. Moreover, the gain and phase bounds of the parameter uncertainty for
second order plus dead time models are found and a new PID tuning scheme for
robust performance is also developed.
C. PI Controller Design for State Time-Delay Systems via ILMI Ap-
proach
A new PI controller design method for general processes with state time delay
is proposed. With the augmented state description, we convert the problem of
PI controller design to that of static output feedback controller design. However,
the difficulty is that the problem of static output feedback controller synthesis
inherently is a bilinear problem which is hard to be formulated into an optimization
problem with LMI constraints. In this thesis, An iterative LMI method is proposed
to solved the problem. Both the stabilizing controller and the suboptimal H∞
Chapter 1. Introduction 8
controller are designed for the processes with state time delay. The sufficient
conditions of existing such controllers are presented and the procedures to find
such controllers are also given. A numerical example is provided to show the
effectiveness of the proposed method.
D. Robust IMC Controller Design via Time Domain Approach
A new approach is proposed to design the IMC controller in order to achieve
the optimal nominal performance under the robust stability constraints. The IMC
design problem is converted into a H2/H∞ multi-objective output feedback control
design problem. With some appropriate manipulations of state space equation of
the closed-lop system, the controller may be obtained via solving a system of LMIs.
Two examples are provided to illustrate the effectiveness of the method.
E. Robust IMC Controller Design via Frequency Domain Approach
A new robust IMC design framework is developed which aims to minimize
the integral square error(ISE) for setpoint step input with the robust stability as a
constraint. The above optimization problem is a nonlinear and nonconvex problem
which is difficult to be solved directly using the description in frequency domain.
For the controller design with fixed poles, it is shown that the optimal solution
is approximated by its upper bound and lower bound which are then solved by
LMI optimization. For the controller design with general form, a branch and
bound method is employed to obtain the upper bound and the lower bound of the
optimal solution. It is shown that the optimal solution can be approximated by
the upper bound and the lower bound with any accuracy. Examples are provided
to show the effectiveness of the proposed method.
F. Modified Smith Predictor Control for Disturbance Rejection with
Unstable Processes
In process control, the Smith predictor (SP) is a well known and very effective
dead-time compensator for stable processes. However, the original Smith predictor
control scheme will be unstable when applied to an unstable process. In this the-
sis, a new modified Smith predictor structure is proposed. A simple but effective
Chapter 1. Introduction 9
control procedure is designed to improve the performance especially for the distur-
bance rejection. The internal stability of the proposed structure is also analyzed
which is not reported in the previous publications. Simulation is given to illustrate
that the proposed method achieves good performance for both setpoint response
and disturbance rejection. Comparisons show that the proposed scheme has better
performance than the best existing method, especially for disturbance rejection.
G. Stability of Decoupled Systems
A new necessary and sufficient solvability condition is developed for decou-
pling problem with internal stability of unity output feedback system. Firstly,
the definition of the minimal decoupling matrix, which has the minimal set of all
common unstable poles and zeros of the plants, is given. Then, the existence and
uniqueness of the minimal decoupling is analyzed and the procedure of finding the
matrix is provided. Based on the properties of the special matrix, the necessary
and sufficient condition for decoupling with stability is developed. The condition
shows that decoupling with stability is solvable if and only if there exists a mini-
mal decoupling matrix such that it has no unstable pole-zero cancellation with the
controlled process.
H. Decoupling Control Design via LMI
A new method is proposed for the design of multi-variable IMC system aiming
at obtaining good loop performance and small loop couplings based on LMI opti-
mization. The decoupling design with performance constraints is formulated into
an optimization problem with LMI constraints, thus the problem can be solved
effectively using LMI toolbox. Robust stability is analyzed and simulation shows
that good control effects can be achieved.
I. TS Fuzzy Modelling and Control for F-16 Aircraft
In the thesis, the problem to design both stabilizing and tracking controller for
F-16 aircraft systems has been addressed via the TS fuzzy modelling approach.
Both basic and augmented TS fuzzy models of F-16 have been obtained using the
best-available F-16 nonlinear model from the literature and validated to be reason-
Chapter 1. Introduction 10
ably accurate in the operating range of interests. Two control design methods have
been proposed and shown to be feasible and better than the normal linear control.
The first method, Lyapunov based control, is developed for stabilization of a TS
fuzzy model and extended to handle attitude tracking problem when the model
is augmented with integrator at each output. The most important feature of the
method is guaranteed stability and step-signal tracking of the closed-loop system
if the TS fuzzy model error is small enough. Besides, The solutions are simple to
get since it involves an LMI feasibility problem only. The second method, gain
scheduling control, shows better performance. It has a very simple control struc-
ture and straightforward design from TS model; Yet it can perform as well as the
best nonlinear controller of high complexity and very sophisticated design. How-
ever, the gain scheduling control cannot guarantee the stability of the closed-loop
system.
1.3 Organization of the Thesis
The thesis is organized as follows. Chapter 2 focuses on the development of PID
controller design, where three control strategies are developed. The first work is
to design robust PID controller for gain and phase margins. The second one is to
develop quantitative robust stability criterion and robust PID tuning scheme. The
last one is to design PI controller for state time delay processes via iterative LMI
optimization.
Chapter 3 is devoted to IMC-based feedback system design. A new formulation
of IMC system design in time domain is proposed and the robust IMC controller
may be found via solving an optimization problem with LMI constraints. Another
scheme is developed based on frequency domain approach for the same problem.
A new method is proposed to find the controller via BMI optimization. For the
unstable processes with time delay, a new modified Smith predictor structure is
proposed to achieve better performance for both setpoint response and disturbance
rejection.
Chapter 4 is concerned with the decoupling control. A new necessary and
Chapter 1. Introduction 11
sufficient condition for solvability of decoupling with internal stability for unity
output feedback for non-singular plants is proposed. A new method is proposed for
the design of multi-variable IMC system aiming at obtaining good loop performance
and small loop couplings based on LMI optimization. The robust stability of the
design is also analyzed.
Chapter 5 investigates the TS fuzzy modelling and controller design for F16
aircraft. The problem to design both stabilizing and tracking controller for F-16
aircraft systems is addressed via the TS fuzzy modelling approach. Two control
designs, namely Lyapunov based control and gain scheduling control, are developed
for the control of the obtained TS fuzzy model.
Finally in Chapter 6, general conclusions are given and suggestions for further
works are presented.
Chapter 2
Three New Approaches to PID
Controller Design
2.1 Preview
The vast majority of the controllers used in industry are of the PID type and
PID control has been an important research topic since 1950’s. According to a
survey paper (Yamamoto and Hashimoto, 1991), more than 90% controllers are of
PID type. Although the PID control is well established in process industries, A
recent survey reports (Cominos and Munro, 2002) that there are still many control
loops poorly tuned. Thus, new tuning schemes for PID controllers are desired and
important for better process operations.
In this chapter, three new PI/PID control schemes are developed. In the first
approach, a new PID control design is proposed to meet gain and phase margins
for a family of processes with norm bounded uncertainty. The robust PID con-
troller design is formulated into an optimization problem with LMI constraints and
solved with interior point method. In the second design, the new robust stability
criterion is developed using both gain and phase information. A new PID tuning
scheme is proposed based on the developed criterion for processed with parameter
uncertainty. The third method aims at designing PI controller for processes with
state time delay. The stabilizing and H∞ suboptimal PI controllers are developed
12
Chapter 2. Three New Approaches to PID Controller Design 13
via iterative LMI optimization.
The chapter is organized as follows. In Section 2.2 a new robust PID method for
gain and phase margins is presented. Section 2.3 proposes a new PID design based
on a new stability criterion. Section 2.4 describes an approach to PI controller
design of stabilizing control and H∞ suboptimal control for processes with state
time delay.
2.2 Robust PID Controller Design for Gain and
Phase Margins
2.2.1 Introduction
PID controllers are the dominant choice in process control and an abundant amount
of research have been reported in the past on the PID controller design (Astrom
and Hagglund, 1995). Gain and phase margins are typical control loop specifi-
cations associated with the frequency response technique (Ho et al., 1995) and
many controller designs have been presented to meet the margins. Ogata (1990)
and Franklin et al. (1994) gave solutions using numerical and graphical methods
respectively. Under some approximations, Ho et al. (1995) presented analytical
formulae to design the PID controller for first-order and second-order plus dead
time plant. A general method to achieve exact gain and phase margin for a general
linear plant is given by Wang et al. (1999a). However, all these methods did not
consider uncertainty of the process. Usually we can only obtain an approximate
linear model because of the complex of the real process. Therefore, it is obvious
the uncertainty of the process should be considered when the controller is designed.
In this section, optimal PID controllers are designed to meet gain and phase
margins for a family of plants. The problem is formulated into an optimization
problem with a set of LMI constraints. The resultant convex optimization can be
solved effectively using interior point method. Simulation is given to demonstrate
the proposed method.
Chapter 2. Three New Approaches to PID Controller Design 14
2.2.2 PID Controller Design Using LMI
Suppose an uncertain process G is described as
∣
∣
∣G(jω)− G(jω)∣
∣
∣
2
< ρ(ω) (2.1)
where the nominal model G and the uncertainty bound ρ are known.
K G+
-
r y
Figure 2.1. Unity feedback system
We adopt the unity feedback system showed in Fig. 2.1 and PID controller with
the transfer function:
K(s) = kp +kis+ kds. (2.2)
Assume that the control system specifications are given in terms of gain margin
Am and phase margin φm. Our objective is to determine the controller parameters
kp, ki and kd such that the desired gain margin Am and phase margin φm can be
best achieved for the uncertain process. These specifications mean that
G(jωp)K(jωp) = −1
Am
and
G(jωg)K(jωg) = −ejφm ,
where ωp and ωg are the phase and gain cross over frequencies of the open loop
system respectively. Suppose both ωg and ωp are available, the controller pa-
rameters can be computed if the exact model of the process is available (Ho et
al., 1995; Wang et al., 1999a). However, it will be much more difficult if the un-
certainty of the model is considered. In the sequel, the constraint optimization
Chapter 2. Three New Approaches to PID Controller Design 15
methods are used to find the optimal controller in order to make the whole family
of the models meet the gain and phase margin specifications as close as possible.
Denote the open-loop frequency response for the uncertain system and the nominal
system by L(jw) and L(jw) respectively. Set the β as the tuning parameter, the
problem can be formulated as follows
P1 : minK
(γ1 + βγ2) (2.3)
subject to∣
∣
∣L(jωp)− L(jωp)∣
∣
∣
2
< γ1, (2.4)
∣
∣
∣G(jωp)− G(jωp)
∣
∣
∣
2
< ρ1, (2.5)
∣
∣
∣L(jωg)− L(jωg)∣
∣
∣
2
< γ2, (2.6)
∣
∣
∣G(jωg)− G(jωg)
∣
∣
∣
2
< ρ2. (2.7)
Remark 2.1. The formulation of the problem can be explained as follows: The
objective is to design PID controller to achieve desired gain and phase margins
for the open-loop frequency response L(jω). That means we wish L(jω) satisfies
the gain margin index at phase cross-over frequency ωp and the phase margin
index at gain cross-over frequency ωg respectively. So we formulate (2.4) and (2.6).
Considering the uncertainty of the process, we cannot make the inequalities become
equality. Therefore, we add two inequalities (2.5) and (2.7) as the constraints. In
short, we wish the whole family of the systems meet the desired gain and phase
margins as close as possible when the process varies in the bound of the uncertain
model.
Chapter 2. Three New Approaches to PID Controller Design 16
Note that
∣
∣
∣L(jωp)− L(jωp)
∣
∣
∣
2
− γ1
=∣
∣
∣G(jωp)K(jωp)− L(jωp)
∣
∣
∣
2
− γ1
=
[real(G(jωp)) + j imag(G(jωp))][kp + j(kiωp− kdwp)]− L(jωp)
2
− γ1
= [kpreal(G(jωp)) + imag(G(jωp))(kiωp− kdωp)− real(L(jωp))]
2
+ [(kdwp −kiωp
)real(G(jωp)) + kpimag(G(jωp))− imag(L0(jwp))]2 − γ1
= real2(L(jωp)) + imag2(L(jωp))− γ1 + 2 real(L(jωp))[
−kp kdωp − kiωp
]
real(G(jωp))
imag(G(jωp))
+ 2 imag(L(jωp))[
kiωp− kdωp −kp
]
real(G(jωp))
imag(G(jωp))
+
real(G(jωp))
imag(G(jωp))
T
k2p 2kp(kdωp − kiωp)
0 ( kiωp− kdωp)
2
real(G(jωp))
imag(G(jωp))
+
real(G(jωp))
imag(G(jωp))
T
( kiωp− kdωp)
2 0
2kp(kdωp − kiωp) k2p
real(G(jωp))
imag(G(jωp))
= real2(L(jωp)) + imag2(L(jωp))− γ1 − 2[real(L(jωp)) imag(L(jωp))]
kp ki/ωp − kdωp
−ki/ωp + kdωp kp
real(G(jωp))
imag(G(jωp))
+
real(G(jωp))
imag(G(jωp))
T
kp ki/ωp − kdωp
−ki/ωp + kdωp kp
T
kp ki/ωp − kdωp
−ki/ωp + kdωp kp
real(G(jωp))
imag(G(jωp))
.
Let
Lp =
Re(L(jωp))
Im(L(jωp))
, L0p =
Re(L(jωp))
Im(L(jωp))
,
Gp =
Re(G(jωp))
Im(G(jωp))
, G0p =
Re(G(jωp))
Im(G(jωp))
,
Chapter 2. Three New Approaches to PID Controller Design 17
A(ωp) =
kp ki/ωp − kdωp
−ki/ωp + kdωp kp
, I =
1 0
0 1
.
then equations (2.4) can be rewritten as
1
Gp
T
LT0pL0p − γ1 −LT0pA(ωp)−AT (ωp)L0p AT (wp)A(ωp)
1
Gp
< 0 (2.8)
and from (2.5),we obtain
1
Gp
T
GT0pG0p − ρ1 −GT
0p
−G0p I
1
Gp
< 0. (2.9)
Equation (2.6) and (2.7) can be treated similarly. Therefore, we can convert P1
into the following equivalent problem.
Given the uncertain system (2.1), find the optimal value of the linear weighting
errors of the desired gain and phase margins:
P2 : minK
(γ1 + βγ2)
subject to
1
Gp
T
LT0pL0p − γ1 −LT0pA(ωp)−AT (ωp)L0p AT (ωp)A(ωp)
1
Gp
< 0,
1
Gp
T
GT0pG0p − ρ1 −GT
0p
−G0p I
1
Gp
< 0,
1
Gg
T
LT0gL0g − γ2 −LT0gA(ωg)−AT (ωg)L0g AT (ωg)A(ωg)
1
Gg
< 0,
1
Gg
T
G0gTG0g − ρ2 −GT
0g
−G0g I
1
Gg
T
< 0.
In order to obtain the solution of the problem, we need two lemmas (Boyd et
al., 1994).
Chapter 2. Three New Approaches to PID Controller Design 18
Lemma 2.1. (S-procedure) Let F0, ...,Fp be quadratic functions of the variable
ς ∈ Rm:
Fi(ς) = ςTTiς + 2uTi ς + vi, i = 1, ..., p.
where Ti = T Ti . The following condition on F0, ...,Fp: F0(ς) ≥ 0 for all ς such
that Fi(ς) ≥ 0, i = 0, ..., p. holds if there exist τ1 ≥ 0, ..., τp ≥ 0 such that
T0 u0
uT0 v0
−∑p
i=1 τi
Ti ui
uTi vi
≥ 0.
Lemma 2.2. (Schur complement) If Q(x) = Q(x)T , R(s) = R(s)T and S(x)depend affinely on x, the LMI
Q(x) S(x)S(x)T R(x)
> 0
is equivalent to
R(x) > 0, Q(x)− S(x)R(x)−1S(x)T > 0.
Now, we can develop the solution of the original system.
Theorem 2.1. If the solution of problem P2 exists, the solution is unique and it
is equivalent to the solution of the LMI problem
P3 : minK
(γ1 + βγ2) (2.10)
subject to
−γ1 + t1(ρ1 − GTp0Gp0) −t1GT
p0 LTp0
−t1Gp0 −t1I AT (−ωp)Lp0 A(ωp) −I
< 0, (2.11)
and
−γ2 + t2(ρ2 − GTg0Gg0) −t2GT
g0 LTg0
−t2Gg0 −t2I AT (−ωg)Lg0 A(ωg) −I
< 0. (2.12)
Chapter 2. Three New Approaches to PID Controller Design 19
Proof. Using the S-procedure, (2.8) and (2.9) can be combined into
LTp0Lp0 − γ1 + t1(ρ1 − GTp0Gp0) LTp0A(ωp)− t1G
Tp0
AT (ωp)Lp0 − t1Gp0 AT (ωp)A(ωp)− t1I
< 0, (2.13)
where t1 is a new variable introduced according to S-procedure Lemma. According
to the Schur complement, (2.13) can be converted into the linear matrix inequality
(LMI) form,
−γ1 + t1(ρ1 − GTp0Gp0) −t1GT
p0 Lp0
−t1Gp0 −t1I AT (−ωp)Lp0 A(ωp) −I
< 0. (2.14)
In a similar way, from (2.6) and (2.7), we have
−γ2 + t2(ρ2 −GTg0Gg0) −t2GT
g0 Lg0
−t2Gg0 −t2I AT (−ωg)Lg0 A(ωg) −I
< 0. (2.15)
Thus, P2 is equivalent to P3. Note that P3 is a standard convex optimization
problem with LMI constraint and it has the unique solution.
Remark 2.2. In this controller design method, the optimal controller design prob-
lem was converted into a standard LMI optimization problem and thus the PID
parameters can be found effectively using interior point method (Boyd et al., 1994).
Remark 2.3. If both ρ1 and ρ2 equal zero, the problem reduce to the nominal design
and the solution will be similar to that of Wang et al. (1999a).
Remark 2.4. Besides finding the optimal PID parameter, we can also obtain the
values of γ1 and γ2 which are the measurement of the errors of the gain and phase
margins for the uncertain system respectively. We can choose the value of β to
make the tradeoff between the error of gain margin and that of the phase margin.
2.2.3 Tuning Guidelines
In order to use the above method to design a PID controller, we must specify gain
margin Am, phase margin φm, phase cross-over frequency ωp and gain cross-over
Chapter 2. Three New Approaches to PID Controller Design 20
frequency ωg. Gain and phase margins are typical control loop specifications which
reflect the performance and stability of the system, so we use it as user specified
parameters. Therefore, the remaining work is to choose reasonable phase cross-over
frequency ωp and gain cross-over frequency ωg.
If the bandwidth of a process is the frequency at which the process gain drops
by 3dB below that at the zero frequency, and it is usually approximated by its
phase cross-over frequency, since frequencies below it constitute the most significant
range in controller design. In controller tuning, the closed-loop bandwidth should
be carefully chosen. If it is too large, the control signal will saturate. If it is too
small, sluggish response results. It is well accepted in engineering practice that the
closed-loop bandwidth should be close to its open-loop bandwidth. For example,
we may set
ωp = αωc, α = [0.5, 2], (2.16)
where ωp is the closed-loop bandwidth and ωc is the open-loop bandwidth. The
default value for α is 1. The value of wc is available from the process frequency
response and is the point that satisfies
]G(jωc) = −π. (2.17)
Once ωp is specified, ωg is not free to be chosen. To be this, one obtain from
kpap − (kdωp − ki/ωp)ap = −1/Am,
kpbp + (kdωp − ki/ωp)bp = 0,
kpag − (kdωg − ki/ωg)bg = −cos(φm),
kpbg + (kdωg − ki/ωg)ag = −sin(φm),
where
ap = Re(G(jωp)), bp = Im(G(jωp)),
ag = Re(G(jωg)), bg = Im(G(jωg)).
Chapter 2. Three New Approaches to PID Controller Design 21
Eliminating kp, ki and kd from the above four equations yields the relationship
between ωp and ωg:
apa2p + b2p
=Am(cosφmag + sinφmbg)
a2g + b2g, (2.18)
which enable us to find ωg from ωp. The frequency ωg may be determined by
searching down from the frequency ω = ωp towards ω = 0 until (2.18) holds.
In the summary, given the nominal process G(jω), the uncertain bound and
the specified gain margin Am and phase margin φm, the PID parameters can be
found as follows:
i) Obtain the nominal process phase cross-over frequency ωc from G(jω).
ii) Specify the ratio of close-loop from open-loop bandwidth α and set ωp = αωc.
iii) Search from ω = ωp down towards ω = 0 for the frequency ωg that satisfies
(2.18).
iv) Compute PID parameter as the solution of P3.
2.2.4 Simulation
We shall now look at some examples and demonstrate the use of the method.
Example 2.1. Consider a second-order plus dead time process
G(s) =e−0.5s
(1 + s)(1 + 0.5s)
and the uncertain system is described as
G(s) = G(s) + δ∆G(s),
where
|δ| < 1
and
∆G(s) =0.1
s+ 1.
Chapter 2. Three New Approaches to PID Controller Design 22
The gain margin and phase margin are set to 3 and 60 degrees. Using the proposed
method, the process phase cross-over frequency wc is obtained as ωc = 3.1416 rad/s.
Choose α = 1, then we have ωp is set as ωp = ωc = 3.1416. Simple search from
ω = ωp towards ω = 0 gives the frequency ωg = 1.0472 at which (2.18) holds. With
the LMI toolbox, we obtain the designed PID controller:
K(s) = 1.5694 +0.8906
s+ 0.7811s.
If the gain and phase margins are specified as 2 and 45 degrees respectively, the
proposed method gives
K(s) = 2.3400 +1.5488
s+ 0.7758s.
The step responses are showed in Fig. 2.2 and Fig. 2.3. The 3 lines correspond to
δ = −1, δ = 0 and δ = 1.
0 5 10 15−0.2
0
0.2
0.4
0.6
0.8
1
1.2
a1
a2
a3
Figure 2.2. Step response of proposed method with Am = 3 and φm = 60
(a1: δ=-1; a2:δ=0; a3:δ=1)
Example 2.2. Consider a high order process
G(s) =e−0.1s
(s2 + s+ 1)(s+ 2)2.
In order to compare with Ho’ method (Ho et al., 1995), suppose the process can
be approximate by a second-order plus dead time model (Wang and Zhang, 1998b)
G(s) =0.25e−1.420s
(0.7207s+ 1)2
Chapter 2. Three New Approaches to PID Controller Design 23
0 5 10 15−0.2
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
a1
a2
a3
Figure 2.3. Step response of proposed method with Am = 2 and φm = 45
(a1: δ=-1; a2:δ=0; a3:δ=1)
and the uncertain system is described as
G(s) = G(s) + δ∆G(s),
where
|δ| < 1
and
∆G(s) =0.2
s+ 5.
For Am = 3 and φm = 60, using the proposed scheme, we obtain the controller
K(s) = 1.9576 +1.5730
s+ 1.2420s.
and Ho’s method gives
K(s) = 2.1239 +1.4749
s+ 1.7646s.
For Am = 2 and φm = 45, the proposed scheme gives
K(s) = 3.4725 +2.1380
s+ 1.4735s.
and Ho’s method gives
K(s) = 3.1858 +2.2124
s+ 1.1469s.
The step responses are showed in Fig. 2.4 and Fig. 2.5. The 3 lines correspond to
δ = −1, δ = 0 and δ = 1.
Chapter 2. Three New Approaches to PID Controller Design 24
0 5 10 15 20 25 30 35 40 45 500
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
Time
a3
a2
a1
Figure 2.4. Step response with Am = 3 and φm = 60
(—— proposed method, - - - Ho’s method)
(a1: δ=-1; a2:δ=0; a3:δ=1)
0 5 10 15 20 25 30 350
0.2
0.4
0.6
0.8
1
1.2
Time
a2
a1
a3
Figure 2.5. Step response with Am = 2 and φm = 45
(—— proposed method, - - - Ho’s method)
(a1: δ=-1; a2:δ=0; a3:δ=1)
2.2.5 Conclusions
A simple method for the design of PID controllers for uncertain process which
achieves gain and phase margin specifications is proposed. The scheme is based
on the uncertain model of the plant. The design problem was formulated into
Chapter 2. Three New Approaches to PID Controller Design 25
a convex optimization problem with quadratic constraint, then it was converted
into an optimization problem with linear matrix inequalities constraint which can
be solved effectively using interior point method. The optimal PID controller is
obtained directly to satisfy the gain and phase margins. Simulation results have
been presented to demonstrate the effectiveness of the method.
2.3 Quantitative Robust Stability Analysis and
PID Controller Design
2.3.1 Introduction
Stability criteria and stabilization methods are important research topics in the
system and control area. Many results have been presented in the area of robust
stability. The small gain theorem is a key result of the robust stability analysis
(Zhou et al., 1996). Essentially, the small gain theorem states that if a feedback
loop consists of stable systems and the loop-gain product is less than unity, then
the closed-loop system is stable. The criterion is widely used in the robust control
area. Unfortunately, it could be very conservative in many cases because only
the gain information of the open-loop system is used and the phase information is
ignored.
In this section, the conservativeness of traditional robust stability results is
studied. With the information on both the gain and phase of the uncertain open-
loop system, a new quantitative robust stability criterion is developed. The gain
and phase bounds of uncertain second-order plus time delay (SOPTD) model are
derived. Based on these new results, a new PID tuning scheme is presented using
the worst gain and phase margins of the open-loop system. Simulation is given to
substantiate the effectiveness of the method.
Chapter 2. Three New Approaches to PID Controller Design 26
)(sH r y
+
+
)(s∆
Figure 2.6. Controlled uncertain system
2.3.2 Review of Robust Control Theory
Consider a controlled uncertain system in Fig. 2.6. According to the small gain
theorem (Zhou et al., 1996), if the nominal system H(s) and the perturbation ∆(s)
in Fig. 2.6 are both stable, the feedback system is robustly stable if the inequality
|H(jω)∆(jω)| < 1, ω ∈ [0,∞) (2.19)
holds. Usually, one assumes that ∆(s) is unknown but bounded in magnitude such
that supω
|∆(jω)| = M . Then, the feedback loop given in Fig. 2.6 will be stable
provided that
‖H(s)‖∞ <1
M.
However, in industrial applications, the structure of the uncertainty ∆(s) may be
known. For instance, a large class of the uncertain process can be modeled as a
transfer function with interval parameters. The above criterion for robust stability
may be very conservative when the structure of ∆(s) is specified.
Example 2.3. Consider a second-order nominal plant with transfer function
G(s) =1
s2 + s+ 1
and it is closed in a negative feedback loop with the controller K(s) = 1. Let the
plant uncertainty be structured and given by
G(s) =1 + δ1
s2 + (1 + δ2)s+ (1 + δ3)(2.20)
Chapter 2. Three New Approaches to PID Controller Design 27
where δ1 ∈ [−0.5, 0.5], δ2 ∈ [−0.5, 0.5] and δ3 ∈ [−0.5, 0.5]. One may easily see that
the closed-loop system is robustly stable, because its characteristic polynomial,
p(s) = s2 + (1 + δ2)s+ 2 + δ1 + δ3
always has positive coefficients and thus its roots lie in th open left half of the
complex plane.
In order to apply the robust stability criterion (2.19), we express the uncertainty
as an additive one:
∆(s) =1 + δ1
s2 + (1 + δ2)s+ (1 + δ3)− 1
s2 + s+ 1,
and the corresponding H in Fig. 2.6 is
H(s) =−K(s)
1 +K(s)G0(s)= −s
2 + s+ 1
s2 + s+ 2.
We have
supω
|H(jω)∆(jω)| > limω→0|H(jω)∆(jω)| = 1
2× |1 + δ1
1 + δ3− 1|.
If we choose δ1 = 0.5 and δ3 = −0.5, or the plant is perturbed to
G(s) =1.5
s2 + s+ 0.5,
there results in
supω
|H(jω)∆(jω)| > 1,
violating criterion (2.19) though the system is stable. This shows the conservative-
ness of criterion (2.19).
It should be pointed out that criterion (2.19) makes use of only gain information
while any phase lag is permitted including infinity. It should be also pointed
out that what we have said so far on the structured uncertainty is different from
the existing robust stability criterion for structured perturbations, or the theory
of structured singular values, which have been developed for the multivariable
systems and its essence is the same as criterion (2.19). However, we focus our
study here on systems with single uncertainty and attempt to develop a new robust
stability criterion which employs the phase information of the perturbations to
reduce conservativeness.
Chapter 2. Three New Approaches to PID Controller Design 28
2.3.3 Quantitative Robust Stability
Let us recall the Nyquist criterion (Zhou et al., 1996): if the controller K(s) and
the plant G(s) have no poles in the right half plane, then the closed-loop system
is stable if and only if the Nyquist curve of G(jω)K(jω) does not encircle the −1point. An uncertain process is actually a set of models. Each model gives rise to
a Nyquist curve and the set thus sweeps a band of Nyquist curves. Then from the
Nyquist criterion, the closed-loop system is robustly stable if and only if the band
of Nyquist cluster for the uncertain process and the controller does not encircle the
−1 point. Conversely, the system is not robustly stable if and only if there exits
a Nyquist curve among the band which encircles −1 point, or there must exist a
specific frequency ω0 whose corresponding uncertain region is such that both
maxG|G(jω0)K(jω0)| > 1
and
minGargG(jω0)K(jω0) 6 −π
hold true.
Note that
maxG|G(jω0)K(jω0)| = |K(jω0)|max
G|G(jω0)|
and
minGargG(jω0)K(jω0) = argK(jω0)+min
GargG(jω0),
thus we establish the following quantitative robust stability theorem.
Theorem 2.2. Supposing that the nominal closed-loop system and the controller
are stable, the uncertain closed-loop system is robustly stable if either of the follow-
ing inequalities holds:
|K(jω)| 6 1
maxG|G(jω)| (2.21)
or
argK(jω) > −π −minGargG(jω). (2.22)
Chapter 2. Three New Approaches to PID Controller Design 29
Alternatively, we may transform into the diagram of Fig. 2.6, where H(s) is
the equivalent nominal controlled system, ∆(s) is the uncertainty of the system.
If we modify the inequalities in (2.21) and (2.22) as follows:
|H(jω)| 6 1
max∆|∆(jω)| (2.23)
and
argH(jω) > −π −min∆arg∆(jω), (2.24)
then we have the corollary below.
Corollary 2.3. If the nominal controlled system H(s) and the uncertainty ∆(s)
are all stable, then the uncertain closed-loop system remains stable if either (2.23)
or (2.24) hold for every ω .
Typically, one may use an additive uncertainty model G = G + ∆a or a mul-
tiplicative uncertainty model G = G(1 + ∆m) to describe the uncertain process.
Then corresponding nominal controlled system H is given respectively,
Ha(jω) =−K(jω)
1 +K(jω)G(jω)
or
Hm(jω) =−K(jω)G(jω)
1 +K(jω)G(jω).
For a designed control system, the robust stability criterion in (2.21) and (2.22)
can be tested graphically as follows:
(i) Draw Bode plots for magnitude 1/max|G(jω)| and phase−π−minargG(jω)respectively, according to the given structure of the uncertain plant;
(ii) Draw Bode plots for K(jω) in the same diagram;
(iii) Check if condition (2.21) or (2.22) is satisfied.
Chapter 2. Three New Approaches to PID Controller Design 30
2.3.4 Second-Order Uncertain Model
In most practical situations, it is reasonable to approximate high order processes
by low order plus delay models (Pintelon et al., 1994; Glover, 1989). Reduced-order
models are often required for simplifying the design and implementation of control
systems (Seborg et al., 1989). For this reason, model reduction has attracted much
attention in engineering sciences, especially in model-based control. However, real
processes are usually not of first order or second order, and sometimes may be
perturbed by the different operating conditions or other disturbances. Thus the
model error is inevitable in practical applications if only the nominal model is used.
One way to avoid the problems arising from the modeling error is to use a model
set to express the desired closed-loop specifications. First-order plus dead time
(FOPDT) model is relatively simple but may not achieve the accuracy required in
many cases as they carry only real poles, hence are unable to generate peaks in
the frequency response of oscillatory processes. Therefore, it is desirable to use a
second-order plus dead time (SOPDT) model to describe uncertain systems.
Consider an uncertain SOPDT model
G(s) =b
s2 + a1s+ a2e−Ls
where the uncertain parameters are b ∈ [b− b+], a1 ∈ [a−1 , a+1 ], a2 ∈ [a−2 , a
+2 ] and
L ∈ [L−, L+], b− > 0, a−1 > 0, a−2 > 0 and L− > 0. For a fixed ω, the gain of the
process is
|G(jω)| = b/√
(a1ω)2 + (a2 − ω2)2. (2.25)
For this process, max|G(jω)|, min|G(jω))| can be derived as
max|G(jω)| =
b+√(a−1 ω)
2+(a−2 −ω2)2, ω ∈
[
0,√
a−2
]
;
b+
a−1 ω, ω ∈
[
√
a−2 ,√
a+2
]
;
b+√(a−1 ω)
2+(a+2 −ω2)2, ω ∈
[
√
a+2 ,∞]
;
(2.26)
and
min|G(jω)| =
b−√(a+1 ω)
2+(a+2 −ω2)2, ω ∈
[
0,
√
a−2 +a+2
2
]
;
b−√(a+1 ω)
2+(ω2−a−2 )2, ω ∈
[√
a−2 +a+2
2,∞]
.
(2.27)
Chapter 2. Three New Approaches to PID Controller Design 31
In a similar way, the phase of the second order system can be obtained as
arg(G(jω)) = −arg(a2 − ω2 + ja1ω)− Lω
Consider the uncertainty of the parameters, we obtain
minarg(G(jω)) =
−arg(a−2 − ω2 + ja+1 ω)− L+ω, ω ∈[
0,√
a−2
]
;
−arg(ja−1 ω + a−2 − ω2)− L+ω, ω ∈[
√
a−2 ,∞]
;
(2.28)
and
maxarg(G(jω)) =
−arg(a+2 − ω2 + ja−1 ω)− L−ω, ω ∈[
0,√
a+2
]
;
−arg(ja+1 ω + a+2 − ω2)− L−ω, ω ∈[
√
a+2 ,∞]
.
(2.29)
10−1
100
101
0
1
2
3
4
5
10−1
100
101
−200
−150
−100
−50
0
Gain
Frequency
Phase
Frequency
Figure 2.7. The plot of max|G(jω)| and minargG(jω)
Example 2.4. In Example 2.3, the small gain theory fails to verify the stability
of the uncertain system. We now use the proposed method to examine if the
uncertain system is stable.
Reconsidering (2.20), we obtain
Chapter 2. Three New Approaches to PID Controller Design 32
max|G(jω)| =
1.5√(0.5ω)2+(0.5−ω2)2
, ω ∈[
0,√0.5]
;
1.50.5ω
, ω ∈[√
0.5,√1.5]
;
1.5√(0.5ω)2+(1.5−ω2)2
, ω ∈[√
1.5,∞]
;
and
minarg(G(jω)) =
−arg(0.5− ω2 + j1.5ω), ω ∈[
0,√0.5]
;
−arg(0.5− ω2 + j0.5ω), ω ∈[√
0.5,∞]
.
As in Example 2.3, the unit feedback controller is used, that is K(jω) = 1.
Then we can use Theorem 2.2 to check the stability of the system. Though the
gain condition (2.21) is violated for ω < 1.66 in the gain plot of Fig. 2.7, one sees
that the phase condition (2.22) will be satisfied for all frequencies in the phase plot
of Fig. 2.7. Thus the uncertain system is robustly stable.
2.3.5 Robust PID controller Design
In the context of controller synthesis, we usually get the controller according to
the nominal plant model. If the uncertain range of the plant model is known, then
the above idea of robust stability may be used to design a controller for robust
performance.
Since PID controllers are widely used in the process control industry and easy to
implement (Astrom and Writtenmark, 1984; Zhuang and Atherton, 1993; Wang et
al., 1997), we use the robust quantitative stability criterion developed to tune a PID
controller such that the desired closed-loop specifications are satisfied. Gain and
phase margins are typical control loop specifications associated with the frequency
response technique (Wang et al., 1999a). Here, we shall develop a PID tuning
algorithm which guarantees the required gain and phase margins of the closed-
loop system for the whole family of the uncertain plant. The transfer function of
the PID controller is given by
K(s) = kp +kis+ kds. (2.30)
Chapter 2. Three New Approaches to PID Controller Design 33
According to the preceding discussion, we can build a model Gw representing the
worst case of the uncertain process using (2.26) and (2.28), that is
Gw(jω) = max|G(jω)|ejminarg(G(jω)) (2.31)
Let the desired closed-loop gain margin be Am and the phase margin be φm, then
the following two equations follow from the definition of gain and phase margins:
Gw(jωp)
[
kp + j(kdωp −kiωp
)
]
= − 1
Am
(2.32)
and
Gw(jωg)
[
kp + j(kdωg −kiωg
)
]
= −ejφm , (2.33)
where wp and wg are the phase and gain cross-over frequency of the cascaded
process-plus-controller system respectively. The tuning objective is to find param-
eters Kp, Ki and Kd so that the given gain and phase margins are achieved, i.e.
(2.32) and (2.33) are satisfied. It is noted that there are altogether five unknowns,
namely,Kp, Ki, Kd, ωg and ωp, in (2.32) and (2.33). Since the equations are com-
plex, they can be broken down into four real equations. Because the number of the
unknowns exceeds the number of equations, there is no unique solution to (2.32)
and (2.33) unless one more constraint is added.
It is well accepted in engineering practice that the closed-loop bandwidth should
be close to its open-loop value. Thus the value of ωp can be assigned near the plant
phase-crossover frequency ωp, that is,
ωp = αωp.
The default value for α is 1. The solution to (2.32) and (2.33) is then (Wang et
al., 1999a)
kp = Re
[ −1AmGw(jωp)
]
(2.34)
and
ki
kd
=
− 1ωp
ωp
− 1ωg
ωg
−1
Im[
−1AmGw(jωp)
]
Im[
−exp(φm)Gw(jωg)
]
(2.35)
Chapter 2. Three New Approaches to PID Controller Design 34
where ωg satisfies
Re
[−exp(jφm)Gw(jωg)
]
= Re
[ −1AmGw(jωp)
]
(2.36)
To test the performance robustness of the closed-loop system for the tuned PID
controller, we simply check both gain and phase margins of the worse case. The
following corollary is from the deviation of Theorem 2.2.
Corollary 2.4. The phase margin of the uncertain process is larger than φm if
argK(jω)+minGargG(jω) > −π + φm
when maxG|G(jω)||K(jω)| > 1 and the gain margin of the uncertain process is
larger than Am if
maxG|G(jω)||K(jω)| 6 1
Am
when argK(jω)+minGargG(jω) < −π.
In summary, given an uncertain plant G(s), the PID controller K(s) in form of
(2.30) can be tuned for the worse case to meet the given gain and phase margins
of Am and φm as follows:
i) Obtain the phase-crossover frequency ωp of the worst model of the uncertain
process from the frequency response.
ii) Set the ωp near that frequency.
iii) Compute kp from (2.34).
iv) Find ωg from (2.36) by searching through the frequency range from ωp down-
wards.
v) Compute ki and kd from (2.35).
Remark 2.5. Although both of the robust PID controllers in this section and last
section are designed for gain and phase margins, there are still much difference
between the approaches. Firstly, the objective of the design in this section is
to satisfy the gain and phase margins for the worst cases of the uncertain model.
Chapter 2. Three New Approaches to PID Controller Design 35
While in last section the design purpose is to achieve smallest error with the desired
gain and phase margins for all the uncertain models. In this point of view, the
approach in the second is more conservative. However, it is difficult to compare the
tween approaches with simulation. The reason is that the two approaches cope with
different kinds of uncertain models. In this section design, the uncertain model is
an second order plus dead time model with uncertain parameters. However, the
uncertainty description used in last section is an unstructured uncertainty which
is different with that used in this section.
Example 2.5. Consider a second order plus dead time uncertain process
G(s) =b
s2 + a1s+ a2e−Ls
with the uncertain parameters as b ∈ [0.8, 1.2], a1 ∈ [1.2, 1.8], a2 ∈ [0.8, 1.2] and
L ∈ [4, 6], we obtain from (2.26) ∼ (2.29),
max|G(jω)| =
1.2√(1.2ω)2+(0.8−ω2)2
, ω ∈[
0,√0.8]
1.21.2ω
, ω ∈[√
0.8,√1.2]
1.2√(1.2ω)2+(1.2−ω2)2
, ω ∈[√
1.2,∞]
min|G(jω)| =
0.8√(1.2ω)2+(1.2−ω2)2
, ω ∈[
0,√
0.8+1.22
]
0.8√(1.2ω)2+(ω2−0.8)2
, ω ∈[√
0.8+1.22
,∞]
minarg(G(jω)) =
−arg(0.8− ω2 + j1.8ω)− 6ω, ω ∈[
0,√0.8]
−arg(j1.2ω + 0.8− ω2)− 6ω, ω ∈[√
0.8,∞]
and
maxarg(G(jω)) =
−arg(1.2− ω2 + j1.2ω)− 4ω, ω ∈[
0,√1.2]
−arg(j1.8ω + 1.2− ω2)− 4ω, ω ∈[√
1.2,∞]
The worst case model is then set according to (2.31). The phase cross-over
frequency of Gw(s) is 0.3890rad/s. If we set the gain and phase margins for the
Chapter 2. Three New Approaches to PID Controller Design 36
nominal plant to be 3 and 60 degrees respectively, and choose ωp the same as the
plant phase-crossover frequency as ωp = 0.3890. Then the proportional gain Kp
can be readily calculated using (2.33) to be
kp = Re[−1
3×Gw(j0.3890)] = 0.2220
The frequency rage from blow ω = ωp is then searched and it is found that ωg =
0.1047, which satisfied equation (2.36). The value of Ki and Kd are then calculated
using (2.35) to be
ki
kd
=
− 10.3890
0.3890
− 10.1047
0.1047
−1
Im[
−13×Gw(j0.3890)
]
Im[
−exp(j1.05)Gw(j0.1047)
]
=
0.0710
0.4784
The designed PID controller is hence
K(s) = 0.222 +0.071
s+ 0.4784s
Then we check the robustness of the designed controller. In the gain plot of Fig. 2.8,
we can see |Gw(jω)K(jω)| > 1 when ω < 0.1047. In the phase plot of Fig. 2.8,
one sees that at the same frequency interval ω < 0.1047, argGw(jω)K(jω) >
−π + φm = −120. According to Corollary 2.3, the uncertain system satisfies the
phase margin specification. We can check that the gain margin specification is also
satisfied using a similar method. Thus the tuned controller meets the condition
for robust performance of the closed-loop system. With different values for b, a1,
a2 and L, the step responses of the uncertain system are showed in Fig. 2.9.
2.3.6 Conclusions
In this section, a new criterion for robust stability is developed which makes use
of both the gain and phase information of the open-loop system in order to reduce
the conservativeness of the small gain theory. Based on it, a robust PID design is
presented for SOPDT uncertain model. Simulation shows the effectiveness of the
method.
Chapter 2. Three New Approaches to PID Controller Design 37
10−2
10−1
100
0
2
4
6
8
10
12
10−2
10−1
100
−400
−350
−300
−250
−200
−150
−100
−50
Gain
Frequency
Phase
Frequency
Figure 2.8. The plot of max|G(jω)K(jω)| and minargG(jω)K(jω)
0 10 20 30 40 50 60 70 80 90 1000
0.2
0.4
0.6
0.8
1
1.2
1.4
Figure 2.9. Step response of the uncertain system
2.4 PI Controller Design for State Time-Delay
Systems via ILMI
2.4.1 Introduction
PID controller is the most popular controller in the industrial world and many ap-
proaches have proposed to design the controller, see Astrom and Hagglund (1995)
and Tan et al. (1999). However, most of the existing methods are developed for sin-
gle input and single output (SISO) systems. Recently, Zheng et al. (2002) proposed
Chapter 2. Three New Approaches to PID Controller Design 38
a scheme to design multivariable PID controllers under various performance spec-
ifications and the problem is solved via iterative linear matrix inequalities (ILMI)
approach. However, the method is limited to processes with no time delay.
Niculescu (1998) proposed an approach to design H∞ state feedback controller
for a class of linear systems with delayed state via LMI optimization. Later, Mah-
moud and Zribi (1999) developed a scheme for H∞ static output feedback (SOF)
control. In their method, under the strict-positive realness condition, the SOF
control design problem is simplified into a state feedback problem. Obviously, the
method cannot be used for general processes. The difficulty of SOF synthesis is
that the problem inherently is a bilinear problem which is hard to be formulated
into an optimization problem with LMI constraints.
In this section, we propose a method to design either SOF or PI controller for
general processes with state time delay via iterative LMI approach. Our idea is to
convert the problem of PI controller design to that of static output feedback (SOF)
controller design. In Cao, Lam and Sun (1998), an ILMI algorithm was developed
to solve the problem of SOF stabilization for processes without time delay. Here,
we extend their method to the processes with state time delay.
2.4.2 Problem Description
Consider a class of linear systems with delayed state:
x(t) = Ax(t) + Adx(t− d) +B1w(t) +B2u(t),
ys(t) = Csx(t), yr(t) = Crx(t),(2.37)
where x(t) ∈ Rn is the state, u(t) ∈ Rl1 is the input manipulatable by the controller,
w(t) ∈ Rl2 is the input which stands for some exogenous disturbances and is
not manipulatable by the controller, ys(t) ∈ Rm1 is the measured output and
yr(t) ∈ Rm2 is the controlled output, and A,Ad, B1, B2, CsCr are matrices with
appropriate dimensions. The problem is to find an SOF controller
u = Kys, (2.38)
Chapter 2. Three New Approaches to PID Controller Design 39
where K ∈ Rl1×m1 is the control gain or a PI controller
u = K1ys +K2
∫ t
0
ysdt, (2.39)
where K1 ∈ Rl1×m1 and K2 ∈ Rl1×m1 are matrices to be designed, to satisfy some
given specifications.
2.4.3 Stabilizing Control
In this subsection, sufficient conditions are developed for stabilizing control for the
process (2.37) with SOF. The algorithm is also given to find such an SOF controller
(2.38) to stabilizing the process (2.37).
Consider the process described in (2.37). The objective of SOF stabilization
problem is to find an SOF controller in (2.38) such that the closed-loop system
x(t) = Acx(t) + Adx(t− d) +B1w(t),
yr(t) = Crx(t), Ac = A+B2KCs
(2.40)
is asymptotically stable.
Lemma 2.3. The closed-loop system (2.40) is asymptotically stable with w(t) = 0
and d ≥ 0 if one of the following equivalent conditions is satisfied:
(i) There exist matrices 0 < PT = P ∈ Rn×n, 0 < QT = Q ∈ Rn×n,
satisfying the algebraic Riccati inequality (ARI)
P(A+B2KCs) + (A+B2KCs)TP +Q+ PAdQ−1AT
dP < 0. (2.41)
(ii) There exist matrices 0 < PT = P ∈ Rn×n, 0 < QT = Q ∈ Rn×n,
satisfying the matrix inequality
W =
P(A+B2KCs) + (A+B2KCs)TP PAd
ATdP −Q
< 0. (2.42)
Proof. (i) Similar to the method used in Mahmoud and Zribi (1999), we define the
quadratic Lyapunov function candidate
V (x, t) = xT (t)Px(t) +∫ t
t−d
xT (v)Qx(v)dv,
Chapter 2. Three New Approaches to PID Controller Design 40
where P ,Q > 0. Note that V (x, t) > 0 when x 6= 0 and V (x, t) = 0 when x = 0.
Suppose that the input signal w(t) = 0, we observe that
V (x, t) =xT (t)Px(t) + xT (t)Px(t) + xT (t)Qx(t)− xT (t− d)Qx(t− d)
=xT (t)(A+B2KCs)TPx(t) + xT (t)P(A+B2KCs)x(t) + xT (t− d)AT
dPx(t)
+ xT (t)PAdx(t− d)− xT (t− d)Qx(t− d)
=ZT (t)WZ(t),
where Z(t) = [xT (t) xT (t− d)]T and W is given by (2.42).
(ii) Using Schur complement, one can obtain (2.41) from (2.42) easily.
Lemma 2.4. The following two matrix inequalities are equivalent:
P(A+B2KCs) + (A+B2KCs)TP +Q+ PAdQ−1AT
dP < 0. (2.43)
PA+ ATP +Q−PB2BT2 P + (BT
2 P +KCs)T (BT
2 P +KCs) + PAdQ−1ATdP < 0.
(2.44)
Proof. Sufficiency. Note that
P(A+B2KCs) + (A+B2KCs)TP +Q+ PAdQ−1AT
dP
≤(A+B2KCs)TP + P(A+B2KCs) + CT
s KTKCs +Q+ PAdQ−1AT
dP
=ATP + PA+Q−PB2BT2 P + (BT
2 P +KCs)(BT2 P +KCs) + PAdQ−1AT
dP
<0
Necessity. If there exists P > 0 and Q > 0 satisfies inequality (2.43), it is easy to
find there exists a scalar ρ > 0 such that
(A+B2KCs)TP + P(A+B2KCs) +Q+ PAdQ−1AT
dP +1
ρ2CTKTKC < 0.
i.e.
ATP+PA+Q−ρ2PB2BT2 P+(ρBT
2 P+1
ρKCs)(ρB
T2 P+
1
ρKCs)+PAdQ−1AT
dP < 0.
(2.45)
Obviously, equation (2.45) is equivalent to
ρ2ATP+ρ2PA+ρ2Q−ρ4PB2BT2 P+(ρ2BT
2 P+KCs)(ρ2BT2 P+KCs)+ρ
2PAd(ρ2Q)−1AT
dPρ2 < 0.
By substituting ρ2P with P and ρ2Q with Q, we obtain inequality (2.44).
Chapter 2. Three New Approaches to PID Controller Design 41
Theorem 2.5. The closed-loop system (2.40) is asymptotically stable with w(t) = 0
and d ≥ 0 if the following inequality is satisfied:
PA + ATP +Q−XB2BT2 P − PB2B
T2 X + XB2B
T2 X (BT
2 P + KCs)T PAd
(BT2 P + KCs) −I 0
ATdP 0 −Q
< 0.
(2.46)
Proof. From Lemmas 2.3 and 2.4, we obtain that the closed-loop system (2.40) is
asymptotically stable with w(t) = 0 and d ≥ 0 if PA + ATP + Q − PB2BT2 P +
(BT2 P+KCs)
T (BT2 P+KCs)+PAdQ−1AT
dP < 0. BecauseXTB2BT2 P+PTB2B
T2 X−
XTB2BT2 X ≤ PTB2B
T2 P for any X and P of the same dimension, we get a suffi-
cient condition for the existence of stabilizing SOF gain matrix K given by
PA+ATP+Q−XB2BT2 P−PB2B
T2 X+XB2B
T2 X+(BT
2 P+KCs)T (BT
2 P+KCs)+PAdQ−1ATdP < 0.
Using the Schur complement, the theorem is proved.
Note that inequality (2.46) is not an LMI. Similar to Cao et. (1998), we may
develop an iterative LMI (ILMI) approach to find K, P > 0 Q > 0 to stabilize the
process (2.37) with SOF.
ILMI algorithm for SOF stabilizing control
Given process (2.37) stabilizable with SOF.
Step 1. Select Y > 0 and solve P from the following algebraic Riccati
equation
ATP + PA− PB2BT2 P + Y = 0.
set i = 1 and Xi = P .Step 2. Solve the following optimization problem for Pi,Qi,F and αi.
OP1: Minimize αi subject to the following LMI constraints
Σ (BT2 Pi +KCs)
T PiAd
(BT2 Pi + kCs) −I 0
ATdPi 0 −Qi
< 0, Pi = PTi > 0, (2.47)
where Σ = PiA+ATPi+Qi−XiB2BT2 Pi−PiB2BT
2 Xi+XiB2BT2 Xi−αiPi. Denote
α∗i as the minimized value of αi.
Chapter 2. Three New Approaches to PID Controller Design 42
Step 3. If α∗i < 0, F is the stabilizing SOF control gain, stop. Otherwise go
to Step 4.
Step 4. Solving the following optimization problem for Pi and K with the
fixed Qi and αi obtained from Step 2.
OP2: Minimize trace(Pi) subject to the LMI constraints (2.47). Denote P∗i as
the optimal Pi.Step 5. If ‖XiB2−P∗
i B2‖ < ε, where ε is a prescribed tolerance, go to Step
6; Otherwise set i = i+ 1, Xi = P∗i , and go to Step 2.
Step 6. It cannot be decided by this algorithm whether the question is
solvable. Stop.
2.4.4 Suboptimal H∞ Control
In this subsection, sufficient conditions are developed for suboptimal H∞ control
for the process (2.37) with SOF. That is, the SOF control gain K is design to
simultaneously stabilize (2.37) and guarantee the H∞ norm bound γ of the closed-
loop transfer function Tyrw. An ILMI algorithm is also given to find such a SOF
controller.
Lemma 2.5. The closed-loop system (2.40) is asymptotically stable and ‖Tyrw‖∞ ≤γ, γ > 0 if there exists matrix P = PT > 0, Q = QT > 0 and K such that the
following condition is satisfied:
P(A+B2kCs) + (A+B2KCs)TP +Q+PAdQ−1AT
dP +CTr Cr + γ2PB1BT
1 P < 0.
(2.48)
Proof. Let us construct the matrices U(s) = [−sI − (A+B2KCs)−Ade−ds]−1 and
V (s) = Q+PAdQ−1ATP−edsATdP−e−dsPAd, one has Tyrw(s) = CrU(s)B1. Intro-
ducing the matrixM = −[P(A+B2KCs)+(A+B2KCs)TP+Q+PAdQ−1AT
dP+
CTr Cr + γ2PB1BT
1 P ], one obtains
PU−1(s) + [UT (−s)]−1P − V (s)− CTr Cr − γ2PB1BT
1 P −M = 0.
Chapter 2. Three New Approaches to PID Controller Design 43
Pre-multiplying by UT (−s) and post-multiplying by U(s), gives
PU(s) + UT (−s)P − γ2UT (−s)PB1BT1 PU(s) = UT (−s)[V (s) + CT
r Cr +M ]U(s)
(2.49)
Further manipulating (2.49), one gets
BT1 PU(s)B1 +BT
1 UT (−s)PB1 − γ2B1rU
T (−s)PB1BT1 PU(s)B1 − γ2I
=− γ2I +BT1 U
T (−s)[V + CTr Cr +M ]U(s)B1.
It follows that
− [γI − γ−1BT1 PU(−s)B1]T [γI − γ−1BT
1 PU(s)B1]
=[BT1 U
T (−s)CTr CrU(s)B1]− γ2I +BT
1 UT (−s)[V +M ]U(s)B1.
Noting that −[γI − γ−1BT1 PU(−s)B1]T [γI − γ−1BT
1 PU(s)B1] ≤ 0, we obtain
T Tyrw
(−s)Tyrw(s)− γ2I +BT1 U
T (−s)[V +M ]U(s)B1 ≤ 0.
According to Mahmoud and Zribi (1999), we get V ≥ 0, M ≥ 0. Thus we conclude
that ‖Tyrw‖∞ < γ.
Theorem 2.6. The closed-loop system (2.40) is asymptotically stable and ‖Tyrw‖∞ ≤γ, γ > 0 if there exist matrices P = PT > 0, Q = QT > 0 and K such that the
following matrix inequality is satisfied:
Σ (BT2 P +KCs)
T PAd CTr PB1
BT2 P +KCs −I 0 0 0
ATdP 0 −Q 0 0
Cr 0 0 −I 0
BT1 P 0 0 0 −γ2I
< 0, (2.50)
where Σ = PA+ ATP +Q−XB2BT2 P − PB2BT
2 X +XB2BT2 X.
Proof. By Lemmas 2.4 and 2.5 and note the fact that XTB2BT2 P + PB2BT
2 X −XB2B
T2 X ≤ PTB2B
T2 P , the theorem follows.
Chapter 2. Three New Approaches to PID Controller Design 44
ILMI algorithm for SOF suboptimal H∞ control
Given process (2.37) stabilizable with SOF and γ > 0.
Step 1. Select Y > 0 and solve P from the following algebraic Riccati
equation
ATP + PA− PB2BT2 P + Y = 0.
set i = 1 and Xi = P .Step 2. Solve the following optimization problem for Pi,Qi,K and αi.
OP1: Minimize αi subject to the following LMI constraints
Σ (BT2 Pi +KCs)
T PiAd CTr PiB1
BT2 Pi +KCs −I 0 0 0
ATdPi 0 −Qi 0 0
Cr 0 0 −I 0
BT1 Pi 0 0 0 −γ2I
< 0, Pi = PTi > 0,
(2.51)
where Σ = PiA+ATPi+Qi−XiB2BT2 Pi−PiB2BT
2 Xi+XiB2BT2 Xi−αiPi. Denote
α∗i as the minimized value of αi.
Step 3. If α∗i < 0, K is the stabilizing SOF control gain, stop. Otherwise
go to Step 4.
Step 4. Solving the following optimization problem for Pi and K with the
fixed Qi and αi obtained from Step 2.
OP2: Minimize trace(Pi) subject to the LMI constraints (2.51). Denote P∗i the
optimal Pi.Step 5. If ‖XiB2−P∗
i B2‖ < ε, where ε is a prescribed tolerance, go to Step
6; Otherwise set i = i+ 1, Xi = P∗i , and go to Step 2.
Step 6. It cannot be decided by this algorithm whether the question is
solvable. Stop.
2.4.5 Control Design with PI Controllers
In this subsection, we investigate the design problem of PI controllers based on the
results in the preceding two subsections.
Chapter 2. Three New Approaches to PID Controller Design 45
Consider the process in equation (2.37) and PI controller in equation (2.39),
let z1 = x, z2 =∫ t
0ysdt. Denote z = [zT1 , z
T2 ]
T , the dynamic of augmented process
model is described by
z1 =x = Ax(t) + Adx(t− d) +B2u(t) +B1w(t),
z2 =Csx(t).
The output of the augmented system is written as
yr(t) = [Cr 0]
z1(t)
z2(t)
, ys(t) =
Cs 0
0 I
z1(t)
z2(t)
.
and the PI controller becomes
u(t) = [K1 K2]ys(t).
The close-loop system with PI controller is described as
z(t) =Az(t) + Adz(t− d) + B2u(t) + B1w(t),
yr(t) =B1z(t), ys(t) = Csz(t),
u(t) =Kys(t),
where A =
A 0
C 0
, Ad =
Ad
0
, B2 =
B2
0
, Cr = [Cr 0], Cs =
Cs 0
0 I
and
K = [K1 K2]. Thus, the problem of PID controller design reduces to that of SOF
controller design problem.
2.4.6 A Numerical Example
Consider a system with state time delay is described by
x(t) =Ax(t) + Adx(t− d) +B2u(t) +B1w(t),
yr(t) =Crx(t), ys(t) = Csz(t),
where A =
−3 2 0
1 −2 −13 0 6
, Ad =
1 0 1
1 −1 0
0 0 2
, B2 =
1 1
4 0
2 0
, B1 =
0.1
0.2
0.6
, Cs =
0 1 0
1 0 1
, Cr =[
1 0 1]
and d = 0.1. Using the proposed algorithm, we can
Chapter 2. Three New Approaches to PID Controller Design 46
obtain SOF and PI controllers corresponding to different performance specifications
accordingly. The results are summarized in Table 2.1.
Remark 2.6. The disturbance signal w(t) is assumed to be zero for stability proof.
However, for the linear time-delay systems discussing in this section, the distur-
bance signal w(t) will not change the stability of closed-loop system if we may
choose stabilizing control. Moreover, the input signal will also not change the
Hinfty of transfer function between the disturbance input and the output. Thus,
for any inputs, our design may satisfy the objective if we may find such a controller.
Table 2.1. SOF and PI controller and their performance
Problem Feedback matrices Performance required Actual Performance Iteration number α
SOF (stab) K =
22.0090 −65.2154
−0.7559 −2.1137
Stability Stable 3 -0.3170
PI (stab) K1 =
9.6587 −30.8859
37.5367 −28.5601
Stability Stable 3 -0.0004
K2 =
2.5154 −9.8625
14.2639 −10.8638
SOF (H∞) K =
16.8052 −51.2591
−0.3142 −3.4181
H∞ < 2 H∞ = 0.1140 3 -0.0481
PI (H∞) K1 =
16.8748 −56.2466
6.9850 −25.4804
H∞ < 2 H∞ = 0.2484 9 -0.0001
K2 =
65.9504 −49.3472
32.4589 −24.2961
2.4.7 Conclusions
In this subsection, algorithms based on the ILMI technique have been developed
to design SOF and PI controller for linear time invariant systems with state time
delay, which guarantee the stability of closed-loop system or H∞ performance spec-
ification. An example is provided to illustrate the effectiveness of the proposed
Chapter 2. Three New Approaches to PID Controller Design 47
Table 2.2. P matrix in SOF and PI controller design
Problem P
SOF (stab)
136.2660 −45.0290 126.2902
−45.0290 15.6715 −42.5981
126.2902 −42.5981 127.7296
PI (stab)
62.8501 −19.9052 53.9336 −82.1373 62.5272
−19.9052 6.9235 −17.4935 26.4612 −20.1282
53.9336 −17.4935 55.0967 −70.6262 53.7021
−82.1373 26.4612 −70.6262 115.4461 −87.8724
62.5272 −20.1282 53.7021 −87.8724 66.9912
SOF (H∞)
97.6294 −31.2270 87.7795
−31.2270 10.7198 −28.8752
87.7795 −28.8752 89.2653
PI (H∞)
93.0279 −28.4674 83.6846 −112.2981 84.0530
−28.4674 9.3379 −25.7822 34.7012 −25.9644
83.6846 −25.7822 85.7134 −102.4361 76.6289
−112.2981 34.7012 −102.4361 142.8461 −106.9121
84.0530 −25.9644 76.6289 −106.9121 80.0845
method.
Chapter 3
Advance in Robust IMC Design
for Step Input and Smith
Controller Design for Unstable
Processes
sectionPreview Most of control designs are based on the model of process. How-
ever, the model is never accurate. Thus, it is of great practical importance that the
controller performs well when the model differs from the real process. The Internal
model control (IMC) provides a simple yet effective structure for robust controller
design (Morari and Zafiriou, 1989), thus IMC structure is employed widely in the
robust control system design. Boulet et al. (2002) developed IMC robust tunable
control based on the performance robustness bounds of the system and knowledge
of the plant uncertainty’s weighting function. Chiu et al. (2000) developed IMC
for transition control with uncertain process. Litrico (2002) proposed robust IMC
flow control for dam-river open-channel systems, in which the robustness is esti-
mated with the use of a bound on multiplicative uncertainty taking into account
the model errors, due to the nonlinear dynamics of the system. However, most of
the existing control design with IMC structure is based on trial and error method.
Moreover, the existing robust controller usually is not optimal for nominal perfor-
48
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 49
mance. It is well known that a Smith predictor can be put into an equivalent IMC
structure. However, the performance of disturbance rejection for existing methods
with unstable processes is not satisfactory.
In this chapter, two new robust IMC control schemes, namely time domain
approach and frequency domain approach, are developed respectively. In the time
domain approach, the IMC design is established through state space description
and LMI optimization. The IMC control design is formulated into an output feed-
back design problem. After appropriate transformation, it is shown that the output
feedback design problem can be solved via available method. In the frequency do-
main method, it is shown that the upper bound and lower bound of the solution
may be found via LMI optimization for the controller with fixed poles and the
upper bound and lower bound of the solution may be found via bilinear matrix
inequalities (BMI) optimization for the controller with general form. Moreover,
the optimal solution can be approximated by the upper bound and lower bound
with arbitrary accuracy. For the modified Smith predictor design, a new struc-
ture is proposed to improve the performance of disturbance rejection for unstable
processes. There are altogether four controllers in the proposed modified Smith
structure, which are designed for disturbance rejection, stability of the time delay
loop, closed-loop dominant pole and setpoint performance.
The chapter is organized as follows. In Section 3.1 a new robust IMC design in
time domain is presented. Section 3.2 proposes an IMC design in frequency domain.
Section 3.3 describes a modified Smith predictor design for unstable processes.
3.1 Robust IMC Design via Time Domain Ap-
proach
3.1.1 Introduction
The IMC is a well-established control scheme in industry control, which is discussed
in the textbook of Morari and Zafiriou (1989) in details. The IMC structure pro-
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 50
vides an effective approach in analyzing control system performance and robustness
(Zhu et al., 1995). Moreover, unity feedback control, linear quadratic optimal con-
troller and Smith predictor can be equivalently transformed into the general IMC
form (Garcia and Morari, 1982). Thus, the IMC structure is an important tool
for robust controller design. When the model is perfect, the IMC system becomes
an open-loop system and the problem of controller design and stability analysis
becomes transparent. However, the model usually is not perfect. The model error
gives rise to feedback in the true sense and leads to possible stability problem. It
gives rise to the problem of robust stability (Morari and Zafiriou, 1989). In the
existing IMC design, the procedure usually consists of two step. The first step is
to yield good nominal performance. The second step is to achieve the desired ro-
bustness by detuning the low-pass filter with a trial and error method. It is noted
that with the filter detuning, the performance is no more optimal in any sense.
In this section, we propose a new approach to design the IMC controller in order
to achieve the optimal nominal performance under the robust stability constraints.
The IMC design problem is converted into a H2/H∞ multiobjective output feed-
back control design problem. The controller is obtained via solving a system of
LMIs. Two examples are provided to illustrate the effectiveness of the method.
3.1.2 The Robust IMC Design
y
r +
Q G
G ˆ
u +
+
Figure 3.1. Internal Model Control
Consider an IMC system shown in Figure. 3.1, where G is a stable process to
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 51
be controlled, G is the model of the process and Q is the controller to be designed.
Without considering the model uncertainty, the H2 optimal control design is to
determine Q such that the Integral-Square-Error (ISE)
‖e(t)‖22 =∫ ∞
0
e2(t)dt.
is minimized for a particular input r. Using Parseval’s theorem, one obtains the
description of ISE in frequency domain
‖e(t)‖22 = ‖E(jω)‖22 =1
2π
∫ ∞
−∞
|E(jω)|2dω
Note that E(jω) = [1−G(jω)Q(jω)]R(jω), one gets
‖e(t)‖22 = ‖[1−G(jω)Q(jω)]R(jω)‖22
For any practical process, the model is never perfect. In this case, the H2 optimal
control is usually not a desired scheme because it cannot guarantee the stability
of the uncertain system. Thus, we aim to develop a control scheme to develop an
approach to achieve optimal ISE for nominal model and guarantee the stability
of the whole closed-loop system. Suppose that the set of the uncertain model is
described by
Π = G : |G(iω)− G(iω)
G(iω)| < lm(ω). (3.1)
In order to achieve robust stability of the closed-loop system in Fig. 3.1, the fol-
lowing condition should be satisfied (Morari and Zafiriou, 1989):
‖GQlm‖∞ < 1 (3.2)
Thus, the robust IMC design problem becomes
minQ‖[1−G(jω)Q(jω)]R(jω)‖22 (3.3)
subject to (3.2).
Here, for the first time, we proposed a new scheme to convert the robust IMC design
problem (3.3) into a multi-objective optimization problem with dynamic output
feedback. There are methods available to solve the multi-objective optimization
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 52
y
r
Gˆ
u
K
+
s
11
2Z
ml
infZ
e
v
Figure 3.2. Robust IMC design
problem with dynamic output feedback via LMI optimization. Suppose the input
of interest r is a unity step and the bound of uncertainty model can be described
by transfer function lm(s), the new system established for robust IMC design is
depicted in Fig. 3.2. We observe that our robust IMC design problem is to find a
dynamic output feedback controller Q in order to minimize the H2 norm between
r and Z2 on the condition that the H∞ norm between r and Zinf is less than 1.
Now let us analysis the proposed new system in Fig. 3.2. Suppose that the state
space description of process model G is
xG = AGxG +BGuG
yG = CGxG +DGuG, (3.4)
and the state space model of lm
xl = Alxl +Blul
yl = Clxl +Dlul. (3.5)
The block 1sin Fig. 3.2 represents the unity step input. However, it makes the
constructed new system unstable, thus it is difficult to compute the controller using
available multiobjective optimization method described in next subsection. Here,
we introduce a small positive number ε and let 1sbecome 1
s+ε. For the element 1
s+ε,
we get the state space description as
xr = εxr + ur
yr = xr. (3.6)
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 53
By observing Fig. 3.2, we obtain the following relationship that uG = u, ul = yG,
ur = e, yr = z2, yl = zinf and r = y. Substitute these relationships into (3.4), (3.5)
and (3.6), we get the state equation of the open-loop system without controller Q
in Fig. 3.2 as
xG
xl
xr
=
AG 0 0
BlCG Al 0
−CG 0 ε
xG
xl
xr
+
0 BG
0 BlDG
1 −DG
r
u
, (3.7)
and the output equation as
Z2
Zinf
y
=
0 0 1
DlCG Cl 0
0 0 0
xG
xl
xr
+
0 0
0 DlDG
1 0
r
u
. (3.8)
Let T2 denote the closed-loop transfer function from r to Z2 and Tinf denote the
closed-loop transfer function from r to Zinf . Our goal is to compute a dynamical
output feedback controller
xQ = AQxQ +BQy
u = CQxQ +DQy
such that min ‖T2‖2 subject to ‖Tinf‖∞ < 1. Here we note that the original robust
IMC design problem is converted into a dynamic output feedback control problem.
Remark 3.1. Usually we may only get the amplitude of the uncertain bound at
our interested frequency points. However, we have to obtain the transfer function
description of uncertain bound in order to use our proposed method to solve the
problem. Fortunately, for our propose approach, we just need the amplitude of
the transfer function is same as that of the uncertain bound at each frequency
point. The reason is that the phase of the uncertain bound is irrelevant with
robust stability of the IMC system. In our design, we may only minimal phase
system to describe the uncertain bound. For minimal phase linear system, if we
have the amplitude of frequency response, we may determine the phase at whole
frequency band.
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 54
3.1.3 LMI Solution
It is shown (Scherer et al., 1997) that the dynamic output feedback control problem
to meet some specifications can be solved via LMI optimization. In this section,
the specifications under consideration include H∞ and H2 performance. The state
space equation of closed-loop system T is given by
x = Acx+Bcr
z = Ccx+Dcr.
According to Bounded Real Lemma (Boyd et al., 1994), Ac is stable and the H∞
norm of T is smaller than γ if and only if there exists a symmetric P with
ATc P + PAc PBc CT
c
BTc P −γI DT
c
Cc Dc −γI
< 0, P > 0.
Regarding H2 norm, the following result is obtained from Scherer et al. (1997). Ac
is stable and ‖T‖22 < ν if and only if there exist symmetric matrix P and Q such
that
ATc P + PAc PBc
BTc P −I
< 0,
P CTc
Cc Q
> 0,
T race(Q) < ν, Dc = 0,
where Tr(·) is the trace of matrix. In practice, the closed-loop system is usually not
much faster than the dynamic of the process (Morari and Zafiriou, 1989). Thus, we
do not wish the poles of closed-loop system are too fast. So we may limit the closed-
loop eigenvalues in some specified LMI regions. The problem around closed-loop
pole placement in LMI regions is studied by Chilali and Gahinet (1996). The matrix
Ac with all its eigenvalues in the LMI region z ∈ C : [lij +mijz +mjiz]i,j < 0 if
and only if there exists a symmetric P such that
[lijP +mijATc P +mjiPAc]i,j < 0.
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 55
In our design, let the Lyapunov matrix P in H∞ norm constraint denote as P1, letthe P in H2 norm optimization as P2 and the P in pole constraints as P3. Accord-ing to the result of Scherer et al. (1997), we must require that all specifications
are forced by a single closed-loop Lyapunov function to satisfy the convexity of
multiobjective optimization problem. Thus, we need to set P1 = P2 = P3 = P in
our computation. Consider (3.7) and (3.8) and denote
A =
AG 0 0
BlCG Al 0
−CG 0 ε
, Bin =
0
0
1
, Bfb =
BG
BlDG
−DG
C2 =[
0 0 1]
, Cinf =[
DlCG Cl 0]
,
Dinf = DlDG,
we obtain the realization of T2 and Tinf as
A =
A BfbCK
0 AK
,B2 =
Bin +BfbDK
BK
C2 =[
C2 0]
,D2 = 0,
and
A =
A BfbCK
0 AK
,B =
Bin +BfbDK
BK
Cinf =[
Cinf 0]
,Dinf = DinfDK ,
respectively. Following the approach in Scherer et al. (1997), partition P and P−1
as
P =
Y N
NT ?
,P−1 =
X M
MT ?
,
and let
A = NAKMT + Y BfbCKM
T + Y AX,
B = NBK + Y BfbDK ,
C = CKMT ,
D = Dk,
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 56
the original robust IMC design problem (3.3) becomes the optimization problem
with LMI constraints described by
minQ
ν,
subject to
AX +XAT +Bbf C + (Bbf C)T ? ? ?
A+ AT ATY + Y A ? ?
(Bin +BfbD)T (Y Bin + B)T −γI ?
CinfX +Dinf C Cinf DinfD −γI
< 0, (3.9)
AX +XAT +Bbf C + (Bbf C)T ? ?
A+ AT ATY + Y A ?
(Bin +BfbD)T (Y Bin + B)T −I
< 0,
X ? ?
I Y ?
C2X C2 Q
> 0, T race(Q) < ν, (3.10)
ljk
X I
I Y
+mjk
AX +BfbC A
A Y A
T
+
mkj
AX +BfbC A
A Y A
< 0, (3.11)
where ? replaces blocks that are readily inferred by symmetry.
3.1.4 Simulation
In this subsection, two examples are given to show the validity of the proposed
approach.
Example 3.1. Consider a multi-lag non-minimum phase process with its nominal
model (Wang and zhang, 1998a)
G(s) =1− 0.3s
(s+ 1)(0.8s+ 1)(0.4s+ 1),
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 57
and the uncertain bound is given by
lm(s) =2s
s+ 1.
ε in (3.6) is chosen as 0.00001. The fastest open loop pole is 2.5. we limit the
closed-loop poles less than 10 times The fastest open loop pole. In the simulation,
we set the feasible closed-loop pole area as the disk with center at original point
and radius as 25.
We are interested in the H∞ performance from r to zinf and H2 performance
from r to z2. The optimal H2 performance from r to z2 is 0.8191 and the controller
is
Q(s) =100s5 + 8175s4 + 44880s3 + 89420s2 + 76270s + 23660
s5 + 34.03s4 + 748.1s3 + 10290s2 + 33230s + 23660, (3.12)
the step response and control signal for nominal process are given in Fig. 3.3.
When using the controller K(s) in (3.12) for the closed-loop system, the H∞ per-
formance from r to z∞ is 2.5510. It does not satisfy the robust stability condition.
If the real process is G(s) = G(s)+G(s)lm(s), the controller in (3.12) really cannot
stabilize the IMC system.
Now consider the controller design with robust stability constraints. The re-
sulting controller with proposed method is
Q(s) =100s5 + 5917s4 + 31880s3 + 63090s2 + 53680s + 16750
s5 + 45.11s4 + 1085s3 + 14220s2 + 40810s + 16750. (3.13)
The obtained H2 norm is 0.9715 and the step response for the nominal process is
shown in Fig. 3.4. The response is more sluggish than the response in Fig. 3.3.
However, the H∞ norm is 0.9764, which is really less than 1. The IMC system
should be stable if the model uncertainty is in the given bound. The step response
with the process G(s) = G(s) + G(s)lm(s) is given in Fig. 3.5. It is shown that
the controller in (3.13) really stabilize such a process.
Example 3.2. Consider a oscillatory process with dead time
G(s) =1
(s2 + 2s+ 3)(s+ 3)e−0.3s.
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 58
0 5 10 15−1
−0.5
0
0.5
1
1.5
time
pla
nt
ou
tpu
t
0 5 10 15−150
−100
−50
0
50
100
150
200
time
co
ntr
ol a
ctio
n
Figure 3.3. Nominal step response for H2 optimal design
0 5 10 15−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
time
pla
nt
ou
tpu
t
0 5 10 15−100
−50
0
50
100
150
time
co
ntr
ol a
ctio
n
Figure 3.4. Nominal step response for robust IMC design
Using first order Pade approximation, we can use the proposed method to design
the controller. Suppose the uncertain bound is also
lm(s) =2s
s+ 1,
and ε in (3.6) is chosen as 0.00001. The optimal H2 norm is 0.6621, and the
controller is
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 59
0 5 10 15−2
−1
0
1
2
3
time
pla
nt
ou
tpu
t
0 5 10 15−100
−50
0
50
100
150
time
co
ntr
ol a
ctio
n
Figure 3.5. Step response of Robust design for process with mismatch
Q(s) =99.97s6 + 300700s5 + 3798000s4 + 16180000s3
s6 + 53.49s5 + 1844s4 + 41160s3
+33350000s2 + 38580000s + 17920000
+541500s2 + 2492000s + 1991000.
The the H∞ performance from r to z∞ is 2.5816. For robust design, the controller
is
Q(s) =99.97s6 + 150400s5 + 1894000s4 + 8061000s3
s6 + 61.8s5 + 2154s4 + 45940s3
+16610000s2 + 19220000s + 8947000
+552900s2 + 2306000s + 994100.
The relevant H2 norm and H∞ norm are 0.8238 and 0.9563 respectively.
3.1.5 Conclusions
A new approach for robust IMC controller design is proposed. It is shown the
IMC optimal H2 performance design with robust stability constraints is converted
into dynamic output feedback control design problem. It is shown that the robust
IMC controller design is solvable via solving a system of LMI. Two examples are
presented with the demonstrated effectiveness of the proposed method.
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 60
3.2 Robust IMC Controller Design via Frequency
Domain Approach
3.2.1 Introduction
In last section, a robust IMC design framework is developed which aims to mini-
mize the integral square error (ISE) for setpoint step input with the robust stability
as a constraint. Using a time domain approach, we have converted the IMC design
problem into output feedback design method and solved using existing LMI opti-
mization. However, the above method cannot cope with the processes with time
delay. Moreover, the obtained control is usually of high order. Thus, we attempt
to solve the IMC design problem in frequency domain.
When the IMC design problem is formulated in frequency domain, it becomes
a nonlinear and nonconvex optimization problem which is difficult to be solved
directly. In this section, we discuss two classes of controllers with fixed pole and
with general form. For the controller design with fixed poles, the optimal solution
is approximated by its upper bound and lower bound which are then solved by
LMI optimization. For the controller design with general form, a branch and
bound method is employed to obtain the upper bound and the lower bound of the
optimal solution. It is shown that the optimal solution can be approximated by
the upper bound and the lower bound with any accuracy. Examples are provided
to show the effectiveness of the proposed method.
3.2.2 IMC Design Review and New Formulation
The schematic of the IMC system is depicted in Fig. 3.1 , where G is the given
stable process to be controlled, G the model of the process, and K(s) the IMC
controller. It follows from Morari and Zafiriou (1989) that the design procedure is
as follows. The model is factorized as
G(s) = G−(s)G+(s), (3.14)
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 61
where G−(s) has the stable and minimal phase part of G(s) only and G+(s) has
all the time delay and right half plane zeros of G(s):
G+(s) = e−Ls(∏
i
1− βis
1 + βis), Re(βi) > 0. (3.15)
The controller K(s) takes the form:
Q = G−1− f, (3.16)
where f is a user-specified filter and is usually chosen as
f =1
(τs+ 1)n, (3.17)
where n is large enough to guarantee that K(s) is proper.
The IMC controller design is to achieve the optimal nominal performance
minQ‖(1− GQ)v‖22 (3.18)
subject to the constraints that the IMC system is robustly stable and K is causal
and stable, where v is a set-point input of interest. Assume that the family of the
stable plants Π is described by
Π = G : |G(iω)− G(iω)
G(iω)| < lm(ω). (3.19)
It is shown that the IMC system is robustly stable if and only if
‖GQlm‖∞ = supω
|GQlm| < 1. (3.20)
In short, one solves (3.18) subject to (3.20). However, the problem is difficult
to be solved directly. In practice, (3.18) is solved first and then τ is used as a
tuning parameter to achieve the appropriate compromise between robustness and
performance. A trial and error method is used to find τ .
In the following, we try to find a new framework to design the controller K(s)
to achieve optimal performance with the robust stability constraint for a given
uncertain bound. With the same objective function as in that of Morari’s design,
the robust IMC design problem is now recast into
minQq
γ, (3.21)
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 62
subject to∣
∣
∣
∣
∣
∣(1− GQq)v∣
∣
∣
∣
∣
∣
2
2< γ, (3.22)
∣
∣
∣
∣
∣
∣lmGQq∣
∣
∣
∣
∣
∣
2
∞< 1. (3.23)
Suppose that the nominal model of the process is described by
G(s) =β(s)
α(s)=
βnsn + βn−1s
n−1 · · ·+ β1s+ β0αnsn + αn−1sn−1 + · · ·+ α1s+ α0
e−Ls, (3.24)
and the controller takes the general form:
Q(s) =b(s)
a(s)=
bmsm + bm−1s
m−1 · · ·+ b1s+ b0amsm + am−1sm−1 + · · ·+ a1s+ 1
. (3.25)
Our task is to find the optimal controller Q(s) in (3.25) to solve the problem (3.21)-
(3.23). In the subsequent two subsections, two schemes are presented respectively.
The first scheme deals with controller design with fixed poles, while the second
treats controller design without this restriction.
3.2.3 Controller Design with Fixed Poles
In the process control practice, the closed-loop bandwidth ωcb can rarely exceed
ten times of the open loop process bandwidth ωpb (Morari and Zafiriou, 1989).
Usually, the desired closed-loop bandwidth is chosen as ωcb = µωpb, µ ∈ [0.1, 10].
Let the denominator of controller be a(s) = (τs+ 1)n, so that
Q(s) =bms
m + bm−1sm−1 · · ·+ b1s+ b0
(τs+ 1)m. (3.26)
We may relate the closed-loop time constant to the closed-loop bandwidth accord-
ing the approach in (Wang et al., 2001):
τ =
√
m√2− 1
µωpb, µ ∈ [0.1, 10]. (3.27)
To ensure the robust stability will naturally reduce the closed-loop performance or
speed, the bandwidth of controller should be more narrow than the process.
Performance Consider the performance requirements (3.21) and (3.22).
Suppose that v is a unit step.
∣
∣
∣
∣
∣
∣(1− GQ)v∣
∣
∣
∣
∣
∣
2
2=
1
π
∫ ∞
0
∣
∣
∣
∣
∣
G(jω)b(jω)− a(jω)
(jω)a(jω)
∣
∣
∣
∣
∣
2
dω. (3.28)
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 63
Write G = Gre + jGim, where Gre and Gim are real and imaginary parts of G
respectively. It follows that
Gb(jω) =Gre[b0 − b2ω2 + b4ω
4 + · · · ]− ωGim[b1 − b3ω2 + b5ω
4 + · · · ]+
jGim[b0 − b2ω2 + b4ω
4 + · · · ] + jωGre[b1 − b3ω2 + b5ω
4 + · · · ],
and
a(jω) = 1 + C1m(jτω) + C2m(jτω)2 + · · ·+ Cm
m (jτω)m,
where Ckm = m!
(m−k)!k!. In order to achieve zero steady-state error for the nominal
IMC system, we set b0 = G−1(0). We thus obtain
∣
∣
∣Gb(jω)− a(jω)
∣
∣
∣
2
=Gre[−b2ω2 + b4ω4 + · · · ]− ωGim[b1 − b3ω
2 + b5ω4 + · · · ] + [C2mτ
2ω2 − C4mτ4ω4 + · · · ]2+
Gim[−b2ω2 + b4ω4 + · · · ] + ωGre[b1 − b3ω
2 + b5ω4 + · · · ] + [−C1mτω + C3mτ
3ω3 + · · · ]2
For simplicity, we let x0 = C2mτ2ω2−C4mτ 4ω4+· · · , y0 = −C1mτω+C3mτ 3ω3+· · · ,
x1 = −ωGim, y1 = ωGre, x2 = ω2Gre, y2 = −ω2Gim, · · · . Thus, we obtain
∣
∣
∣Gb(jω)− a(jω)∣
∣
∣
2
=[x0(ω) + b1x1(ω) + b2x2(ω) + · · · ]2 + [y0(ω) + b1y1(ω) + b2y2(ω) + · · · ]2
=[
1 b1 · · ·]
x20(ω) + y20(ω) x0(ω)x1(ω) + y0(ω)y1(ω) · · ·x0(ω)x1(ω) + y0(ω)y1(ω) x21(ω) + y21(ω) · · ·
......
. . .
1
b1...
.
Let θ = [1 b1 · · · bm]T and
P =
x20(ω) + y20(ω) x0(ω)x1(ω) + y0(ω)y1(ω) · · ·x0(ω)x1(ω) + y0(ω)y1(ω) x21(ω) + y21(ω) · · ·
......
. . .
,
which is symmetric and positive matrix. (3.28) is then written as
∣
∣
∣
∣
∣
∣(1− GQ)v∣
∣
∣
∣
∣
∣
2
2=
1
π
∫ ∞
0
θT P(ω)θω2(τ 2ω2 + 1)m
dω = θT
[
1
π
∫ ∞
0
P(ω)ω2(τ 2ω2 + 1)n
dω
]
θ.
(3.29)
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 64
Let
P =1
π
∫ ∞
0
P(ω)ω2(τ 2ω2 + 1)n
dω, (3.30)
(3.22) is converted, by Schur complement (Boyd et al., 1994), into
γ b
bT P−1
> 0. (3.31)
Thus, the optimal performance design is to solve (3.21) subject to (3.31). For a
fixed τ , (3.31) becomes an LMI.
Robustness For the uncertain process described by (3.19), if both G(s)
and G(s) are asymptotically stable, the robust stability condition is (3.23). With
the controller in (3.26) and b0 = G−1(0), we obtain
|b(jω)|2 = |G−1(0) + b1(jω) + b2(jω)2 + · · · |2
= |G−1(0)− b2ω2 + · · · ]2 + ω2[b1 − b3ω
2 + · · · ]2
=[
1 b1 · · ·]
G−2(0) −G−1(0)ω2 · · ·−G−1(0)ω2 ω4 · · ·
......
. . .
1
b1...
.
Let θ = [1 b1 · · · bm]T and
Q(ω) =
G−2(0) −G−1(0)ω2 · · ·−G−1(0)ω2 ω4 · · ·
......
. . .
, (3.32)
where Q(ω) is a symmetric and positive definite matrix. Thus, one has
∣
∣
∣
∣
∣
∣lm(ω)G(jω)Q(jω)∣
∣
∣
∣
∣
∣
2
∞= max
ω∈(0,∞)
∣
∣
∣
∣
∣
lm(ω)G(jω)b(jω)
a(jω)
∣
∣
∣
∣
∣
2
= maxω∈(0,∞)
θT |lm(ω)G(jω)|2Q(ω)θ
ω2(τ 2ω2 + 1)m. (3.33)
Let Q = |lm(ω)G(jω)|2Q(ω)ω2(τ2ω2+1)m
, (3.33) is
maxω∈(0,∞)
θTQ(ω)θ. (3.34)
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 65
We notice that (3.34) is a complex nonlinear function with respect to ω and θ even
if τ is fixed. Thus, the exact solution with constraint (3.23) is hard to be found.
However, we can approximate the constraint (3.23) with finite frequency points ωi,
i = 1, · · · , l. Then (3.34) becomes
θTQ(ωi)θ < 1, i = 1, · · · , l. (3.35)
With such a finite point approximation, we notice that the constraint (3.23) is less
strict, thus the lower bound of the optimal solution may be found with the relaxed
constraint (3.35). If we increase l, the lower bound will get closer to the global
optimal solution. However, in this way, we do not know how large the l should be
and if the approximate solution is good enough. Fortunately, we can also find the
upper bound of the optimal solution with some assumption. If bi ≥ 0, i = 1, · · · ,m,
maxω∈(0,∞)
[bTQ(ω)b] < bT [Qmax(ω)]b, where
Qmax(ω) =
maxω∈(0,∞)
Q11 maxω∈(0,∞)
Q12 · · ·
maxω∈(0,∞)
Q21 maxω∈(0,∞)
Q22 · · ·...
.... . .
.
Notice that bT [Qmax(ω)]b < 1 is equivalent to bT [Qmaxi (ω)]b < 1, i = 1, · · · , l + 1,
where
Qmaxi (ω) =
maxω∈(ωi−1,ωi)
Q11 maxω∈(ωi−1,ωi)
Q12 · · ·
maxω∈(ωi−1,ωi)
Q21 maxω∈(ωi−1,ωi)
Q22 · · ·...
.... . .
,
with ω0 = 0 and ωm+1 =∞. Thus, the constraint (3.23) can be also approximated
by
bT Qmaxi (ω) b < 1. i = 1, · · · , l, l + 1. (3.36)
Since it is a more strict constraint than (3.23). Thus we can obtain the upper
bound of the problem (3.21). Similarly, if we increase l, the the upper bound will
get closer to the global optimal solution. If the error between the upper and lower
bound is small enough, we obtain the good approximate to the optimal solution
with a desired accuracy.
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 66
According to Schur complement, (3.35) and (3.45) are converted respectively
into LMIs:
1 b
bT Q(ωi)−1
> 0, i = 1, · · · , l, (3.37)
and
1 b
bT [Qmaxi (ω)]−1
> 0, i = 1, · · · , l, l + 1. (3.38)
Algorithm 3.1. Given G(s) in (3.24), seek Q(s) in (3.26) to minimize (3.21)
subject to (3.22) and (3.23).
Step 1 Compute τ in the denominator of the controller from (3.27).
Step 2 Calculate P and Q(ω) from (3.30) and (3.32) respectively.
Step 3 Select a set of frequency points ωi, i = 1, · · · , l.Step 4 Find the lower bound γ: minimize γ subject to (3.31) and (3.37).
Step 5 Find the upper bound γ: minimize γ subject to (3.31) and (3.38).
Step 6 If γ − γ < ε, where ε is given error bound, stop. Otherwise, increase
l, choose more frequency points and go to Step 4.
3.2.4 Controller Design with General Form
In the preceding subsection, the denominator of controller is chosen based on the
open loop bandwidth. It is obvious that it may yield a good controller but usually
not the optimal one. In this section, we take the controller Q in a general form in
(3.25). Thus if the global optimal solution is found, the controller will be the best
among all the controller of the same order for the IMC system.
Performance In order to achieve optimal performance, we should minimize
γ subject to (3.22). Suppose that v is a unit step, we have
∣
∣
∣
∣
∣
∣(1− GQ)v∣
∣
∣
∣
∣
∣
2
2=
1
π
∫ ∞
0
∣
∣
∣
∣
∣
G(jω)b(jω)− a(jω)
(jω)a(jω)
∣
∣
∣
∣
∣
2
dω. (3.39)
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 67
Write G = Gre + jGim, where Gre and Gim are real and imaginary parts of G
respectively. It follows that
Gb(jω) =Gre[b0 − b2ω2 + b4ω
4 + · · · ]− ωGim[b1 − b3ω2 + b5ω
4 + · · · ]+
jGim[b0 − b2ω2 + b4ω
4 + · · · ] + jωGre[b1 − b3ω2 + b5ω
4 + · · · ],
and
a(jω) = 1 + a1(jω) + a2(jω)2 + · · ·
In order to achieve zero steady-state error for the nominal IMC system, we set
b0 = G−1(0). We thus obtain
∣
∣
∣Gb(jω)− a(jω)
∣
∣
∣
2
=Gre[−b2ω2 + b4ω4 + · · · ]− ωGim[b1 − b3ω
2 + b5ω4 + · · · ] + [a2ω
2 − a4ω4 + · · · ]2+
Gim[−b2ω2 + b4ω4 + · · · ] + ωGre[b1 − b3ω
2 + b5ω4 + · · · ] + [−a1ω + a3ω
3 + · · · ]2
=ω2[Gimb1 + Greωb2 − Gimω2b3 − Greω
3b4 + Gimω4b5 + · · · − ωa2 + ω3a4 + · · · ]2+
[Greb1 + Gimωb2 − Greω2b3 − Gimω
3b4 + Greω4b5 + · · ·+ a1 − ω2a3 + · · · ]2
For simplicity, let xb1 = Gim, xb2 = Greω, · · · , xa1 = 0, xa2 = −ω, · · · , yb1 = Gre,
yb2 = Gimω, · · · , ya1 = 1, ya2 = 0, · · · , we obtain
∣
∣
∣Gb(jω)− a(jω)∣
∣
∣
2
ω2
=[a1xa1(ω) + a2x
a2(ω) + · · ·+ b1x
b1(ω) + b2x
b2(ω) + · · · ]2+
[a1ya1(ω) + a2y
a2(ω) + · · ·+ b1y
b1(ω) + b2y
b2(ω) + · · · ]2
=[
a1, a2, · · · , b1, b2 · · ·]
xa12 + ya1
2 xa1xa2 + ya1y
a2 · · · xa1x
b1 + ya1y
b1 xa1x
b2 + ya1y
b2 · · ·
xa2xa1 + ya2y
a1 xa2
2 + ya22 · · · xa2x
b1 + ya2y
b1 xa2x
b2 + ya2y
b2 · · ·
......
. . ....
... · · ·xb1x
a1 + yb1y
a1 xb1x
a2 + yb1y
a2 · · · xb1
2+ yb1
2xb1x
b2 + yb1y
b2 · · ·
xb2xa1 + yb2y
a1 xb2x
a2 + yb2y
a2 · · · xb2x
b1 + yb2y
b1 xb2
2+ yb2
2 · · ·...
.... . .
...... · · ·
a1
a2...
b1
b2...
.
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 68
Let φ = [a1, a2, · · · , b1, b2, · · · ]T , ϕ = [1, a1, a2, · · · ]T ,
P(ω) =
xa12 + ya1
2 xa1xa2 + ya1y
a2 · · · xa1x
b1 + ya1y
b1 xa1x
b2 + ya1y
b2 · · ·
xa2xa1 + ya2y
a1 xa2
2 + ya22 · · · xa2x
b1 + ya2y
b1 xa2x
b2 + ya2y
b2 · · ·
......
. . ....
... · · ·xb1x
a1 + yb1y
a1 xb1x
a2 + yb1y
a2 · · · xb1
2+ yb1
2xb1x
b2 + yb1y
b2 · · ·
xb2xa1 + yb2y
a1 xb2x
a2 + yb2y
a2 · · · xb2x
b1 + yb2y
b1 xb2
2+ yb2
2 · · ·...
.... . .
...... · · ·
(3.40)
and
P(ω) =
1 0 −ω2 · · ·0 ω2 0 · · ·−ω2 0 ω4 · · ·...
......
. . .
, (3.41)
where both P(ω) and P(ω) are symmetric and positive definite matrices. (3.29) is
then written as
∣
∣
∣
∣
∣
∣(1− GQ)v
∣
∣
∣
∣
∣
∣
2
2= φT
[
1
π
∫ ∞
0
P(ω)ϕT P(ω)ϕ
dω
]
φ. (3.42)
Let P =[
1π
∫∞
0P(ω)
ϕT P(ω)ϕdω]
, (3.42) is converted, by Schur complement, into
γ φ
φT P−1
> 0. (3.43)
Thus, the optimal performance design is to solve (3.21) subject to (3.43). It is
noticed that (3.43) is an LMI if we set ϕ = ϕ0 = [1, a01, a02, · · · ]T .
Robustness For the uncertain process described by (3.19), if both G(s)
and G(s) are asymptotically stable, the robust stability condition is (3.23). With
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 69
the controller in (3.26) and b0 = G−1(0), we obtain
|lm(ω)G(jω)b(jω)|2
=|lm(ω)G(jω)|2|G−1(0) + b1(jω) + b2(jω) + · · · |2
=|lm(ω)G(jω)|2[G−1(0)− b2ω2 + · · · ]2 + ω2[b1 − b3ω
2 + · · · ]2
=[
1 b1 · · ·]
G−2(0)|lm(ω)G(jω)|2 −G−1(0)ω2|lm(ω)G(jω)|2 · · ·−G−1(0)ω2|lm(ω)G(jω)|2 ω4|lm(ω)G(jω)|2 · · ·
......
. . .
1
b1...
.
Let θ = [1, b1, b2, · · · ]T and
Q =
G−2(0)|lm(ω)G(jω)|2 −G−1(0)ω2|lm(ω)G(jω)|2 · · ·−G−1(0)ω2|lm(ω)G(jω)|2 ω4|lm(ω)G(jω)|2 · · ·
......
. . .
,
Q(ω) =
1 ω −ω2 · · ·ω −ω2 ω3 · · ·−ω2 ω3 ω4 · · ·...
......
. . .
,
one has
∣
∣
∣
∣
∣
∣lm(ω)G(jω)Q(jω)
∣
∣
∣
∣
∣
∣
2
∞= max
ω∈(0,∞)
∣
∣
∣
∣
∣
lm(ω)G(jω)b(jω)
a(jω)
∣
∣
∣
∣
∣
2
= maxω∈(0,∞)
θT Q(ω)θ
ϕT Q(ω)ϕ. (3.44)
We notice that (3.44) is a complex nonlinear function respect to ω, θ and ϕ.
Thus, the exact solution with constraint (3.23) is hard to be found. However, we
can approximate the constraint (3.23) with finite frequency point ωi, i = 1, · · · , l.Then (3.44) becomes
θT Q(ωi)θ < ϕT Q(ωi)ϕ, i = 1, · · · , l. (3.45)
With such a finite point approximation, we notice that the constraint (3.23) is less
strict, thus the lower bound of optimal solution may be found with the relaxed
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 70
constraint (3.45). If we increase l, the the lower bound will get closer to the global
optimal solution. However, in this way, we do not know how large the l should be
and if the approximate solution is good enough. Fortunately, we can also find the
upper bound of the optimal solution with some assumption. If bi ≥ 0, i = 1, · · · ,m,
maxω∈(0,∞)
[θT Q(ω)θ] ≤ θT [Qmax(ω)]θ, where
Qmax(ω) =
maxω∈(0,∞)
Q11 maxω∈(0,∞)
Q12 · · ·
maxω∈(0,∞)
Q21 maxω∈(0,∞)
Q22 · · ·...
.... . .
.
Moreover, if ai ≥ 0, i = 1, · · · ,m, minω∈(0,∞)
[ϕT Q(ω)ϕ] ≥ ϕT [Qmin(ω)]ϕ, where
Qmin(ω) =
minω∈(0,∞)
Q11 minω∈(0,∞)
Q12 · · ·
minω∈(0,∞)
Q21 minω∈(0,∞)
Q22 · · ·...
.... . .
.
Notice that θT [Qmax(ω)]θ < ϕT [Qmin(ω)]ϕ is equivalent to θT [Qmaxi (ω)]θ < ϕT [Qmin
i (ω)]ϕ, i =
1, · · · , l + 1, where
Qmaxi (ω) =
maxω∈(ωi−1,ωi)
Q11 maxω∈(ωi−1,ωi)
Q12 · · · maxω∈(ωi−1,ωi)
Q21 maxω∈(ωi−1,ωi)
Q22 · · ·...
.... . .
,
and
Qmini (ω) =
minω∈(ωi−1,ωi)
Q11 minω∈(ωi−1,ωi)
Q12 · · ·
minω∈(ωi−1,ωi)
Q21 minω∈(ωi−1,ωi)
Q22 · · ·...
.... . .
,
with ω0 = 0 and ωl+1 =∞. Thus, the constraint (3.23) can be also approximated
by
θT Qmaxi (ω) θ < ϕT [Qmin
i (ω)]ϕ. i = 1, · · · , k, k + 1, (3.46)
which is more strict constraint than (3.23). Thus we can obtain the upper bound of
the problem (3.21). Similarly, if we increase model order m, the the upper bound
will get closer to the global optimal solution. If the error between the upper and
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 71
lower bound is small enough, we obtain the good approximation optimal solution
with a desired accuracy.
According to Schur complement, (3.45) and (3.46) are converted respectively
into
ϕT Q(ωi)ϕ θ
θT Q(ωi)−1
> 0, i = 1, · · · , l, (3.47)
and
ϕT [Qmini (ω)]ϕ θ
θT [Qmaxi (ω)]−1
> 0, i = 1, · · · , l, l + 1. (3.48)
Notice that (3.47) and (3.48) are bilinear matrix inequalities (BMI), which cannot
be solved by LMI optimization directly. Thus, we introduce branch and bound
method to convert the BMI problem into LMI problem.
Branch and Bound
The basic idea of branch and bound is widely used in many forms to handle
difficult optimization problem (Reiner and Tuy, 1992). However, the basic idea is
very simple. Firstly, the lower bound γ(M0) and (if possible) upper bound γ(M0)
of γ(M0) should be found on a feasible set M0 respectively. Secondly, partition
M0 into finite many subsets Mi, i = 1, · · · ,m and find the lower bound γ(Mi) and
upper bound γ(Mi) of each section. Let γ = miniγ(Mi) and γ = min
iγ(Mi) be
overall bounds. If γ is sufficiently close to γ then the algorithm terminates. If not,
then the domain is subdivided and the procedure is repeated for each section. If
the lower bound for a region γ(Mi) is greater than the current upper bound γ then
that region may be discarded. The convergence of the algorithm is guaranteed as
eventually the upper bound and lower bound will be equal to each other within
the specified tolerance.
The most important step to develop an efficient branch and bound algorithm is
to derive tight upper and lower bounds from the objective function over any given
part of the domain. Introducing a set of variables wij, which we wish wij = aiaj.
Thus, the constraint (3.47) is converted into linear matrix constraints as
Q(ωi)wij θ
θT Q(ωi)−1
> 0, i = 1, · · · , l. (3.49)
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 72
Suppose that li < ai < ri, i = 1, · · · ,m, wij should satisfy (Fiacco and G.P.McCormick,
1990)
minlilj, lirj, rirj < wij < maxlilj, lirj, rirj
wij 6 liai + rjaj − ljri
wij 6 liai + rjaj − rjli
wij 6 liai + rjaj − lilj
wij 6 liai + rjaj − rirj (3.50)
The lower bound of the optimal problem with the constraint (3.47) can be found
by LMI optimal problem with constraints (3.49) and (3.50). The upper bound can
found by fixing some variables ai = li. Thus the optimal solution can be found by
branch and bounding. Similarly, the optimization problem with constraint (3.48)
can also be solved.
Algorithm 3.2. Given G(s) in (3.24), seek Q(s) in (3.26) to minimize (3.21).
Step 1 Select ϕ0.
Step 2 Calculate P , Q(ω) and Q(ω) from (3.42) and (3.44).
Step 3 Select a set of frequency points ωi, i = 1, · · · , l.Step 4 Find the lower bound of the optimal solution γ: minimize γ subject to
(3.43) and (3.47) using branch and bound method.
Step 5 Find the upper bound of the optimal solution γ: minimize γ subject to
(3.43) and (3.48) using branch and bound method.
Step 6 If γ − γ < ε, where ε is given error bound, go to Step 7. Otherwise,
increase l, choose more frequency points and go to Step 4.
Step 7 If ‖ϕstar −ϕ0‖ < ε1, where ε1 is given error bound, stop. Otherwise, let
ϕ0 = ϕstar, go to Step 2.
3.2.5 Examples
In this subsection, two examples are provided to illustrate the effective of the
proposed design.
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 73
0 5 10 150
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
time
plan
t out
put
Figure 3.6. Step response
(- - - Proposed controller with fixed pole,
—- Proposed controller with general form,
...... Morari’s controller.)
Example 3.3. Consider a first order plus dead time model
G =e−s
s+ 1,
with uncertain bound described by∣
∣
∣
∣
∣
G− G
G
∣
∣
∣
∣
∣
< lm(ω) =
√10ω√w2 + 1
.
Let G(0)Q(0) = 1, the first order controller is
Q =bs+ 1
τs+ 1.
Step 1 Following Algorithm 3.1, we calculate τ from (3.27) and obtain
τ =
√
n√2− 1
µωpb=
√2− 1
0.3× 1= 2.894.
Step 2 Set v as unit step input, P is computed as
P =
2.1161 −0.5421−0.5421 0.3201
,
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 74
and Q(ω) is
Q(ω) =
4ω4
(ω2+1)2(τ2ω2+1)0
0 4ω2
(ω2+1)2(τ2ω2+1)
.
Step 3 Choose ωi = [0.8, 1.2, 1.5, 2.0, 5], the lower bound of the optimal value
can be computed by
γ = minb
γ
subject to
γ 1 b
1 0.8347 1.4136
b 1.4136 5.5178
> 0,
1 1 b
1 1.0e7 × 2.5006 0
b 0 1.0e7 × 0.0003
> 0,
1 1 b
1 4.0999 0
b 0 1.4760
> 0,
1 1 b
1 2.7808 0
b 0 1.3626
> 0,
1 1 b
1 2.0768 0
b 0 1.3291
> 0,
1 1 b
1 0.9944 0
b 0 1.3291
> 0
1 1 b
1 1.0e3 × 0.0108 0
b 0 1.0e3 × 1.0819
> 0.
The lower bound of the optimal problem is γ = 1.546 and optimal solution b =
1.198.
Step 4 The upper bound of the optimal problem can be computed by
γ = minb
γ
subject to
γ 1 b
1 0.8676 1.2787
b 1.2787 4.6203
> 0,
1 1 b
1 1.0e6 × 1.0002 0
b 0 1.0e6 × 0.0001
> 0,
1 1 b
1 0.1522 0
b 0 0.0548
> 0,
1 1 b
1 0.1008 0
b 0 0.0494
> 0,
1 1 b
1 0.0734 0
b 0 0.0470
> 0,
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 75
1 1 b
1 0.0272 0
b 0 0.0465
> 0,
1 1 b
1 0.1988 0
b 0 19.8817
> 0.
The upper bound of the optimal problem is γ = 1.548 and optimal solution b =
1.206.
Step 5 The optimal controller (for upper bound) obtained is
Q =1.206s+ 1
1.s+ 1,
and the optimal H2 norm is γ = 1.548.
With the proposed method for general form controller, we obtain
Q =0.194s+ 1
1.337s+ 1
and the H2 norm is 1.528.
Using Morari’s trial and error method, the controller is given as
Q =s+ 1
2.21s+ 1
and the H2 norm is 1.827.
The step responses of nominal plant for three methods are given in Fig. 3.6.
Example 3.4. Consider a second order non minimal phase model
G =s− 5
(s+ 1)(s+ 5)
with uncertain bound described by
∣
∣
∣
∣
∣
G− G
G
∣
∣
∣
∣
∣
< lm(ω) =2ω√w2 + 1
.
Let G(0)Q(0) = −1, the first order controller is
Q =bs+ 1
τs+ 1.
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 76
0 5 10 15 20 25 30−0.2
0
0.2
0.4
0.6
0.8
1
1.2
time
plan
t out
put
Figure 3.7. Step response
(- - - Proposed controller with fixed pole,
—- Proposed controller with general form,
...... Morari’s controller.)
With the proposed method, the controller with fixed pole is
Q =−2.084s− 1
2.89s+ 1
and the H2 norm is 1.092.
With the proposed method, the controller with general form is
Q =−0.469s− 1
0.820s+ 1
and the H2 norm is 1.061.
With Morari’s method, the controller is
Q =−s− 1
2.21s+ 1
and the H2 norm is 1.227.
The step responses of nominal plant for three methods are given in Fig. 3.7.
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 77
3.2.6 Conclusions
In this section, a new IMC controller design method is proposed to achieve the
optimal performance and robust stability. The method is based on LMI/BMI
optimization. It is shown that better optimal performance is achieved compared
with Morari’s method with the same robust stability constraints. It also should be
noted that the proposed scheme is a systemic method while most of existing IMC
tuning schemes use trial and error method.
3.3 Modified Smith Predictor Control for Dis-
turbance Rejection with Unstable Processes
3.3.1 Introduction
In process control, the Smith predictor (SP) (Smith, 1959) is a well known and
very effective dead-time compensator for stable processes. However, the origi-
nal Smith predictor control scheme will be unstable when applied to an unstable
process (Wang et al., 1999b). To overcome this obstacle, many modifications to
the original Smith scheme have been proposed. Astrom et al. (1994) presented
a modified Smith predictor for integrator plus dead-time process and can achieve
faster setpoint response and better load disturbance rejection with decoupling of
the setpoint response from the load disturbance response. Matausek and Micic
(1996) considered the same problem with similar results but their scheme is easier
to tune. Majhi and Atherton (1999) proposed a modified Smith predictor control
scheme which has high performance particularly for unstable and integrating pro-
cess. This method achieves optimal integral squared time error criterion (ISTE)
for setpoint response and employs an optimum stability approach with a propor-
tional controller for an unstable process. Another paper of Majhi and Atherton
(2000) extends their work for better performance and easy tuning procedure for
first order plus dead time (FOPDT) and second order plus dead time (SOPDT)
processes. To our best knowledge, Majhi and Atherton (2000) achieve best the
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 78
performance for both setpoint response and disturbance rejection with unstable
dead time processes with modified Smith predictor structure.
In this section, a new modified Smith predictor structure is proposed. A simple
but effective control procedure is designed to improve the performance especially
for the disturbance rejection. The internal stability of the proposed structures is
also analyzed which is not reported in the previous publications. Simulation is
given to illustrate the effectiveness of the proposed method.
3.3.2 The Proposed New Structure
Ls 0 (s)e G -
) ( ˆ s 0 G
Ls e -
u
d
y
y m
r
ˆ
Figure 3.8. Proposed smith predictor control scheme
Let the process and model be represented by G0(s)e−Ls and G0(s)e
−Ls respec-
tively, where G0 and G0 are rational functions with at least one unstable pole. A
new modified Smith predictor control scheme is depicted in Fig.3.8.
Suppose that the model used perfectly matches the plant dynamics, i.e. G0 =
G0 and L = L. Using Mason’s formula (Franklin et al., 2002; Goodwin et al., 2000),
we obtain the system determinant ∆ as:
∆ = 1 +G0K2 +G0K1 +K3G0e−Ls + (1 +G0K2 +G0Q1)K3G0e
−Ls
= (1 +G0K12)(1 +K3G0e−Ls).
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 79
The closed-loop response to setpoint and disturbance inputs are given by
Tr =G0K4
1 +G0K12e−Ls,
and
Td =G0e
−Ls(1 +K12G0 −K2G0e−Ls)
(1 +G0K12)(1 +K3G0e−Ls),
where K12 = K1 +K2.
In this new structure, there are four controllers to be designed for different
objectives. We may note that the denominator of the transfer function Tr has
no time delay. Thus, we select controller K12 to place the closed-loop poles to
desired points using pole placement method. Moreover, similar to the two degree
of freedom design, the prefilter K4 may be designed to make the setpoint response
achieve better performance. For the disturbance path, the K3 is employed to
stabilize the time delay part. The controller K2 is chosen to minimize 1+K12G0−K2G0e
−Ls. If we obtain 1 + K12G0 − K2G0e−Ls = 0, the disturbance will be
completely rejected. However, we cannot find such a realizable controller K2 for
most processes. We can only choose appropriate K2 to minimize the disturbance
in critical frequency range.
3.3.3 Internal Stability
Input/output stability does not necessarily acceptable for control system design.
The unstable pole-zero cancellations in the system may cause unbounded signals
in the system and the system may be damaged by such signals. Thus, internal
stability is a prerequisite for any control systems. In our modified Smith predictor
structure, the interconnected system consists of seven subsystems and each of them
is of single input and single output (SISO). It is shown by Wang et al. (1999b) that
an interconnected system consisting only of SISO plants is internally stable if and
only if pc(s) , ∆∏
i pi(s) has all its roots in the open left half of the complex
plane, where pi(s) are the denominators of the subsystem transfer functions and
the ∆ is the system determinant as defined in the Mason’s formula. To apply this
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 80
theorem, one sees that the system in Fig.3.8 consists of seven subsystems:
K1(s), K2(s), K3(s), K4(s), G0, G0e−Ls and e−Ls.
Let K1(s) =c1(s)d1(s)
, K2(s) =c2(s)d2(s)
, K3(s) =c3(s)d3(s)
, K4(s) =c4(s)d4(s)
and G0(s) = G0(s) =
α(s)β(s)
. Their respective pi are
p1 = d1(s), p2 = d2(s), p3 = d3(s), p4 = d4(s), p5 = p6 = β(s) and p7 = 1.
Therefore, we have
pc(s) = ¯∆(s)5∏
i=1
pi(s)
= [1 + (K1 +K2)G0][1 +K3G0e−Ls] · d1(s) · d2(s) · d3(s) · d4(s) · β(s) · β(s) · 1
= d4(s)[d1(s)d2(s)β(s) + c1(s)d2(s)α(s) + c2(s)d1(s)α(s)][d3(s)β(s) + c3(s)α(s)e−Ls].
The polynomial, d1(s)d2(s)β(s)+c1(s)d2(s)α(s)+c2(s)d1(s)α(s), reflects stabiliza-
tion of delay-free G0 by the controller K12 = K1 + K2, which is always possible
designed, say, by pole placement. K4 is a stable filter to improve the setpoint re-
sponse. Thus, the overall system is internally stabilizable if and only if the process
with time delay G0e−Ls is stabilizable by controller K3. For a general unstable de-
lay system, an approach to design a stabilizing controller K3 is reported by Bonnet
and Partington (1999).
In view of the above analysis, it can be concluded that unlike the original Smith
system where the characteristic equation is delay-free, the modified Smith scheme
gets no simplification as far as stabilization is concerned. It is indeed the case for
all the existing modified Smith schemes for unstable processes as we found (Majhi
and Atherton, 1999; Astrom et al., 1994; Matausek and Micic, 1996), that is, their
characteristic equations are all delay-dependent.
3.3.4 Controller Design
Most of the industrial processes may be approximated by first order plus time
delay (FOPTD) or second order plus time delay (SOPTD) processes. Moreover,
there are many integral plus time delay (IPTD) processes are reported (Astrom et
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 81
al., 1994). In this subsection, controllers are designed for IPTD, unstable FOPTD
and SOPTD processes with the proposed modified Smith predictor structure.
IPTD Processes Consider the following IPTD model:
G(s) = G0(s)e−Ls =
k
se−Ls,
where all coefficients are non-negative. The closed-loop transfer function for set-
point tracking is written as
Tr =kK4
s+ kK12e−Ls =
1
λs+ 1e−Ls,
where λ > 0 is an adjustable closed-loop design parameter. The higher value of λ,
the more robust of the closed-loop system. The small value of λ, the faster of the
response. The controller K12 and K4 are computed as
K12 =1
kλ,
and
K4 = K12.
The closed-loop transfer function for disturbance is given by
Td =ke−Ls(s+ kK12 − kK2e
−Ls)
(λs+ 1)(s+K3ke−Ls).
The controller K3 is designed to stabilize the second part of last equation s +
K3ke−Ls). According to Majhi and Atherton (2000), K3 is computed to satisfy the
optimum phase margin criterion (60o) which gives
K3 =π − 2× π
3
2Lk.
In order to achieve good disturbance rejection performance, K2 is selected to min-
imize s+ kK12 − kK2e−Ls. We approximate e−Ls by 1− Ls and choose
K2 = as+ b.
We wish
s+ kK12 − kK2e−Ls
=s+ kK12 − k[b+ (a− bL)s− aLs2]
=(kK12 − kb) + (1 + kbL− ka)s+ akLs2
=0
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 82
Let kK12 − kb = 0 and 1 + kbL − ka=0, we obtains b = K12 and a = 1/k + bL.
Finally, we may get K1 = K12 −K2.
FOPTD Processes Consider the following unstable FOPTD model:
G(s) = G0(s)e−Ls =
k
Ts− 1e−Ls,
where all coefficients are non-negative. The closed-loop transfer function for set-
point tracking is written as
Tr =kK4
Ts− 1 + kK12e−Ls =
1
λs+ 1e−Ls,
where λ > 0 is an adjustable closed-loop design parameter. The controller K12 and
K4 are computed as
K12 =1 + T/λ
k,
and
K4 = K12 −1
k.
The closed-loop transfer function for disturbance is given by
Td =G0e
−Ls(Ts− 1 + kK12 − kK2e−Ls)
(λs+ 1)(1 +K3G0e−Ls).
The controller K3 is designed to stabilize the second part of last equation 1 +
K3G0e−Ls. De Paor and O’Malley (1989) suggested a proportional controller to
stabilize an unstable FOPTD process based on the optimum phase margin criterion
which gives
K3 =
√
T
Lk2.
In order to achieve good disturbance rejection performance, the K2 is selected to
minimize Ts− 1 + kK12 − kK2e−Ls. We approximate e−Ls by 1− Ls and choose
K2 = as+ b.
We wish
Ts− 1 + kK12 − kK2e−Ls
=Ts− k[b+ (a− bL)s− aLs2] + kK12 − 1
=(kK12 − kb− 1) + (T + kbL− ka)s+ akLs2
=0
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 83
Let kK12 − kb − 1 = 0 and T + kbL − ka=0, we obtains b = K12 − 1/k and
a = T/k + bL. Finally, we may get K1 = K12 −K2.
SOPTD Processes Consider the following unstable SOPTD model:
G(s) = G0(s)e−Ls =
k
(T1s− 1)(T2s+ 1)e−Ls,
where all coefficients are non-negative. The closed-loop transfer function for set-
point tracking is written as
Tr =kK4/(T2s+ 1)
T1s− 1 + kK12/(T2s+ 1)e−Ls =
1
λs+ 1e−Ls,
where λ > 0 is an adjustable closed-loop design parameter. The controller K12 and
K4 are computed as
K12 = (T2s+ 1)1 + T/λ
k,
and
K4 = (T2s+ 1)(K12
T2s+ 1− 1
k).
The closed-loop transfer function for disturbance is given by
Td =G0e
−Ls(Ts− 1 + kK12/(T2s+ 1)− kK2/(T2s+ 1)e−Ls)
(λs+ 1)(1 +K3G0e−Ls).
The controller K3 is designed to stabilize the second part of last equation 1 +
K3G0e−Ls. Using De Paor and O’Malley (1989)’s method
K3 =
√
T1Lk2
(T2s+ 1).
In order to achieve good disturbance rejection performance, the K2 is selected to
minimize Ts− 1 + kK12/(T2s+ 1)− kK2/(T2s+ 1)e−Ls. We approximate e−Ls by
1− Ls and choose
K2 = (as+ b)(T2s+ 1).
We wish
Ts− 1 + kK12/(T2s+ 1)− kK2/(T2s+ 1)e−Ls
=Ts− k[b+ (a− bL)s− aLs2] + kK12/(T2s+ 1)− 1
=(kK12/(T2s+ 1)− kb− 1) + (T + kbL− ka)s+ akLs2
=0
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 84
Let kK12/(T2s+ 1)− kb− 1 = 0 and T + kbL− ka=0, we obtains b = K12/(T2s+
1)− 1/k and a = T/k + bL. Finally, we may get K1 = K12 −K2.
3.3.5 Examples
0 20 40 60 80 100 120 140 160 180 2000
0.2
0.4
0.6
0.8
1
1.2
1.4
Figure 3.9. Step responses for IPTD process
(- - - - Majhi’s method; —— proposed method)
Example 3.5. Consider an IPTD process (Majhi and Atherton, 2000)
G(s) =1
se−5s,
we obtain the controllers from the proposed method in last section as K1 = −6s,K2 = 6s + 1, K3 = 0.105 and K4 = 1. The amplitude of setpoint and load
disturbance are set as 1 and 0.1 respectively. Note that K1 and K2 in a form
of PD controller, we implement them with an industrial PD controller with the
form kds+kpkdNs+1
. In the examples, we choose N = 100. The responses of the proposed
method and Majhi’s method are shown in Fig. 3.9. It is illustrated that the
proposed method has better performance for setpoint response and disturbance
rejection.
Example 3.6. Consider an unstable FOPTD process (Majhi and Atherton, 2000)
G(s) =4
10s− 1e−5s,
we obtain the controllers from the proposed method in last section as K1 =
−8.75s + 0.25, K2 = 8.75s + 1.25, K3 = 0.35 and K4 = 1.25. The amplitude
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 85
0 20 40 60 80 100 120 140 160 180 2000
0.2
0.4
0.6
0.8
1
1.2
1.4
Figure 3.10. Step responses for unstable FOPTD process
(- - - - Majhi’s method; —— proposed method)
of setpoint and load disturbance are set as 1 and 0.1 respectively. The responses of
the proposed method and Majhi’s method are shown in Fig. 3.10 . It is illustrated
that the proposed method has better disturbance rejection performance. For this
example, we may check the closed-loop poles with first order Pade approximation
for the time delay. We have the closed-loop poles
pc(s) = [10s− 1 + (−8.75s+ 0.25) ∗ 4 + (8.75s+ 1.25) ∗ 4]× [(10s− 1)(1 + 2.5s) + 4 ∗ 0.35 ∗ (1− 2.5s)]
= 5(2s+ 1)(25s2 + 7.5s− 1 + 1.4− 3.5s)
= 5(2s+ 1)(25s2 + 4s+ 0.4),
thus the closed-loop system is stable.
0 20 40 60 80 100 120 140 160 180 2000
0.2
0.4
0.6
0.8
1
1.2
1.4
Figure 3.11. Step responses for unstable SOPTD process
(- - - - Majhi’s method; —— proposed method)
Chapter 3. Advance in Robust IMC Design for Step Input and Smith ControllerDesign for Unstable Processes 86
Example 3.7. Consider an unstable SOPTD process
G(s) =4
(10s− 1)(2s+ 1)e−5s,
we obtain the controllers from the proposed method in last section asK1 = (−15s+0.25)(2s+1), K2 = (15s+2.5)(2s+1), K3 = 0.707s+0.354 and K4 = 2.5(2s+1).
The amplitude of setpoint and load disturbance are set as 1 and 0.1 respectively.
The responses of the proposed method and Majhi’s method are shown in Fig.
3.11. It is illustrated that the proposed method has better disturbance rejection
performance as well as setpoint response.
Remark 3.2. The idea in the proposed method is to use the zero of controller to
compensate the low frequency dynamics of the delay part. The simulations illus-
trate the proposed approach may achieve much better performance for disturbance
rejection especially for the low frequency band, which is the main part of distur-
bance in many cases. Moreover, if we need to achieve better performance, we may
approximate the time delay with higher order zero which results in higher order
controller. Further, we only employ the approximation in the closed-loop zero de-
sign. Thus, the proposed method may also achieve the closed-loop stability for
time delay and maintain the advantage of the original Smith predictor design.
3.3.6 Conclusions
Simple and effective design is proposed based on our new modified Smith predic-
tor structure for unstable process with dead time. Moreover, the internal stabil-
ity of the proposed system is analyzed. Simulations illustrate that the proposed
method yields much improvement of load disturbance responses over the best ex-
isting method.
Chapter 4
Decoupling with Stability and
Decoupling Control Design
4.1 Preview
The problem of decoupling linear time-invariant multi-variable systems received
considerable attention in the system theoretical literature for several decades.
There are many important processes in which multiple control loops interact each
other. The interaction between loops may cause deleterious effects on the sys-
tem stability and performance. Although decentralized control is employed in
most practical multi-variable processes due to its simplicity, the decoupling control
scheme is a desired choice for complex multi-variable system, especially for the
processes which have strongly coupled loops. The trend in the process industry is
towards increased process integration and the highly complex multi-variable sys-
tem is more popular. Thus the decoupling design becomes an important area in
control field.
In this chapter, the solvability of the decoupling problem with internal stability
for unity output feedback systems is discussed. A necessary and sufficient condition
is developed with polynomial approach. Moreover, a robust decoupling control
scheme is developed via LMI optimization. The control aims at gaining good loop
performance and largely removing the coupling simultaneously for multi-variable
87
Chapter 4. Decoupling with Stability and Decoupling Control Design 88
processes. The robust stability of the proposed method is analyzed.
This chapter is organized as follows. In Section 4.2, a new necessary and suffi-
cient condition for the solvability of the decoupling problem with internal stability
for unity output feedback systems is given. In Section 4.3, a decoupling design via
LMI is presented.
4.2 Decoupling Problem with Stability via Trans-
fer Function Matrix Approach
4.2.1 Introduction
In the area of decoupling design, much attention was directed towards decoupling
by state feedback (Morgan, 1964; Falb and Wolovich, 1967; Hautus and Heymann,
1983; Decusse et al., 1988). More general block decoupling was investigated by
Williams and Antsaklis (1986). The necessary and sufficient conditions were given
by Wang and Davison (1975) for static output feedback. Decoupling through a
combination of precompensation and output feedback was considered by Pernebo
(1981) and Eldem and Ozguler (1989). A algebraic theory for decoupling via two-
degree-freedom design was formulated in Desoer and Gundes (1986). Eldem (1996)
presented the solution to the problem of decoupling through specified input-output
channels with internal stability. It is reported that the solvability condition and
the set of all solutions may be obtained. The regular block decoupling problem by
static state feedback was discussed by Camarta using geometric tools (Camarta
and Martnez-Garca, 2001).
In the case of unity output feedback systems, the existence condition of con-
troller for the plants which have full row rank is very simple and well known. But
the problem of decoupling in such a configuration while maintaining internal sta-
bility of the decoupled system appears to be much more difficult. A necessary and
sufficient solvability condition for the diagonal decoupling problem with internal
stability is given by Linnemann and Maier (1990) for 2 × 2 plants, and by Wang
Chapter 4. Decoupling with Stability and Decoupling Control Design 89
(1992) for square plant. The necessary and sufficient condition for this problem
was also derived based on transfer function matrices and residue by Lin (1997).
Here we will propose another necessary and sufficient solvability condition which
is simpler, more understandable and make use of transfer function matrices only.
4.2.2 Minimal C+-Decoupler
An m×m proper rational transfer function matrix D is said to be decoupled if D
is diagonal and nonsingular; Gd a decoupler for G if GGd is decoupled. Given an
m×m nonsingular proper rational transfer function matrix G, let the set of all the
common right-half plane (RHP) poles and zeros of G be Γ = γ1, γ2, . . . , γn. Letδp(G) denote the McMillan degree of a transfer function G at the RHP pole p, i.e.,
multiplicity of the pole at p. For simplification, we denote δp(G) = δp(G)+δp(G−1).
Definition 4.1. Gd,min is called a minimal C+- decoupler for G if (i) GGd,min
is decoupled; (ii) For any decoupler Gd for G, there holds δγ(Gd,min) 6 δγ(Gd) for
any γ ∈ Γ.
Lemma 4.1. (Existence and uniqueness of Gd,min) For any nonsingular G , (i)
there exists at least a Gd,min; (ii) All Gd,min have the same δγ(Gd,min).
Proof. Obviously, Gd = G−1 is a decoupler for any m×m nonsingular G. One can
see that GdD is still a decoupler for any decoupled D. For each γ ∈ Γ, we perform
the following search to find Gd,min from an initial decoupler Gd. Let lp (resp. lz)
be the greatest multiplicity of the pole of all the elements of Gd (resp. G−1d ) at
γ ∈ Γ. Suppose that lj, j = 1, 2, · · · ,m, are integers, D = diag(s − γ)l1 , (s −γ)l2 , · · · , (s − γ)lm and L = (l1, l2, · · · , lm)| − lz ≤ li ≤ lp, i = 1, 2, · · · ,m. If
(l1, l2, · · · , lm) /∈ L, then there exists at least a l∗j such that (a) l∗j > lp or (b)
l∗j < −lz. Suppose the case (a) (the case (b) can be treated with obvious changes).
Then, Gdiag(s− γ)l1 , · · · , (s− γ)lj−1 , (s− γ)lp , (s− γ)lj+1 , · · · , (s− γ)lm cannot
have any pole-zero cancellation with diag1, · · · , 1, (s− γ)l∗j−lp , 1, · · · , 1 and thus
δr(GD) with (l1, l2, · · · , lm) = (l1, · · · , lj−1, l∗j , lj+1, · · · , lm) is strictly greater than
δr(GD) with (l1, l2, · · · , lm) = (l1, · · · , lj−1, lp, lj+1, · · · , lm). It thus follows that a
Chapter 4. Decoupling with Stability and Decoupling Control Design 90
minimum δr(GD) will only come from some D with (l1, l2, · · · , lm) ∈ L if it exists.
Further, the combinations of such lj in the range L are finite and the unique
minimum δγ(GdD) must be achieved with one or a few combinations of lj in the
range and the corresponding D will yield Gd,min = GdD. If we start the search with
another initial decoupler Gd, then GGd = D and GGd = D, for some decoupled
D and D, which lead to Gd = G−1D = GdD−1D = GdD for a decoupled D whose
RHP pole and zero structure at γ must correspond to some D above. Thus, the
search will result in the same δγ(Gd,min) as before, in other words, δγ(Gd,min) is
independent of initial Gd and it is the property of G only.
The proof of Lemma 4.1 naturally gives rise to the following constructive pro-
cedure for finding Gd,min from G.
Step 1: Calculate Gd = G−1 and determine the set Γ from pole locations of
elements of G andG−1. If Γ = φ, i.e. empty set, then take Gd,min = Gd;
otherwise, proceed to step 2.
Step 2: For each γ ∈ Γ, find lpγ and lzγ from pole multiplicity of elements of Gd
and G−1d . Let Dγ = diag(s − γ)l1 , (s − γ)l2 , · · · , (s − γ)lm, tabulate
δγ(GdDγ) for all possible combinations of lj such that −lzγ 6 lj 6 lpγ
and determine the minimum δγ(GDγ) and the corresponding D∗γ.
Step 3: Set Gd,min = GdD∗γ1D∗γ2· · ·D∗
γn.
Remark 4.1. Note that the locations of poles of a multivariable system must be
the same as those of its elements (Maciehowski, 1989). Then the zeros of G are the
poles of G−1 and we can find the common RHP zeros and poles of G by examining
the poles of the elements of G and G−1.
Example 4.1. (Wang, 1992) Let
G =
1s+1
1s+2
1(s−1)(s+1)
s(s−1)(s+2)
.
It follows that
G−1 =
s(s+1)s−1
−(s+ 1)
− s+2s+1
(s+ 2)
Chapter 4. Decoupling with Stability and Decoupling Control Design 91
for which we can easily find that the only common RHP pole and zero of G is
at 1 i.e., Γ = 1. Take Gd = G−1 and obviously we have lpγ = 1 and lzγ = 1.
For D = diag(s − 1)l1 , (s − 1)l2, where −1 6 l1, l2 6 1, the search for Gd,min is
tabulated in Table 4.1, from which two Gd,min are found as
Gd,min1 =
s(s+1)s−1
− s+1s−1
− s+2s−1
s+2s−1
and
Gd,min2 =
s(s+ 1) −(s+ 1)
−(s+ 2) (s+ 2)
.
If we start our procedure with a very different Gd:
Gd =
1 − s+1s+2
− s+2s(s+1)
1
,
with lpγ = 0. And we have
G−1d =
ss−1
s(s+1)(s+2)(s−1)
s+2(s−1)(s+1)
ss−1
,
so that lzγ = 1. For D = diag(s− 1)l1 , (s− 1)l2 with −1 6 l1, l2 6 0, the search
for Gd,min is tabulated in Table 4.2, from which two Gd,min are found as
Gd,min1 =
1 − s+1s+2
− s+2s(s+1)
1
(4.1)
and
Gd,min2 =
1s−1
− s+1(s+2)(s−1)
− s+2s(s+1)(s−1)
1s−1
. (4.2)
Tables 4.1 and 4.2 verify that δγ(Gd,min) is unique for a given G although the Gd,min
could be different for different Gd.
4.2.3 Decoupling with Stability
The problem under consideration is to find the conditions on anm bym nonsingular
proper rational transfer function matrix G under which there is a K such that the
Chapter 4. Decoupling with Stability and Decoupling Control Design 92
Table 4.1. Search for Gd,min in Example 4.1
(l1, l2) K = Gd
(s− 1)l1 0
0 (s− 1)l2
δ1(K)
(−1,−1)
s(s+1)
(s−1)2− s+1
s−1
− s+2(s+1)2
s+2s−1
3
(−1, 0)
s(s+1)
(s−1)2−(s+ 1)
− s+2(s−1)2
s+ 2
3
(−1, 1)
s(s+1)
(s−1)2−(s+ 1)(s− 1)
− s+2(s−1)2
(s− 1)(s+ 2)
4
(0,−1)
s(s+1)s−1
− s+1s−1
− s+2s−1
s+2s−1
1
(0, 0)
s(s+1)s−1
−(s+ 1)
− s+2s−1
s+ 2
2
(0, 1)
s(s+1)s−1
−(s+ 1)(s− 1)
− s+2s−1
(s− 1)(s+ 2)
3
(1,−1)
s(s+ 1) − s+1s−1
−(s+ 2) s+2s−1
2
(1, 0)
s(s+ 1) −(s+ 1)
−(s+ 2) s+ 2
1
(1, 1)
s(s+ 1) −(s+ 1)(s− 1)
−(s+ 2) (s+ 2)(s− 1)
2
resulting unity output feedback system is internally stable and (I + GK)−1GK
is decoupled. This problem is here abbreviated to the DPSUF (the decoupling
problem with internal stability for unity output feedback system). For later use,
let us recall some notations. For any nonsingular rational matrix G, G = G+G− is
a stability decomposition if G+ is completely unstable and G− is completely stable
(Wang, 1992). The G+ can always be factorized as G+ = N+(D+)−1 where N+
Chapter 4. Decoupling with Stability and Decoupling Control Design 93
Table 4.2. Search for Gd,min in Example 4.1
(l1, l2) K = Gd
(s− 1)l1 0
0 (s− 1)l2
δ1(K)
(−1,−1)
1s−1
− s+1(s+2)(s−1)
− s+2s(s+1)(s−1)
1s−1
1
(−1, 0)
1s−1
− s+1s+2
− s+2s(s+1)(s−1)
1
2
(0,−1)
1 − s+1(s−1)(s+2)
− s+2s(s+1)
1s−1
2
(0, 0)
1 − s+1s+2
− s+2s(s+1)
1
1
and D+ are nonsingular polynomial matrices and right coprime. For the pair of
square polynomial matrices N and D, N and D are called externally skew prime
and N and D internally skew prime, if DN = N D with det(N) = αdet(N) with a
real α and N and D left coprime (Wolovich, 1978). Let Pra and Pla be a right and
left strict adjoints of a nonsingular polynomial matrix P respectively (Hammer
and Khargonekar, 1984).
Lemma 4.2. (Wang, 1992) (i) the DPSUF is solvable if and only if there is a
K such that GK is decoupled and has no RHP pole-zero cancellation; (ii) If the
DPSUF is solvable, then all solutions are given by
K = (G−)−1N+
ra(D+
la)−1K2 (4.3)
where K2 is decoupled such that K1 = N+N+ra(D+
laD+)−1 has no RHP pole-zero
cancellations with K2.
Chapter 4. Decoupling with Stability and Decoupling Control Design 94
Lemma 4.3. If Gd is a solution to DPSUF for G, then there exists a Gd,min such
that (i) Gd = Gd,minD, where D is decoupled; (ii) δγ(Gd) = δγ(Gd,min) + δγ(D),
where γ ∈ Γ.
Proof. Take Gd,min = (G−)−1N+
ra(D+
la)−1 as in (4.3) and the lemma follows from
Lemma 2 (ii).
Theorem 4.1. For a nonsingular G, the DPSUF is solvable if and only if there
exists a Gd,min which has no pole-zero cancellation with G at any γ ∈ Γ ⊂ C+.
Proof. Sufficiency: By the definition of Gd,min, GGd,min is decoupled. The theorem
will follow from Lemma 4.2 (i) if a Gd,min has no RHP pole-zero cancellation with
G at any ρ ∈ C+ but ρ /∈ Γ. Then, such ρ can be a pole or a zero of G but not
both. If ρ be a pole (resp. zero) of G, then ρ must be a zero (resp. pole) of Gd,min
for the cancellation to occur. Form Gd = 1(s−ρ)l
Gd,min (resp. (s − ρ)lGd,min) for
a sufficiently large integer l such that Gd no longer has any zero (resp. pole) at
ρ. This Gd has no pole-zero cancellation with G at ρ. Repeat this elimination of
pole-zero cancellation for all possible ρ and the resultant Gd will have no pole-zero
cancellation with G at ρ ∈ C+ but ρ /∈ Γ.
Necessity: It follows from Lemma 4.2 (i) that if the DPSUF is solvable, then
there is a Gd such that GGd is decoupled and meets
δγ(G) + δγ(Gd) = δγ(GGd), γ ∈ Γ. (4.4)
By Lemma 3, for Gd there exists a Gd,min such that Gd = Gd,minD and
δγ(Gd) = δγ(Gd,min) + δγ(D). (4.5)
Substituting (4.5) into (4.4) yields
δγ(G) + δγ(Gd,min) + δγ(D) = δγ(GGd,minD), (4.6)
implying thatG andGd,min have no pole-zero cancellation at γ ∈ Γ. This completes
the proof.
Chapter 4. Decoupling with Stability and Decoupling Control Design 95
Remark 4.2. The condition in Theorem 4.1 is much simpler to apply than that in
Wang (1992). The new condition here involves rational matrix manipulations only
and no complicated polynomial operations such as coprime factorization and skew
primeness checking are required.
4.2.4 Examples
In this subsection, several examples are presented to illustrate our results.
Example 4.1. (continued) (Wang, 1992) Consider the same G as in Example
4.1 with Γ = 1. For Gd,min1 in (4.1), one sees
GGd,min1 =
1s+1
1s+2
1(s−1)(s+1)
s(s−1)(s+2)
1 − s+1s+2
− s+2s(s+1)
1
=
s−1s(s+1)
0
0 1s+2
δ1(G) = 1, δ1(Gd,min1) = 0, and δ1(GGd,min1) = 0, δ1(G)+δ1(Gd,min1) 6= δ1(GGd,min1).
There is a pole-zero cancellation between G and Gd,min1 at γ = 1.
For Gd,min2 in (4.2), it follows that
GGd,min2 =
1s+1
1s+2
1(s−1)(s+1)
s(s−1)(s+2)
1s−1
− s+1(s+2)(s−1)
− s+2s(s+1)(s−1)
1s−1
=
1s(s+1)
0
0 1(s+2)(s−1)
,
δ1(G) = 1, δ1(Gd,min2) = 1, and δ1(GGd,min1) = 1, so that δ1(G) + δ1(Gd,min1) 6=δ1(GGd,min1). There is a pole-zero cancellation between G and Gd,min2 at γ = 1 too.
Thus, none of Gd,min can meet the condition in Theorem 4.1 and so the DPSUF
for this plant is not solvable.
Example 4.2. Let
G(s) =
ss+1
1s+2
1s(s+1)
1s(s+2)
.
Choose
Gd = G−1 =
s+1s−1
− s(s+1)s−1
− s+2s−1
s2(s+2)s−1
.
It is readily seen that the only common RHP pole and zero of G(s) is 1, lzγ = 0
and lpγ = 1. From Table 4.3, one can see that Gd,min is not unique. For
Gd,min1 =
s+1s−1
− s(s+1)s−1
− s+2s−1
s2(s+2)s−1
,
Chapter 4. Decoupling with Stability and Decoupling Control Design 96
Table 4.3. Search for Gd,min in Example 4.2
(l1, l2) K = Gd
(s− 1)l1 0
0 (s− 1)l2
δ1(K)
(0, 0)
s+1s−1
s(s+1)s−1
− s+2s−1
s2(s+2)s−1
1
(1, 0)
s+ 1 −s(s+1)s−1
−(s+ 2)s2(s+2)s−1
2
(0, 1)
s+1s−1
−s(s− 1)
− s+2s−1
s2(s+ 2)
2
(1, 1)
s+ 1 −s(s+ 1)
−(s+ 2) s2(s+ 2)
1
we have
GGd,min1 =
ss+1
1s+2
1s(s+1)
1s(s+2)
s+1s−1
− s(s+1)s−1
− s+2s−1
s2(s+2)s−1
=
1 0
0 1
,
δ1(G) = 0, δ1(Gd,min1) = 1, and δ1(GGd,min1) = 0, so δ1(G) + δ1(Gd,min1) 6=δ1(GGd,min1). There is a pole-zero cancellation between G and Gd,min1 at γ = 1.
For
Gd,min2 =
s+ 1 −s(s− 1)
−(s+ 2) s2(s+ 2)
,
we have
GGd,min2 =
ss+1
1s+2
1s(s+1)
1s(s+2)
s+ 1 −s(s+ 1)
−(s+ 2) s2(s+ 2)
=
s− 1 0
0 s− 1
,
δ1(G) = 0, δ1(Gd,min2) = 0, and δ1(GGd,min2) = 0, so δ1(G) + δ1(Gd,min2) =
δ1(GGd,min2). There is no pole-zero cancellation between G and Gd,min2 at γ = 1.
So the DPSUF of G(s) is solvable.
Chapter 4. Decoupling with Stability and Decoupling Control Design 97
Table 4.4. Search for Gd,min in Example 4.3
(l1, l2, l3) K = Gd
(s− 1)l1 0 0
0 (s− 1)l2 0
0 0 (s− 1)l3
δ1(K)
(−1,−1,−1)
1s−1
1s(s−1)
0
1s−1
1s−1
ss−1
0 0 1s−1
2
(−1,−1, 0)
1s−1
1s(s−1)
0
1s−1
1s−1
s
0 0 1
1
(−1, 0,−1)
− 1s−1
1s
0
1s−1
1 s
0 0 1
3
(−1, 0, 0)
1s−1
1s
0
1s−1
1 s
0 0 1
2
(0,−1,−1)
1 1s(s−1)
0
1 1s−1
ss−1
0 0 1s−1
3
(0,−1, 0)
1 1s(s−1)
0
1 1s−1
s
0 0 1
2
(0, 0,−1)
1 1s
0
1 1 ss−1
0 0 1s−1
2
(0, 0, 0)
1 1s
0
1 1 s
0 0 1
1
Chapter 4. Decoupling with Stability and Decoupling Control Design 98
Example 4.3. Let
G(s) =
ss−1
− 1s−1
ss−1
−1 1 −s0 0 s− 1
.
From Table 4.4, there are two Gd,min. For
Gd,min1 =
1 1s
0
1 1 s
0 0 1
,
we calculate
GGd,min1 =
ss−1
− 1s−1
ss−1
−1 1 −s0 0 s− 1
1 1s
0
1 1 0
0 0 1
=
1 0 0
0 s−1s
0
0 0 s− 1
,
δ1(G) = 1, δ1(Gd,min1) = 0, and δ1(GGd,min1) = 0, δ1(G)+δ1(Gd,min1) 6= δ1(GGd,min1).
There is a pole-zero cancellation between G and Gd,min1 at γ = 1.
For
Gd,min2 =
1s−1
1s(s−1)
0
1s−1
1s−1
s
0 0 1
,
we have
GGd,min2 =
ss−1
− 1s−1
ss−1
−1 1 −s0 0 s− 1
1s−1
1s(s−1)
0
1s−1
1s−1
s
0 0 1
=
1s−1
0 0
0 1s
0
0 0 s− 1
,
δ1(G) = 1, δ1(Gd,min2) = 1, and δ1(GGd,min2) = 1, δ1(G)+δ1(Gd,min2) 6= δ1(GGd,min2).
There is a pole-zero cancellation between G and Gd,min2 at γ = 1, too. So the DP-
SUF for the plant is not solvable.
4.2.5 Conclusions
In this section, a new necessary and sufficient condition for the solvability of the
decoupling problem with internal stability for unity output feedback systems with
square plants is given. The conditions are only concerned with Gd,min of the plant
G, thus it is easy to check.
Chapter 4. Decoupling with Stability and Decoupling Control Design 99
4.3 Decoupling Control Design via LMI Approaches
4.3.1 Introduction
The problem of decoupling of multivariable systems was first intensively studied by
three researchers, namely, Falb and Wolovich (1967) and Gilbert (1969), by means
of state feedback. These works initiated studies on structure of linear systems in the
1970s. Wang (1992) discussed the internal stability of the decoupled unity output
feedback system. Furthermore, the decoupling control has been tried to apply
to many practical industrial control problems for the last decades. Wang et al.
(1997) proposed a multivariable PID tuning method, where the objectives of control
design, as many applications require, are to make the closed-loop control system
decoupled and the resultant independent loops have good transient and steady state
accuracy in a usual sense of single-variable systems. Therefore, the controllers in
diagonal elements are designed with respect to the “equivalent processes” and
the controllers in off-diagonal elements are designed to achieve decoupling. If
the coupling between the loops is modest, simulations showed that the design
generally gives significant improvement in both system decoupling and diagonal
loop performances compared with conventional multi-loop design. However, if the
coupling is large, PID controller leads to bad decoupling performance and then
yielding poor loop performance. Thus, PID control may fail to yield satisfactory
performance even for simple multivariable models.
In this section, a new control scheme is proposed for the design of general
multivariable system via LMI technique aiming at obtaining good loop performance
and small loop coupling with a fixed order controller. Stability and Robustness are
analyzed and simulation shows that good control effects can be achieved.
4.3.2 Problem Formulation
Consider an IMC system shown in Fig. 3.1, where G = [Gij(s)] is an m×m plant,
G = [Gij(s)] is the model of the plant and Q = [Qij(s)] is the controller to be
designed so as to acquire not only good loop performance but also comparatively
Chapter 4. Decoupling with Stability and Decoupling Control Design 100
low couplings.
Let S and T be the sensitivity function and complementary sensitivity function
of the IMC system. Then we can obtain the sensitivity transfer matrix S = I−GQand complementary sensitivity transfer matrix T = GQ if the model is perfect. A
particular transfer function, say i-input/j-output, of S and T can be written as
Sij = LiSRj and Tij = LiTRj, where the matrices Li and Rj select the appropriate
channel. If the state space realization of T is (A,B,C,D), then the state space
realization of Tij is (A,BRi, LjC,LjDRi). Similarly, we can obtain the state space
realization of Sij. For decoupling, we wish that ||Tij||∞, i 6= j, is as small as
possible, i.e, the coupling of each loop is minimized. At the same time, we may
consider the loop performance by choosing weighting functions Wi and Vi of Tii
and Sii respectively. One notes that if the single loop is successful in attenuating
the effects of disturbance acting on the plant, Sii should be small, and if the output
is to be insensitive to measurement errors, we need small Tii. Moreover, if we want
to reduce control energy, we need small Tii, and if tracking references is important,
then Sii should be small. In short, if we wish good loop performance, then we need
both Tii and Sii to be small in some sense. However, we know Tii(s) + Sii(s) = 1.
We thus have an unavailable trade-off between different performance specifications.
In practice, the conflict can be resolved by making one small at some frequencies
and the other small at other frequencies. Usually, the solution is to make |Tii(s)|small at high frequency and |Sii(s)| small at low frequency (Maciehowski, 1989).
Using this method, each loop can be tuned to satisfy some low sensitivity and high
noise attenuation performance index. Moreover, we wish the static gain of each
loop is 1, so we set Tii(0) = 1. In short, the design can be described as
P1 : minQij
∑
i,j
aijεij +∑
i
(γTi+ γ
Si) (4.7)
subject to
||WiTii||∞ < γTi, (4.8)
||ViSii||∞ < γSi, (4.9)
||Tij||∞ < εij, i 6= j, (4.10)
Chapter 4. Decoupling with Stability and Decoupling Control Design 101
Tii(0) = 1. (4.11)
4.3.3 Controller Design via LMI
In order to design controller to achieve the given specification in the last section,
we need the following lemma.
Lemma 4.4. Given any system G = D + C(sI − A)−1B, ||G||∞ < γ and G is
asymptotically stable if and only if the following LMI in X is feasible:
ATX +XA XB CT
BTX −γI DT
C D −γI
< 0, (4.12)
where X > 0.
The above lemma is known as the Bounded Real lemma (Boyd,1994). Note that
ifA andB is fixed, the problem can be easily solved as a semi-definite programming.
Let Qrs = Nrs(s)/Drs(s) be the transfer function of the individual SISO com-
ponent systems in Q, where
Nrs = βn−1rs sn−1 + · · ·+ β1rss+ β0rs,
and
Drs = sn + αn−1rs sn−1 + · · ·+ α1rss+ α0rs.
Then the realization of Qrs can be express as
AQrs=
0 1 · · · 0 0
0 0 · · · 0 0...
.... . .
......
0 0 0 0 1
−α0rs −α1rs · · · −αn−2rs −αn−1rs
, BQrs=
0
0...
0
1
,
CQrs=[
β0rs β1rs · · · βn−1rs
]
, DQrs= 0.
The transfer function Tij can be written as Tij =∑m
l=1QilGlj, where Qil are the
individual SISO entries of Q.
Chapter 4. Decoupling with Stability and Decoupling Control Design 102
Suppose that the realization of Gij is AGij, BGij
, CGij, DGij
, we obtain the state
space realization of GijQrs in terms of state space matrices of Gij and Qrs.
Aij,rs =
AGij0
BQrsCGij
AQrs
, Bij,rs =
BGij
BQrsDij
,
Cij,rs =[
DQrsCGij
CQrs
]
, Dij,rs = DQrsDGij
.
Finally, we get the desired state space representation for Tij,
Aij =
Ai1,1j 0 · · · 0
0 Ai2,2j · · · 0...
.... . .
...
0 0 · · · Aim,mj
, Bij =
Bi1,1j
Bi2,2j
...
Bim,mj
,
Cij =[
Ci1,1j Ci2,2j · · · Cim,mj
]
, Dij =∑m
l=1Dim,mj .
From the above, we see that if the denominator of controller elements Qij is spec-
ified, all the coefficients of Qij appear in the Cij of Tij. In the same way, we can
also find a similar state space realization of Sij.
As mentioned above, we need to specify the denominator of the controller Q
as a filter, thus the problem can be easily solved using LMI toolbox. Moreover,
we wish the static gain of each loop is 1. Choose Q(0) = G(0)−1, the objective
can be achieved. That means, we can add the following linear constraint into the
optimization.m∑
j=1
gij(0)β0jiα0ji
= 1, i = 1, 2, · · · ,m.
The corresponding algorithm is proposed as follows:
Step 1: According to the performance requirement of each loop, choose appropri-
ate Wi and Vi for each loop.
Step 2: Choose filter
fij =1
(τis+ 1)n
as the denominator of each element in the controller matrix for each i.
Chapter 4. Decoupling with Stability and Decoupling Control Design 103
Step 3: Form the LMI constraints using the above mentioned method and solve
the the LMI optimization problem using the LMI toolbox.
Remark 4.3. Note that the introduction of the filter fij only aims at maintaining
the properness of controller Cii, so τi is chosen not to influence the decoupling
property of the system. It is only used to tune the performance and the robustness
of each loop.
Remark 4.4. Wi and Vi should be chosen in such a way that W−1i and V −1
i give an
upper bound for the sensitivity and complementary sensitivity functions respec-
tively (Maciehowski, 1989).
4.3.4 Stability and Robustness Analysis
The IMC system in Fig. 3.1 is referred to as the nominal case if G = G. It
follows from the known results (Morari and Zafiriou, 1989) that the IMC system
is nominally stable if both process G and the controller Q are stable. However,
in the real application, the model never represents the process exactly. Thus,
nominal stability is not enough and the robust stability has to be investigated.
As is common in robust analysis of IMC system, it is assumed that the nominal
IMC system is stable (i.e. both the G and Q are stable). Suppose that the actual
process G is also stable and is described by
G ∈∏
= G : σ(G−1(jω)[G(jω)− G(jω)]) = σ(∆m(jω)) ≤ lm(ω),
where lm(ω) is the bound on the multiplicative uncertainty. Suppose that the
following inequalities hold. (This is guaranteed by the control design presented in
last subsection.)
‖WiTii(jω)‖∞ ≤ γTi
i = 1, · · · ,m, (4.13)
and
‖Tij‖∞ ≤ εij i = 1, · · · ,m, j = 1, · · · ,m, j 6= i. (4.14)
Denote
δi(ω) = |γTiW−1
i (jω)|+√
Σmi=1Σ
mj=1,j 6=i|εij|2
Chapter 4. Decoupling with Stability and Decoupling Control Design 104
We can obtain the following robust stability theorem.
Theorem 4.2. Suppose that Wi(jω) 6= 0,∀ω, the nearly decoupled IMC system isrobustly stable if
δi(ω) ≤ l−1m (ω),∀ω. (4.15)
Proof. The nominal complementary sensitivity function of the IMC system T (jω)
can be written as:
T (jω) = Tdiag(jω) + Toff (jω),
where
Tdiag(jω) =
T11 0 · · · 0
0 T22 · · · 0...
.... . .
...
0 0 · · · Tmm
, Toff (jω) =
0 T12 · · · T1m
T21 0 · · · T2m...
.... . .
...
Tm1 Tm2 · · · 0
.
From the properties of singular value, we have
σ(T (jω)) = σ(Tdiag(jω) + Toff (jω))
≤ σ(Tdiag(jω)) + σ(Toff (jω))
≤ maxi|Tii(jω)|+ [Σm
i=1Σmj=1,j 6=i|Tij(jω)|2]
12 (4.16)
From (4.13) and (4.14), we have
|Tii(jω)| ≤ |γTiW−1
i |,∀ω (4.17)
and
|Tij(jω)| ≤ εij,∀ω. (4.18)
Combining the inequalities (10), (11) and (12), we obtain
σ(T (jω)) ≤ maxi|γ
TiW−1
i (jω)|+√
Σmi=1Σ
mj=1,j 6=i|εij|2,∀ω.
It follows from the known robust stability results (Morari and Zafiriou, 1989) that
the IMC systems is robustly stable if and only if
σ(T (jω)) = σ(G(jω)Q(jω)) < l−1m (jω),∀ω.
Thus, proof is completed.
Chapter 4. Decoupling with Stability and Decoupling Control Design 105
4.3.5 Simulation
In this section, two examples are given to show the effectiveness of the proposed
design.
0 20 40 60 80 100 120
0
0.5
1
0 20 40 60 80 100 120
−0.5
0
0.5
1
0 20 40 60 80 100 120−1
0
1
2
0 20 40 60 80 100 120−0.2
0
0.2
Time
y1
y2
u2
u1
t
t
t
t
Figure 4.1. Step Tests of the Plant in Example 4.4
( —— proposed method, - - - Luyben’s method)
Example 4.4. Consider the Wood/Berry binary distillation column plant with
first order Pade approximation (Wood and Berry, 1973).
G(s) =
12.8(1−0.5s)(1+16.7s)(1+0.5s)
−18.9(1−1.5s)(1+21s)(1+1.5s)
6.6(1−3.5s)(1+10.9s)(1+3.5s)
−19.4(1−1.5s)(1+14.4s)(1+1.5s)
. (4.19)
For desired loop performance, we choose
Chapter 4. Decoupling with Stability and Decoupling Control Design 106
W1(s) =s+ 30
s+ 1000,
W2(s) =s+ 10
s+ 1000,
V1(s) =s+ 10
s+ 0.6,
V2(s) =s+ 10
s+ 0.6,
and
f1 =1
(s+ 0.35)3,
f2 =1
(s+ 0.6)3.
Using the proposed method, we obtain the controller as
Q =
q11 q12
q21 q22
,
where
q11 =−0.3856s3 + 0.1115s2 + 0.087s+ 0.0067
(s+ 0.35)3,
q12 =−0.0122s3 + 0.2929s2 − 0.1535s− 0.0314
(s+ 0.6)3,
q21 =−0.1733s3 − 0.0781s2 + 0.0284s+ 0.0023
(s+ 0.35)3,
q22 =0.017s3 + 0.0106s2 − 0.2059s− 0.0218
(s+ 0.6)3.
The step tests is shown in Fig. 4.1 with the solid line. During step tests, two step
reference signals, r1 and r2, each with a step amplitude 1, are introduced at t = 0
and t = 80, respectively. The step tests using Luyben’s method (Luyben, 1986)
is also shown in Fig. 4.1 with dashed line. It can be seen that the response of
the proposed method achieves better performance and less loop interactions than
those of Luyben’s method.
Chapter 4. Decoupling with Stability and Decoupling Control Design 107
0 10 20 30 40 50 600
0.2
0.4
0.6
0.8
1
1.2
1.4
0 10 20 30 40 50 60−0.5
0
0.5
1
1.5
Time
Figure 4.2. Step Tests of the plant in Example 4.5
( —— proposed method, - - - Wang’s method)
10−3
10−2
10−1
100
101
102
103
10−4
10−3
10−2
10−1
100
101
102
a1
a2
a3
Figure 4.3. Robust Stability bound in Example 4.5
(a1: δ−11 (ω), a2: δ−12 (ω) a3: lm(ω) )
Chapter 4. Decoupling with Stability and Decoupling Control Design 108
0 10 20 30 40 50 600
0.2
0.4
0.6
0.8
1
1.2
1.4
0 10 20 30 40 50 60−0.5
0
0.5
1
1.5
Time
Figure 4.4. Step response for perturbed process in Example 4.5
( —— proposed method, - - - Wang’s method )
Example 4.5. To demonstrate the effectiveness of the proposed design, we
consider the plant adopted in Palmor et al. (1993):
G(s) =
0.5(0.1s+1)2(0.2s+1)2
−1(0.1s+1)(0.2s+1)2
1(0.1s+1)(0.2s+1)2
2.4(0.1s+1)(0.2s+1)2(0.5s+1)
. (4.20)
For desired loop performance, we choose
W1(s) =s+ 0.01
s+ 4,
W2(s) =s+ 0.01
s+ 1,
V1(s) =s+ 3
s+ 0.02,
V2(s) =s+ 3
s+ 0.02,
and
f1 =1
(s+ 0.7)3,
Chapter 4. Decoupling with Stability and Decoupling Control Design 109
f2 =1
(s+ 0.8)3.
Set a12 = a21 = 100, using the proposed LMI method, we can obtain the controller
as
Q =
q11 q12
q21 q22
,
where
q11 =0.7916s2 + 1.1940s + 0.3742
(s + 0.7)3,
q12 =−0.0345s3 + 0.7996s2 + 0.7429s + 0.2369
(s + 0.8)3,
q21 =−0.1649s3 − 0.5786s− 0.5755s− 0.1559
(s + 0.7)3,
q22 =−0.0388s3 + 0.3386s2 + 0.3736s + 0.1147
(s + 0.8)3.
The step tests is depicted in Fig. 4.2. During step tests, two step reference
signals, r1 and r2, each with a step amplitude 1, are introduced at t = 0 and t = 30.
The step response of Wang’s method (Wang et al., 1997) is also shown in Fig. 4.2.
Next, we will check the robustness of the proposed method. Let the uncertainty
of the process described by:
G ∈∏
= G : σ(G−1(jω)[G(jω)− G(jω)]) = σ(∆m(jω)) ≤ lm(ω),
where
∆m(s) =1.5s
s+ 4I2×2,
and I is an unity matrix. We can obtain
lm(ω) =1.5ω
√16 + ω2
16 + ω2
The resultant γT1
= 0.0262, γT2
= 0.6129, ε21 = 3.7096e − 006 and ε12 = 0.0074.
The frequency curves of lm(ω), δ−11 (ω) and δ−12 (ω) are shown in Fig. 4.3. The step
responses of the perturbed process are also given in Fig. 4.4.
Remark 4.5. From the simulation results, the proposed decoupling design avoids
overshoot but leads to a sluggish response. Moreover, it reduces the coupling of
Chapter 4. Decoupling with Stability and Decoupling Control Design 110
each loop. The gain of proposed controllers is smaller compared to that of the
Wang’s controller (Wang et al., 1997). The proposed design is more conservative
with the uncertain and disturbance, so it causes sluggish response.
4.3.6 Conclusions
A method for IMC controller design with decoupling has been proposed to gain
good loop performance and largely remove the effect of process coupling simultane-
ously for multi-variable plants. Two examples are presented with the demonstrated
effectiveness of the proposed method.
Chapter 5
Fuzzy Modelling and Control for
F-16 Aircraft
5.1 Introduction
In this chapter, the fuzzy modelling and controller design for F-16 aircraft is dis-
cussed. A TS fuzzy modelling and two control design approaches are developed.
5.1.1 F-16 Aircraft and Control
F-16 aircraft is designed for all electric control in 1970s. The classical flight con-
trol technology can be used via local linearization method. However, the aircraft
model is inherently nonlinear, exact input-output linearization is often used to
control specific output variable sets in nonlinear flight control problems (Lane and
Stengel, 1988; Krsti and Tsiotras, 1999). Feedback linearization is a theoretically
established and widely used method in controlling nonlinear systems. Unfortu-
nately, this direct application of feedback linearization requires the second or third
derivatives of uncertain aerodynamic coefficients and does not guarantee internal
stability for non-minimum phase systems.
Another approach to designing flight control laws with feedback linearization
is to separate the flight dynamics into fast and slow dynamics by using timescale
properties (Menon et al., 1987; Snell et al., 1992). This method allows the con-
111
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 112
troller design process to be performed without state transformation. Because each
separated subsystem is square, the number of control inputs is equal to the number
of states. The design process of this method can be divided into two steps. In the
outer loop, the controller for the slow states α, β, and φ is designed to facilitate
tracking of the given commands by assuming that the fast states p, q, and r are
control inputs, which achieve their commanded values instantaneously. With the
slow states controller designed in the outer loop, a separated inner-loop controller
is designed to make the fast states p, q, and r follow the outer loop’s control in-
put trajectories using the real control inputs: aileron, rudder, and elevator. This
method can be justified only if there is sufficient timescale separation between the
inner- and outer-loop dynamics because the fast states p, q, and r are used as
control inputs in the outer-loop system. Hence p, q, and r in the inner loop should
be much faster than the states α, β, and φ in the outer loop. The stability of
this timescale separation approach may be analyzed by the singular perturbation
theory. However, in most nonlinear flight control research, the gain of the inner-
loop controller is set much larger than that of the outer-loop controller, and it is
assumed that the aircraft dynamics satisfies this property. This, therefore, does
not guarantee closed-loop stability.
Schumacher and Khargonekar analyzed theoretically the stability of the flight
control system with the two-timescale separation assumption (Schumacher and
Khargonekar, 1998). Using the Lyapunov theory, they determined the minimal
inner-loop gain guaranteeing closed-loop stability. However, this approach is so
complicated and conservative that the calculated value of minimal inner-loop gain
is too large to be applied in the flight controller. It may excite un-modelled dy-
namics or saturate the control inputs and, therefore, cause robustness problems.
Another difficulty related to the application of feedback linearization to a flight
control system is that a complete and accurate aircraft dynamic model including
aerodynamic coefficients is required. It is difficult to identify accurately aerody-
namic coefficients because they are nonlinear functions of several physical variables.
An inverse optimal design is developed by Sepulchre et al. (1997) to derive the
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 113
optimal stabilizing controller. This approach was first proposed by Kalman (1964)
to establish the gain and phase margins of linear quadratic regulators and was
recently revised by Freeman and Kokotovic (1996) to develop a design methodology
of robust nonlinear controllers.
The control problem of a F-16 aircraft model has been addressed by many re-
searchers for the purpose of the control of spacecraft and aircraft (Debs and Athans,
1969; Dabbous and Ahmed, 1982; Porcelli, 1968; Sohoni and Guild, 1970; Vadali et
al., 1984). Also, there have been several works that consider performance indices
such as time and/or fuel in the formulation of the optimal control problems (Athans
et al., 1963; Freeman and Kokotovic, 1970; Junkins et al., 1981; Etter, 1989; Bil-
imoria and Wie, 1991; Scrivener and Thomson, 1994). These studies have mainly
addressed the regulation problem for the angular velocity subsystem and for some
quadratic costs (Athans et al., 1963; Windeknecht, 1963; Kumar, 1965; Tsiotras
et al., 1996). Recently, the attitude control problem of the complete system that
includes the dynamics as well as the kinematics has been investigated by many
researchers. Carrington and Junkins (1986) have used a polynomial expansion ap-
proach to approximate the solution to the Hamilton-Jacobin-Bellman (HJB) equa-
tion. Rotea et al. (1998) have shown that, for some special cases of performance
outputs, Lyapunov functions that include a logarithmic term in the kinematic pa-
rameters result in linear controllers with a finite quadratic cost. For the general
quadratic cost, they have also presented sufficient conditions that guarantee the
existence of a linear, suboptimal, stabilizing controller. Tsiotras et al. (1996) has
derived a new class of globally asymptotically stabilizing feedback control laws for
the complete attitude motion of a nonsymmetric rigid body and has also presented
a family of exponentially stabilizing optimal control laws for the complete system.
Tsiotras (1997), by using the natural decomposition of the complete system into
its kinematics and dynamics subsystems and the inherent passivity properties of
these two subsystems, has presented a partial solution to the optimal regulation
of the symmetry axis of a spinning rigid body. Bharadwaj et al. (1998) have de-
rived a couple of new globally stabilizing attitude control laws using the inverse
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 114
optimal approach of Freeman and Kokotovic (1996) where minimal, exponential
coordinates are used to represent the kinematic equations. To design a stabiliz-
ing control law for F-16 aircraft, the method of backstepping (Krsti et al., 1995)
can be considered, which was used by Sontag and Sussmann (1988) for the first
time to design feedback controls laws for an underachieved rigid body. Tsiotras
and Longuski (1994) have employed this method for the attitude stabilization of
an axis symmetric spacecraft with two control torques. In their design, a con-
trol Lyapunov function along with a stabilizing controller is derived and thus the
stabilization of closed-loop system is guaranteed.
5.1.2 Objective of the Design
In this study, we consider the complete attitude motion of a rigid spacecraft
(Shuster, 1963). Moreover, both stabilizing and attitude tracking controller are
needed to be developed based on a nonlinear F-16 model.
Since Takagi and Sugeno (1985) opened a new direction of research in the area
of control by introducing the Takagi-Sugeno (TS) fuzzy model, there have been
several studies concerning the systematic design of stabilizing fuzzy controllers
(Tanaka and Sugeno, 1992; Tanaka et al., 1996b; Wang et al., 1996; Tanaka et
al., 1996a; Wang and Tanaka, 1996; Thathachar and Viswanath, 1997; Tanaka et
al., 1997; Tanaka et al., 1998). In this thesis, we propose two design procedures
yielding the stabilizing controller for the nonlinear system described by a TS fuzzy
model. In the TS fuzzy model (Takagi and Sugeno, 1985), the overall system is
described by several fuzzy IF-THEN rules, each of which represents a local linear
state equation x = Aix+Biu. To derive the first stabilizing controller, we employ
the Lyapunov stabilizing theory and LMIs method. We show that the parameters
of the stabilizing controller can be found by solving an LMI problem. The LMI
formulation of the controller synthesis problems is of great practical value because
it can be solved by using reliable and efficient convex optimization techniques
(Boyd et al., 1994), for example, the LMI Control Toolbox of MATLAB (Gahinet
et al., 1995). Another method is based on gain scheduling, LQR optimal design
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 115
and TS fuzzy modelling. The resulting controller is a fuzzy controller.
Our proposed methods are based on the design of the control law for the TS
fuzzy model. To the author’s best knowledge, the proposed approach is the first
attempt to design both stabilizing and tracking control with application to the
complete attitude motion of a rigid spacecraft.
5.1.3 Organization of the Chapter
The rest of this chapter is organized as follows. First, the F-16 aircraft model is
presented. Next, the modelling of TS fuzzy system of F-16 aircraft is described.
Further, synthesis of the stabilizing and tracking controllers for the TS fuzzy sys-
tems is considered. The controller design is studied with two methods, namely,
Lyaponov based control and gain scheduling control. Finally, the discussion and
conclusions are given.
5.2 F-16 Aircraft Model
5.2.1 Need for Modelling
Model building is a fundamental process. An aircraft designer who has a mental
model of the type of aircraft, uses physical models to gather wind-tunnel data,
and designs with mathematical models that incorporate the experimental data
(Taeyoung and Youdan, 2001). The modelling process is often iterative; a math-
ematical model based on the laws of physics will suggest what experimental data
should be taken, and the model may then undergo considerable adjustment in or-
der to fit the data. In building the mathematical model we recognize the onset
of the law of diminishing returns and build a model that is good enough for our
purposes but has known limitations. Some of the limitations of the models involve
uncertainty in the values of their parameters. Later, we attempt to characterize
this uncertainty mathematically and allow for it in control system design.
Because of the high cost of building and flight testing a real aircraft, the impor-
tance of aircraft mathematical models goes far beyond control system design. The
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 116
mathematical model is used, in conjunction with computer simulation, to evalu-
ate the performance of the prototype aircraft and hence improve the design. It
can also be used to drive training simulators, to reconstruct the flight conditions
involved in accidents, and to study the effects of modifications to the design. Fur-
thermore, other mathematical models are used in all aspects of the aircraft design
(e.g., structural models for studying stress distribution and predicting fatigue life).
In the aerospace industry it is necessary for a wide range of specialists to work
together, thus flight control engineers must be able to work with the aerodynam-
icists as well as with structural and propulsion engineers. Each have some un-
derstanding of the terms and mathematical models used by the other. This is
becoming increasingly important as designers seek to widen aircraft performance
envelopes by integrating many parts of the whole design process. Furthermore, at
the prototype stage the control system designer must work closely with the test
pilots to make the final adjustments to the control systems. This may take many
hours of simulator time and flight test, and the flight controls engineer involved in
the final stages of the design is also likely to be a qualified pilot.
5.2.2 Modelling Method
The aerodynamic forces and moments on an aircraft are produced by the rela-
tive motion with respect to the air and depend on the orientation of the aircraft
with respect to the airflow. In a uniform airflow these forces and moments are
unchanged after a rotation around the free-stream velocity vector. Therefore, only
two orientation angles (with respect to the relative wind) are needed to specify the
aerodynamic forces and moments. The angles that are used are the angle of attack
(alpha) and the sideslip angle (beta). They are known as the aerodynamic angles
and will now be defined by means of coordinate rotations in three dimensions. Note
that the aerodynamic force and moments are also dependent on angular rates, but
for the moment we are concerned only with their dependence on orientation.
The angle of attack usually specified in the aerodynamic data for an aircraft is
measured with respect to a fuselage reference line and denoted by αflr. We shall
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 117
assume that our aircraft body-fixed axes are aligned with this fuselage reference
line; otherwise, the alpha used in the body-axes equations of motion would differ
by a constant from the alpha used in the aerodynamic data base.
Figure 5.1. Definition of aircraft axes and angles.
Fig. 5.1 shows an aircraft with the relative wind on its side (i.e., sideslipping)
and with the conventional right-handed (forward, starboard, and down) set of
body-fixed axes illustrated. The relative wind vector is equal and opposite to
the cg relative velocity vector vB used in the equations of motion. The angles
of attack and sideslip are defined by performing a plane rotation about the body
y-axis, followed by a plane rotation about the new z-axis, such that the final x-axis
is aligned directly into the relative wind. The first rotation defines the stability
axes, and the angle of attack is the angle between the body-fixed x-axis and the
stability x-axis. Alpha is positive if the rotation about the body-fixed y-axis was
negative; thus a positive alpha is shown in the figure. The second rotation leads to
a set of wind axes, and the sideslip angle is the angle between the stability x-axis
and the wind x-axis. Beta is positive if the rotation about the stability z-axis was
positive, and a positive beta is shown in the figure.
An aircraft of conventional shape must fly more or less directly into the appar-
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 118
ent wind in order to have low drag; therefore, beta is usually very small in steady
fight. Alpha must be large enough to generate the required lift but is also usually
quite small. Therefore, although the stability axes and wind axes have a variable
orientation depending on the flight condition, they essentially point forward, star-
board, and down, the same as the body-fixed axes. Note that technically all three
sets of axes are “body axes” but only one is body-fixed; we shall drop the term
“fixed” and simply refer to them as body, stability,and wind axes.
A left-handed wind-axes system, aligned backwards, and “up” relative to the
aircraft, has often been used in the past for wind tunnel data. Lift L, drag D, and
cross-wind force C were defined naturally in these axes as the aerodynamic force
components along the respective positive axes. In the right-handed wind axes that
we have defined, we use −L and −D for the z and x force components, and define
a component Y for the aerodynamic side-force measured along the positive y-axis.
The symbol Y is also commonly used for the aerodynamic side-force component
along the body y-axis, but it will be clear from the context which axes are being
used.
5.2.3 Workable Model
The body-fixed axes, nonlinear equations of motion for an aircraft over a flat Earth
are given by (Stevens and Frank, 1992)
V =cosα cos β
m[T + Fx] +
sin β
m[Fy] +
sinα cos β
m[Fz]
+g[− cosα cos β sin θ + sin β sinφ cos θ + sinα cos β cos θ], (5.1a)
α = −p cosα tan β + q − r sinα tan β − sinα
mV cos β[T + Fx]
+cosα
mV cos βFz +
g
V cos β[sinα sin θ + cosα cosφ cos θ], (5.1b)
β = p sinα− r cosα− cosα sin β
mV[T + Fx] +
cos β
mV[Fy]−
sinα cos β
mV[Fz]
+g
V[cosα sin β sin θ + cos β cos θ sinφ− sinα sin β cosφ cos θ], (5.1c)
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 119
p = I2pq + I1qr + I3L+ I4N, (5.2a)
q = I5pr − I6(p2 − r2) + I7M, (5.2b)
r = −I2qr + I8pq + I4L+ I9N, (5.2c)
φ = p+ tan θ(q sinφ+ r cosφ), (5.3a)
θ = q cosφ− r sinφ, (5.3b)
ψ =q sinφ+ r cosφ
cos θ, (5.3c)
where the moments of inertia Ii, i = 1, 2, ..., 9, are defined as follows:
I1 =Iz(Iz − Iy) + I2xz
IxIz − I2xz, I2 =
Ixz(Ix − Iy + Iz)
IxIz − I2xz, I3 =
IzIxIz − I2xz
, I4 =Ixz
IxIz − I2xz,
I5 =Iz − IxIy
, I6 =IxzIy, I7 =
1
Iy, I8 =
Ix(Ix − Iy) + I2xzIxIz − I2xz
, I9 =Ix
IxIz − I2xz.
It is assumed that the aerodynamic forces and moments are expressed as functions
of angle of attack, sideslip angle, angular rates, and control surface deflection. For
example, Fx, L and the other symbols are expressed as follows:
Fx = [Cx(α, δe) + (cq/2V )Cxq(α)]qS, (5.4a)
Fy = [Cy(β, δα, δr) + (bp/2V )Cyp(α) + (br/2V )Cyr ]qS, (5.4b)
Fz = [Cz(α, β, δe) + (cq/2V )Czq(α)]qS, (5.4c)
L = [Cl(α, β) + Clδα(α, β)δα + Clδr
(α, β)δr + (p/2V )Clp(α) + (r/2V )Clr(α)]qSb, (5.5a)
M = [Cm(α, δe) + (q/2V )Cmq(α)]qSc (5.5b)
N = [Cn(α, β) + Cnδα(α, β)δα + Cnδr
(α, β)δr + (p/2V )Cnp(α) + (r/2V )Cnr(α)]qSb,(5.5c)
For the F-16 aircraft, the force coefficients Cx, Cxq , Cy, Cyp , Cyr , Cz, Czq and mo-
ment coefficients Cl, Clδα, Clδr
, Clp , Clr , Cm, Cmq, Cn, Cnδα
, Cnδr, Cnp , Cnr and other
necessary parameters can be obtained from Stevens and Frank (1992). The control
input are thrust T , elevator angle (el) δa, aileron angle (ail) δa and rudder angle
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 120
(rdr) δr. The nine states are speed V , angle of attack (alpha) α, sideslip angle
(beta) β, roll rate p, pitch rate q, yawing rate r, roll angle (phi) φ, pitch angle
(theta) θ and yaw angle (psi) ψ.
5.3 TS Fuzzy Modelling
From the development of F-16 aircraft nonlinear model in the last section, we
observe that the nonlinear model is rather complex and coefficients can be only
obtained from the look-up tables. Thus, the control design for such model is quite
difficult. We need a simpler model which is easier to be used for control design.
This section reviews the Takagi and Sugeno Fuzzy modelling approach and develop
TS model of F-16 nonlinear model.
5.3.1 The Technique
The fuzzy model proposed by Takagi and Sugeno consists of several fuzzy IF-THEN
rules, each of which represents the local linear state equation of a nonlinear system.
In this report, we consider the fuzzy rules described as follows.
Rule i:
IF x1(t) is Mi1 and · · · and xn(t) is Min,
THEN
x(t) = Aix(t) +Biu(t), i = 1, 2, ..., N, (5.6)
Here, xi(t), i = 1, 2, ..., n, and Mij, i = 1, 2, ..., N , j = 1, 2, ..., n, are state variables
and fuzzy sets, respectively, and N is the number of IF-THEN rules; u ∈ Rm is the
input vector and Ai ∈ Rn×n and Bi ∈ Rn×m, i = 1, 2, ..., N . Following the usual
inference method of the TS fuzzy model, the state equation at time t is represented
in the form of weighted average along the trajectory x(t) ∈ Rn:
x(t) =N∑
i=1
wi(x(t))Aix(t) +Biu(t)/
N∑
i=1
wi(x(t)), (5.7)
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 121
In (5.7), the weight functions are defined as
wi(x(t)) =n∏
j=1
Mij(xj(t)),
where Mij(xj(t)) is the grade of xj(t) in the fuzzy set Mij. The weight functions
wi, which are nonnegative and measurable, usually satisfy
N∑
i=1
wi(x(t)) > 0, ∀ t > 0. (5.8)
Throughout this study, it is assumed that (5.8) always holds and that the vector
x(t) can be measured in real time. With the normalization of weight functions
hi(x(t)) = wi(x(t))/
N∑
i=1
wi(x(t)), i = 1, 2, ..., N, (5.9)
the state equation (5.7) can be written in the polytopic form
x(t) =N∑
i=1
hi(x(t))(Aix(t) +Biu(t)), (5.10)
where the normalized weights hi satisfy hi(x(t)) ≥ 0, i = 1, 2, ..., N , and
N∑
i=1
hi(x(t)) = 1, ∀ t ≥ 0.
Suppose that the truth model of a plant has the form
x = F (x, u). (5.11)
Expanding F by means of a Taylor series around (x0, u0) yields
x = F (x0, u0) +∂F
∂x|x=x0,u=u0(x− x0) +
∂F
∂u|x=x0,u=u0(u− u0) (5.12)
The point (x0, u0) is called an equilibrium point of (5.11) if at (x0, u0) we have
x = F (x0, u0) = 0. Let δx = x − x0 and δu = u − u0, the linearized model about
the equilibrium (x0, u0) is obtained by neglecting higher order terms. Observing
that for F (x0, u0) = 0, the linearized model has the form
d
dtδx = Aδx+Bδu, (5.13)
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 122
where A = ∂F∂x|x=x0,u=u0 and B = ∂F
∂u|x=x0,u=u0 . We start with generating linear
local models describing plant’s behavior about the equilibrium state (x0, u0) =
(0, 0), we obtain the first linear model by using the above described linearization
technique. The resulting model is given by x = A1x + B1u. Next, we need to
construct local linear models describing plant’s behavior at the remaining operating
points. Suppose that x = xj is the next operating state of interest. The result
of linearization of a nonlinear model using Taylor series expansion operating at a
non-equilibrium point is an affine rather than linear model. Even if the operating
point is an equilibrium point, Taylor series linearization, in general, will not yield
a local model that is linear in x and u. Indeed, suppose that the operating point
(xj, uj) is an equilibrium point, the resulting linearized model is
d
dt(x− xj) = Aj(x− xj) +Bj(u− uj). (5.14)
We can represent model (5.14) in the form
x = Ajx+Bju− (Ajxj +Bjuj). (5.15)
The term (Ajxj + Bjuj) does not have to be equal to zero and, hence, model
(5.15) is not a linear model, but rather an affine model. However, we note that
Taylor’s linearization will yield a linear system in x and u if the equilibrium point
(x0, u0) = (0, 0). Suppose that we are given an operating state x0 that does not
have to be an equilibrium state. Our goal is to construct a linear model in x and u
that approximates the behavior of (5.11) in the vicinity of the operating state x0;
that is, we wish to find constant matrices A and B such that in a neighborhood of
x0
F (x, u) ≈ Ax+Bu (5.16)
and
F (x0, u0) = Ax0 +Bu0. (5.17)
Let A = [a1 a2 · · · an] and B = [b1 b2 · · · bn]. We wish to have
Fi(x0, u0) = aTi x0 + bTi u0 (5.18)
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 123
and
bi =∂Fi∂u|x=x0,u=u0 . (5.19)
where Fi is the ith row of F . Moreover, we also wish that
∂Fi∂x|x=x0,u=u0 ≈ ai. (5.20)
Using the method of Marcelo and Stanislaw (1999), we obtain that
ai =∂Fi∂x|x=x0,u=u0 +
Fi(x0, u0)− xT0∂Fi∂x|x=x0,u=u0
‖x0‖2x0, x0 6= 0. (5.21)
ai =∂Fi∂x|x=x0,u=u0 , x0 = 0.
5.3.2 TS Model of F-16
We now apply the proposed fuzzy modelling method to the F-16 aircraft modelling,
where the F-16 nonlinear model is described by (5.1a)-(5.3c). The model can be
written as
x = F (x, u),
y = Cx, (5.22)
where x = [V, α, β, p, q, r, φ, θ, ψ]′ and u = [T, δe, δa, δr]′. In the following, we will
present the details of constructing a F-16 TS model: 1) Selection of equilibrium
state of the nonlinear model. 2) Selection of the operating points around the
equilibrium state. 3) Computation of the linear models at the chosen operating
points. 4) Determination of the weighting function of the TS model.
It has been shown by Stevens and Frank (1992) that one of the equilibrium
states of the F-16 aircraft is
[x0, u0] = [502 0.03936 0 0 0 0 0 0.03936 0.1485, 9.6427− 1.931 0.0 0.0]′
For the convenience of analysis and simulation, we move this point to the original
point x0 = 0. Substituting [x, u] = [x, u] + [x0, u0] into (5.22), we have a new
nonlinear model with the equilibrium point [x0, u0] = [0, 0].
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 124
From our extensive simulation, we choose nine operating points to build the TS
fuzzy model. If more points are chosen, the error of the model does not decrease
rapidly and less points result in generally bad approximation. For the states of
the nonlinear F-16 system near the equilibrium point with α ∈ [−0.2, 0.2] andφ ∈ [−0.2, 0.2], the TS model constructed by the proposed method is described by
x =9∑
i=1
hi(α, φ)(Aix+Biu), (5.23)
where hi(α, φ) = MαiMφi . Mαi and Mφi are the fuzzy membership functions of α
and φ respectively. In our simulation, the nine operating points evenly distributed
in the state space. That is, we choose
(α1, φ1) = (−0.2,−0.2),
(α2, φ2) = (−0.2, 0),
(α3, φ3) = (−0.2, 0.2),
(α4, φ4) = (0,−0.2),
(α5, φ5) = (0, 0),
(α6, φ6) = (0, 0.2),
(α7, φ7) = (0.2,−0.2),
(α8, φ8) = (0.2, 0),
(α9, φ9) = (0.2, 0.2)
as the nine operating points.
At each operating point, we can obtain the linear model at these operating
points from (5.19) and (5.21). For example, at the operating point (α, φ) =
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 125
(−0.2,−0.2), from (5.21), we get
A1 =
-3.0489 0.0003 -0.9873 0 0 -3.7977 0 0
0 -1.3815 -0.0029 0.0009 -1.1356 0 -0.0335 0
0.0312 0.0025 -0.5181 0 0 9.7538 0 0
0 2.5473 0 -0.0832 191.1245 -6.3862 -80.3479 -31.5328
0 0.9639 0 0.0003 -1.4637 0 0.1846 -0.0127
-0.1609 0 -0.9798 0.0000 0.0004 -0.2779 0.0632 0.0005
1.0000 -0.0078 0.0386 0 0 0 0 0
0 0.9801 0.1987 0 0 0 0 0
,
and from (5.19), we obtain
B1 =
0 0 -0.5983 0.1588
0 -0.1925 0 0
0 0 -0.0337 -0.0675
0 0.4584 0 0
0 -0.0020 0 0
0 0 0.0003 0.0008
0 0 0 0
0 0 0 0
.
Similarly, we can obtain the linear models (Ai, Bi), i = 2, · · · , 9, at each operating
point.
For weightings, we select triangle function as the fuzzy membership functions,
which are
Mα1 =
1 if α ≤ −0.22.5 ∗ α− 0.5 if − 0.2 < α < 0.2
0 if α ≥ 0.2
Mα2 =
0 if α ≤ −0.25 ∗ α + 1 if − 0.2 < α ≤ 0
−5 ∗ α + 1 if 0 < α < 0.2
0 if α ≥ 0.2
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 126
Mα3 =
0 if α ≤ −0.22.5 ∗ α + 0.5 if − 0.2 < α < 0.2
1 if α ≥ 0.2
Mφ1 =
1 if φ ≤ −0.22.5 ∗ φ− 0.5 if − 0.2 < φ < 0.2
0 if φ ≥ 0.2
Mφ2 =
0 if φ ≤ −0.25 ∗ φ+ 1 if − 0.2 < φ ≤ 0
−5 ∗ φ+ 1 if 0 < φ < 0.2
0 if φ ≥ 0.2
Mφ3 =
0 if φ ≤ −0.22.5 ∗ φ+ 0.5 if − 0.2 < φ < 0.2
1 if φ ≥ 0.2
The function is depicted in Fig. 5.2.
−0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.30
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
alpha(theta)
memb
ership
funct
ion
Figure 5.2. Fuzzy triangle membership functions
The complete model with all the data is given in Appendix A.
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 127
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.50
20
40
60
80
100
120
140
alpha(rad)
error
( —– δFLS, - - - δFL )
Figure 5.3. Approximate error of TS-fuzzy and linear model (different α)
5.3.3 Model Validation
To show the effectiveness of the obtained TS fuzzy model to approximate the
original nonlinear model, we compare the error between the TS fuzzy model and
linear model. The results are shown in Fig. 5.3 (with different α) and Fig.
5.4 (with different φ), where δFTS = ‖F (x, u) − FTs‖2 and δFL = ‖F (x, u) −
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.50
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
phi(rad)
error
( —– δFLS, - - - δFL )
Figure 5.4. Approximate error of TS-fuzzy and linear model (different φ)
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 128
FL‖2. In Fig. 5.3, we obtain FTS =∑9
i hi(α, φ0)(Aix + Biu) from TS fuzzy
model, FL = A5x + B5u is from the linear model at equilibrium point, where
x = [p0, q0, r0, V0, α, β0, φ0, θ0, ψ0]′ and u = [T0, δe0, δa0, δr0]
′, α ∈ [−0.2, 0.2]. In
Fig. 5.4, we have FTS =∑9
i hi(α0, φ)(Aix + Biu), FL = A5x + B5u, where
x = [p0, q0, r0, V0, α0, β0, φ, θ0, ψ0]′ and u = [T0, δe0, δa0, δr0]
′, φ ∈ [−0.2, 0.2]. It
is shown that the TS model has smaller error than the linear model around the
equilibrium point. At the interval α ∈ [−0.2, 0.2] , the maximal model error is
δFmaxTS = 2.0214 for TS fuzzy model and the maximal error is δFmax
L = 20.6833 for
linear model in Fig. 5.3. At the interval φ ∈ [−0.2, 0.2], the maximal model error
is δFmaxTS = 0.0001 for TS fuzzy model and the maximal error is δFmax
L = 0.0252
for linear model in Fig. 5.4.
5.4 Lyapunov Based Control
F-16 Aircraft is internally a nonlinear system. The nonlinear system control prob-
lem is generally difficult to solve. Moreover, even the nonlinear controller is found,
it is often very complicated and hard to be implemented. On the other hand, the
linear control is usually designed based on linear model. Because of the severe non-
linearity of the real F-16 aircraft, the linear model cannot represent the dynamics
of the system in large scale. Thus, we try to design controller based on TS fuzzy
model. From the simulation in last subsection, we observe that TS fuzzy model
has much less model error than linear model. In our approaches, Lyaponov based
controller is with linear controller and the gain schedule control is with fuzzy con-
troller. It is obvious that both of the controllers are of simpler structure than most
nonlinear controller. The method discussed in this section is based on Lyapunov
stability theory and called Lyapunov based control.
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 129
5.4.1 Stabilization
Suppose that the TS model of the nonlinear systems is described by
x =N∑
i=1
hi(Aix+Biu), (5.24)
y = Cx. (5.25)
The stability conditions for TS fuzzy system can be readily obtained via Lya-
punov stability theory. When u(t) = 0 for all t ≥ 0, the TS fuzzy system becomes
an input-free polytopic system given by
x(t) =N∑
i=1
hiAix(t). (5.26)
It is well known from the stability theory, an autonomous dynamic system
is stable if there exists a positive definite quadratic function V (x) = xTPx that
decreases along every nonzero trajectory of the system. A system having such a
Lyapunov function is called quadratically stable. In the polytopic system (5.26),
the time derivative of V along a nonzero trajectory x(·) is given by
dV
dt=
d
dt[xT (t)Px(t)] = xT (t)
N∑
i=1
hi[x(t)]ATi P + P
N∑
i=1
hi[x(t)]Ai
x(t)
=N∑
i=1
hi[x(t)]xT (t)AT
i P + PAix(t).
Then, we can see that the polytopic system (5.26) is quadratically stable if there
exists a symmetric matrix P satisfying the following inequalities:
P > 0, ATi P + PAi < 0, i = 1, 2, ..., N. (5.27)
Note that the left sides of these inequalities are all linear in the matrix variable P .To find P satisfying Eq.(5.27) or to determine if there does not exist such P is a
convex problem called the LMI feasibility problem. An LMI is any constraint of
the form
A(x) = A0 + x1A1 + · · ·+ xNAN < 0 (5.28)
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 130
where x=[x1, ..., xN ]T is the variable and A0, ..., AN are given symmetric matrices.
Since A(y) < 0 and A(z) < 0 implies A[(y + z)/2] < 0, the LMI (5.28) is a convex
constraint on the variable x. It is well known that LMI feasibility problems as well
as LMI-based minimization problems can be solved by interior-point algorithms
with polynomial time, and a toolbox of MATLAB that is dedicated to convex
problems involving LMIs is now available.
For the closed-loop system with state feedback controller
u = Kx, (5.29)
the TS fuzzy system can be described by
x(t) =N∑
i=1
hi(Aix(t) +Biu(t)) (5.30)
=N∑
i=1
hi(Ai +BiK)x(t). (5.31)
Then, we can see that the system (5.30) is quadratically stable if there exists a
symmetric matrix P and a state feedback controller K satisfying the following
inequalities:
P > 0, (Ai +BiK)TP + P(Ai +BiK) < 0, i = 1, 2, ..., N. (5.32)
Now multiplying the inequality on the left and right by P−1 and defining two new
variables Σ = P−1 and Ω = KP−1, we rewrite the conditions as
Σ > 0, ΣTATi + ΩTBT
i + AiΣ +BiΩ < 0, i = 1, 2, ..., N. (5.33)
The above conditions are LMIs with respect to variables Σ and Ω. We can find a
positive definite matrix Σ and Ω satisfying the LMIs or determinate that no such
Σ and Ω exist. A common P and the feedback gains K can be obtained as
P = Σ−1, K = ΩΣ−1 (5.34)
from the solutions of Σ and Ω.
In the control design, we often need to put the constraints on inputs and outputs
to obtain different performance in different conditions. Assume that the constraint
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 131
on input is ‖u‖2 ≤ µ, we obtain
uT (t)u(t) = xT (t)KTKx(t) ≤ µ2. (5.35)
Suppose that xT (0)Px(0) = xT (0)Σ−1x(0) ≤ 1 for t ≥ 0, if
1
µ2xT (t)KTKx(t) ≤ xT (t)Σ−1x(t), (5.36)
then (5.35) holds. Therefore, we have
1
µ2KTK − Σ−1 ≤ 0. (5.37)
By defining P = Σ−1 and Ω = KΣ, we can obtain
1
µ2ΩTΩ− Σ ≤ 0. (5.38)
Assume that the initial condition x(0) is known, the constraint ‖u‖2 ≤ µ is enforced
at all times t ≥ 0 if the LMIs
1 xT (0)
x(0) Σ
≥ 0 (5.39)
and
Σ ΩT
Ω µ2I
≥ 0 (5.40)
holds.
In a similar way, we can obtain the following result for output constraints
problem. Suppose the output of the TS system is
y = Cx. (5.41)
Assume that the initial condition x(0) is known, the constraint ‖y‖2 ≤ λ is enforced
as all times t ≥ 0 if the LMIs
1 xT (0)
x(0) Σ
≥ 0 (5.42)
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 132
and
Σ ΣCT
CΣ λ2I
≥ 0 (5.43)
holds.
Moreover, the speed of response is related to decay rate. The condition that
V (x(t)) ≤ −2αV (x(t)) for all trajectories is equivalent to
ATi P + PAi + 2αP < 0, i = 1, 2, ..., N. (5.44)
Let P = Σ−1, left multiplied and right multiplied by Σ, the corresponding LMI is
ΣATi + AiΣ + 2αΣ < 0, i = 1, 2, ..., N. (5.45)
5.4.2 Tracking
In control system design, we often need the process to track some reference signals.
In order to obtain zero steady tracking error, we often add a set of new states on
the system, which are the integrals of the error signals. If tracking error is not
zero, the differential of the augmented state, i.e, the error between the specified
signal and the reference signal, will no be zero, that is, the augmented system is
not in steady state or can not be stabilized. Thus, with the augmented states, the
specified state will track the reference signal at the steady state if the augmented
system is stabilized by a controller. The augmented system can be described by
x
xa
=
F (x, u)
x− xr
,
where xa(t) =∫ t
0(xr(τ) − x(τ))dτ , xr are reference signals, and F (x, u) is the
original system dynamics.
In aircraft control, we are often interested not only in regulating the state near
zero, but also in following a nonzero reference command signal. For example, we
may be interested in designing a control system for optimal step-response shaping.
This reference-input tracking or servo design problem is important in the design of
command augmentation system, where the reference signal may be, for instance,
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 133
the roll and pitch angle. According to the proposed design, we augment the F-
16 aircraft model with corresponding states in order to obtain zero steady-state
tracking error. Then we can use similar stabilizing control algorithm to stabilize
the augmented system. If the augmented system is stabilized, the differential of
the augmented stats approach to zero when time approaches infinity, that is, the
output is tracking the reference-input. Suppose that we wish the F-16 nonlinear
system to track the step signal of θr and φr, while keeping other states around
zero. The states are augmented for the nonlinear F-16 model. The two augmented
states θa(t) =∫ t
0(θr(τ)− θ(τ))dτ and φa(t) =
∫ t
0(φr(τ)− φ(τ))dτ are incorporated
into the nonlinear F-16 model, where θr and φr are reference signal of θ and φ
respectively. The nonlinear system can be written as
x
θa
φa
=
F (x, u)
θ − θr
φ− φr
Note that the augmented nonlinear model is different from the original F-16 model.
However, with the new model, the similar method can be employed to design the
tracking controller.
5.4.3 Simulation
We now apply the proposed method to F-16 control design. The detail of TS fuzzy
model with F-16 aircraft (5.23) is given in Appendix A. Without any input and
output constraints, we may solve the optimization problem with LMI constraint
(5.33) and find the stabilizing controller
K =
0 0 0 0 0 0 0 0
0 611.2 −0.1 −24.7 −2177.8 0 0.1 7706.3
22.7 −0.1 170.3 0 0.7 387 2276.4 −1.9−15.2 −0.1 338.5 0 0 −201.4 −507.5 −1.1
,
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 134
0 1 2 3−50
0
50
100
150
time(s)
rd
r(d
eg
)
0 1 2 3−200
0
200
400
600
time(s)
el(d
eg
)
0 1 2 3−15
−10
−5
0
5
time(s)
alp
ha
(d
eg
)
0 1 2 3−500
0
500
time(s)
ail(d
eg
)
0 1 2 3−15
−10
−5
0
5
time(s)
ph
i(d
eg
)
Figure 5.5. Lyapunov based stabilizing control
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 135
0 0.5 1 1.5 2−10
−5
0
5
time(s)
rd
r(d
eg
)
0 0.5 1 1.5 20
50
100
time(s)
el(d
eg
)
0 0.5 1 1.5 2−100
−50
0
time(s)
alp
ha
(d
eg
)
0 0.5 1 1.5 2−6
−4
−2
0
2
time(s)
ail(d
eg
)
0 0.5 1 1.5 2−150
−100
−50
0
time(s)
ph
i(d
eg
)
Figure 5.6. Linear stabilizing control
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 136
and the corresponding P matrix is
P =
0.0000 0.0000 −0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0001 −0.0000 −0.0000 −0.0002 −0.0000 0.0000 0.0008
−0.0000 −0.0000 0.0000 0.0000 0.0000 −0.0000 0.0000 −0.00000.0000 −0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 −0.00000.0000 −0.0002 0.0000 0.0000 0.0034 0.0000 0.0000 −0.00550.0000 −0.0000 −0.0000 0.0000 0.0000 0.0001 0.0001 −0.00000.0000 0.0000 0.0000 0.0000 0.0000 0.0001 0.0004 −0.00000.0000 0.0000 −0.0000 −0.0000 −0.0055 −0.0000 −0.0000 0.0130
.
The initial state is chosen as α = −0.2 rad (−11.46 degree) and φ = −0.2rad (−11.46 degree). Other initial states are all zero. The response of α, φ and
the control signals are shown in Fig. 5.5. For comparison, we also find the con-
troller using linear model on the equilibrium. The obtained controller using same
conditions gives
K =
0 0 0 0 0 0 0 0
−0.0073 5.9730 0.0039 −2.5628 −41.0560 −0.0010 −0.0025 162.8011
−1.6331 0.0046 0.3952 0.0015 0.0002 −16.9832 2.3450 0.0032
11.7954 0.0167 2.5714 0.0185 0.0034 139.4477 3.0502 0.0030
,
and the corresponding P matrix is
P =
0.0021 −0.0000 −0.0000 −0.0000 −0.0000 0.0001 0.0007 0.0000
−0.0000 0.0010 −0.0000 −0.0004 −0.0008 0.0000 0.0000 0.0023
−0.0000 −0.0000 0.0019 −0.0000 −0.0000 −0.0003 −0.0000 0.0000
−0.0000 −0.0004 −0.0000 0.0019 −0.0000 0.0000 −0.0000 −0.0014−0.0000 −0.0008 −0.0000 −0.0000 0.0039 0.0000 −0.0000 −0.00620.0001 0.0000 −0.0003 0.0000 0.0000 0.0018 0.0000 0.0000
0.0007 0.0000 −0.0000 −0.0000 −0.0000 0.0000 0.0021 0.0000
0.0000 0.0023 0.0000 −0.0014 −0.0062 0.0000 0.0000 0.0249
.
We can check that it really gives stabilizing control for the linear system lin-
earized at the equilibrium point. However, it cannot stabilize the nonlinear F-16
model. The response of α, φ and the control signals are shown in Fig. 5.6.
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 137
In the above simulation, we may notice that the proposed method can stabilize
the process while linear control fails. However, the control signal of the proposed
method is too large. In the following design, we wish to use constraints to limit
the input signal.
With the same TS fuzzy model of f-16 aircraft, we add the input constraint,
i.e., set µ = 100 and solve the LMIs feasible problem with constraints (5.33), (5.39)
and (5.40). The obtained controller is
K =
0 0 0 0 0 0 0 0
0.0003 4.9962 0.0072 −0.0330 −1.2235 0.0054 0.0031 11.2421
1.8132 0.0012 4.6381 0.0000 0.0001 −9.9594 10.6000 −0.0058−0.3218 0.0023 9.1266 0.0000 0.0032 −8.0655 −0.7151 −0.0071
,
and the corresponding P matrix is
P =
0.0006 0.0000 0.0007 −0.0000 −0.0000 −0.0004 0.0018 0.0000
0.0000 0.0029 0.0000 −0.0000 0.0001 −0.0000 0.0000 0.0048
0.0007 0.0000 0.0113 −0.0000 −0.0000 −0.0093 0.0043 0.0000
−0.0000 −0.0000 −0.0000 0.0000 0.0000 0.0000 −0.0000 −0.0001−0.0000 0.0001 −0.0000 0.0000 0.0135 −0.0000 −0.0000 −0.0104−0.0004 −0.0000 −0.0093 0.0000 −0.0000 0.1018 −0.0058 −0.00000.0018 0.0000 0.0043 −0.0000 −0.0000 −0.0058 0.0125 −0.00000.0000 0.0048 0.0000 −0.0001 −0.0104 −0.0000 −0.0000 0.0213
.
The simulation results are shown in Fig. 5.7. We may observe that the system
gives sluggish response. However, the magnitude of control signals is much smaller.
In attitude tracking design, the TS fuzzy model of the augmented system is
described by
x =9∑
i=1
hi(φ, θ)(Aix+Biu), (5.46)
where hi(φ, θ) = MφiMθi . Mφi and Mθi are the fuzzy membership functions of φ
and θ respectively. In our simulation, the nine operating points evenly distributed
in the state space. That is, we choose
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 138
0 2 4 6 8 10−0.4
−0.2
0
0.2
0.4
time(s)
rd
r(d
eg
)
0 2 4 6 8 10−0.5
0
0.5
1
time(s)
el(d
eg
)
0 2 4 6 8 10−15
−10
−5
0
5
time(s)
alp
ha
(d
eg
)
0 2 4 6 8 10−3
−2
−1
0
1
time(s)
ail(d
eg
)
0 2 4 6 8 10−15
−10
−5
0
5
time(s)
ph
i(d
eg
)
Figure 5.7. Lyapunov based stabilizing control with control signal constraints
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 139
(φ1, θ1) = (−0.15,−0.2),
(φ2, θ2) = (−0.15, 0),
(φ3, θ3) = (−0.15, 0.2),
(φ4, θ4) = (0,−0.2),
(φ5, θ5) = (0, 0),
(φ6, θ6) = (0, 0.2),
(φ7, θ7) = (0.15,−0.2),
(φ8, θ8) = (0.15, 0),
(φ9, θ9) = (0.15, 0.2)
as the nine operating points. At each operating point, we can obtain the lin-
ear model at these operating points from (5.19) and (5.21). For example, at the
operating point (φ, θ) = (−0.2,−0.2), from (5.21), we get
A1 =
-3.6722 0.0003 0.6741 0 0 -30.9113 0 0 0 0 0 0
0 -1.3859 -0.0029 0.0000 -2.4977 0 -0.0003 -0.0002 0 0 0 0
-0.0263 0.0025 -0.4986 0 0 9.4747 0 0 0 0 0 0
0 -0.6501 0 -0.0202 6.8833 -6.3521 0.0559 -31.9571 0 0 0 0
0 0.9048 0 -0.0002 -1.0285 0 0.0064 0.0047 0 0 0 0
0.0389 0 -0.9916 0.0000 -0.0000 -0.3315 0.0636 -0.0005 0 0 0 0
1.000 0.0221 -0.1089 0 0 0 0 0 0 0 0 0
0 0.9801 0.1987 0 0 0 0 0 0 0 0 0
0 -0.1999 0.9861 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 1.0000 -0.0000 0 0 0 0
0 0 0 0 0 0 -0.0000 1.0000 0 0 0 0
0 0 0 0 0 0 0 0 1.0000 0 0 0
,
and from (5.19), we obtain
B1 =
0 0 -0.7336 0.1310
0 -0.1824 0 0
0 0 -0.0327 -0.0643
0 0.1704 0 0
0 -0.0022 0 0
0 0 0.0003 0.0008
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
.
The resulting (Ai, Bi), i = 1, 2, · · · , 9, are given in Appendix B. Moreover, we select
triangle function in last section as the fuzzy membership functions, which are the
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 140
functions depicted in Fig. 5.2. The complete augmented model with all the data
is given in Appendix B.
In order to obtain better tracking speed, we incorporate the decay rate con-
straint into the controller design. In the simulation, we choose the delay rate
α = 0.0224. The obtained controller is
K = 103
0 0 0 0 0 0 0 0 0 0
0.0000 0.2643 0.0000 0.0002 −0.0516 0.0000 −0.0000 1.4067 −0.0000 0.8095
−0.0004 −0.0000 0.0488 0.0000 0.0000 0.0488 0.0090 −0.0002 −0.0088 −0.0002
0.0016 −0.0003 0.2624 0.0000 0.0000 0.3774 −0.0221 −0.0013 −0.0674 −0.0009
.
The step amplitude for θr and φr are 0.2 rad (11.46 degree) and 0.15 rad (8.59
degree) and the step time is at t = 0s and t = 100s respectively. The result
shows that the tracking problem is solvable by this scheme. For comparison, the
controller ia also designed on the basis of the linear model derived from the same
condition. It is shown that the controller based on the linear model fails to track
the desired trajectory. The simulation results are shown in Fig. 5.8 and Fig. 5.9.
In the simulation, we can observe that the Lyapunov based control can stabilize
the F-16 aircraft as well as tracking the specified altitude signals.
5.5 Gain Scheduling Control
In the last section, we proposed Lyapunov based methods for F-16 control. How-
ever, the method gives slow response in the altitude tracking control. we may
notice that the scheme only employ the linear controller and they inherently linear
control system except that the model is TS fuzzy model. In this section, we try to
use fuzzy controller and gain scheduling method to get faster response of tracking
control.
5.5.1 Gain Scheduled Linear Quadratic Regulator Design
Suppose that the TS model is described by
x(t) =N∑
i=1
hiAix(t) +Biu(t),
y(t) = Cx(t),
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 141
0 100 200 300−20
0
20
40
60
time(s)
rd
r(d
eg
re
e)
0 100 200 300−4
−3
−2
−1
0
time(s)e
l(d
eg
re
e)
0 100 200 300−0.1
0
0.1
0.2
time(s)
ph
i(ra
d)
0 100 200 300−30
−20
−10
0
10
time(s)
ail(d
eg
re
e)
0 100 200 3000
0.1
0.2
0.3
0.4
time(s)
the
ta(ra
d)
Figure 5.8. Lyapunov based tracking control
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 142
0 50 100 150 200−10
0
10
20
30
time(s)
rd
r(d
eg
re
e)
0 50 100 150 200−40
−20
0
20
40
time(s)
el(d
eg
re
e)
0 50 100 150 200−0.1
0
0.1
0.2
0.3
time(s)
ph
i(ra
d)
0 50 100 150 200−15
−10
−5
0
5
time(s)
ail(d
eg
re
e)
0 50 100 150 2000
0.1
0.2
0.3
0.4
time(s)
the
ta(ra
d)
Figure 5.9. Linear tracking control
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 143
and the controller is described by
u =N∑
i=1
hiKix.
Our objective in this section is to develop gain scheduled linear quadratic regulator
(LQR) fuzzy controller for the TS model to obtain desirable closed-loop response
characteristics. In the design, we can select performance criterion at different
design points, which reflect our concern with closed-loop stability and good time
response.
At each design point, we can obtain a linear time-invariant state-variable model
x = Aix+Biu, (5.47)
with x(t) ∈ Rn the state, u(t) ∈ Rm the control input. The controls will be state
feedbacks of the form
u = Kix, (5.48)
whereKi is anm×nmatrix of constant feedback coefficient to be determined by the
design procedure. The objective of state regulation for the aircraft is to drive any
initial condition error to zero, thus maintaining stability. This may be achieved by
selecting the control input u(t) to minimize a quadratic cost or performance index
(PI) of the type
Ji =1
2
∫ ∞
0
(xTQix+ uTRiu)dt, (5.49)
where Qi and Ri are symmetric positive semi-definite weighting matrices. The
definiteness assumption on Qi and Ri guarantee that Ji is nonnegative. The rel-
ative magnitudes of Qi and Ri may be selected to trade off requirements on the
smallness of the state against requirements on the smallness of the input. We can
observe that larger Ri penalizes the control signal, so that they will be smaller in
norm relative to the state vector. On the other hand, to make x(t) go to zero more
quickly with respect to time, we may select a larger Qi.
By substituting the control (5.48) into (5.47), the closed-loop system equations
are found to be
x = (Ai +BiKi)x = Acix, (5.50)
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 144
the performance index may be expressed in terms of Ki as
Ji =1
2
∫ ∞
0
xT (Qi +KTi RiKi)xdt, (5.51)
The design problem is now to select the gain Ki so that Ji is minimized subject
to the dynamical constraint (5.49). The dynamical optimization problem may be
converted into an equivalent static one that is easier to solve. Suppose that we can
find a constant, symmetric, positive semi-definite matrix Pi so that
d
dt(xTPix) = −xT (Qi +KT
i RiKi)x, (5.52)
then Ji may be written as
Ji =1
2xT (0)Pix(0)−
1
2limt→∞
xT (t)Pix(t). (5.53)
Assuming that the closed-loop system is asymptotically stable so that x(t) vanishes
with time, this becomes
Ji =1
2xT (0)Pix(0). (5.54)
If P satisfies (5.52), we may use (5.49) to see that
−xT (Qi +KTi RiKi)x =
d
dt(xTPix) = xTPix+ xT ...Pix = xT (AT
ciPi + PiAci)x.(5.55)
Since this must hold for all initial conditions, and hence for all state trajectories
x(t), we may write
ATciPi + PiAci +KT
i RiKi +Qi = 0. (5.56)
If Ki and Qi are given, Pi can be obtained by solving (5.56). We can also obtain
Ki = R−1i BT
i Pi. (5.57)
To obtain the output feedback gain minimizing the performance index, we need
to solve the two coupled equations (5.56) and (5.57). Once the performance index
weighting matrices Qi and Ri have been selected, the determination of the optimal
feedback gain Ki is a formal procedure relying on the solution of nonlinear coupled
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 145
matrix equations. For nonlinear control, the process changes with the operating
conditions and the control gains should be gain scheduled. Gain scheduling is a
nonlinear feedback of special type; it has a linear controller whose parameters are
changed as a function of operating conditions in a preprogrammed way. When
scheduling variables have been determined, the controller parameters are calcu-
lated at a number of operating points by using some suitable design method. The
controller is thus tuned for each operating condition. Notice that the stability
and performance of the system are typically evaluated by simulation. In this re-
port, we use gain scheduling method to design tracking control based on TS fuzzy
modelling. In this scheme, we set the fuzzy controller
K =N∑
i=1
hi(x)Ki(t), (5.58)
where each Ki is obtained from the LQR optimal design for each linear model in
TS modelling and hi(x) is got from the coefficients of the TS fuzzy model in section
(5.3).
5.5.2 Simulation
Now, we employed the proposed gain scheduling approach in F-16 aircraft design.
Consider the TS model (5.46) in last section and reference signals are given step
signals on φ and θ. For Qi = diag(1, 1E3, 1, 1, 1, 1, 1, 1, 1, 1E6, 1E7) and Ri =
diag(1, 1, 1, 1), using the gain scheduling optimal design, the fuzzy controller is
given in Appendix D (controller I) and the simulation results are shown in Fig.
5.10.
Moveover, we may try to track α, β and φ, the TS fuzzy model of this system
is described by
x =9∑
i=1
hi(α, φ)(Aix+Biu), (5.59)
where hi(α, φ) = MαiMφi . Mαi and Mφi are the fuzzy membership functions of α
and φ respectively. In our simulation, the nine operating points evenly distributed
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 146
0 2 4 6−2
0
2
4
6
time(s)
rd
r(d
eg
)
0 2 4 6−10
−5
0
5
time(s)
el(d
eg
)
0 2 4 60
5
10
time(s)
ph
i(d
eg
)
0 2 4 6−30
−20
−10
0
10
time(s)
ail(d
eg
)
0 2 4 60
5
10
15
time(s)
the
ta(d
eg
)
Figure 5.10. The gain scheduled tracking control with φ θ
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 147
0 5 10 15 20−50
0
50
time(s)
rd
r(d
eg
)
0 5 10 15 20−100
0
100
200
time(s)
el(d
eg
)
0 5 10 15 20−5
0
5
10
time(s)
alp
ha
(d
eg
)
0 5 10 15 20−10
0
10
20
time(s)
ail(d
eg
)
0 5 10 15 20−5
0
5
time(s)
be
ta(d
eg
)
0 5 10 15 20−20
0
20
40
60
time(s)
ph
i(d
eg
)
Figure 5.11. The gain scheduled tracking control I with α β and φ
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 148
0 5 10 15 20−200
−100
0
100
200
time(s)
rd
r(d
eg
)
0 5 10 15 20−100
0
100
200
time(s)
el(d
eg
)
0 5 10 15 20−5
0
5
10
time(s)
alp
ha
(d
eg
)
0 5 10 15 20−20
0
20
40
60
time(s)
ail(d
eg
)
0 5 10 15 20−0.05
0
0.05
time(s)
be
ta(d
eg
)
0 5 10 15 20−20
0
20
40
60
time(s)
ph
i(d
eg
)
Figure 5.12. The gain scheduled tracking control II with α β and φ
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 149
0 5 10 15 20−20
−10
0
10
20
time(s)
rd
r(d
eg
)
0 5 10 15 20−20
−10
0
10
20
time(s)
el(d
eg
)
0 5 10 15 20−5
0
5
10
time(s)
alp
ha
(d
eg
)
0 5 10 15 20−10
−5
0
5
10
time(s)
ail(d
eg
)
0 5 10 15 20−0.4
−0.2
0
0.2
0.4
time(s)
be
ta(d
eg
)
0 5 10 15 20−20
0
20
40
60
time(s)
ph
i(d
eg
)
Figure 5.13. The gain scheduled tracking control III with α β and φ
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 150
Figure 5.14. Lee’s Backstepping control
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 151
in the state space. That is, we choose
(α1, φ1) = (−0.1,−0.8),
(α2, φ2) = (−0.1, 0),
(α3, φ3) = (−0.1, 0.8),
(α4, φ4) = (0,−0.8),
(α5, φ5) = (0, 0),
(α6, φ6) = (0, 0.8),
(α7, φ7) = (0.1,−0.8),
(α8, φ8) = (0.1, 0),
(α9, φ9) = (0.1, 0.8)
as the nine operating points. At each operating point, we can obtain the linear
model at these operating points from (5.19) and (5.21). The resulting (Ai, Bi), i =
1, 2, · · · , 9, are given in Appendix E. Moreover, we select triangle function in last
section as the fuzzy membership functions, which are the function is depicted in Fig.
5.2. ForRi = diag(1000, 1000, 100, 100 andQi = diag(1, 1e9, 1, 1, 1, 1, 1, 1, 1e10, 1e9, 1e9),
the controller is given in Appendix E (controller I) and simulation results are given
in Fig. 5.11. Compared with the nonlinear control law of Taeyoung and Youdan
(2001) in Fig. 5.14, we achieve similar output performance. However, the ampli-
tude of the beta response is much larger than that of the nonlinear control. We
detune the weightings in our control design. For Ri = diag(1000, 1000, 100, 100
and Qi = diag(1, 1e8, 1, 1, 1, 1, 1, 1, 1e10, 1e15, 1e9), the simulation result is shown
in Fig. 5.12 and controller is given in Appendix E (controller II). We can ob-
serve that the maximal amplitude of beta is less than 0.04 deg which is much
smaller than around 4 deg in the original design in Fig. 5.11. The response of
alpha and theta has little change. Moreover, we redesign our controller to re-
duce the amplitude of the control signal. For Ri = diag(1000, 1000, 1000, 1000)
and Qi = diag(1, 1e9, 1, 1, 1, 1, 1, 1, 1e9, 1e12, 1e8), the simulation result is shown
in Fig. 5.13. We can observe that the amplitude of control signal is smaller and
the response of outputs is more sluggish.
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 152
Table 5.1. The φ tracking specifications of the two control methods
RT ST OV |δe|max |δa|max |δr|max
Lyaponuv based control 40 80 0.06 1.4388 1.9203 11.4588
Gain scheduling control I 0.9508 2.5799 0.0523 2.4171 1.6574 14.3380
Gain scheduling control II 0.3562 0.4550 0.0288 2.7179 8.2767 11.5618
5.6 Discussions and Conclusions
5.6.1 Comparative Studies
We now study performance of the altitude tracking control of the proposed two
approaches. The reference signals of φ and θ are chosen as step signals with altitude
0.1 rad (5.73 degree). For the Lypunov based control, the controller is given by
(5.47); for the gain scheduling control I and II, we can find the fuzzy controller
from Appendix D. The comparison of the specification index is given in Table. 5.1
and Table. 5.2 .
In gain scheduling control I, we chooseQi = diag(1, 1E3, 1, 1, 1, 1, 1, 1, 1, 1E6, 1E7)
and Ri = diag(1000, 1000, 1000, 1000), the responses of output and control signals
are shown in Fig. 5.15 and Fig. 5.16. The detail of response in the initial phase is
also given in Fig. 5.17.
In the second design, we try to get even better output performance. For
Qi = diag(1, 1E3, 1, 1, 1, 1, 1, 1, 1, 1E6, 1E7) and Ri = diag(1, 1, 1, 1), the result
of outputs and control signals are given in Fig. 5.18 and Fig. 5.19. The detail
of response in the initial phase is also given in Fig. 5.20. In this design, we can
observe that the output response are speeded up. Inevitably, the control signals
have larger overshoot and (or) undershoot in initial part.
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 153
0 5 10 15 20 25 300
1
2
3
4
5
6
7
time(s)
the
ta(d
eg
)
0 5 10 15 20 25 300
1
2
3
4
5
6
7
time(s)
ph
i(d
eg
)
Figure 5.15. Outputs of the two proposed methods
(gain scheduling —, Lyapunov based - - -)
One can see that the gain scheduling controller perform much better than Lya-
punov based controller. This is because that the former is nonlinear and can adapt
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 154
0 100 200 300 400 500 600 700−5
0
5
10
15
time(s)
rdr(
de
g)
0 100 200 300 400 500 600 700−3
−2
−1
0
1
time(s)
el(d
eg
)
0 100 200 300 400 500 600 700−2
−1
0
1
time(s)
ail(d
eg
)
Figure 5.16. Control signals of the two proposed methods
(gain scheduling —, Lyapunov based - - -)
to nonlinearity of F-16 aircraft, while the latter is of fixed gain controller and in
fact linear, and they can only ensure stability and tracking by design.
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 155
0 2 4 6−0.2
0
0.2
0.4
0.6
time(s)
rdr(
de
g)
0 2 4 6−0.6
−0.4
−0.2
0
0.2
time(s)
el(d
eg
)
0 2 4 60
5
10
time(s)
ph
i(d
eg
)
0 2 4 6−2
−1
0
1
time(s)
ail(d
eg
)
0 2 4 60
2
4
6
8
time(s)
the
ta(d
eg
)
Figure 5.17. Initial part of gain scheduling control
5.6.2 Conclusions
In this chapter, we have carried out and achieved the following:
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 156
0 5 10 15 20 25 300
1
2
3
4
5
6
7
time(s)
the
ta(d
eg
)
0 5 10 15 20 25 300
1
2
3
4
5
6
7
time(s)
ph
i(d
eg
)
Figure 5.18. Outputs of the two proposed methods
(gain scheduling —, Lyapunov based - - -)
(i) The problem to design both stabilizing and tracking controller for F-16
aircraft systems has been addressed via the TS fuzzy modelling approach. To our
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 157
0 100 200 300 400 500 600 700−5
0
5
10
15
time(s)
rdr(
de
g)
0 100 200 300 400 500 600 700−3
−2
−1
0
1
time(s)
el(d
eg
)
0 100 200 300 400 500 600 700−10
−5
0
5
time(s)
ail(d
eg
)
Figure 5.19. Control signals of the two proposed methods
(gain scheduling —, Lyapunov based - - -)
best knowledge, the TS fuzzy approach has not been used to find the approximated
TS fuzzy model and design both stabilizing and tracking controller for nonlinear
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 158
0 2 4 6−1
0
1
2
3
time(s)
rdr(
de
g)
0 2 4 6−3
−2
−1
0
1
time(s)
el(d
eg
)
0 2 4 6−2
0
2
4
6
time(s)
ph
i(d
eg
)
0 2 4 6−10
−5
0
5
time(s)
ail(d
eg
)
0 2 4 60
2
4
6
time(s)
the
ta(d
eg
)
Figure 5.20. Initial part of gain scheduling control
F-16 aircraft systems.
(ii) Both basic and augmented TS fuzzy models of F-16 have been obtained
Chapter 5. Fuzzy Modelling and Control for F-16 Aircraft 159
Table 5.2. The θ tracking specifications of the two control methods
RT ST OV |δe|max |δa|max |δr|max
Lyapunov based control 8 10 0.01 1.4388 1.9203 11.4588
Gain scheduling control I 0.9508 2.0021 0.0653 2.4171 1.6574 14.3380
Gain scheduling control II 0.3562 0.4550 0.0377 2.7179 8.2767 11.5618
using the best-available F-16 nonlinear model the literature and validated to be
reasonable accurate in the operating range of interests.
(iii) Two control methods have been studied through the available TS fuzzy
modelling and fuzzy control design methods (Takagi and Sugeno, 1985) and shown
to be feasible and better than the normal linear control.
(iv) The Lyapunov based control is developed for stabilization of a TS fuzzy
model and extended to handle attitude tracking problem when the model is aug-
mented with integrator at each output. The most important feature of the method
is guaranteed stability and step-signal tracking of the closed-loop system if the TS
fuzzy model error is small enough, while most other nonlinear controllers cannot.
Besides, The solutions are simple to get since it involves an LMI feasibility problem
only.
(v) The gain scheduling control shows the better performance. It has a very
simple control structure and straightforward design from TS model; yet it can
perform as well as the best nonlinear controller of high complexity and very so-
phisticated design. However, the gain scheduling control cannot guarantee the
stability of the closed-loop system.
(vi) These results could be useful to understand and design a flight control
system for rapid maneuver of F-16 aircraft.
Chapter 6
Conclusions
6.1 Main Findings
In this thesis, several new results are obtained around control system design for
better performance and robustness. Briefly, the results are summarized as follows:
A. PID Controller Design
In this thesis, three new PID controller design approaches are developed. In
the first method, a simple scheme for the design of PID controllers for uncertain
processes which achieve gain and phase margin specifications is proposed. The
approach is based on the uncertain model of the plant. The design problem is
formulated into a convex optimization problem with quadratic constraint, then it is
converted into an optimization problem with LMI constraints which can be solved
effectively using interior point method. The optimal PID controller is obtained
directly to satisfy the gain and phase margins. Simulation and comparisons are
presented to demonstrate the effectiveness of the method. For the second approach,
a new criterion for robust stability is developed which made use of both the gain and
phase information of the open-loop system in order to reduce the conservativeness
of the small gain theory. Based on it, a robust PID design approach is developed for
second order plus dead time uncertain processes. Simulation shows the effectiveness
of the method. In the third work, algorithms based on the iterative LMI technique
are developed to design PI controller for linear time invariant processes with state
160
Chapter 6. Conclusions 161
time delay, which guarantee the stability of closed-loop system or H∞ performance
specification. An example is provided to illustrate the applicable of the proposed
method.
B. IMC and Smith Controller Design
Two new approaches for robust IMC controller design, one in time domain
and another in frequency domain, are proposed. For the IMC design in time
domain, it is shown that the IMC optimal H2 performance design with robust
stability constraints can be converted into dynamic output feedback control design
problem. With the new formulation, it is shown that the robust IMC controller
design is solvable via solving a system of LMI. Two examples are presented with
the demonstrated effectiveness of the proposed method. For the IMC design in
frequency domain, a new IMC controller design method is proposed to achieve
the optimal performance and robust stability. The method is based on LMI/BMI
optimization. It is shown that better optimal performance is achieved compared
with Morari’s method with same robust stability constraints. The advantage of
the frequency domain method is that it can cope with the processes with time
delay. Moreover, the controllers obtained through frequency domain approach are
fixed order ones while the resultant controllers via time domain method usually
are of high order. It is also noted that the two proposed schemes are systemic
methods while most of existing IMC tuning schemes use trial and error method.
As we move on to the Smith control with open-loop unstable processes, a simple and
effective design is proposed based on our new modified Smith predictor structure for
unstable process with dead time. Moreover, the internal stability of the proposed
system is analyzed. Simulation illustrates that the proposed method yields much
improvement of load disturbance responses over the best existing method.
C. Decoupling Control Design
In this thesis, a new necessary and sufficient condition for the solvability of the
decoupling problem with internal stability for unity output feedback systems with
square plants is given. The conditions are only concerned with Gd,min of the plant
Chapter 6. Conclusions 162
G. Compared with the existing method, the proposed criterion is simpler, more
understandable and make use of transfer function matrices only. For decoupling
controller design, a new method for IMC system is proposed to gain good loop
performance and largely remove the effect of process coupling simultaneously for
multi-variable plants. The controller may be obtained via LMI optimization. The
robustness of the proposed method is analyzed. Two examples are presented with
the demonstrated effectiveness of the proposed method.
D. TS Fuzzy Control Design
A TS fuzzy modelling method and two fuzzy control strategy are discussed in
the thesis. Both basic and augmented TS fuzzy models of F-16 have been obtained
using the best-available F-16 nonlinear model the literature and validated to be
reasonable accurate in the operating range of interests. Two control methods, Lya-
punov based control and gain scheduling control, have been proposed and shown
to be feasible and better than the normal linear control. The Lyapunov based
control are developed for stabilization of a TS fuzzy model and extended to handle
attitude tracking problem when the model is augmented with integrator at each
output. The most important feature of the method is guaranteed stability and
step-signal tracking of the closed-loop system if the TS fuzzy model error is small
enough, while most other nonlinear controllers cannot. Besides, The solution is
simple to get since it involves an LMI feasibility problem only. The gain schedul-
ing control shows better performance. It has a very simple control structure and
straightforward design from TS model; Yet it can perform as well as the best non-
linear controller of high complexity and very sophisticated design. However, the
gain scheduling control cannot guarantee the stability of the closed-loop system.
The stability of the closed-loop system can only be checked through simulation.
6.2 Suggestions for Further Work
The thesis has taken the full route from initial ideas, via theoretical developments,
to methodologies that can be applied to relevant practical problems. Several new
Chapter 6. Conclusions 163
results have been obtained but some topics remain open and are recommended for
further work.
A. Multi-variable Robust PID Controller Design
Two schemes of single variable robust PID controller design have been proposed
in this thesis. For the single variable case, frequency domain stability margins such
as gain and phase margins are very popular, and used for performance assessment,
design specifications and robustness measure of PID control systems. There are
some attempts to define multi-variable system stability margins in frequency do-
main, but none of them is well known. One may look for better definitions which
are meaningful, useful and easily checkable with clear link to the single variable
case. They can then be used as performance assessment, design specifications and
robustness measure of Multi-variable PID control systems and may lead to a large
branch of tuning rules similar to the single variable case. Basically, there are two
types of Multivariable PID controllers: decentralized (or multi-loop, diagonal) and
centralized (full matrix). The problems should be studied for each type of the
controllers, respectively, as the difference in controller structures calls for different
solution methods in general.
B. Multi-variable Robust IMC Controller Design
In this thesis, the robust IMC designs are developed both in time domain and
frequency domain for SISO systems. Based on the results of single loop design,
more studies may be given to investigate if the ideas is extendible to multi-variable
processes. The robust stability, decoupling and loop performance should be tackled
in the multi-variable IMC system design. Since there are usually more states
and modes in the multi-variable systems, simple design procedures and efficient
computation methods are needed to be investigated.
C. Multi-variable Smith Predictor for Unstable Processes
A modified Smith predictor design for single variable unstable processes has
been proposed. For multi-variable processes, the proposed approach may encounter
problems because of the coupling and different time delay of each element in the
Chapter 6. Conclusions 164
processes. One possible method is to develop decoupling controller to make the
system decoupled. Thus, the scheme used for single variable processes can be used
for the decoupled processed. However, exact decoupling is usually sensitive to
the process model used. Thus, it is desirable to design robust decoupling Smith
predictor such that the interaction of the resultant system is kept within a certain
tolerance for the whole family of the uncertain processes. Another possible scheme
is to find controller with full matrix to make the time delay not appear in the
denominator matrix as well as to maintain the internal stability of the closed-loop
system. Algorithm is needed to be developed to find such controllers.
D. Decoupling Design for Nonlinear Processes
The processes of decoupling design considered in this thesis are assumed to be
linear. However, in practice, all physical systems are nonlinear in nature. Some
complex processes may not be described by linear model with enough accuracy.
Thus, the nonlinear model may be needed in some practical control system design.
In this thesis, the TS fuzzy modelling approach is developed with F-16 aircraft.
However, in the controller design, the decoupling approach is not employed. Works
will be needed to investigate a new method to decouple complex nonlinear control
systems and achieve better performance.
Bibliography
Astrom, K. J. and B. Writtenmark (1984). Computer Control Systems: Theory
and Design.. Prentice Hall.. Upper Saddle River, New Jersey.
Astrom, K. J. and T. Hagglund (1995). PID controller: Theory, Design and Tun-
ing, second edition. Instrument Society of America. Research Triangle Park,
NC.
Astrom, K. J., C. C. Hang and B. C. Lim (1994). A new Smith predictor for
controlling a process with an integrator and long dead time. IEEE Trans.
Automatic Control 39(2), 343–345.
Astrom, K.J., T. Hugglund, C. C. Hang. and W. K. Ho (1993). Automatic tuning
and adaptation for PID controllers - a survey. Control Eng. Practice 1(4), 669–
714.
Athans, M., P. L. Falb and R. T. Lacoss (1963). Time-, fuel-, and energy-optimal
control of nonlinear norm-invariant systems. IRE Trans. on Automatic Control
8, 196–202.
Ball, J. A. and N. Cohen (1987). Sentsitivity minimization in an H∞ norm: Pa-
rameterization of suboptimal solutions. Int. J. Control 46, 785–816.
Bharadwaj, S., M. Osipchuk, K. D. Mease and F. C. Park (1998). Optimal regu-
lation and passivity results for axisymmetric rigid bodies using two controls.
Journal of Guidance, Control, and Dynamics 21, 930–939.
165
Bibliography 166
Bilimoria, K. D. and B. Wie (1991). Time-optimal reorientation of a rigid ax-
isymmetric spacecraft. Proceedings of the AIAA Guidance, Navigation, and
Control Conference pp. 422–431.
Black, H.S. (1977). Inventing the negative feedback amplifier. IEEE Spectrum
14(4), 55–60.
Bode, H. W. (1964). Feedback-The Histry of an Idea. in Seclected Papers on Math-
ematical Trends in Control Theory.. Dover, NY.
Bonnet, Catherine and Jonathan R. Partington (1999). Bezout factors and l1-
optimal controllers for delay systems using a two-parameter compensator
scheme. IEEE Trans. Automatic Control 44(8), 1512–1521.
Boulet, B., Yingxuan Duan and H. Michalska (2002). Some preliminary results for
imc-based robust tunable control. Proceedings of the 2002 American Control
Conference 6, 4894–4895.
Boyd, Stephen, Laurent EI Ghaoui and Venkataramanan Blaakrishnan (1994).
Linear Matrix Inequalities in System and Control Theory. SIAM Studies in
Applied Mathematics.
Camarta, J. F., M. Malabre and J. C. Martnez-Garca (2001). Fixed poles of si-
multaneous disturbance rejection and decoupling: a geometric approach. Au-
tomatica 37(2), 297–302.
Carrington, C. K. and J. L. Junkins (1986). Optimal nonlinear feedback control for
spacecraft attitude maneuvers. Journal of Guidance, Control, and Dynamics
9, 99–107.
Chilali, M. and P. Gahinet (1996). h∞ design with pole placement constraints: An
lmi approach. IEEE Transcations on Automatic Control 41, 358–367.
Chiu, Min-Sen, Shan Cui and Qing-Guo Wang (2000). Internal model control de-
sign for transition control. AIChE Journal 46(2), 309–320.
Bibliography 167
Cominos, P. and N. Munro (2002). PID controllers: recent tuning methods and
design to specification. IEE Proceedings-Control Theory and Applications
149, 46–53.
Dabbous, T. E. and N. U. Ahmed (1982). Nonlinear optimal feedback regulation
of satellite angular momenta. IEEE Trans. on Aerospace Electronics Systems
18, 2–10.
De Paor, A.M. and M. O’Malley (1989). Controllers of ziegler nichols type for
unstable processes. International Journal of Control 49, 1273–1284.
Debs, A. S. and M. Athans (1969). On the optimal angular velocity control of
asymmetrical space vehicles. IEEE Trans. on Automatic Control 14, 80–83.
Decusse, J., J. F. Lafay and Malabre M. (1988). Solution to morgan’s problem.
IEEE Trans. Autom. Control AC-33, 732–739.
Desoer, C. A. and A. N. Gundes (1986). Decoupling linear myltiinput multiout-
put plants by dynamic output feedback: An algebraic theory.. IEEE Trans.
Autom. Control AC-31, 744–750.
Doyle, J. C., K. Glover, P. P. Khargonekar and B. A. Francis (1989). State space
solutions to standard H2 and H∞ control problems. IEEE Trans. Autom.
Control AC-34, 831–847.
Eldem, V. (1996). Decoupling through specified input-output channels with inter-
nal stability. Automatica 32(2), 379–384.
Eldem, V. and A. B. Ozguler (1989). A solution to the diagonalization problem by
constant precompensator and dynamic output feedback.. IEEE Trans. Autom.
Control 34, 1061–1067.
Etter, J. R. (1989). A solution of the time-optimal euler rotation problem. Proceed-
ings of the AIAA Guidance, Navigation, and Control Conference pp. 1441–
1449.
Bibliography 168
Falb, P. L. and W. A. Wolovich (1967). Decoupling in design and synthesis of
multivariable control systems.. IEEE Trans. Autom. Control AC-12, 651–
669.
Fiacco, A. V. and G.P.McCormick (1990). Nonlinear programming - sequential un-
constrained minimization techniques. SIAM Studies in Applied Methematics.
Franklin, G. F., J. D. Powell and A. Emami-Naeini (1994). Feedback control of
dynamic systems. Addison-Wesley Publishing Company. Workingham, U.K.
Franklin, Gene. F., J. David Powell and Abbas Emami. Naeini (2002). Feedback
Control of dynamic systems. Prentice Hall. Upper Saddle River, NJ.
Freeman, R. A. and P. V. Kokotovic (1970). Inverse optimality in robust stabiliza-
tion. Journal of Spacecraft and Rockets 7, 1345–1351.
Freeman, R. A. and P. V Kokotovic (1996). Inverse optimality in robust stabiliza-
tion. SIAM Journal on Control and Optimization 34, 1365–1391.
Gahinet, P., A. Nemirovski, A. J. Laub and M. Chilali (1995). LMI Control Toolbox.
MathWorks. Natick,MA.
Garcia, C.E. and M. Morari (1982). “Internal model control - 1. a unifying review
and some new results”. IECPDD 21, 308–323.
Gilbert, E. G. (1969). The decoupling of multivarialble systems by state feedback.
SIAM J. Control 7(1), 50–63.
Glover, K. (1989). A tutorial on model reduction. In J. C. Willems, editor, From
Data to Model. Springer-Verlag. Berlin.
Goodwin, G.C., L. Wang and D.Miller (1999). Bias-variance trade-off issues in
robust controller design using statistical confidence bounds. Proceedings of
14th IFAC Word Congress, Beijing, 1999.
Goodwin, G.C., S. F. Graebe and M.E. Sagado (2000). Control System Design.
Prentice Hall International. Inc.
Bibliography 169
Gundes, A. N. (1990). Parameterization of all decoupling compensators and all
achievable diagonal maps for unity-feedback system.. Proc. 29th IEEE CDC
AC-12, 2492–2493.
Haddad, Wassim M., VijaySekhar Chellaboina, Joseph R. Corrado and Dennis S
Bernstein (2000). Robust fixed-structure controller synthesis using the implicit
small-gain bound. Journal of the Franklin Institute 337(1), 85–96.
Hammer, J. and P. P. Khargonekar (1984). Decoupling of linear systems by dy-
namic output feedback.. Math. Syst. Theory 17, 135–157.
Hautus, M. L. J. and M. Heymann (1983). Linear feedback decoupling- transfer
funciton analysis.. IEEE Trans. Autom. Control 26(8), 1558–1564.
Ho, W. K., C. C. Hang and L.S. Cao (1995). Tuning of PID controllers based on
gain and phase margin specifications. Automatica 31(3), 497–502.
Jiang, Zhong-Ping and I.M.Y. Marcels (1997). A small-gain control method for
nonlinear cascaded systems with dynamic uncertainties. IEEE Trans. Autom.
Control 42(3), 292–308.
Junkins, J. L., C. K. Carrington and C. E. Williams (1981). Time-optimal mag-
netic attitude maneuvers. SIAM Journal of Guidance, Control, and Dynamics
4, 363–368.
Kalman, R. E. (1964). When is a linear control system optimal?. Journal of Basic
Engineering 86, 51–60.
Krsti, M. and P. Tsiotras (1999). Inverse optimal stabilization of a rigid spacecraft.
IEEEAC 44(5), 1042–1049.
Krsti, M., I. Kanellakopoulos and P. V. Kokotovic (1995). Nonlinear and adaptive
control design.
Kumar, K. S. P. (1965). On the optimum stabilization of a satellite. IEEE Trans.
on Aerospace Electronics Systems 1, 82–83.
Bibliography 170
Lane, S. H. and R. F. Stengel (1988). Flight control design using non-linear inverse
dynamics. Automatica 24(4), 471–483.
Lin, Ching-An (1997). Necessary and sufficient conditions for existence of decou-
pling controllers. IEEE Trans. Autom. Control 42(8), 1157–1161.
Linnemann, A. and R. Maier (1990). Decoupling by precompensation while main-
taining stabilizability. Proc. 29th IEEE CDC 29, 2921–2922.
Litrico, X. (2002). Robust imc flow control of simo dam-river open-channel systems.
IEEE Transactions on Control Systems Technology 10(3), 432–437.
Luyben, W. L. (1986). Simple method for tuning SISO controllers in multivariable
system. Industrial and Engineering Chemistry: Process Design and Develop-
ment 25, 654–660.
Maciehowski, J. M. (1989). Multivariable feedback design.. Addison-Wesley. Work-
ingham, U.K.
Mahmoud, M.S. and M. Zribi (1999). Hinfty controllers for time-delay system us-
ing linear matrix inequalities. Journal of optimization theory and application
100(1), 89–122.
Majhi, S and D. P. Atherton (1999). Modified Smith predictor and controller for
processes with time delay. IEE Proc. -Control Theory Appl. 146(5), 359–366.
Majhi, Somanath and Derek P. Atherton (2000). Obtaining controller parameters
for a new Smith predictor using autotuning. Automatica 36, 1651–1658.
Marcelo, C. M. and H. Zak Stanislaw (1999). Stabilzing controller design for un-
certain nonlinear system using fuzzy models. IEEE Transactions on Fuzzy
Systems 7(2), 133–142.
Matausek, M. R. and A. D. Micic (1996). A modified Smith predictor for controlling
a process with an integrator and long dead-time. IEEE Trans. Automatic
Control 41(8), 1199–1203.
Bibliography 171
Menon, P., Badgett P. and R. Walker (1987). Nonlinear flight test trajectory con-
trollers for aircraft. Journal of Guidance, Control, and Dynamics 10(1), 67–72.
Morari, M. and E. Zafiriou (1989). Robust Process Control. Prentice Hall. Eagle-
wood cliffs, NJ.
Morgan, B. S. (1964). The systhesis of linear multivariable systems by state variable
feedback. IEEE Trans. Autom. Control 9, 404–411.
Niculescu, Silviu-Iulian (1998). H∞ memoryless control with an α-stability con-
straint for time-delay system: An LMI approach. IEEE Transactions on Au-
tomatic Control 43(5), 739–743.
Nyquist, H. (1932). Regeneration theory. Bell Syst. Tech. J. 11, 126–147.
Ogata, K. (1990). Mordern control Engineering, 2nd edition. Prentice Hall. Engle-
wood Cliffs, NJ.
Palmor, Z. J., Y. Halevi and N. Krasney (1993). Automatic tuning of decentralized
PID controllers for TITO processes. In: Proc. of 12th IFAC World Congress.
Sydney. pp. 311–314.
Pernebo, L. (1981). An algebraic theory for the design of controllers for linear
multivariable systems.. IEEE Trans. Autom. Control AC-26, 171–193.
Pintelon, R., P. Guillaume, Y. Rolain, J. Schoukens and H. Van hamme (1994).
Parametric identification of transfer functions in the frequency domain – a
survey. IEEEAC 39, 2245–2260.
Porcelli, P. (1968). Optimal attitude control of a dual-spin satellite. Journal of
Spacecraft and Rockets 5, 881–888.
Praly, Laurent; Wang, Yuan (1996). Stabilization in spite of matched unmodeled
dynamics and an equivalent definition of input-to-state stability. Mathematics
of Control, Signals, and Systems 9(1), 1–33.
Bibliography 172
Reiner, Horst and Hoang Tuy (1992). Global optimization: deterministic ap-
proaches. Springer-Verlag.
Rotea, M., P. Tsiotras and M. Corless (1998). Suboptimal control of rigid body
motion with a quadratic cost. Dynamics and Control 8, 55–81.
Scherer, Cardten, Pascal Gahinet and Mahmoud Chilali (1997). Multiobjective
output-feedback control via LMI optimization. IEEE Transcations on Auto-
matic Control 42(7), 896–911.
Schumacher, C. and P. P. Khargonekar (1998). Stability analysis of a missile control
system with a dynamic inversion controller. Journal of Guidance, Control, and
Dynamics 21(3), 508–515.
Scrivener, S. L. and R. C. Thomson (1994). Survey of time-optimal attitude ma-
neuvers. Journal of Guidance, Control, and Dynamics 17, 225–233.
Seborg, D. E., T. F. Edgar and D. A. Mellichamp (1989). Process Dynamics and
Control. Wiley. New York.
Sepulchre, R., M. Jankovi and P. V. Kokotovic (1997). Constructive Nonlinear
Control. Springer-Verlag. New York.
Shuster, M. D. (1963). A survey of attitude representations. SIAM Journal of
Astronautical Sciences 41, 439–517.
Smith, O. J. (1959). A controller to overcome dead time. ISA J. 6(2), 28–33.
Snell, S. A., Enns D. F. and W. L. Garrard (1992). Nonlinear inversion flight
control for a supermaneuverable aircraft. Journal of Guidance, Control, and
Dynamics 15(4), 976–984.
Sohoni, V. S. and D. R. Guild (1970). Optimal attitude control systems for spinning
aerospace vehicles. Journal of Astronautical Sciences 18, 86–100.
Sontag, E. and H. Sussmann (1988). Further comments of the stabilizability of the
angular velocity of a rigid body. Systems and Control Letters 12, 213–217.
Bibliography 173
Stevens, L. Brian and L. Lewis Frank (1992). Aircraft Control and Simulation.
John Wiley & Sons. New York.
Stoorvogel, A. A. (1991). The H∞ Control Problem: A state space approach. Pren-
tice Hall. Eaglewood cliffs, NJ.
Taeyoung, Lee. and K Youdan (2001). Nonlinear adaptive flight control using back-
stepping and neural networks controller. Journal of Guidance, Control, and
Dynamics 24, 675–682.
Takagi, T. and M. Sugeno (1985). Fuzzy identification of systems and its applica-
tions to modeling and control. IEEE Trans. on Systems, Man and Cybernetics
15(1), 116–132.
Tan, K. K., Qing-Guo Wang and C. C. Hang (1999). Advances in PID control.
Springer. London.
Tanaka, K. and M. Sugeno (1992). Stability analysis and design of fuzzy control
systems. Fuzzy Sets and Systems 45(2), 135–156.
Tanaka, K., T. Ikeda and H. Wang (1996a). Design of fuzzy control systems based
on lmi stability conditions. Proceedings of the 35th Conference on Decision
and Control pp. 598–603.
Tanaka, K., T. Ikeda and H. Wang (1996b). Robust stabilization of a class of
uncertain nonlinear systems via fuzzy control: Quadratic stabilizability, 〈∞
control theory and linear matrix inequalities. IEEE Trans. on Fuzzy Systems
4(1), 1–13.
Tanaka, K., T. Ikeda and H. Wang (1997). An LMI approach to fuzzy controller
designs based on relaxed stability conditions. Proceedings of the 6th Interna-
tional Conference on Fuzzy Systems pp. 171–176.
Tanaka, K., T. Ikeda and H. Wang (1998). Fuzzy regulators and fuzzy observers:
Relaxed stability conditions and lmi-based designs. IEEE Trans. on Fuzzy
Systems.
Bibliography 174
Thathachar, M. and P. Viswanath (1997). On the stability of fuzzy systems. IEEE
Trans. on Fuzzy Systems.
Tsiotras, P. (1997). Optimal regulation and passivity results for axisymmetric
rigid bodies using two controls. Journal of Guidance, Control, and Dynamics
20, 457–463.
Tsiotras, P. and J. M. Longuski (1994). Spin-axis stabilization of symmetric space-
craft with two control torques. Systems and Control Letters 23, 395–402.
Tsiotras, P., M. Corless and M. Rotea (1996). Optimal control of rigid body angular
velocity with quadratic cost. Proceedings of the 35th Conference on Decision
and Control pp. 1630–1635.
Vadali, S. R., L. G. Kraige and J. L. Junkins (1984). New results on the optimal
spacecraft attitude maneuver problem. SIAM Journal of Guidance, Control,
and Dynamics 7, 378–380.
Wang, H. and K. Tanaka (1996). An LMI-based stable fuzzy control of nonlin-
ear systems and its application to control of chaos. Proceedings of the 5th
International Conference on Fuzzy Systems pp. 1433–1438.
Wang, H., K. Tanaka and M. Griffin (1996). An approach to fuzzy control of
nonlinear systems: Stability and design issues. IEEE Trans. on Fuzzy Systems
4(1), 14–23.
Wang, L. (1999). Robust control of multivariable systems using statistical confi-
dence bounds. Proceedings of 38th IEEE Conference on Decision and Control,
Phoenix, U.S.A.
Wang, L. and G.C. Goodwin (2000). Integrating identification with robust control:
a mixed H2/H∞ approach. Proceedings of 39th IEEE Conference on Decision
and Control, Sydney, Australia.
Wang, Q. G. (1992). Decoupling with internal stability for unity output feedback
systems. IEEE Trans. Autom. Control 28(2), 411–415.
Bibliography 175
Wang, Q. G. and Y. zhang (1998a). A method to incorporate control input con-
straints into controller design. ISA Transactions 37, 323–327.
Wang, Q. G., C.C. Hang and X. P. Yang (2001). Single loop design through imc
priciple. Automatica 37(12), 2041–2048.
Wang, Qing-Guo and Yu Zhang (1998b). A fast algorithm for reduced-order mod-
elling. ISA Transactions 38, 225–230.
Wang, Qing-Guo, Biao Zou, Tong-Heng Lee and Qiang Bi (1997). Auto tuning
of multivariable PID controllers from decentralized relay feedback. IEEEAC
33(3), 319–330.
Wang, Qing-Guo, Ho-Wang Fung and Yu Zhang (1999a). PID tuning with exact
gain and phase margins. ISA Transactions 38, 243–249.
Wang, Qing-Guo, Tong-Heng Lee and Jian-Bo He (1999b). Internal stability of
interconnected systems. IEEE Trans. Automatic Control 44(3), 593–596.
Wang, S. H. and E. J. Davison (1975). Design of decoupled control systems: A
frequency domain approach.. Int. J. Control 21, 529–536.
Williams, T. W. C. and P. J. Antsaklis (1986). A unifying approach to the decou-
pling of linear multivariable systems. IJC 44, 181–201.
Windeknecht, T. G. (1963). Optimal stabilization of rigid body attitude. Journal
of Mathematical Analysis and Applications 6, 325–335.
Wolovich, W. A. (1978). Skew prime polynomial matrices. IEEEAC 23(5), 880–
887.
Wood, R. K and M. W. Berry (1973). Terminal composition control of binary
distillation column. Chem. Engng. Sci. 28, 1707–1717.
Wood, W.A. (1986). Linear Multivariable Control: A Geomatric Approach, 3rd
ed.. Springer-Verlag. New York.
Bibliography 176
Yamamoto, S. and I. Hashimoto (1991). Present status and future needs: The
view from japanese industry. Chemical Process Control - CPCIV: Proc. of 4th
International Conference on Chemical Procee Control.
Zames, G. (1981). Feedback and optimal sensitivity: Model reference transferma-
tion, multiplicative seminormes and approximate inverses. IEEEAC 26, 301–
320.
Zheng, Feng, Qing-Guo Wang and Tong Heng Lee (2002). On the design of multi-
varaible PID controller via LMI approach. Automatica 38, 517–526.
Zhou, K., Doyle J. C. and Glover K. (1996). Robust and optimal control. Prentice
Hall. Eaglewood cliffs, NJ.
Zhu, H.A., G.S. Hong, C.L. Teo and A.N. Poo (1995). Internal model control with
enhanced robustness. International Journal of Systems Science 26, 277–293.
Zhuang, M. and D. P. Atherton (1993). Automatic tuning of optimum PID con-
trollers. Proc. IEE, Part D 140(3), 216–224.
Appendix A
TS Fuzzy Basic Model
The model with α ∈ [−0.2, 0.2] and φ ∈ [−0.2, 0.2] assumes the form:
x =9∑
i=1
hi(α, φ)(Aix+Biu).
Selection of the operating points and the fuzzy rules
Rule 1: If α = −0.2 and φ = −0.2, then the linearized model is x = A1x+B1u.
Rule 2: If α = 0 and φ = −0.2, then the linearized model is x = A2x+B2u.
Rule 3: If α = 0.2 and φ = −0.2, then the linearized model is x = A3x+B3u.
Rule 4: If α = −0.2 and φ = 0, then the linearized model is x = A4x+B4u.
Rule 5: If α = 0 and φ = 0, then the linearized model is x = A5x+B5u.
Rule 6: If α = 0.2 and φ = 0, then the linearized model is x = A6x+B6u.
Rule 7: If α = −0.2 and φ = 0.2, then the linearized model is x = A7x+B7u.
Rule 8: If α = 0 and φ = 0.2, then the linearized model is x = A8x+B8u.
Rule 9: If α = 0.2 and φ = 0.2, then the linearized model is x = A9x+B9u.
Resulting linear models:
177
Appendix A TS Fuzzy Basic Model 178
A1 =
-3.0489 0.0003 -0.9873 0 0 -3.7977 0 0
0 -1.3815 -0.0029 0.0009 -1.1356 0 -0.0335 0
0.0312 0.0025 -0.5181 0 0 9.7538 0 0
0 2.5473 0 -0.0832 191.1245 -6.3862 -80.3479 -31.5328
0 0.9639 0 0.0003 -1.4637 0 0.1846 -0.0127
-0.1609 0 -0.9798 0.0000 0.0004 -0.2779 0.0632 0.0005
1.0000 -0.0078 0.0386 0 0 0 0 0
0 0.9801 0.1987 0 0 0 0 0
, B1 =
0 0 -0.5983 0.1588
0 -0.1925 0 0
0 0 -0.0337 -0.0675
0 0.4584 0 0
0 -0.0020 0 0
0 0 0.0003 0.0008
0 0 0 0
0 0 0 0
A2 =
-3.6722 0.0003 0.6741 0 0 -30.9113 0 0
0 -1.3859 -0.0029 0.0000 -2.4977 0 -0.0005 0
-0.0263 0.0025 -0.4986 0 0 9.4747 0 0
0 -0.6501 0 -0.0202 7.2411 -6.3862 0.1259 -32.169
0 0.9048 0 -0.0003 -1.0189 0 0.0064 0.0001
0.0389 0 -0.9916 0.0000 0 -0.3220 0.0636 0.0005
1.0000 -0.0078 0.0386 0 0 0 0 0
0 0.9801 0.1987 0 0 0 0 0
, B2 =
0 0 -0.7336 0.1310
0 -0.1824 0 0
0 0 -0.0327 -0.0643
0 0.1704 0 0
0 -0.0022 0 0
0 0 0.0003 0.0008
0 0 0 0
0 0 0 0
A3 =
-3.2067 0.0003 1.8464 0 0 -65.6296 0 0
0 -1.6530 -0.0029 -0.0016 -0.7348 0 1.2180 0
-0.0826 0.0025 -0.5456 0 0 8.8395 0 0
0 -7.4582 0 -0.0917 -92.9348 -6.3862 -33.758 -31.5228
0 0.9031 0 -0.0006 -1.0248 0 0.0054 0.0128
0.2390 0 -0.9633 0.0000 -0.0004 -0.2984 0.0632 0.0005
1.0000 -0.0078 0.0386 0 0 0 0 0
0 0.9801 0.1987 0 0 0 0 0
, B3 =
0 0 -0.6839 0.1258
0 -0.2000 0 0
0 0 -0.0254 -0.0641
0 -0.1139 0 0
0 -0.0021 0 0
0 0 0.0003 0.0008
0 0 0 0
0 0 0 0
A4 =
-3.0489 0.0003 -0.9873 0 0 -3.7977 0 0
0 -1.3815 -0.0029 0.0009 -1.1690 0 0 0
0.0312 0.0025 -0.5181 0 0 9.7538 0 0
0 2.5473 0 -0.0832 111.2891 0 0 -31.5328
0 0.9639 0 0.0003 -1.2854 0 0 -0.0127
-0.1609 0 -0.9798 0.0000 0.000 -0.2779 0.0640 0.00
1.0000 0 0.0394 0 0 0 0 0
0 1.0000 0 0 0 0 0 0
, B4 =
0 0 -0.5983 0.1588
0 -0.1925 0 0
0 0 -0.0337 -0.0675
0 0.4584 0 0
0 -0.0020 0 0
0 0 0.0003 0.0008
0 0 0 0
0 0 0 0
A5 =
-3.6722 0.0003 0.6741 0 0 -30.9113 0 0
0 -1.3859 -0.0029 0.0000 -2.4977 0 0 0
-0.0263 0.0025 -0.4986 0 0 9.4747 0 0
0 -0.6501 0 -0.0202 7.8814 0 0 -32.169
0 0.9048 0 -0.0003 -1.0190 0 0 0
0.0389 0 -0.9916 0.0000 0 -0.3220 0.064 0
1.0000 0 0.0394 0 0 0 0 0
0 1.0000 0 0 0 0 0 0
, B5 =
0 0 -0.7336 0.1310
0 -0.1824 0 0
0 0 -0.0327 -0.0643
0 0.1704 0 0
0 -0.0022 0 0
0 0 0.0003 0.0008
0 0 0 0
0 0 0 0
A6 =
-3.2067 0.0003 1.8464 0 0 -65.6296 0 0
0 -1.6530 -0.0029 -0.0016 -1.9527 0 0 0
-0.0826 0.0025 -0.5456 0 0 8.8395 0 0
0 -7.4582 0 -0.0917 -58.4168 0 0 -31.5287
0 0.9031 0 -0.0006 -1.0240 0 0 0.0127
0.2390 0 -0.9633 0.0000 -0.0000 -0.2987 0.0640 0
1.0000 0.00 0.0394 0 0 0 0 0
0 1.0000 0 0 0 0 0 0
, B6 =
0 0 -0.6839 0.1258
0 -0.2000 0 0
0 0 -0.0254 -0.0641
0 -0.1139 0 0
0 -0.0021 0 0
0 0 0.0003 0.0008
0 0 0 0
0 0 0 0
A7 =
-3.0489 0.0003 -0.9873 0 0 -3.7977 0 0
0 -1.3815 -0.0029 0.0009 -1.1356 0 0.0335 0
0.0312 0.0025 -0.5181 0 0 9.7538 0 0
0 2.5473 0 -0.0832 191.1245 6.3862 80.3479 -31.5328
0 0.9639 0 0.0003 -1.4637 0 0.1846 -0.0127
-0.1609 0 -0.9798 -0.0000 -0.0004 -0.2779 0.0632 -0.0005
1.0000 0.0078 0.0386 0 0 0 0 0
0 0.9801 -0.1987 0 0 0 0 0
, B7 =
0 0 -0.5983 0.1588
0 -0.1925 0 0
0 0 -0.0337 -0.0675
0 0.4584 0 0
0 -0.0020 0 0
0 0 0.0003 0.0008
0 0 0 0
0 0 0 0
Appendix A TS Fuzzy Basic Model 179
A8 =
3.6722 0.0003 0.6741 0 0 -30.9113 0 0
0 -1.3859 -0.0029 0.0000 -2.4977 0 0.0005 0
-0.0263 0.0025 -0.4986 0 0 9.4747 0 0
0 -0.6501 0 -0.0202 7.2411 6.3862 -0.1259 -32.169
0 0.9048 0 -0.0003 -1.0189 0 -0.0064 0.0001
0.0389 0 -0.9916 -0.0000 0 -0.3220 0.0636 -0.0005
1.0000 0.0078 0.0386 0 0 0 0 0
0 0.9801 -0.1987 0 0 0 0 0
, B8 =
0 0 -0.7336 0.1310
0 -0.1824 0 0
0 0 -0.0327 -0.0643
0 0.1704 0 0
0 -0.0022 0 0
0 0 0.0003 0.0008
0 0 0 0
0 0 0 0
A9 =
-3.2067 0.0003 1.8464 0 0 -65.6296 0 0
0 -1.6530 -0.0029 -0.0016 -0.7348 0 -1.2180 0
-0.0826 0.0025 -0.5456 0 0 8.8395 0 0
0 -7.4582 0 -0.0917 -92.9348 6.3862 33.7585 -31.5228
0 0.9031 0 -0.0006 -1.0248 0 -0.0054 0.0128
0.2390 0 -0.9633 -0.0000 0.0004 -0.2984 0.0632 -0.0005
1.0000 0.0078 0.0386 0 0 0 0 0
0 0.9801 -0.1987 0 0 0 0 0
, B9 =
0 0 -0.6839 0.1258
0 -0.2000 0 0
0 0 -0.0254 -0.0641
0 -0.1139 0 0
0 -0.0021 0 0
0 0 0.0003 0.0008
0 0 0 0
0 0 0 0
Resulting Weighting Functions:
hi(α, φ) =MαiMφi , i = 1, 2, 3,
where
Mα1 =
1 if α ≤ −0.22.5 ∗ α− 0.5 if − 0.2 < α < 0.2
0 if α ≥ 0.2
Mα2 =
0 if α ≤ −0.25 ∗ α + 1 if − 0.2 < α ≤ 0
−5 ∗ α + 1 if 0 < α < 0.2
0 if α ≥ 0.2
Mα3 =
0 if α ≤ −0.22.5 ∗ α + 0.5 if − 0.2 < α < 0.2
1 if α ≥ 0.2
Mφ1 =
1 if φ ≤ −0.22.5 ∗ φ− 0.5 if − 0.2 < φ < 0.2
0 if φ ≥ 0.2
Appendix A TS Fuzzy Basic Model 180
Mφ2 =
0 if φ ≤ −0.25 ∗ φ+ 1 if − 0.2 < φ ≤ 0
−5 ∗ φ+ 1 if 0 < φ < 0.2
0 if φ ≥ 0.2
Mφ3 =
0 if φ ≤ −0.22.5 ∗ φ+ 0.5 if − 0.2 < φ < 0.2
1 if φ ≥ 0.2
Appendix B
TS Fuzzy Augmented Model I
The model with φ ∈ [−0.2, 0.2] and θ ∈ [−0.15, 0.15] assumes the form:
˙x =9∑
i=1
hi(φ, θ)(Aix+ Biu)
Selection of the operating points and the fuzzy rules
Rule 1: If φ = −0.2 and θ = −0.15, then the linearized model is ˙x = A1x+ B1u.
Rule 2: If φ = 0 and θ = −0.15, then the linearized model is ˙x = A2x+ B2u.
Rule 3: If φ = 0.2 and θ = −0.15, then the linearized model is ˙x = A3x+ B3u.
Rule 4: If φ = −0.2 and θ = 0, then the linearized model is ˙x = A4x+ B4u.
Rule 5: If φ = 0 and θ = 0, then the linearized model is ˙x = A5x+ B5u.
Rule 6: If φ = 0.2 and θ = 0, then the linearized model is ˙x = A6x+ B6u.
Rule 7: If φ = −0.2 and θ = 0.15, then the linearized model is ˙x = A7x+ B7u.
Rule 8: If φ = 0 and θ = 0.15, then the linearized model is ˙x = A8x+ B8u.
Rule 9: If φ = 0.2 and θ = 0.15, then the linearized model is ˙x = A9x+ B9u.
181
Appendix B TS Fuzzy Augmented Model I 182
Resulting linear models:
A1 =
-3.6722 0.0003 0.6741 0 0 -30.9113 0 0 0 0 0 0
0 -1.3859 -0.0029 0.0000 -2.4977 0 -0.0003 -0.0002 0 0 0 0
-0.0263 0.0025 -0.4986 0 0 9.4747 0 0 0 0 0 0
0 -0.6501 0 -0.0202 6.8833 -6.3521 0.0559 -31.9571 0 0 0 0
0 0.9048 0 -0.0002 -1.0285 0 0.0064 0.0047 0 0 0 0
0.0389 0 -0.9916 0.0000 -0.0000 -0.3315 0.0636 -0.0005 0 0 0 0
1.000 0.0221 -0.1089 0 0 0 0 0 0 0 0 0
0 0.9801 0.1987 0 0 0 0 0 0 0 0 0
0 -0.1999 0.9861 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 1.0000 -0.0000 0 0 0 0
0 0 0 0 0 0 -0.0000 1.0000 0 0 0 0
0 0 0 0 0 0 0 0 1.0000 0 0 0
B1 =
0 0 -0.7336 0.1310
0 -0.1824 0 0
0 0 -0.0327 -0.0643
0 0.1704 0 0
0 -0.0022 0 0
0 0 0.0003 0.0008
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
A2 =
-3.6722 0.0003 0.6741 0.0000 0.0000 -30.9113 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -1.3859 -0.0029 0.0000 -2.4977 0.0000 0.0000 -0.0007 0.0000 0.0000 0.0000 0.0000
-0.0263 0.0025 -0.4986 0.0000 0.0000 9.4747 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -0.6501 0.0000 -0.0202 7.5202 0.0000 0.0000 -32.0497 0.0000 0.0000 0.0000 0.0000
0.0000 0.9048 0.0000 -0.0003 -1.0286 0.0000 0.0000 0.0048 0.0000 0.0000 0.0000 0.0000
0.0389 0.0000 -0.9916 0.0000 0.0000 -0.3316 0.0637 0.0000 0.0000 0.0000 0.0000 0.0000
1.0000 0.0000 -0.1111 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 1.0062 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000
B2 =
0.0000 0.0000 -0.7336 0.1310
0.0000 -0.1824 0.0000 0.0000
0.0000 0.0000 -0.0327 -0.0643
0.0000 0.1704 0.0000 0.0000
0.0000 -0.0022 0.0000 0.0000
0.0000 0.0000 0.0003 0.0008
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
Appendix B TS Fuzzy Augmented Model I 183
A3 =
-3.6722 0.0003 0.6741 0.0000 0.0000 -30.9113 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -1.3859 -0.0029 0.0000 -2.4977 0.0000 0.0003 -0.0002 0.0000 0.0000 0.0000 0.0000
-0.0263 0.0025 -0.4986 0.0000 0.0000 9.4747 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -0.6501 0.0000 -0.0202 6.8833 6.3521 -0.0559 -31.9571 0.0000 0.0000 0.0000 0.0000
0.0000 0.9048 0.0000 -0.0002 -1.0285 0.0000 -0.0064 0.0047 0.0000 0.0000 0.0000 0.0000
0.0389 0.0000 -0.9916 -0.0000 0.0000 -0.3315 0.0636 0.0005 0.0000 0.0000 0.0000 0.0000
1.0000 -0.0221 -0.1089 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.9801 -0.1987 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.1999 0.9861 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000
B3 =
0.0000 0.0000 -0.7336 0.1310
0.0000 -0.1824 0.0000 0.0000
0.0000 0.0000 -0.0327 -0.0643
0.0000 0.1704 0.0000 0.0000
0.0000 -0.0022 0.0000 0.0000
0.0000 0.0000 0.0003 0.0008
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
A4 =
-3.6722 0.0003 0.6741 0.0000 0.0000 -30.9113 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -1.3859 -0.0029 0.0000 -2.4977 0.0000 -0.0005 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0263 0.0025 -0.4986 0.0000 0.0000 9.4747 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -0.6501 0.0000 -0.0202 7.2411 -6.3862 0.1259 -32.1690 0.0000 0.0000 0.0000 0.0000
0.0000 0.9048 0.0000 -0.0003 -1.0189 0.0000 0.0064 0.0001 0.0000 0.0000 0.0000 0.0000
0.0389 0.0000 -0.9916 0.0000 0.0000 -0.3220 0.0636 0.0005 0.0000 0.0000 0.0000 0.0000
1.0000 -0.0078 0.0386 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.9801 0.1987 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -0.1988 0.9808 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000
B4 =
0.0000 0.0000 -0.7336 0.1310
0.0000 -0.1824 0.0000 0.0000
0.0000 0.0000 -0.0327 -0.0643
0.0000 0.1704 0.0000 0.0000
0.0000 -0.0022 0.0000 0.0000
0.0000 0.0000 0.0003 0.0008
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
Appendix B TS Fuzzy Augmented Model I 184
A5 =
-3.6722 0.0003 0.6741 0.0000 0.0000 -30.9113 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -1.3859 -0.0029 0.0000 -2.4977 0.0000 -0.0005 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0263 0.0025 -0.4986 0.0000 0.0000 9.4747 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -0.6501 0.0000 -0.0202 7.2411 -6.3862 0.1259 -32.1690 0.0000 0.0000 0.0000 0.0000
0.0000 0.9048 0.0000 -0.0003 -1.0189 0.0000 0.0064 0.0001 0.0000 0.0000 0.0000 0.0000
0.0389 0.0000 -0.9916 0.0000 0.0000 -0.3220 0.0636 0.0005 0.0000 0.0000 0.0000 0.0000
1.0000 -0.0078 0.0386 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.9801 0.1987 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -0.1988 0.9808 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000
B5 =
0.0000 0.0000 -0.7336 0.1310
0.0000 -0.1824 0.0000 0.0000
0.0000 0.0000 -0.0327 -0.0643
0.0000 0.1704 0.0000 0.0000
0.0000 -0.0022 0.0000 0.0000
0.0000 0.0000 0.0003 0.0008
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
A6 =
-3.6722 0.0003 0.6741 0.0000 0.0000 -30.9113 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -1.3859 -0.0029 0.0000 -2.4977 0.0000 0.0005 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0263 0.0025 -0.4986 0.0000 0.0000 9.4747 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -0.6501 0.0000 -0.0202 7.2411 6.3862 -0.1259 -32.1690 0.0000 0.0000 0.0000 0.0000
0.0000 0.9048 0.0000 -0.0003 -1.0189 0.0000 -0.0064 0.0001 0.0000 0.0000 0.0000 0.0000
0.0389 0.0000 -0.9916 -0.0000 0.0000 -0.3220 0.0636 -0.0005 0.0000 0.0000 0.0000 0.0000
1.0000 0.0078 0.0386 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.9801 -0.1987 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.1988 0.9808 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000
B6 =
0.0000 0.0000 -0.7336 0.1310
0.0000 -0.1824 0.0000 0.0000
0.0000 0.0000 -0.0327 -0.0643
0.0000 0.1704 0.0000 0.0000
0.0000 -0.0022 0.0000 0.0000
0.0000 0.0000 0.0003 0.0008
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
Appendix B TS Fuzzy Augmented Model I 185
A7 =
-3.6722 0.0003 0.6741 0.0000 0.0000 -30.9113 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -1.3859 -0.0029 0.0000 -2.4977 0.0000 -0.0003 0.0002 0.0000 0.0000 0.0000 0.0000
-0.0263 0.0025 -0.4986 0.0000 0.0000 9.4747 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -0.6501 0.0000 -0.0202 6.8909 -6.2769 0.2860 -31.8332 0.0000 0.0000 0.0000 0.0000
0.0000 0.9048 0.0000 -0.0002 -1.0093 0.0000 0.0063 -0.0047 0.0000 0.0000 0.0000 0.0000
0.0389 0.0000 -0.9916 0.0000 -0.0000 -0.3124 0.0634 0.0011 0.0000 0.0000 0.0000 0.0000
1.0000 -0.0381 0.1878 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.9801 0.1987 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -0.2023 0.9979 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000
B7 =
0.0000 0.0000 -0.7336 0.1310
0.0000 -0.1824 0.0000 0.0000
0.0000 0.0000 -0.0327 -0.0643
0.0000 0.1704 0.0000 0.0000
0.0000 -0.0022 0.0000 0.0000
0.0000 0.0000 0.0003 0.0008
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
A8 =
-3.6722 0.0003 0.6741 0.0000 0.0000 -30.9113 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -1.3859 -0.0029 0.0000 -2.4977 0.0000 0.0000 0.0007 0.0000 0.0000 0.0000 0.0000
-0.0263 0.0025 -0.4986 0.0000 0.0000 9.4747 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -0.6501 0.0000 -0.0202 7.5202 0.0000 0.0000 -32.0493 0.0000 0.0000 0.0000 0.0000
0.0000 0.9048 0.0000 -0.0003 -1.0094 0.0000 0.0000 -0.0048 0.0000 0.0000 0.0000 0.0000
0.0389 0.0000 -0.9916 0.0000 0.0000 -0.3124 0.0629 0.0000 0.0000 0.0000 0.0000 0.0000
1.0000 0.0000 0.1917 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 1.0182 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000
B8 =
0.0000 0.0000 -0.7336 0.1310
0.0000 -0.1824 0.0000 0.0000
0.0000 0.0000 -0.0327 -0.0643
0.0000 0.1704 0.0000 0.0000
0.0000 -0.0022 0.0000 0.0000
0.0000 0.0000 0.0003 0.0008
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
Appendix B TS Fuzzy Augmented Model I 186
A9 =
-3.6722 0.0003 0.6741 0.0000 0.0000 -30.9113 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -1.3859 -0.0029 0.0000 -2.4977 0.0000 0.0003 0.0002 0.0000 0.0000 0.0000 0.0000
-0.0263 0.0025 -0.4986 0.0000 0.0000 9.4747 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -0.6501 0.0000 -0.0202 6.8909 6.2769 -0.2860 -31.8332 0.0000 0.0000 0.0000 0.0000
0.0000 0.9048 0.0000 -0.0002 -1.0093 0.0000 -0.0063 -0.0047 0.0000 0.0000 0.0000 0.0000
0.0389 0.0000 -0.9916 -0.0000 0.0000 -0.3124 0.0634 -0.0011 0.0000 0.0000 0.0000 0.0000
1.0000 0.0381 0.1878 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.9801 -0.1987 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.2023 0.9979 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 -0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 -0.0000 1.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000
B9 =
0.0000 0.0000 -0.7336 0.1310
0.0000 -0.1824 0.0000 0.0000
0.0000 0.0000 -0.0327 -0.0643
0.0000 0.1704 0.0000 0.0000
0.0000 -0.0022 0.0000 0.0000
0.0000 0.0000 0.0003 0.0008
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
Resulting Weighting Functions:
hi(φ, θ) =MφiMθi , i = 1, 2, 3,
where
Mφ1 =
1 if φ ≤ −0.22.5 ∗ φ− 0.5 if − 0.2 < φ < 0.2
0 if φ ≥ 0.2
Mφ2 =
0 if φ ≤ −0.25 ∗ φ+ 1 if − 0.2 < φ ≤ 0
−5 ∗ φ+ 1 if 0 < φ < 0.2
0 if φ ≥ 0.2
Mφ3 =
0 if φ ≤ −0.22.5 ∗ φ+ 0.5 if − 0.2 < φ < 0.2
1 if φ ≥ 0.2
Appendix B TS Fuzzy Augmented Model I 187
Mθ1 =
1 if θ ≤ −0.152.5 ∗ θ − 0.5 if − 0.15 < θ < 0.15
0 if θ ≥ 0.15
Mθ2 =
0 if θ ≤ −0.155 ∗ θ + 1 if − 0.15 < θ ≤ 0
−5 ∗ θ + 1 if 0 < θ < 0.15
0 if θ ≥ 0.15
Mθ3 =
0 if θ ≤ −0.152.5 ∗ θ + 0.5 if − 0.15 < θ < 0.15
1 if θ ≥ 0.15
Appendix C
TS Fuzzy Augmented Model II
The model with α ∈ [−0.1, 0.1] and φ ∈ [−0.8, 0.8] assumes the form:
˙x =9∑
i=1
hi(α, φ)(Aix+ Biu),
Selection of the operating points and the fuzzy rules
Rule 1: If α = −0.1 and φ = −0.8, then the linearized model is ˙x = A1x+ B1u.
Rule 2: If α = −0.1 and φ = −0.8, then the linearized model is ˙x = A2x+ B2u.
Rule 3: If α = 0 and φ = −0.8, then the linearized model is ˙x = A3x+ B3u.
Rule 4: If α = 0.1 and φ = −0.8, then the linearized model is ˙x = A4x+ B4u.
Rule 5: If α = −0.1 and φ = 0, then the linearized model is ˙x = A5x+ B5u.
Rule 6: If α = 0 and φ = 0, then the linearized model is ˙x = A6x+ B6u.
Rule 7: If α = 0.1 and φ = 0.8, then the linearized model is ˙x = A7x+ B7u.
Rule 8: If α = 0 and φ = 0.8, then the linearized model is ˙x = A8x+ B8u.
Rule 9: If α = 0.1 and φ = 0.8, then the linearized model is ˙x = A9x+ B9u.
188
Appendix C TS Fuzzy Augmented Model II 189
Resulting linear models
A1 =
-3.2408 0.0003 -0.0600 0.0000 0.0000 -15.6934 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -0.3779 -0.0029 -0.0008 2.5382 0.0000 -0.0784 0.0000 0.0000 0.0000 0.0000
0.0159 0.0025 -0.4696 0.0000 0.0000 8.9211 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 2.6825 0.0000 -0.0236 76.3394 -23.0616 -4.9446 -32.0320 0.0000 0.0000 0.0000
0.0000 0.9157 0.0000 -0.0000 -1.0911 0.0000 0.0275 -0.0057 0.0000 0.0000 0.0000
-0.0641 0.0000 -0.9908 0.0001 0.0016 -0.3147 0.0572 0.0017 0.0000 0.0000 0.0000
1.0000 -0.0265 0.0257 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.6967 0.7174 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 -0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000
B1 =
0.0000 0.0000 -0.7377 0.1355
0.0000 -0.1553 0.0000 0.0000
0.0000 0.0000 -0.0344 -0.0620
0.0000 0.2820 0.0000 0.0000
0.0000 -0.0021 0.0000 0.0000
0.0000 0.0000 0.0003 0.0008
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
A2 =
-3.6776 0.0003 0.6645 0.0000 0.0000 -30.6426 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -1.0772 -0.0029 0.0000 0.8221 0.0000 -0.0000 0.0000 0.0000 0.0000 0.0000
-0.0254 0.0025 -0.4765 0.0000 0.0000 8.5416 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -0.5750 0.0000 -0.0193 -0.9269 -23.0616 0.4497 -32.1567 0.0000 0.0000 0.0000
0.0000 0.9051 0.0000 -0.0002 -1.0182 0.0000 0.0243 0.0007 0.0000 0.0000 0.0000
0.0364 0.0000 -0.9917 0.0001 -0.0000 -0.3213 0.0574 0.0017 0.0000 0.0000 0.0000
1.0000 -0.0265 0.0257 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.6967 0.7174 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000
B2 =
0.0000 0.0000 -0.7331 0.1315
0.0000 -0.1755 0.0000 0.0000
0.0000 0.0000 -0.0319 -0.0620
0.0000 0.1737 0.0000 0.0000
0.0000 -0.0021 0.0000 0.0000
0.0000 0.0000 0.0003 0.0008
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
Appendix C TS Fuzzy Augmented Model II 190
A3 =
-3.3910 0.0003 1.3693 0.0000 0.0000 -45.5472 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -1.1801 -0.0029 0.0001 -0.1877 0.0000 -0.0631 0.0000 0.0000 0.0000 0.0000
-0.0688 0.0025 -0.4598 0.0000 0.0000 8.5921 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -4.0169 0.0000 -0.0427 -71.6334 -23.0616 -3.8999 -31.9601 0.0000 0.0000 0.0000
0.0000 0.9010 0.0000 -0.0004 -1.0101 0.0000 0.0251 0.0071 0.0000 0.0000 0.0000
0.1381 0.0000 -0.9826 0.0001 -0.0016 -0.3140 0.0572 0.0017 0.0000 0.0000 0.0000
1.0000 -0.0265 0.0257 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.6967 0.7174 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 -0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000
B3 =
0.0000 0.0000 -0.7087 0.1261
0.0000 -0.1755 0.0000 0.0000
0.0000 0.0000 -0.0290 -0.0613
0.0000 0.0450 0.0000 0.0000
0.0000 -0.0022 0.0000 0.0000
0.0000 0.0000 0.0003 0.0008
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
A4 =
-3.2408 0.0003 -0.0600 0.0000 0.0000 -15.6934 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -0.3779 -0.0029 -0.0008 1.9109 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0159 0.0025 -0.4696 0.0000 0.0000 8.9211 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 2.6826 0.0000 -0.0236 42.9300 0.0000 0.0000 -32.0093 0.0000 0.0000 0.0000
0.0000 0.9157 0.0000 -0.0000 -1.0651 0.0000 0.0000 -0.0064 0.0000 0.0000 0.0000
-0.0641 0.0000 -0.9908 0.0000 0.0000 -0.3135 0.0640 0.0000 0.0000 0.0000 0.0000
1.0000 0.0000 0.0369 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000
B4 =
0.0000 0.0000 -0.7377 0.1355
0.0000 -0.1553 0.0000 0.0000
0.0000 0.0000 -0.0344 -0.0620
0.0000 0.2820 0.0000 0.0000
0.0000 -0.0021 0.0000 0.0000
0.0000 0.0000 0.0003 0.0008
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
Appendix C TS Fuzzy Augmented Model II 191
A5 =
-3.6776 0.0003 0.6645 0.0000 0.0000 -30.6426 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -1.0772 -0.0029 0.0000 0.8221 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0254 0.0025 -0.4765 0.0000 0.0000 8.5416 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -0.5750 0.0000 -0.0193 8.8168 0.0000 0.0000 -32.1700 0.0000 0.0000 0.0000
0.0000 0.9051 0.0000 -0.0003 -1.0189 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0364 0.0000 -0.9917 0.0000 0.0000 -0.3220 0.0640 0.0000 0.0000 0.0000 0.0000
1.0000 0.0000 0.0369 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000
B5 =
0.0000 0.0000 -0.7331 0.1315
0.0000 -0.1755 0.0000 0.0000
0.0000 0.0000 -0.0319 -0.0620
0.0000 0.1737 0.0000 0.0000
0.0000 -0.0021 0.0000 0.0000
0.0000 0.0000 0.0003 0.0008
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
A6 =
-3.3910 0.0003 1.3693 0.0000 0.0000 -45.5472 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -1.1801 -0.0029 0.0001 0.3169 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0688 0.0025 -0.4598 0.0000 0.0000 8.5921 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -4.0169 0.0000 -0.0427 -27.1270 0.0000 0.0000 -32.0093 0.0000 0.0000 0.0000
0.0000 0.9010 0.0000 -0.0005 -1.0181 0.0000 0.0000 0.0064 0.0000 0.0000 0.0000
0.1381 0.0000 -0.9826 0.0000 0.0000 -0.3166 0.0640 0.0000 0.0000 0.0000 0.0000
1.0000 0.0000 0.0369 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000
B6 =
0.0000 0.0000 -0.7087 0.1261
0.0000 -0.1755 0.0000 0.0000
0.0000 0.0000 -0.0290 -0.0613
0.0000 0.0450 0.0000 0.0000
0.0000 -0.0022 0.0000 0.0000
0.0000 0.0000 0.0003 0.0008
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
Appendix C TS Fuzzy Augmented Model II 192
A7 =
-3.2408 0.0003 -0.0600 0.0000 0.0000 -15.6934 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -0.3779 -0.0029 -0.0008 2.5382 0.0000 0.0784 0.0000 0.0000 0.0000 0.0000
0.0159 0.0025 -0.4696 0.0000 0.0000 8.9211 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 2.6825 0.0000 -0.0236 76.3394 23.0616 4.9446 -32.0320 0.0000 0.0000 0.0000
0.0000 0.9157 0.0000 -0.0000 -1.0911 0.0000 -0.0275 -0.0057 0.0000 0.0000 0.0000
-0.0641 0.0000 -0.9908 -0.0001 -0.0016 -0.3147 0.0572 -0.0017 0.0000 0.0000 0.0000
1.0000 0.0265 0.0257 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.6967 -0.7174 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 -0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000
B7 =
0.0000 0.0000 -0.7377 0.1355
0.0000 -0.1553 0.0000 0.0000
0.0000 0.0000 -0.0344 -0.0620
0.0000 0.2820 0.0000 0.0000
0.0000 -0.0021 0.0000 0.0000
0.0000 0.0000 0.0003 0.0008
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
A8 =
-3.6776 0.0003 0.6645 0.0000 0.0000 -30.6426 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -1.0772 -0.0029 0.0000 0.8221 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0254 0.0025 -0.4765 0.0000 0.0000 8.5416 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -0.5750 0.0000 -0.0193 -0.9269 23.0616 -0.4497 -32.1567 0.0000 0.0000 0.0000
0.0000 0.9051 0.0000 -0.0002 -1.0182 0.0000 -0.0243 0.0007 0.0000 0.0000 0.0000
0.0364 0.0000 -0.9917 -0.0001 0.0000 -0.3213 0.0574 -0.0017 0.0000 0.0000 0.0000
1.0000 0.0265 0.0257 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.6967 -0.7174 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000
B8 =
-3.6776 0.0003 0.6645 0.0000
0.0000 -1.0772 -0.0029 0.0000
-0.0254 0.0025 -0.4765 0.0000
0.0000 -0.5750 0.0000 -0.0193
0.0000 0.9051 0.0000 -0.0002
0.0364 0.0000 -0.9917 -0.0001
1.0000 0.0265 0.0257 0.0000
0.0000 0.6967 -0.7174 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
Appendix C TS Fuzzy Augmented Model II 193
A9 =
-3.3910 0.0003 1.3693 0.0000 0.0000 -45.5472 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -1.1801 -0.0029 0.0001 -0.1877 0.0000 0.0631 0.0000 0.0000 0.0000 0.0000
-0.0688 0.0025 -0.4598 0.0000 0.0000 8.5921 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 -4.0169 0.0000 -0.0427 -71.6334 23.0616 3.8999 -31.9601 0.0000 0.0000 0.0000
0.0000 0.9010 0.0000 -0.0004 -1.0101 0.0000 -0.0251 0.0071 0.0000 0.0000 0.0000
0.1381 0.0000 -0.9826 -0.0001 0.0016 -0.3140 0.0572 -0.0017 0.0000 0.0000 0.0000
1.0000 0.0265 0.0257 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.6967 -0.7174 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000 -0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000
B9 =
-3.3910 0.0003 1.3693 0.0000
0.0000 -1.1801 -0.0029 0.0001
-0.0688 0.0025 -0.4598 0.0000
0.0000 -4.0169 0.0000 -0.0427
0.0000 0.9010 0.0000 -0.0004
0.1381 0.0000 -0.9826 -0.0001
1.0000 0.0265 0.0257 0.0000
0.0000 0.6967 -0.7174 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 0.0000 0.0000
Resulting Weighting Functions:
hi(α, φ) =MαiMφi , i = 1, 2, 3,
where
Mα1 =
1 if α ≤ −0.12.5 ∗ α− 0.5 if − 0.1 < α < 0.1
0 if α ≥ 0.1
Mα2 =
0 if α ≤ −0.15 ∗ α + 1 if − 0.1 < α ≤ 0
−5 ∗ α + 1 if 0 < α < 0.1
0 if α ≥ 0.1
Mα3 =
0 if α ≤ −0.12.5 ∗ α + 0.5 if − 0.1 < α < 0.1
1 if α ≥ 0.1
Mφ1 =
1 if φ ≤ −0.82.5 ∗ φ− 0.5 if − 0.8 < φ < 0.8
0 if φ ≥ 0.8
Mφ2 =
0 if φ ≤ −0.85 ∗ φ+ 1 if − 0.8 < φ ≤ 0
−5 ∗ φ+ 1 if 0 < φ < 0.8
0 if φ ≥ 0.8
Mφ3 =
0 if φ ≤ −0.82.5 ∗ φ+ 0.5 if − 0.8 < φ < 0.8
1 if φ ≥ 0.8
194
Appendix D
Gain Scheduling Control for
Tracking (φ, θ)
The Fuzzy controller I is described by
u = Kx,
K =9∑
i=1
hi(x)Ki,
where hi(x) are the same as those from the TS model in Appendix B and
K1 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.3687 -88.9811 -18.4567 0.0187 13.6331 -10.5508 2.5302 -785.9696 3.5997 6.6243 -3148.3817 0.0934
-19.6926 -1.8319 -0.6296 -0.0295 -0.0409 37.8419 -218.7384 -19.7345 -7.7237 -978.4534 -81.3664 -0.2049
3.8462 -6.8619 -9.9549 -0.1443 -0.2180 -41.8981 46.5754 -84.7639 -36.9614 206.3615 -284.7313 -0.9743
,
K2 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0003 -90.1064 0.0168 0.0050 13.6127 0.0021 -0.0020 -786.3111 -0.0005 -0.0055 -3162.2777 -0.0000
-19.7074 0.0019 0.0217 0.0000 -0.0003 38.8846 -218.8523 0.0107 -6.9331 -978.7325 0.0272 -0.2051
3.7759 0.0062 -7.0133 0.0000 -0.0013 -36.7352 46.0509 0.0407 -33.2557 205.1407 0.0454 -0.9787
,
K3 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.3692 -88.9746 18.4896 0.0187 13.6329 10.5554 -2.5340 -785.9653 -3.6012 -6.6347 -3148.3697 -0.0934
-19.6925 1.8355 -0.6319 0.0295 0.0403 37.8412 -218.7380 19.7584 -7.7239 -978.4526 81.4256 -0.2049
3.8465 6.8739 -9.9658 0.1443 0.2154 -41.9025 46.5769 84.8575 -36.9619 206.3649 284.8479 -0.9743
,
K4 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.4269 -88.9861 -18.4282 0.0193 13.6407 -10.4395 3.3450 -785.9401 3.6987 11.2019 -3148.3571 0.0931
-19.7031 -1.5915 -4.0668 -0.0284 -0.0530 35.9706 -219.1541 -22.6468 -7.3719 -981.4157 -90.9952 -0.1897
3.5044 -6.8992 -8.9633 -0.1507 -0.2986 -42.1319 43.0282 -84.0728 -38.2344 191.5668 -282.0773 -0.9774
,
K5 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0004 -90.1064 0.0168 0.0050 13.6128 0.0020 -0.0030 -786.3134 -0.0005 -0.0100 -3162.2777 -0.0000
-19.7180 0.0015 -3.4165 0.0000 -0.0003 37.0166 -219.2708 0.0133 -6.6003 -981.7266 0.0400 -0.1903
3.4336 0.0062 -6.0098 0.0000 -0.0013 -36.7581 42.4945 0.0391 -34.2506 190.2970 0.0403 -0.9817
,
K6 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.4277 -88.9797 18.4611 0.0193 13.6405 10.4439 -3.3509 -785.9359 -3.7003 -11.2215 -3148.3451 -0.0931
-19.7030 1.5945 -4.0690 0.0284 0.0525 35.9699 -219.1537 22.6759 -7.3721 -981.4146 91.0801 -0.1897
3.5047 6.9113 -8.9741 0.1507 0.2960 -42.1363 43.0298 84.1632 -38.2349 191.5708 282.1837 -0.9774
,
195
K7 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.4849 -88.9933 -18.3769 0.0193 13.6401 -10.3015 4.1588 -785.8823 3.7354 15.8037 -3148.2859 0.0926
-19.7041 -1.3474 -7.5631 -0.0264 -0.0439 34.0512 -219.4880 -25.6273 -6.8772 -984.1375 -100.7489 -0.1745
3.1617 -6.9253 -8.1083 -0.1525 -0.2730 -42.7041 39.4722 -83.6222 -38.9124 176.7022 -279.5447 -0.9803
,
K8 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0004 -90.1065 0.0167 0.0050 13.6147 0.0018 -0.0040 -786.3172 -0.0005 -0.0146 -3162.2777 -0.0000
-19.7197 0.0012 -6.9156 0.0000 -0.0002 35.0816 -219.6131 0.0160 -6.1687 -984.5057 0.0531 -0.1754
3.0907 0.0062 -5.1514 0.0000 -0.0013 -37.2067 38.9350 0.0377 -34.8552 175.3526 0.0359 -0.9845
,
K9 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.4858 -88.9870 18.4096 0.0193 13.6399 10.3057 -4.1668 -785.8781 -3.7369 -15.8325 -3148.2737 -0.0927
-19.7041 1.3497 -7.5653 0.0264 0.0435 34.0507 -219.4876 25.6616 -6.8773 -984.1362 100.8599 -0.1745
3.1620 6.9376 -8.1188 0.1525 0.2704 -42.7083 39.4739 83.7100 -38.9129 176.7069 279.6423 -0.9803
.
The Fuzzy controller II is described by
u = Kx,
K =9∑
i=1
hi(x)Ki,
where hi(x) are the same as those from the TS model in Appendix B and
K1 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0429 -19.9092 -2.6266 0.0003 12.3847 -10.6580 0.3252 -76.2711 0.5021 0.2666 -99.6939 0.0025
-4.2508 -0.2817 -9.0478 -0.0005 0.0622 27.4442 -23.8241 -0.9093 -0.7013 -31.4126 -1.7274 -0.0036
0.0122 -0.9190 -1.7946 -0.0067 0.1970 -7.4881 0.9075 -5.5614 -6.4708 3.6299 -7.6254 -0.0313
,
K2 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0007 -20.1568 0.0089 -0.0002 12.4197 0.0388 -0.0041 -76.0442 -0.0018 -0.0046 -100.0000 -0.0000
-4.2518 -0.0013 -8.9981 -0.0000 -0.0002 27.5275 -23.8278 0.0124 -0.7195 -31.4152 0.0177 -0.0036
0.0071 0.0035 -1.5758 -0.0000 -0.0011 -7.0617 0.8871 0.0184 -6.6230 3.6171 0.0251 -0.0314
,
K3 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0443 -19.9068 2.6437 0.0003 12.3844 10.7364 -0.3335 -76.2721 -0.5056 -0.2760 -99.6893 -0.0025
-4.2508 0.2790 -9.0493 0.0005 -0.0626 27.4402 -23.8240 0.9343 -0.7012 -31.4125 1.7629 -0.0036
0.0122 0.9257 -1.7981 0.0067 -0.1991 -7.4944 0.9076 5.5988 -6.4704 3.6301 7.6764 -0.0313
,
K4 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0563 -19.9074 -2.6121 0.0003 12.4028 -10.7174 0.4147 -76.2638 0.5181 0.4379 -99.6853 0.0025
-4.2525 -0.2249 -9.6707 -0.0004 0.0530 24.9603 -23.8203 -1.3358 -0.5999 -31.4698 -2.1331 -0.0030
-0.0403 -0.9233 -1.7735 -0.0068 0.1907 -7.3174 0.5850 -5.6031 -6.6648 3.0754 -7.6343 -0.0314
,
K5 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0007 -20.1554 0.0089 -0.0002 12.4372 0.0383 -0.0043 -76.0535 -0.0018 -0.0051 -100.0000 -0.0000
-4.2530 -0.0014 -9.6209 -0.0000 -0.0002 25.0637 -23.8211 0.0134 -0.6179 -31.4745 0.0186 -0.0031
-0.0456 0.0035 -1.5526 -0.0000 -0.0010 -6.8960 0.5629 0.0183 -6.8259 3.0588 0.0250 -0.0315
,
K6 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0577 -19.9050 2.6291 0.0003 12.4024 10.7950 -0.4233 -76.2647 -0.5217 -0.4481 -99.6806 -0.0025
-4.2525 0.2220 -9.6721 0.0004 -0.0533 24.9560 -23.8202 1.3627 -0.5997 -31.4697 2.1704 -0.0030
-0.0402 0.9300 -1.7770 0.0068 -0.1928 -7.3237 0.5851 5.6402 -6.6644 3.0756 7.6851 -0.0314
,
K7 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0694 -19.9062 -2.5907 0.0003 12.4204 -10.7558 0.5025 -76.2520 0.5286 0.6100 -99.6747 0.0025
-4.2529 -0.1680 -10.3002 -0.0003 0.0447 22.4445 -23.8097 -1.7683 -0.4872 -31.5161 -2.5413 -0.0025
-0.0924 -0.9251 -1.7787 -0.0070 0.1892 -7.2874 0.2663 -5.6736 -6.7936 2.5226 -7.6487 -0.0314
,
K8 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0008 -20.1540 0.0088 -0.0002 12.4546 0.0379 -0.0045 -76.0626 -0.0018 -0.0055 -100.0000 -0.0000
-4.2545 -0.0016 -10.2535 -0.0000 -0.0001 22.5833 -23.8198 0.0143 -0.5041 -31.5241 0.0195 -0.0025
-0.0979 0.0035 -1.5555 -0.0000 -0.0010 -6.8690 0.2426 0.0183 -6.9596 2.4959 0.0249 -0.0315
,
196
K9 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0710 -19.9037 2.6076 0.0003 12.4201 10.8324 -0.5116 -76.2528 -0.5322 -0.6210 -99.6698 -0.0025
-4.2529 0.1647 -10.3015 0.0003 -0.0449 22.4399 -23.8095 1.7972 -0.4870 -31.5159 2.5804 -0.0025
-0.0924 0.9318 -1.7822 0.0070 -0.1913 -7.2937 0.2664 5.7106 -6.7932 2.5229 7.6993 -0.0314
.
197
Appendix E
Gain Scheduling Control for
Tracking (α, β, φ)
The Fuzzy controller I is described by
u = Kx,
K =9∑
i=1
hi(x)Ki,
where hi(x) are the same as those from the TS model in Appendix C and
K1 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0148 -319.8224 0.0681 0.0035 -1163.5845 0.8604 -0.0505 1.9771 -3161.3123 1.6918 24.6493
-30.4307 0.2319 -26.0497 0.0026 -58.6892 189.3081 -451.9232 0.0475 -234.3342 794.6246 -3059.9151
9.9318 0.0590 -137.7025 0.0137 23.7448 794.3268 177.1732 0.1982 78.3093 3060.8076 794.2567
,
K2 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0169 -315.0360 0.0036 0.0765 -1177.1797 -0.2862 -0.4803 -0.5077 -3161.5351 -0.1421 21.6692
-30.0387 -0.0073 -26.9467 0.0036 -55.9749 183.8534 -470.4688 0.0825 -213.7533 539.8986 -3115.1148
9.2292 0.0584 -138.0877 0.0158 8.9910 797.3002 85.1577 0.1750 35.5965 3115.8481 539.7814
,
K3 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0147 -314.3092 -0.0603 0.1460 -1180.6002 -1.2998 -0.0353 -2.8390 -3161.3692 -1.9262 23.8904
-30.8471 0.1321 -28.2105 0.0042 -65.1202 179.9095 -489.3780 0.2854 -239.6715 279.8238 -3148.9608
8.8470 0.0081 -138.4025 0.0167 -3.9576 796.2902 -9.4219 0.1731 1.9596 3149.8669 279.8894
,
K4 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0001 -319.7585 0.0140 0.0172 -1168.7053 -0.0128 -0.0026 2.1826 -3162.2777 0.0254 -0.0121
-30.4300 0.0267 -26.2300 -0.0000 0.0986 187.7532 -451.9922 -0.0002 0.1805 788.3355 -3062.4381
9.9154 0.0543 -137.6344 -0.0000 0.2210 795.2099 176.9594 -0.0006 0.2163 3062.4381 788.3355
,
K5 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0000 -315.0222 0.0123 0.0911 -1177.2192 -0.0133 -0.0007 -0.3111 -3162.2777 0.0164 -0.0091
-30.0390 0.0196 -27.1345 -0.0000 0.0741 182.1268 -470.4097 0.0000 0.1174 533.5920 -3116.9343
9.2140 0.0422 -138.0574 -0.0000 0.1837 797.5140 84.9014 0.0000 0.1462 3116.9343 533.5920
,
K6 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0001 -314.3445 0.0123 0.1604 -1178.8704 -0.0140 0.0009 -2.7210 -3162.2777 0.0143 -0.0082
-30.8474 0.0176 -28.4079 -0.0000 0.0664 177.8635 -489.1065 0.0002 0.0943 273.5094 -3150.4274
8.8325 0.0417 -138.4122 -0.0000 0.1817 795.7978 -9.5868 0.0006 0.1354 3150.4274 273.5094
,
198
K7 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0146 -319.8224 -0.0400 0.0035 -1163.5841 -0.8852 0.0452 1.9771 -3161.3107 -1.6386 -24.6734
-30.4307 -0.1785 -26.0494 -0.0026 58.8869 189.3112 -451.9230 -0.0479 234.7015 794.6340 -3059.9099
9.9318 0.0495 -137.7025 -0.0137 -23.3008 794.3259 177.1738 -0.1992 -77.8563 3060.8054 794.2694
,
K8 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0169 -315.0360 0.0210 0.0765 -1177.1794 0.2597 0.4789 -0.5077 -3161.5338 0.1750 -21.6875
-30.0387 0.0466 -26.9467 -0.0036 56.1238 183.8562 -470.4689 -0.0824 213.9895 539.9058 -3115.1120
9.2292 0.0261 -138.0877 -0.0158 -8.6237 797.2997 85.1578 -0.1748 -35.3042 3115.8468 539.7907
,
K9 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0146 -314.3092 0.0849 0.1460 -1180.5998 1.2715 0.0372 -2.8390 -3161.3678 1.9538 -23.9068
-30.8471 -0.0969 -28.2103 -0.0042 65.2536 179.9125 -489.3784 -0.2850 239.8591 279.8324 -3148.9587
8.8470 0.0752 -138.4026 -0.0168 4.3192 796.2900 -9.4215 -0.1718 -1.6996 3149.8660 279.9001
.
The Fuzzy controller II is described by
u = Kx,
K =9∑
i=1
hi(x)Ki,
where hi(x) are the same as those from the TS model in Appendix C and
K1 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0160 -319.8224 0.0866 0.0035 -1163.5966 2.8082 -0.1165 1.9771 -3161.3542 113.2555 24.1655
-61.7805 0.1494 -452.2293 0.0416 -51.3543 44345.2652 -411.5411 0.7900 -198.5068 1791246.5206 -2605.2822
-21.1176 -0.0498 -673.2023 0.0624 37.1016 56428.0118 334.3478 1.1332 137.8165 2606038.3294 1790.7163
,
K2 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0167 -315.0360 -0.0045 0.0765 -1177.1799 -0.3313 -0.4806 -0.5077 -3161.5357 -9.5502 21.6613
-24.5328 -0.0135 -155.7431 0.0150 -54.7177 16631.6208 -452.8960 0.3100 -209.1669 823339.3640 -3052.4969
40.6684 0.0220 -879.7431 0.0812 14.8781 73255.1039 178.8092 1.4856 56.3057 3053213.4368 823.1468
,
K3 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0132 -314.3093 -0.0953 0.1460 -1180.6184 -2.8532 -0.1286 -2.8389 -3161.4289 -113.4439 23.1670
-68.9775 0.1435 188.6482 -0.0161 -63.9774 -16783.9465 -556.7211 -0.0941 -230.2015 -340483.2103 -3143.0514
110.8765 0.0146 -819.4763 0.0767 -8.2526 71176.9056 -2.1045 1.4340 -26.0716 3143894.2500 -340.3836
,
K4 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0004 -319.7585 0.0057 0.0172 -1168.7053 0.3638 0.0033 2.1826 -3162.2777 7.9588 0.0257
-61.7686 -0.0103 -451.9914 0.0000 -0.0691 44342.9116 -408.4676 0.0001 -0.1670 1791556.5457 -2605.8252
-21.0894 0.0058 -673.3556 -0.0000 0.0813 56429.7647 339.1063 -0.0001 0.2114 2605825.2326 1791.5565
,
K5 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0003 -315.0222 0.0038 0.0911 -1177.2192 0.3667 -0.0000 -0.3111 -3162.2777 7.2716 -0.0092
-24.5479 0.0123 -155.3781 -0.0000 0.0485 16631.8404 -451.8088 0.0000 0.1074 823883.5124 -3053.0666
40.7040 -0.0057 -879.8141 -0.0000 0.0136 73252.9325 184.7772 0.0000 0.0463 3053066.6481 823.8835
,
K6 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0006 -314.3445 0.0061 0.1604 -1178.8704 0.3068 -0.0026 -2.7210 -3162.2777 8.6844 -0.0350
-69.0080 0.0293 188.8630 -0.0000 0.1528 -16787.9568 -557.7728 0.0003 0.3387 -340247.1483 -3143.9198
110.9017 0.0020 -819.4280 -0.0000 0.0479 71173.4297 3.6736 0.0001 0.1240 3143919.8268 -340.2471
,
K7 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0153 -319.8225 -0.0752 0.0035 -1163.5977 -2.0801 0.1230 1.9771 -3161.3581 -97.3450 -24.1142
-61.7805 -0.1701 -452.2293 -0.0416 51.2175 44345.2677 -411.5408 -0.7897 198.1745 1791246.4609 -2605.2848
-21.1176 0.0615 -673.2023 -0.0624 -36.9405 56428.0101 334.3475 -1.1334 -137.3950 2606038.3768 1790.7195
,
K8 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0172 -315.0360 0.0122 0.0765 -1177.1796 1.0645 0.4805 -0.5077 -3161.5344 24.0847 -21.6799
-24.5328 0.0380 -155.7431 -0.0150 54.8153 16631.6218 -452.8961 -0.3100 209.3843 823339.3165 -3052.4954
40.6684 -0.0335 -879.7431 -0.0812 -14.8511 73255.1036 178.8092 -1.4856 -56.2138 3053213.4489 823.1474
,
199
K9 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0145 -314.3092 0.1075 0.1460 -1180.6170 3.4665 0.1234 -2.8389 -3161.4238 130.8076 -23.2371
-68.9776 -0.0849 188.6482 0.0161 64.2843 -16783.9444 -556.7217 0.0949 230.8797 -340483.2893 -3143.0464
110.8765 -0.0105 -819.4763 -0.0767 8.3488 71176.9058 -2.1046 -1.4338 26.3196 3143894.2346 -340.3818
.
The Fuzzy controller III is described by
u = Kx,
K =9∑
i=1
hi(x)Ki,
where hi(x) are the same as those from the TS model in Appendix C and
K1 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0270 -992.3972 0.2408 0.0012 -706.0301 3.0279 0.3083 6.2426 -998.5364 -0.7241 17.1026
-19.2330 0.1824 -146.2877 0.0136 -17.5950 2820.4547 -67.5476 0.2485 -36.3580 23419.7846 -212.1772
-0.2718 -0.1340 -224.2869 0.0212 20.4004 2774.3344 101.1619 0.3584 40.0386 21248.8515 233.8553
,
K2 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0268 -988.5698 -0.1052 0.2614 -735.5971 -6.7014 -0.1193 -9.8859 -998.7195 -0.2662 15.9980
-10.9429 0.2101 -44.8607 0.0096 -26.1542 567.0138 -101.8420 0.3044 -50.1632 4104.7996 -313.1508
12.6292 -0.0002 -307.5862 0.0284 3.4218 4263.5254 32.0601 0.4727 6.5585 31355.2327 40.9956
,
K3 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0210 -987.8328 -0.4633 0.4812 -760.8022 -11.4936 1.5729 -19.6861 -997.8876 2.2954 20.5433
-24.1816 0.2720 66.2055 0.0039 -33.4559 -1811.9725 -107.0155 1.5549 -56.1619 -15924.4931 -272.6273
23.8109 0.1637 -251.3802 0.0302 -20.2271 3415.6758 -47.6251 1.3891 -32.6516 27320.5145 -158.9098
,
K4 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
-0.0001 -992.1874 0.0087 0.0429 -715.6615 0.0864 0.0015 5.4011 -1000.0000 0.7647 0.0078
-19.1997 0.0011 -145.9548 -0.0000 0.0005 2816.7570 -66.5739 -0.0000 0.0014 23406.5985 -212.6338
-0.2262 0.0028 -224.6691 -0.0000 0.0254 2778.8986 102.7283 -0.0002 0.0345 21263.3757 234.0660
,
K5 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0003 -988.5346 0.0049 0.3309 -734.9544 0.1132 0.0012 -11.8979 -1000.0000 0.9023 0.0006
-10.9563 0.0021 -44.1816 -0.0000 0.0036 559.0578 -101.5542 0.0001 0.0019 4077.1723 -313.5884
12.7014 0.0007 -307.6620 -0.0000 0.0209 4264.2442 34.0962 0.0003 0.0285 31358.8371 40.7717
,
K6 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0006 -988.0003 0.0076 0.5436 -747.9396 0.0682 0.0008 -23.5632 -1000.0000 0.6384 -0.0056
-24.2233 0.0034 66.5865 -0.0000 0.0084 -1817.5296 -107.2320 0.0003 0.0051 -15939.8845 -273.1154
23.8456 0.0016 -251.1299 -0.0000 0.0228 3411.5403 -45.7355 0.0007 0.0264 27311.5375 -159.3988
,
K7 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0267 -992.3972 -0.2234 0.0012 -706.0311 -2.8536 -0.3052 6.2424 -998.5390 2.2539 -17.0874
-19.2330 -0.1803 -146.2877 -0.0136 17.5964 2820.4563 -67.5475 -0.2486 36.3615 23419.7864 -212.1772
-0.2718 0.1397 -224.2869 -0.0212 -20.3510 2774.3327 101.1618 -0.3588 -39.9705 21248.8494 233.8565
,
K8 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0275 -988.5698 0.1150 0.2614 -735.5972 6.9282 0.1216 -9.8861 -998.7196 2.0709 -15.9971
-10.9429 -0.2059 -44.8606 -0.0096 26.1620 567.0155 -101.8421 -0.3042 50.1678 4104.8014 -313.1507
12.6292 0.0016 -307.5862 -0.0284 -3.3800 4263.5248 32.0601 -0.4721 -6.5017 31355.2324 40.9965
,
K9 =
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0222 -987.8328 0.4784 0.4812 -760.8010 11.6291 -1.5712 -19.6861 -997.8853 -1.0189 -20.5547
-24.1816 -0.2651 66.2055 -0.0039 33.4732 -1811.9703 -107.0158 -1.5545 56.1726 -15924.4915 -272.6271
23.8109 -0.1604 -251.3802 -0.0302 20.2737 3415.6766 -47.6252 -1.3879 32.7045 27320.5155 -158.9087
.
200
Author’s Publications
Journal publications
[1] Qing-Guo. Wang, C.C Hang, Yongsheng Yang and J.B.He ”Quantitative Ro-
bust Stability Analysis and PID Controller Design”, IEE Proceeding in Control
Theory Application. Vol. 149, No.1, January, 2002.
[2] Yang, Yong-Sheng, Wang, Qing-Guo, Dai, WZ, ”Robust PID controller design
for gain and phase margins”, Journal of Chemical Engineering of Japan, Vol.35,
No.9, pp.874-879, September, 2002.
[3] Wang, Qing-Guo and Yang, Yongsheng, ”Transfer Function Matrix Approach
to Decoupling Problem with Stability”, Systems and Control Letters, Vol. 47, Issue
2, Pages 103-110, October, 2002.
[4] Yongsheng Yang and Qing-Guo Wang, ”Decoupling Control Design via LMI”,
submitted to IEE Proceeding in Control Theory Application.
[5] Yong-Sheng Yang, Qiangguo Pu and Qing-Guo Wang, ” PI Controller Design
for State Time-Delay Systems: An ILMI Approach”, submitted to Automatica.
[6] Yong-Sheng Yang, Qing-Guo Wang and C. C. Hang, ”Modified Smith Predictor
Control for Disturbance Rejection with Unstable Processes”, submitted to Journal
201
Author’s Publications 202
of Process Control.
[7] Wang, Qing-Guo, Yong-sheng Yang and Xiang Lu, ”Robust IMC Design via
LMI/BMI Optimization in Frequency Domain”, submitted to Industrial & Engi-
neering Chemistry Research.
[8] Wang, Qing-Guo, Yong-sheng Yang and Yong Zhang ”Robust IMC Design via
LMI optimization”, submitted to Journal of Chemical Engineering of Japan.
Conference Publications
[1] Yongsheng Yang and Q.G. Wang, ”Robust PID Controller Design for Uncertain
Process”, International Conference on Computational Intelligence, Robotics and
Autonomous Systems, Singapore, Nov. 2001.
[2] Bin Huang; Qing-Guo Wang; Yong-Sheng Yang, ”An H∞ approach to decou-
pling control design”, Proceedings of the 4th World Congress on Intelligent Control
and Automation 2002, Volume 4, Page 2878-2882.
[3] Wang, Qing-Guo; Ru, He; Yang, Yongsheng, ”Enhancement of PID control for
complex processes”, IEEE International Symposium on Intelligent Control - Pro-
ceedings, 2002, Pages 592-595.
[4] Yong-Sheng Yang, Xue-Ping Yang, Qing-GuoWang, C. C. Hang and Chong Lin,
”IMC based Control System Design for Unstable Process”, International Sympo-
sium on Design, Operation and Control of Chemical Plants, December 4 - 6, 2002,
Taipei, Taiwan.
Author’s Publications 203
[5] Qing-Guo Wang, Yong-Sheng Yang, Xiang Lu, ”Robust IMC Controller Design
In Frequency Domain”, First Humanoid, Nanotechnology, Information Technology,
Communication and Control Environment and Management (HNICEM) Interna-
tional Conference, March 29-31, 2003, Manila, Philippines.
[6] Wang, Qing-Guo, Yong-sheng Yang and Qiangguo Pu, ”Robust IMC Design via
multiobjective optimization”, submitted to the Second International Conference on
Computational Intelligence, Robotics and Autonomous Systems, 2003, Singapore.