JAMRIS 2010 Vol 4 No 3

Editor-in-Chief

Co-Editors:

Janusz Kacprzyk

Dimitar Filev

Kaoru Hirota

Witold Pedrycz

Roman Szewczyk

(Systems Research Institute, Polish Academy of Sciences , Poland)

(Research & Advanced Engineering, Ford Motor Company, USA)

(Interdisciplinary Graduate School of Science and Engineering,

Tokyo Institute of Technology, Japan)

(ECERF, University of Alberta, Canada)

(PIAP, Warsaw University of Technology )

; PIAP

, Poland

(Polish Academy of Sciences; PIAP, Poland)

Editorial Office:

Al. Jerozolimskie 202, 02-486 Warsaw, POLANDTel. +48-22-8740109,

Chairman:

Industrial Research Institute for Automationand Measurements PIAP

Janusz KacprzykPlamen AngelovZenn BienAdam BorkowskiWolfgang BorutzkyOscar CastilloChin Chen ChangJorge Manuel Miranda DiasBogdan GabryśJan JabłkowskiStanisław KaczanowskiTadeusz KaczorekMarian P. KaźmierkowskiJózef KorbiczKrzysztof KozłowskiEckart KramerAndrew KusiakMark LastAnthony MaciejewskiKrzysztof Malinowski

[email protected]

Editorial Board:

(Lancaster University, UK)

(Korea Advanced Institute of Science and Technology, Korea)

(Polish Academy of Sciences, Poland)

(Fachhochschule Bonn-Rhein-Sieg, Germany)

(Tijuana Institute of Technology, Mexico)

(Feng Chia University, Taiwan)

(University of Coimbra, Portugal)

(Bournemouth University, UK)

(PIAP, Poland)

(PIAP, Poland)

(Warsaw University of Technology, Poland)


(University of Zielona Góra, Poland)

(Poznań University of Technology, Poland)

(Fachhochschule Eberswalde, Germany)

(University of Iowa, USA)

(Ben–Gurion University of the Negev, Israel)

(Colorado State University, USA)


Executive Editor:

Associate Editors:

Webmaster:

Proofreading:

Copyright and reprint permissionsExecutive Editor

Anna Ładan

Mariusz AndrzejczakKatarzyna Rzeplińska-Rykała

Tomasz Kobyliński

Urszula Wiączek

Andrzej MasłowskiTadeusz MissalaFazel NaghdyZbigniew NahorskiAntoni NiederlińskiWitold PedryczDuc Truong PhamLech PolkowskiAlain PruskiLeszek RutkowskiKlaus SchillingRyszard Tadeusiewicz

Stanisław TarasiewiczPiotr TatjewskiWładysław TorbiczLeszek TrybusRené WamkeueJanusz ZalewskiMarek ZarembaTeresa Zielińska

[email protected]

[email protected]

(PIAP, Poland)

(PIAP, Poland)

(PIAP, Poland)

(PIAP, Poland)

(University of Wollongong, Australia)

(Polish Academy of Science, Poland)

(Silesian University of Technology, Poland)

(University of Alberta, Canada)

(Cardiff University, UK)

(Polish-Japanese Institute of Information Technology, Poland)

(University of Metz, France)

(Częstochowa University of Technology, Poland)

(Julius-Maximilians-University Würzburg, Germany)

(AGH University of Science and Technology

in Kraków, Poland)

(University of Laval, Canada)


(Polish Academy of Sciences, Poland)

(Rzeszów University of Technology, Poland)

(University of Québec, Canada)

(Florida Gulf Coast University, USA)

(University of Québec, Canada)


JOURNAL of AUTOMATION, MOBILE ROBOTICS& INTELLIGENT SYSTEMS

All rights reserved © 1

Publisher:Industrial Research Institute for Automation and Measurements PIAP

If in doubt about the proper edition of contributions, please contact the Executive Editor. , excluding advertisements and descriptions of products.The Editor does not take the responsibility for contents of advertisements, inserts etc. The Editor reserves the right to make relevant revisions, abbreviations

and adjustments to the articles.

Articles are reviewed

DEPARTMENTS

IN THE SPOTLIGHT

EVENTS

2

JOURNAL of AUTOMATION, MOBILE ROBOTICS& INTELLIGENT SYSTEMSVOLUME 4, N° 3, 2010

CONTENTS

REGULAR PAPER

A comparative study of incremental algorithms forcomputing the inverse kinematics of redundantarticulated systems

A novel flexible micro assembly system:implementation and performance analysis

Intelligent PI controller and its applicationto dissolved oxygen tracking problem

Fish-like swimming prototype of mobile underwaterrobot

A mobile system for measurements of partialdischarges controlled by electroencephalographicwaves

Automatic generation of fuzzy inference systemsusing heuristic possibilistic clustering

Capacitive human presence sensor for safetyapplications

Sliding mode speed control for multi-motors system

Modelling and optimization of the force sensornetwork

D. Um, D. Ryu, B. Dong, D. Foor, K. Hawkins

T. Zubowicz, M.A. Brdys, R. Piotrowski

M. Malec, M. Morawski, J. Zając

A. Błachowicz, S. Paszkiel

D.A. Viattchenin

P. Frydrych, R. Szewczyk

B. Bouchiba, A. Hazzab, H. Glaoui, F. Med-Karim,I.K. Bousserhane

G. Bialic, M. Zmarzły, R. Stanisławski

A. Moussaoui, R. Otmani, A. Pruski

3

10

16

31

25

36

45

50

55

60

62

Abstract:

1. Introduction

2. Background

-

-

The comparative study that we present in this paperconcerns algorithms to determine the inverse kinematicsof redundant articulated systems with very high numberdegrees of freedom. These systems are highly non-linearand require special methods of resolution. The principleis to calculate the value of joint variables such as:

With the joint variables and [X] the constraintvector.

Many methods are proposed in the literature that wecan classify into three categories:

- Analytical methods;

The field of control of redundant articulated mechanical systems requires the use of algorithms to compute theinverse kinematics. The fields of animation, virtual realityor game in particular, are very interested in these algorithms. We propose in this paper a comparison betweenseveral algorithms of incremental type. The consideredapplication concerns the accessibility evaluation of anenvironment used by a handicapped person (an apartment,a house, an institution…). The physical disability involvesa particular characteristic of the human articulated structure that gives rise to constraints that must be taken intoaccount in computing the inverse kinematics.

-

-

The field of robotics has been the first that was interested in the problem of inverse kinematics. The articu-lated systems have long been made up of six degrees offreedom and their controls were based on linearizationmethods, which gave solutions consistent with expecta-tions. Currently the humanoid robots with high redun-dancy require more computing time. On the other hand,the animation of avatars requires the use of the sametypes of computing. We note that the type of algorithm ishighly dependent on the type of application envisaged.In this paper we contribute to propose algorithms forcomputing the inverse kinematics with constraints rela-ted to an application in the field of accessibility asses-sment. We analyse the relevance of each in a specificcontext.

Keywords: inverse kinematics, accessibility, physicaldisability.

�=f [(X])-1

�

- Linearization methods;- Optimization methods.

Each method has advantages and drawbacks. In general each depends by the application that you want to manage. The selection criteria are mostly the computingtime and/or the accuracy obtained. A redundant systemhas no single solution and other criteria are frequentlyused to converge toward a particular solution.

The analytical methods are used when the number ofvariables is not too important. [1] and [2] have developed methods adapted to a humanoid articulated structure at the arm level. This structure has led to a method ofsolving the inverse kinematics specific for obtaininga solution quickly. By cons, it is linked to this particularstructure and cannot extend to other structures.

Linearization methods are commonly used when theproblem is complex and non-linear. The method is toapproach the solution by successive increments considering that the system is linear around the operating point.The problem is solved by inverting a Jacobian matrix thatis usually singular implying to calculate its inverse by thepseudo inverse techniques whose computational cost isimportant. [3]

The optimization methods are the most interestingand often used when the number of variables is importantif we wish to obtain solutions meeting certain criteria.The principle involves formulating the problem as a costfunction minimisation problem. Many approaches havebeen developed which include the gradient descent withspecial adaptations [4], [5] and genetic algorithms [6].These algorithms are effective but can often lead to localminima. They can be very fast especially using BFGS-typemethods, which approximate calculation of the inversematrix.

The work presented in this paper concerns algorithmsthat are part of the class of optimization methods.

--

--

-

The work presented in this article is made in the con-text of a specific application that is to test the accessi-bility manipulation (Fig. 1) of a living place for a disabledperson. For this purpose the articulated system consists ofa humanoid-type structure that moves with a mobile basethat models a walker or a wheelchair. The displacementdevice is important since we must consider the physicalplacement of the articulated structure with its mobilebase in the environment. This application requires takinginto account several parameters:

3. Context

3.1. Introduction

A COMPARATIVE STUDY OF INCREMENTAL ALGORITHMS FOR COMPUTING

THE INVERSE KINEMATICS OF REDUNDANT ARTICULATED SYSTEMS

Abdelak Moussaoui, Rafaa Otmani, Alain Pruski

Received 9 ; accepted 26 May 2010.th April 2010 th

Journal of Automation, Mobile Robotics & Intelligent Systems VOLUME 4, N° 3 2010

Articles 3

- The number of degrees of freedom taking into accountthe human trunk, one arm and the mobility systemwhich is 24 degrees of freedom, as will see later;

- In essence, the human articulated system has limita-tions that we consider;

- We want verify only if a solution exists. We do not takeinto account other characteristics such as comfort orenergy expended in computing the solution. We be-lieve that if a solution exists then the person canreach the considered point;

- The accessibility evaluation is conducted in statics.We do not consider the motion type (e.g. the wheel-chair non holonomy is not taken into account). Onlythe admissible geometric placement, which is to saywithout intersection with the physical environment, isconsidered;

- The required computing accuracy is not very importantsince we can consider that the compliance of thehuman body compensates for errors.

The incremental algorithms that we study in this arti-cle require knowledge of the direct kinematics model. Thesystem is described by a biomechanical model of thehuman being with limits in the range of joint movement.These are taken into account in the algorithm of the in-verse kinematics, as we will detail later. The model we useis that proposed by [7] from which we extracted a modelwith 21 degrees of freedom from waist toward the tip of

Fig. 1. Accessibility of the environment with a wheelchair.

Fig. 2. The used kinematic joint structure.

3.2. Modelling

his right hand. We believe that mobility is achieved bya rectangular base corresponding to a wheelchair. Themotion device is modelled using three degrees of free-dom: a rotation and two translations. The frame positionof the root structure is situated at the height of the waistof a seated person and the area swept by the wheelchair isa rectangle. Figure 2 shows the kinematics chain of theglobal user with his mobile base.

The mathematical model is established from the mul-tiplication of Denavit-Hartenberg matrices [8] whose pro-totype is given below:

(1)And for n joints:

(2)

For any vector in the frame in homogeneouscoordinates, we will have in the world frame:

(3)

Two of the three algorithms that we analyse, work inan iterative manner that is to say that the joints are mo-dified one after the other in order to verify if the end-effector approaches the goal. This sequential aspect ofthe algorithm allows to speeds up the computation of thedirect model without having to reconsider all the joints.When we change one variable, only the variable corres-ponding matrix is modified.

We can write , by considering two matrices.

(4)

If we want to change the matrix corresponding to thevariable i we can write that

(5)or

(6)

With for the iteration computing. This method re-quires only three matrix multiplications instead of . Theinverse is directly given by the following expression:

(7)

The two equations (5) and (6) are used, the first whenthe variables are changed sequentially from to (from

V nR

DH n

qn

DH

n

0

0

-1

0

Journal of Automation, Mobile Robotics & Intelligent Systems

Articles4

VOLUME 4, N° 3 2010

��

�

�

��

�

��

��

1000

cossin0

sincossincoscossin

cossinsinsincoscos

iii

iiiiiii

iiiiiii

id

a

a

DH

nin DHDHDHDH ......0,0

nRnR VDHV ,00

ni,i0,n0, DHDH=DH ��1

1)( �� q

ni,qi

1+q

i

q

1i0,

1+q

n0,

1

,1

1

,0 * ��

� q

ni

q

i DHDH

DHDHDHDH=DH

1

1)( �� q

ni

1+q

iqi

q

i0,

1+q

n0,

1

,1,0

�� q

ni

1+q

i DHDH

DHDHDHDH=DH

��

�

�

��

�

�

��

��

�

1000

coscoscossinsinsin

sinsincoscossincos

0sincos

iiiiiii

iiiiiii

iii

d��

d��

a��

�1i=DH

The principle of this algorithm is to modify each valueof variables from root to tip in order to minimize themagnitude of the cost function . Unlike the CCD algo-rithm, the value of the increment Inc applied to each vari-able is not calculated but imposed.

The resulted value of is preserved if it is within theallowed range. The increment is calculated for eachjoint as :

(10)

with and the minimum and maximumlimits of the joint . IncrementRate can adjust the speedof the algorithm convergence. The parameters isvery important in two aspects of its sign and amplitudethat contribute to the speed of convergence. In the me-thods of gradient descent as Newton-Raphson, gradientmatrices and the inverse of the Hessian fulfil these roles.The optimization of these values can accelerate the con-vergence. In our case we modify the basic algorithm bystoring the sign of . For each variable we use thesame sign at the next iteration. Convergence is rapid initi-ally and then the variation becomes smaller with the pro-ximity of the solution. We propose a modification of thealgorithm by adjusting the step of the incrementdepending on the magnitude of the cost function ina non-linearly manner as in equation (11). Other adapta-tion functions could be applied:

ifthen (11)

A linear adjustment does not improve the speed ofconvergence. If the increment is sufficiently large,the distance change is rapidly becoming zero around thesolution. We use the cancellation of De to decrease thevalue of the increment . This algorithm is adaptableto any articulated structure. In this case, it is equation(5), which is used to calculate the direct model.

1. Initialise randomly the joint variables2. Do

2.1. Define the increment2.2. Do for each variable

2.2.1.Compute the distance between CurrentSolution and Goal such asif then keepElse

Computeif then keepElse(keep the original value)

3. While Stop Conditions not verified

Unlike the other two algorithms, this one treats allvariables simultaneously. In a range of values defi-ned by equation (10), we choose randomly incrementsthat are added to the current values of variables. If the

3.3.2. Incremental approximation algorithm (IAA)

3.3.3. The Random algorithm ApproximationAlgorithm (RAA)

�

�

�

��

��

� � �

��

i

i

i

i i

i

i i

i

i i

Inc

Inc(i)=(Max(i)-Min(i))*IncrementRate

Max(i) Min(i)

Inc(i)

Inc

Inc(i)

( =0)IncrementRate=IncrementRate/2

Inc(i)

Inc(i)

Inc (i)

= + Inc (i)

=f([ ])-[X]( ) <0)

= 2* Inc (i)=f([ ])-[X]

( <0)= + Inc (i)

Inc(i)

i

i

i

��

��

��

��

�

�

�

�

�

�

�

�

�

�

�

root to tip) and the second when the variables are chan-ged from to (root to tip). The proposed algorithms usethe two equations.

We propose in this section three algorithms whoseperformance depends on usage. All these algorithms areof incremental type with optimising a cost function,which corresponds to the error between the point to bereached, and the current point. This point is defined bythe position of the hand and the orientation of the twovectors and (Fig. 3). The error is the sum of the dif-ferences between the projections of current vectorsand and the desired vector in world frame.

Thus we determine:(Current Position - Desired Position)

(Current Orientation - Desired Orientation) (8)

And generally: (9)

With the instantaneous position and thedesired position and orientation.

This algorithm has been developed over many yearsand the following paper [9] details the fundamental prin-ciples. It is based on an incremental computing of thevariables from the system end toward the root by mini-mizing the error between the desired point (and direc-tion) and the current point (and direction). Some draw-backs are well known as an inhomogeneous variables ada-ptation in the chain joints and a slow convergence nearthe target. Where no limits imposed on variables, then thealgorithm converges without local minima. It is possibleto impose limits on the joints as we do in our application,in that case local minima may occur.


For each variable from toFind the optimum value of that minimisethe error


The joint limits are taken into account by checking ifthe optimum value of is included in the admissiblelimits. Otherwise, the calculated value is not taken intoaccount and we move to the next variable. In Phase 2 ofthe algorithm, it is necessary to compute the direct modelin particular to obtain the error between the goal and thatcurrent point corresponding to the cost function . Incomputing the loop, only one variable is changed therebyapplying equation (6).

n

X ZX

Z

i=n

0

= w *+ w *

=|f ([ ])-[X]|

f([ ]) [X]

1

3.3. The algorithms

21

21

0

1

Fig. 3. Definition of vectors used in defining orientation.

�

�

�

��

��

3.3.1. Algorithm Cyclic Coordinate Descent (CCD)

�

�

�

�

�

i

i

i

i

i


Articles 5

VOLUME 4, N° 3 2010

X21

Z21

choice lowers the cost function then the setof newvalues is maintained if not it is rejected. The procedure isrepeated until the stop conditions are satisfied. In thiscase, it is equation (2), which is used to determine themodel because all variables are modified during eachiteration. The definition of the range increments isperformed as previously by a non-linearly adjustment ofIncrementRate (11) by varying the cost function.


2.1. Define the increment2.2. For all variable

( returns a random value in therange )

2.3. Compute the distance between Current Solu-tion and Goal such as

2.4. if then keep Else keep theoriginal value of


When a set of values reduces the cost function thenthe same set of increment is applied at the nextiteration. This accelerates the convergence.

The algorithms presented above have different beha-viours depending on the type of use and constraints rela-ted to the application. The performance study of each al-gorithm is performed on a PC of Pentium 4 CPU running at3.4 GHz. The algorithms stop conditions consists of twoelements:- A minimum value of the error (cost function) equal

to with and from equation (8)- A maximum number of iterations equal to 1000.

The RAA and IAA algorithms require imposing a ratevalue IncrementRate we placed in 0.015. This value,defined empirically, lead to the best results. It is linked tothe size of the components of the articulated structure,which are defined in appendix.

The three algorithms show no local minimum when thejoint variables are not restricted. In our case we limit theamplitude of the joints based on physical abilities of theperson. Thus in some cases it is possible that the algo-rithms do not find a solution even if it exist. This is due to

�

� � �

��

�

Inc(i)

Inc (i)Do = + Rand(Inc (i))

Rand(X)X

=f([ ])-[X]( ) <0)

Inc(i)

1 w = 1 w = 5000

�

�

�

�

�

�

�

i

i i i

i

i

��

��

3.4 Comparative analysis of the algorithmsbased on their application.

3.4.1. Introduction

0 1

bad choice of joint variables values in initialisationduring algorithms phase 1. We give our results in a successrate on a set of 10,000 trials. We consider that the algo-rithm has found a solution if the error is less than a mini-mum allowed value .

In Figure 4 we see that the algorithms convergequickly during the first iterations and take longer then.This overall behaviour remains the same but according tothe initial value of variables and the goal to achieve, thealgorithms speed may vary. Our application requires theaccessibility verification in many the environment pointsrequiring for each target point in space, to calculate theinverse kinematics. The comparison is performed on a sta-tistical study on a large number of trials (10,000). We cancompare the number of iterations between the CCD algo-rithm and IAA since the same number of matrices pro-ducts is performed at each iteration, it is to say(where is the number of dof). By cons, RAA algorithmrequires multiplications per iteration, which is lowerthan the other two algorithms. A larger number of itera-tions is necessary in order to find the solution and thecomputing time will be longer. We consider that an itera-tion is achieved when all variables of the articulatedstructure has been treated.

We propose to decompose the analysis such as we donot consider the mobility at the first time, without the de-gree of freedom in rotation and without the two degreesof freedom in translation and . The system has 21 de-grees of freedom corresponding to a human joint structurefrom the waist to a hand. The target choice is achieved asfollows. The joint variables values are initialised randomly

3 * nn

n

ZX Y

Fig. 4. Global behaviour of algorithms.

3.4.2. Comparative analysis of the computingof the inverse kinematics without mobility.

0


Articles6

VOLUME 4, N° 3 2010

0

5000

10000

15000

20000

25000

30000

35000

1 24 47 70 93 116 139 162 185 208 231 254 277 300 323 346 369 392

Nb Increments

Erro

r

CCD

IAA

RAA

Table 1. Comparison between the described algorithms.

Algorithms

Position and

Position and orientationand

w =1 w =0

w =1 w =5000

0 1

0 1

CCD

Averageiterationnumber

18

146


131

4770


12

211

AverageComputingtime in ms

7

46.0


6.9

301


2.7

60.9

IAA RAA

that allows to determine associated position and orienta-tion. It is these values that we target to reach.

Table 1 shows the comparative elements between thealgorithms. We note that the CCD algorithm is the fasterwhen the position is imposed. When we impose both theposition and the orientation, the algorithm IAA is slightlymore efficient. The results are averages on 10,000 trials.We note that the RAA algorithm is the least. Sometimes,a local minimum may occur and a solution is not available.In this case we start the algorithm again with other initialvalues given in phase 1. For that reason all solutions aregiven with 100% success.

In this case, it is necessary to take into account boththe additional degrees of freedom and the bulk of themobile base. We assume that the mobile base moves alongthe floor, which allows us to define the mobility area bya polygon E that we call, envelop polygon. Some obstaclescreate exclusion areas in which the base cannot moves.All possible positions inside the Envelop polygon is calledconfiguration polygon C. The shape of the polygon C isrelated to the orientation of the mobile. It correspondsto the Minkowski difference of which the reader will finddetails in [10]. The configuration polygon is determinedby scanning the envelope polygon for any value of thevariable since it is dependent on the orientation of themobile (Fig. 5).

Now, the problem is to find a solution to the followingproblem

(12)

3.4.3. Comparative analysis for the computingof the inverse kinematics of articulatedstructure based on mobile.

Z

Z

f([ ])=[X]- (x,y,z)

0

0

Fig. 5. Different polygons definition.

� �

It is to check whether there is a set of variables suchas may reach the point whereas the mobilityspace . The position to achieve does not longercorrespond to a point but a surface. Thus the cost func-tion becomes

(13)

We can write that:(Current Position - Desired Surface)(Current Orientation - Desired Orientation) (14)

Factor of the cost function remains unchanged butthe other part requires the distance computation betweena point and a surface. This is easy to calculate. The CCDalgorithm requires a point to reach and not a surface tofind the optimal variables value at each iteration. Thusthis method is not applicable to the CCD algorithm.

We note that the inverse kinematics computing issignificantly faster than for the fixed base. The reasoncomes from the number of potential points to reach is mo-re important since we have to compute the distance fromone point to a surface. The algorithm must compute, foreach value of the mobile base orientation, the relatedconfiguration polygon. In order to accelerate the compu-ting, we have pre compute these polygons. Adding themobile base accelerates the computing.

In our application, which is to verify the existence ofa solution in the case of accessibility, we test a set ofpoints in the environment. We must check for each envi-ronment point if a solution exists. The environment po-ints are achieved by taking each target point as situatedin the neighbourhood of the previous one. Thus, we donot achieve a random joint variables initialisation as defi-ned in the algorithms. We maintain the joint variablesvalues as previously. We initialise the joint variables ran-domly only at the beginning of the process of computingwhen the first application of the algorithm. To model thisapplication we perform the computation of a set of targetpoints situated on a circle consisting of 200 points.

As we saw earlier, the application of the method thattakes into account the mobility can not be applied to theCCD algorithm where the shaded area in Table 3. We seethat the various algorithms require only a few iterations toachieve the goal where the weak computing time.

��

� �

f([ ]) [X](x,y,z)

=|f ([ ])-[X] + (x,y,z)|

= w *+ w *

w

0

1

1

3.4.4. Comparative analysis for the computationof the inverse kinematics of articulatedstructure on fixed base with small variations


Articles 7

VOLUME 4, N° 3 2010

The envelopPolygon E

Exclusion obstacle polygone

The configurationpolygon C

The mobile base

Table 2. Comparison between the algorithms IAA and RAA by considering the bulk of the mobile base.

Position and

Position and orientationand

w =1 w =0

w =1 w =5000

0 1

0 1

IAA


26

2028


11.3

35.9


1.6

125


0.7

15.7

RAA


Articles8

In the case of mobile base, we find, as in Table 2, theincreasing speed of execution with a limited number ofiterations. The execution speed on our computer becomesless than a millisecond for a point target for the algorithmIAA. We recall that the error allowed is 1 for equation (14).

In this article, we compare three algorithms forcomputing the joint variable values in the case of the in-verse kinematics of an articulated mechanical structure.The context of the use of these algorithms consists tocheck the accessibility of living environments for personswith disabilities. This application requires taking intoaccount the constraints that have been detailed. Thecomputation time is a key criterion that led the work con-ducted in this area. The CCD algorithm, well known in thefield of animated avatars, seems particularly helpfulwhen the base is fixed and when the target point is de-fined by a position. When we introduce the orientationconstraint on the target point then the algorithm IAA wepropose is slightly faster for a higher accuracy. The diffe-rence in results is not really significant and does not leadto a clear choice of best algorithm to use. When we con-sider that the mechanical structure is based on a mobilebase so we get a greater difference in the results. The pro-posed algorithm takes into account the area swept by themobile base is not adaptable to the CCD algorithm. Itrequires knowledge of a goal point for compute the jointvariable values. Especially for the calculation of scalarand vector products. The proposed algorithm considersthe point to reach is materialized by a surface and thenumber of potential points to reach is more importantwhich has the effect of reducing the execution time andthe number of iterations. We have tried to take the near-est point from the polygon solution to the articulatedsystem. No significant improvements are noted and forthe position and orientation constraint the results arelower in terms of iteration number.

4. Conclusion

AUTHORSAbdelak Moussaoui, Rafaa Otmani, Alain Pruski*

Appendix

-Lasc, ISEA, 7 rue Marconi, University of Metz, France,tel. + 33 3 87 31 52 81. E-mails:{abdelhak.moussaoui, otmani, alain.pruski}@univ-metz.fr.* Corresponding author

L1=10; L2=10; L3=10; L4=5; L5=10; L6=10; L7=30; L8=30;

Table 4. The Denavit-Hartenberg table of the consideredkinematic chain.

VOLUME 4, N° 3 2010

Table 3. Comparison between the algorithms if a weak variation of the objective point is performed and if the algorithms step 1of variable initialisation is not done. Two cases are considered: when the base is fixed and when it moves.

Algorithms

Positionand

Positionand orientation

and

Positionand

Positionand orientation

and

w =1 w =0

w =1 w =5000

w =1 w =0

w =1 w =5000

0 1

0 1

0 1

0 1

CCD


1.8

6.2

0.7

4.2


39

369

80

132


1.4

8.9


0.7

2.2

0.5

1.7


2.1

17.1

5.2

20


0.3

2.2

IAA RAA

Fixe

d re

fere

nce

Refe

renc

eon

mob

ileba

se

0123456789101112131415161718192021

�PI/2PI/2PI/2PI/2PI/2PI/2PI/2PI/2PI/2PI/2PI/2PI/2-PI/2

000

PI/200

PI/2PI/2

0

D000L100L200L300L40000L70L800

PI/2PI/2PI/2PI/2PI/2PI/2PI/2PI/2PI/2PI/2PI/2PI/2PI/2PI/2

- PI/2PI/2PI/2

- PI/2PI/2PI/2PI/2

0

a000000000000L50L60000000


Articles 9

References[1] Tolani D., Goswami A., Badler N., “Real-Time Inverse

Kinematics Techniques for Anthropomorphic Limbs”,, vol. 62, 2000, pp. 353-388.

[2] Kallmann M., “Analytical inverse kinematics with bodyposture control”,

, 2008, 19, pp. 79-91.[3] Baerlocher P.,

Federalschool of Lausanne Swaziland, PhD Thesis, 2001.

[4] Mukundan R., A fast Inverse Kinematics Solution for ann-link Joint Chain,

, 2008, pp. 349-354.[5] Muller-Cajar R., Mukundan R., “Triangulation: A new

algorithm for Inverse Kinematics”, Proc., New Zealand, 2007, pp. 181-186.

[6] Abdel-Malek K., Yu W., Yang J., Nebel K., “A mathema-tical method for ergonomic-based design placement”,

, 2004, pp. 375-394.[7] Yang J., Pitarch E.P., ”

”, Technical Report n° VSR-04.02;the University of Iowa, Contract/PR NO.DAAE07-03-D-L003/0001, 2004.

[8] Denavit J., Hartenberg R.S., “A Kinematic Notation forLower Pair Mechanisms Based on Matrices”,

, vol . 77, 1955, pp. 215-221.[9] Wang L.T., Chen C.C., “A combined Optimization

Method for Solving the Inverse Kinematics Problem ofMechanical Manipulators”,

, vol. 7, no. 4, 1991, pp. 489-499.[10] Lozano-Perez T., “Spatial planning: a configuration

Space Approach”., , vol. C 32,n°2, 1983, pp. 108-120.

Graphical Models

Computer animation and virtualworlds

Inverse kinematics techniques for theinteractive posture control of articulated figures.

5 Int. Conf. on Information Techno-logy and Applications

Image andVision Computing

Int. Journal of Industrial ErgonomicsDigital Human Modeling and

Virtual Reality for FCS

Journal ofApplied Mechanics

IEEE Transactions of Roboticsand Automation

IEEE Trans. on Computer

th

VOLUME 4, N° 3 2010

Abstract:

1. IntroductionDemands for micro/nano products and assembly sys-

tems have been raised significantly to meet the ever com-plex technical needs for modern society. Among many isthe MEMS (MicroElectroMechanicalSystem) technologythat has demonstrated, in nearly every sector, miniaturi-zation of mechanical parts or systems [1]. However,although many studies have shown significant advancesin the micro/nano manufacturing technology during thelast decades, a full blown solution with flexible manufac-turing capability in mind still falls short of industrialimplementation in terms of mass production. For exam-ple, Fatikow developed a flexible micro-robot forcomplex microsystems assembly, though the vision sys-tem became much complex for easy implementation inindustry [2]. In [3], Aoyama proposed a in-situ microrobots for flexible micro assembly, but the lack of meansfor realtime operations may hinder immediate implemen-tations in industry.

Demands for micro/nano products and assembly sys-tems have been raised significantly to meet the ever com-plex technical needs for modern society. In this paper, weshare the experiences and results of the study on theflexible micro assembly workcell focused primarily on a no-vel system implementation and system performanceanalysis. For flexible and autonomous assembly opera-tions, we investigated a novel model based 3D depth mea-surement technology for faster and cost effective means topromote autonomous micro assembly systems in variousindustries. Micro parts, by its nature, are known of theirshapes in advance for the majority of micro applications.We take advantage of the previously known shape of microparts and hence apply a model based approach for a fasterand cost effective localization and 3D depth measure ofrandomly loaded micro-parts on the workcell. The proposed3D depth measuring method is based on the pattern reco-gnition and multi-focus technique enabling it to extractonly information useful for micro parts assembly for fasterrecognition. For demonstration purpose, silicon basedoxide gears are fabricated by bulk micromachining and areused to study performance indices and to prove the use-fulness of the proposed micro 3D depth measurementtechnology.

et al.

Keywords: flexible manufacturing, micro assembly, micro3D vision.

In this paper, we share the experiences and results ofthe study on the flexible micro assembly workcell espe-

cially in the implementation and performance analysis.The term “flexible” is used for multiple model assemblyloaded randomly in position and orientation on the work-station. For demonstration purpose of the proposedmicro-assembly system, we fabricated silicon dioxide ba-sed micro gears and assembly base bulk microma-chining technology. To avoid crystallographic edge for-mation [4], several options are tried out for anisotropicwet chemical etch of silicon parts with the photolitho-graphic masks in Figure 1. The wafers used were 100 mmthick p-type doped Si with a Miller indices crystal orienta-tion of <100>. The SiO layer was grown using a wet oxi-dation process followed by Photolithography and BOE(Buffered Oxide Etch), and several parts release pro-cesses.

via

Fig. 1. Photomask for parts (top), and assembly area(bottom).

2

A NOVEL FLEXIBLE MICRO ASSEMBLY SYSTEM:IMPLEMENTATION AND PERFORMANCE ANALYSIS

Dugan Um, Dongseok Ryu, Bo Dong, Dave Foor, Kenneth Hawkins

Received 8 ; accepted 11 May 2010.th April 2010 th


Articles10

In order to minimize the crystallographic edge effectof the bulk micromachining, boron is diffused on thesilicon dioxide surface to slow down the sharp edgeformation [5]. The boron doping process allowed thesilicon to be removed while strengthening the resistanceof the SiO parts to the KOH etch. When used in con-junction with polyimide coating, the developed proces-ses produced a part with good surface quality whichmatched the initial CAD model. Figure 2 shows highquality micro-gear parts produced by combined methodsof boron doping and polyimide coating.

In order to assemble fabricated SiO gears, a novelmodel based 3D depth measurement technology for microparts is introduced in this paper. Unlike chemically relea-sed parts and assembled integrity by MEMS technology,flexible micro assembly is still a daunting task due todifficulties in parts visualization and assembly autono-my. Among many 3D visualization technologies, the mostpopular for micro scale parts is the confocal mapping [6].The processing time of confocal mapping to obtain 10 to100 photos for each pixel and expensive device value,though, hinders commercialization for various microapplication industries. To overcome such a barrier, we in-vestigate a novel model based 3D depth measurementtechnology for faster and cost effective means to promo-te micro assembly technology in various applicationfields. Micro parts, by its nature, are known of their sha-pes in advance for the majority of micro assembly appli-cations. We take advantage of the previously knownshapes of micro parts and hence apply a model basedapproach for a faster and cost effective 3D depth mea-sure. The proposed 3D depth measuring method is basedon pattern recognition and multi-focus technique ena-bling it to extract only information useful for micro partsassembly such as location, size, and height of a micropart. In addition, a functioning micro assembly system isdeveloped and used to prove the usefulness of the pro-posed 3D depth measurement technology.

2

2

2

Fig. 2. SiO gears in various sizes.

2. Assembly workcell integrationIn this section, we present an integration methodolo-

gy of a flexible micro gear assembly workcell with magne-tic grippers, a micro precision robot, and a model based 3Ddepth measure system. A complete feedback loop bet-ween visual sensing and positioning/grasping of a micropart is the enabling technology of a flexible micro assem-bly workcell. As the name implies, micro parts can be lo-aded in the workcell randomly, and the vision system andthe grippers will find their way to locate and grasp themicro parts. In our case, we use the micro gears and latch-es fabricated by bulk micro-machining for demonstration.

1. Micro precision robotic station

2. Model based location and 3D depthmeasurement system

The assembly station used in the assembly workcellimplementation is a precision micro robotic system byNational Aperture with the uni-directional repeatabilityof 2 μm. Figure 3 is the picture of the 3 Degrees ofFreedom (DOF) robotic system integrated in one assemblywith a magnetic grippers.

The base station that holds the wafer has two-axisrobotic module for x-y motion (Figure 3). The third axis isassembled in a tilted angle for grasping and up-downmotion in one assembly. The magnetic grippers open andclose the tip to grip and release a micro gear down to thesize of 200 μm. Filted design of the 3 axis in conjunctionwith the lengthy grippers allows not only precision pickand place operations, but maximizes the limited workspace under the microscope.

The key component to close the loop of the flexiblemicro assembly is the visual feedback system, with themicro 3D visualization capability. 3D measuring of microgears is accomplished through four consecutive steps.The first stage involves the extraction of the contours sothat the candidate regions of gears can be addressed. Thesecond process entails sorting out the actual gear regionamong the candidate regions by comparing the extractedcontours with the known information of gears. After thisprocess, the contour that conforms the known gear shaperemains, and all other contours are ignored. In the thirdstage, the height of each contour, which corresponds toan actual gear region, is measured with 10 differentfocused images. Finally, the simplified height map isdeveloped by associating the gear regions with theobtained height information. Details of each process arediscussed in the following sections.

Although the pixel-based recognition method isdominant in object recognition technology, we use thecontour-based vector tracing approach for localizationand 3D depth measuring of micro parts. With the pixel-based approach, an error of one pixel does not signifi-cantly affect the final result of the height map. On theother hand, one missed contour in the contour based ap-proach causes a fatal error, though the result is signifi-

Fig. 3. 3-Axes Robotic wafer platform with a micro gripper.

rd

A. Contour detection and gear recognition


Articles 11

The magnitude and direction of this vector aredenoted as

(2)

(3)

The partial derivatives in the above equations denotethe direction and rate of change of the grayscale of eachpixel. There are many ways to quantify the gradientvalues, such as Robel, Prewitt [8], Sobel, Laplacian, etc.The Laplacian operator is commonly used as a secondorder operator, represented by

(5)

In order to prevent the change in , directions fromcancelling each other, Nayar and Nakagawa [9] proposeda modified version of the Laplacian operator with a dis-crete approximation as

(6)

They proposed the sum-modified-Laplacian (SML)[10], as a focus measure at a point . It was denotedas

for (7)

Where is the measuring window size around the point. The parameter is a distinct threshold value. In

this research, the SML value of each pixel in a contour iscalculated for height, and one representative heightvalue for each contour region is defined by averaging SMLvalues in that region. Although SML method works wellwith the micro parts, we experienced much delays on the3D depth measure of the fabricated gears, which, in fact,becomes the bottleneck toward the near realtime assem-bly operation [11].

In order to overcome the latency of the SML method,we propose

, namely “Simplified Height Map (SHM)” measu-rement method to facilitate the gear identification pro-cess using the object’s geometry information. In this me-thod, we assume that the height of each gear are knowna priori and stored in the database. In addition, for de-monstration purpose, we used gears with the equivalentheight but without excluding overlapping possibilities.To measure the height of each region, 10 pictures withdifferent focal planes are taken from the same view. Figu-res 5 (a)-(d) show four examples of different focal planes.Finally, a simplified height map is composed of the reco-gnized gear regions with several levels of height, asshown in Figure 6.

Figure 6a) shows the extracted contours from an ac-tual microscopic image in 3D format. The detected con-tours are compared with the known information of thegears. The size and shape of the gears are well-definedduring the fabrication process, and the information ofgears used in assembly is stored in a database before the

x y

x,y

Nx,y T

( )

( )

a model based position and height measuretechnique

cantly faster. Thus, the contouring method is the mostimportant process in this step, and thus, it was conse-quently executed with three different binary images in aneffort to obtain more reliable results.

First, the contour is extracted from a binary image.The focus of the image does not significantly affect thebinary image, but the binary process is sensitive to thelight condition. The micro assembly process is, therefore,performed in a well-controlled environment, and all lightconditions are set as predefined adequate values. Foreven distribution of light intensity and to minimize thedisturbance by ambient lights, infrared camera, filter,and fiber optic cables are used for the vision assembly(See Figure 4).

By analyzing the center of the recognized gear inplanar coordinates, the wafer holding station moves thewafer to the exact location of the geometric center sothat the gripper properly approaches the correspondinggear. The radius of the detected gear is calculated fromthe contour, and the gripper opens the proper amount topick up the appropriate gear. However, the height of thegear still remains unknown. Without the height informa-tion, the gripper could be stuck by crashing into thewafer, or the gripper could miss the gear.

Conformal mapping technique is based on measuringthe focused and defocused area to create a 3D depth ima-ge. That is, due to the limited depth-of-focus of opticallenses, an object located out of the focal plane cannot beseen clearly in the image [6]. This implies measuringblurriness provides the height information of the objects[7]. The blurriness can be defined as an image gradient.When an image is described by a continuous brightnessfunction, , its gradient at position can berepresented by a vector:

(1)

Fig. 4. Vision system and robotic platform assembly.

B. Simplified height map measurement methodusing Sum-Modified-Laplacian

I x,y x,y( ) ( )


Articles12

assembly process. Contours which don’t appear to be ge-ars are ignored, and only matched contours with the da-tabase remain as regions of interest for measure-ment. Figure 6b) shows the recognized gears.

Using the extracted gear location and height informa-tion by SHM method, our robotic grippers in conjunctionwith the x-y precision platform demonstrated successfulpick and place operations, though current success ratioof the pick and place is fairly low to prove the usefulnessof the proposed technology (See Table 1). With the tiltedangle design, we achieved up to 0.5 μm z-axis accuracy ofthe gripping motion. The accuracy of close and openoperation of the grippers, however, is not consistentovertime. In addition, although the height measure ofthe micro object is near realtime, processing the assem-bly of the micro gears on the axial shaft is yet performedwith no realtime visual feedback, reducing the assemblysuccess ratio to some extent.

A precision pick and place grippers is designed andrapid prototyped as shown in Figure 7. The lengthy twe-ezer-type manipulator serves two purposes. First, it ena-bles precision manipulation in the 1.5 cm of clearancebetween the assembly wafer and the microscope lens andlighting assembly. Secondly, the actuator opens andcloses the grippers with the clearance on the order of

height

(a) focus on wafer (b) focus on lower gear

(c) focus on upper gear (d) focus on above gears

Fig. 5. Height measure using conformal mapping technique.

(a) without model (b) with model

Fig. 6. SHM with a model based gear recognition.

3. Precision grippers

10–100 micrometer for precision operations. Thereforethe arms of the lever would magnify the deflection up to5 times to grip various sizes of micro-gears.

The first choice of the grippers’ actuator was EAP(Electro Active Polymer) driven by controlled currents.However, it turned out that a EAP actuator has reliabilityproblems over time as shown in Figure 8a). The amount ofdisplacement reduces significantly, thus frequent repla-cement was inevitable. In addition, unexpected secon-dary displacement occurred. For instance, when the tipof the EAP strip begins from rest with no potential ap-plied at the starting position marker, the EAP displaces tothe primary displacement with the potential applied(Figure 8b)).

While the potential remains on and constant, the EAPthen begins secondary displacement in the opposite di-

Fig. 7. Magnetic gripper assembly.

(a)

(b)

Fig. 8. EAP (Electro Active Polymer) actuator performanceanalysis.


Articles 13

rection of the primary displacement (in the direction ofthe negative terminal). As a result, the actuator is chan-ged to an electromagnetic driver for pick and place opera-tions. Piezo-electric grippers demonstarted consistentperformance with better controllability in grasp control.

3. Assembly experimentsThe complete system has been implemented with

a precision 3 axis robot, magnetic grippers and the 3Dmicro vision system (see Figure 9a)). For fully autono-mous micro assembly operation, integration software hasbeen developed to close the assembly loop between thevision sensor and the assembly grippers (see Figure 9b)).

A complete assembly operation starts with recogni-zing randomly placed micro gears in the first quadrant ofthe assembly wafer followed by placing them at thecorresponding location at the 4 quadrant. A completeassembly cycle includes gear location identification, gearheight measure, robotic gripper control for pick and placeoperations (Figure 10). One example of the gear grasping

the magnetic grippers is shown in Figure 11. Figure 12

(a)

(b)

Fig. 9. Assembly workcell (a) & integration software (b).

via

th

depicts the gears placed at designated locations in 4quadrant of the wafer.

St.Aveg. Dev. Success(min) (sec) ratio (%)

Search for gears (x,y) 2.5 56 91Identify gear height (z) 1.4 25 N/AGrip gear 0.75 12 78Move gear 0.5 15 97Place gear 1.2 18 45

In Table 1, the results of assembly time and root causeanalysis for assembly performance study are shared. Thesuccess ratio of the gear identification for coordi-nates was not being analyzed due to the absence of theexact location information of each gear. The SHM methodintroduced in the paper improved the gear height identifi-cation speed up to 4 times faster than the conventional

th

Fig. 11. Magnetic grippers holing a SiO gear.

Fig. 12. Gears on the assembly workcell.

Table 1. Assembly time & root cause analysis.

2

x,y,z


Fig. 10. Assembly sequence.

Articles14


SML method. As shown in the table, gear height identifi-cation stage is no longer the bottleneck of the assemblyprocess due to the improved process speed by the SHMmethod in time analysis.

In terms of success ratio, the gear placement stage isthe root cause of the low overall success ratio. The overallassembly success ratio was about 77% and has yet to beimproved for usefulness in industry. In order to improvethe success ratio, realtime visual feedback has to beimplemented for pick and place operations.

The technical challenge tackled in this research wasto develop an autonomous and flexible micro partsassembly workcell. Among several components thatconstitute a flexible assembly workcell, the localizationand 3D depth measure of micro parts were the mostchallenging tasks. To provide a low-cost and fast gearrecognition method, we proposed the simplified heightmap generation for visualizing micro parts in 3D. Theproposed method was designed to utilize the knownmodels of micro parts such as size and shape, thus modelbased, so that it increases the reliability of the objectidentification and localization in near realtime fashion.Precision robotic platform and magnetic gripper havebeen put together with the micro 3D depth measuresystem.

Finally, micro gear assembly experiments were perfor-med to prove the usefulness of the proposed system. Theoverall assembly success ratio was about 77% and has yetto be improved for usefulness in industry. The bottleneckprocess is turned out to be the gear identification stagein time analysis. However, the gear placement stage isthe root cause of the low overall success ratio. In order toimprove the success ratio, realtime visual feedback has tobe implemented for pick and place operations.

- Texas A&M University –Corpus Christi, Science and Technology Suite 222, 6300Ocean Drive, Unit 5797, Corpus Christi, Texas 78412.Tel. 361-825-5849, Fax 361-825-3056. E-mail:[email protected].

- The College of Williams & Mary, Williamsburg,VA.

- Austin Community College, Austin, TX.- San Marcos High School, San Marcos,

TX.* Corresponding author

4. Conclusion

ACKNOWLEDGMENTS

AUTHORSDugan Um*, Dongseok Ryu

Bo Dong

Dave FoorKenneth Hawkins

This material is based upon the project titled, “Micro/Nanoassembly workcell via micro visual sensing and haptic feedback”supported by the National Science Foundation grant #0755355awarded to Texas A&M University–Corpus Christi and thesupport of Texas State University-San Marcos. Any opinions,findings, and conclusions or recommendations expressed inthis material are those of the author(s) and do not necessarilyreflect the views of NSF, Texas A&M University–Corpus Christi,and Texas State University.

References[1] Woods D., “The fabrication of silicon microsystems”,

, vol. 9 (129),2000, pp. 129-136.

[2] Fatikow S., Seyfried J., Fahlbusch S., Buerkle A.,Schmoeckel F., “A Flexible Microrobot-Based Microas-sembly Station”,

, vol. 27, no. 1-2, Jan. 2000, pp. 135-169.[3] Aoyama H., Fuchiwaki O., “Flexible Micro-Processing by

Multiple Micro Robots in SEM”, In:,

Seoul, Korea, May 2001, pp.3429-3434.[4] McGregor M. T., Mahlke H. A., Dozier S.M., Asiabanpour

B., Um D., “Producing micro scale silicon dioxide gearsby bulk micro machining process” ,

, vol. 37, 2009.[5] Madou M.J.,

, 2002.[6] Kim T., Kim T., Lee S., Gweon D., “Optimum conditions

for high-quality 3D reconstruction in confocal scanningmicroscopy”. In: , vol. 6090, Feb. 2006.

[7] Zlotnik A., Ben-Yaish S., Zalevsky Z., “Extending thedepth of focus for enhanced three-dimensional imagingand profilometry: an overview”, , vol. 48,no. 34, Oct. 2009, pp. 105-112.

[8] Prewitt J.M.,, Edited by: Lipkin

B.S., Rosenfeld A. New York: New York: Academic, 1970,pp. 75-149.

[9] Nayar S.K., Nakagawa Y., “Shape from focus”,

, vol. 16, no. 8, Aug. 1994, pp. 824-831.[10] Zhao H., Lia Q., Fenga H., “Multi-focus color image

fusion in the HSI space using the sum-modified-laplacian and a coarse edge map”,

, vol. 26, no. 9, Sep. 2008, pp. 1285-1295.[11] Um D., Asiabanpour B., Jimenez J., “A Flexible Micro

Manufacturing System for Micro Parts Assembly viaMicro Visual Sensing and EAP based Grasping”,

, vol. 8, no. 2, Dec.2009.

Engineering Science and Education Journal

Journal of Intelligent and RoboticSystems

Proc. of IEEEInternational Conference on Robotics & Automation

Transactions of theNAMRI/SME

Fundamentals of microfabrication: thescience of miniaturization. 2. CRC Press

Proc. SPIE

Applied Optics

Object enhancement and extraction inPicture Processing and Psychopictoris

IEEETransactions on Pattern Analysis and MachineIntelligence

Image and VisionComputing

Journalof Advanced Manufacturing System

VOLUME 4, N° 3 2010

Articles 15

Abstract:

1. IntroductionControl algorithms for the WWTP have been investiga-

ted intensively, particularly for DO control. The DO dyna-mics is nonlinear and of high dimension. Dissolved oxy-gen is the most important control parameter for biolo-gical processes in WWTP. Increasing performance of theDO control system is a basic action leading to improve theeffectiveness and efficiency of the regarded system. Itbecomes more interesting when it can be done withoutinterference into the system structure and only by simplemanipulation of the control algorithm. The DO trackingproblem is one of the most complex and still fundamentalissue of biological WWTP involving activated sludgetechnology due to its influence on dynamics of bioche-mical processes. Its complexity derives from strong nolinearity and time dependence of the process variables.

In this paper a very effective controller design appro-ach is presented, which uses both classical and artificialintelligence methods to find a solution dissolved oxygen-tracking problem.

The paper is organized as follows. The problem state-ment is described in Section 2. Section 3 presents the PIcontroller design. The controller calibration process isdescribed in Section 4. The engineering and scientific

The paper addresses design, calibration, implementa-tion and simulation of the intelligent PI controller used fordissolved oxygen (DO) tracking at wastewater treatmentplant (WWTP). The calibration process presented in this pa-per utilizes both engineering and scientific methods. Veri-fication of the control system design method was obtainedvia simulation experiments.

Keywords: aeration process, artificial intelligence, controlsystems, dissolved oxygen, tracking problem, fuzzy logiccontroller, genetic algorithms, intelligent control, TakagiSugeno Kang method, TSK, soft switching, wastewatertreatment.

methods are used. Application of the controller to Kar-tuzy WWTP is described in Section 5 and the simulationresults are presented. Section 6 concludes the paper.

The previous papers [Brdys, , 2002; Piotrowski,, 2004; Piotrowski, , 2008] propose a two le-

vel controller to track prescribed dissolved oxygen trajec-tory. The upper level control unit prescribes trajectoriesof desired airflows to be delivered into the aerobic biolo-gical reactor zones. The lower level controller forces theaeration system to follow these set point trajectories.A non-linear model predictive control algorithm is appli-ed to design this controller unit.

The goal of this paper is to provide the design metho-dology of intelligent multiregional controller [Domanski,

, 1999], with enhanced regional controllers, to en-able highly nonlinear system, in this case aeration sys-tem, to work in a whole control space with the same highquality performance under heavy disturbances. It alsoshows how to utilize optimization method to calibratethe resulting controller parameters.

Activated sludge wastewater treatment plant (ASW-WTP) is a very complex and demanding structure for clo-sed-loop control due to its internal processes nonlineari-ties, inconsistency, and time and parameter variability.The scheme of the biological WWTP is presented in Fig. 1.

The advanced biological processes with nutrient re-moval is accomplished in the activated sludge reactordesigned and operated according the University of CapeTown (UCT) process.

The first zone where the phosphorus is released isanaerobic.

The second zone where the denitrification process isconducted is anoxic.

The internal recirculation 2 of mixed liquor originatesfrom the anoxic zone. The returned activated sludge fromthe bottom of the clarifiers and the internal recirculation1 from the end of the aerobic zone (containing nitrates)are directed to the anoxic zone.

2. Problem statementet. al.

et. al. et. al.

et al.

INTELLIGENT PI CONTROLLER AND ITS APPLICATION

TO DISSOLVED OXYGEN TRACKING PROBLEM

Tomasz Zubowicz, Mieczysław A. Brdys, Robert Piotrowski

Received 23 ; accepted 27 .rd September 2009 April 2010th


Fig. 1. Scheme of the biological WWTP.

Articles16

The last part of the reactor (aerobic) is aerated bya diffused aeration system. This zone is divided into fourcompartments of various intensity of aeration.

The biologically treated wastewater and biomass(activated sludge) are separated in two parallel horizon-tal (rectangular) secondary clarifiers.

In order to ensure a high level of phosphorus removal,iron sulphate (PIX) is added to the aerobic zone to preci-pitate most of the remaining soluble phosphorus (simul-taneous precipitation). There is also the opportunity toprecipitate phosphorus in the grit chamber (pre-precipi-tation).

The excess biological sludge is stored in a thickener,then dewatered in two centrifuges and finally chemicallystabilized in lime. It is expected that the sludge will bedisposed of for agricultural applications.

For the purpose of the research the activated sludgemodel type ASM2d [Henze, , 1995] was used. Thisselection was a good compromise between accuracy andspeed in predicting dynamics of the biochemical proces-ses [Brdys, , 2004].

The ASM2d type model derives from the activated slu-dge model family (ASM), which also consists of the typesASM1, ASM2, ASM3, ASM3 Bio-P. The differences betweeneach model are generally situated in their mathematicalrepresentation accuracy of biochemical processes. TheASM family evolved according to the stated order withASM2d model placed in between ASM2 and ASM3.

The WWTP model was designed using commercial pac-kage Simba [Simba, 2005]. After the quality and quantityparameter identification the model calibration processwas applied. For the purpose of calibration the genetic al-gorithms (GA) were used. Final step, the validation pro-cess verified and confirmed the usefulness of the cons-tructed model.

This section addresses directly the DO tracking pro-blem. In [Yoo, , 2002] for the purpose of solving theDO tracking problem the PI controller with on-line para-meter adaptation was proposed. First of the mechanismswas invoked to enable in the controller ability to rejectdisturbances and the second to carry out control of thenonlinear process with the same performance in wholecontrol area.

In this paper the more effective (regarding both con-troller design methodology and computational require-ments) method was proposed. For the purpose of distur-bance rejection the classical PI regulator was applied re-gionally. In order to enhance the regional fixed para-meter controllers to work as a one in whole control spacethe control signal switching system was utilized. Due todecreasing of performance during the hard switchingtechniques, the artificial intelligence (AI)Takagi-Suge-no-Kangs (TSK) soft switching method is used (Fig. 2).This also has allowed to eliminate troubling parameteradaptation procedure presented in [Yoo, , 2002].

For the purpose of the design process the followingassumptions have been made:

the aeration system is regarded as ideal thus the ge-nerated and reference airflows fulfill the expression:

et al.

et al.

et al.

et al.

3. Controller design

�

;

The intelligent controller (Fig. 2) design process canbe easily divided into four main stages:

The first step that needs to be taken is to representknowledge of the process in the form of static charac-teristic of as a function of airflow , generated bythe aeration system. As mentioned before (see Section 2)the considered process is under heavy influence of timevarying disturbances: wastewater inflow , chemicaloxygen demand , total nitrogen and totalphosphorus . The model based sensitivity analysisper-formed to acquire the steady state characteristic ofthe process resulted in obtaining the whole family of

curves. In Fig. 3, the mean charac-teristic (determined via the studies) and its boundarieshave been shown.

As determined in Stage I the relationship betweenand is highly nonlinear (Fig 3). In the second

stage the mean steady state characteristic is piecewiselinearized (Fig. 4) for the purpose of nonlinearity appro-ximation thus divided into affine, linear spaces (regionindication: for ). This method has manyadvantages such as: simplicity, ability to indicate linear-ly representative regions of process, associated error andabove all in some cases can be done graphically (mainlywhile regarding SISO systems). This last feature enablesthe designer to decide if the balance between accuracyand complexity is maintained at proper level.

Piecewise linearization also can be done an opti-mization algorithm with properly chosen cost function,but the first approach represents, in authors opinion,a good balance between simplicity and effectiveness andthus recommended for this and similar cases.

�

�

�

�

�

�

plant being under control is a single aerationchamber.

stage I gathering information;stage II region indication;stage III local controller design;stage IV fuzzy logic applying;stage V calibration (optional).

Fig. 2. Scheme of the control system.

Stage I:

Stage II:

via

DO Q

QCOD TN

TP

DO(Q )

l , i

air

((

(

)( ) ))

= 1,..,5

DO(Q )

DO Q

air air

air

i


( ) ( )�ref

air airQ k Q k

Articles 117

Fig. 3. Static characteristic of the aeration process and itsboundaries.

Fig. 4. Mean and piecewise linear static characteristic ofthe aeration process.

Stage III:Indicating the regions with lack or small enough er-

ror, which can be treated as linear, allows looking for alsolinear regional controller.

PI controller is used for a local (regional) controller(Fig. 5). Additionally magnitude and rate limiter of thegenerated control signal have been added along with theanti wind-up filter controlling the performance of themagnitude limiter. As a result the nonlinear PI controlleris used.

Incorporating stated nonlinearities into PI model re-sults in many advantages. The magnitude and rate limi-ter enables controller to generate control signal meetingthe actuator system static and dynamic limitations. Bothinfluence the actuators lifespan and also time neededbetween maintenance by reducing stress generated inthe system. The negative influence of saturation statesentailed mostly by the magnitude limiter is reduced bythe anti wind-up filter [Bohn and Atherton, 1995] utili-zed in the controller.

The PI controller (Fig. 5) can be described as:

where: - error; - control signal; - signal insummation line; - signal before magnitudelimitation; - proportional gain; - summation gain[1/s]; - anti wind-up filter gain; - output scalingfactor.

In the system (1) the error signal is defined as:

(2)

where: - measured dissolved oxygen value;- reference value of ; - input scaling

factor.

The saturation function is as follows:

The calibration of the local controllers parameters cansimply be done using well-known engineering approachboth on plant experiments or model-based simula-tions. Parameters for the five local PI controllers (see Fig.5) chosen arbitrary to balance the performance in eachregion have been presented in Table 1.

(1)

(3)

e k u k u k

K KK g

DO(k)

( ) ( ) ( )i

pre

p i

aw

ref max

u k

e k

DO k DO DO

( )

( )

( )

via


Fig. 5. Local discrete nonlinear PI controller.

u k g k sat u k( ) = ( )(0.5 ( ( ))+0.5)� p

u k e k K u k sat u ki aw p p( ) = ( ) ( ( ) 0.5 ( ( )) 0.5)� � � �

refe k DO k DO k( ) = ( ( ) ( ))�

sat sign min{ .(.) = (.) , 1}� �

u k sat u ip u( ) = 0.5 ( ( ))� ��

u k K e k K K u ju p p i i( ) = ( )+ ( )�

p

pi=1

j=1

DOmax

1

Articles18

In this step the TSK fuzzy logic mechanism (4) isbeing applied [Yaochu, 2002; Tanaka and Sugeno, 1992].The sole of the TSK can be seen as the weighted average(blended) control signals of regional controllers:

(4)

where: - multiregional controller output;output value of -th regional controller; -

number of regions in this case .

The weights is on-line tuned according to theprocess state, which in this case is denoted by the DOconcentration. The possible weight values are fuzzy setsof a priori defined membership functions (MFs). Each MFis directly correlated with corresponding regional con-troller; hence the number of regions determines the num-ber of MFs. In this case the Sigmoidal and Gauss condi-tional functions were utilized [Yaochu, 2002].

Defining the MF simply means to chose its type andshape. It should be noted that the choice of the MFs iscrucial for acquiring stability and good performance ofthe closed loop system, however it can be done while ha-ving basic knowledge of the process. It is also useful toknow the shape of the assessed process static characteri-stic, which in this case is already achieved (see Stage I).The procedure for choosing MFs is as follows:

Stage IV

Q ki pp

w DO

( )

= 5

( )

air

air, i

i

ref

refQ k

i i

( ) -

each regional controller should be given a correspon-ding MF;the core of -th MF should overlap with the part of -thregion that possesses relative error of approximationless then 10% (which is an arbitrary chosen value);

�

�

�

�

the value of membership of -th MF should be equal to0 when it reaches the neighboring MFs core;created fuzzy partition should be normal and consis-tent.

i

Resulting fuzzy partition is presented in Fig. 6.

The formulas for each MF and parameters in the nume-rical form have been presented in Table 2.

Calibration of the MFs parameters is the final stage ofthe controller design process. This stage can be omittedwith no loss of functionality of the designed control sys-tem, however it certainly improves the overall systemperformance. As a tool for that purpose the GA was cho-sen and applied. The whole calibration process is descri-bed in details in Section 4 due to its rather extensivecharacter.

Fig. 6. Initial MFs.

Stage V


Table 1. Values of local PI controller's parameters.

Name

Region

TK

KK

s

i, i

p, i

aw, i

[s][1/s]

No

1

2

Controller I

l

300.0416.839.67

1

Region

l

l

1

2

Controller II

l

300.0210.549.6

2

Function name

Sigmoidal

Gauss

Controller III

l

300.016.449.79

3

Formula

Controller IV

l

300.0314.9815.88

4

Controller V

l

300.0211.2916.22

5

a = -10c = 1.483

sig = 0.422c = 1.98sig = 0.145c = 2.85

1

1

2

2

Parameters

Q kairref ( ) =

Q kair, iref ( )� ( ( ))w DO ki �

p

i=1

� ( ( ))w DO ki

p

i=1

Table 2. MFs and their parameters.

A� ( ( )) =DO k

A� ( ( )) = 1;DO k

exp

exp ; left - most curve

; right - most curve

whenever <c c1 2

1+exp( ( ( ) )� �a DO k c

�12

�22

1

� �( ( ) )DO k c12

� �( ( ) )DO k c22

Articles 19

4. MFs parameters calibration

In order to calibrate parameters of the MF, the GA wasused due to their large number, task computational diffi-culty and lack of knowledge about system model deriva-tives.

The main features of the utilized algorithm are asfollows [O'Reilly, , 2002]:

genes in a form of real numbers (16 genes see Table 2);initial population contains one fixed chromosome withgenes equal to the parameters presented in Table 2;algorithm uses operands: mating (crossing and blen-ding) and mutation both with gene value and dupli-city control;elitism;generation counter as algorithm stop condition.

The cost function used to evaluate each chromosomeis given by the formula:

(5)

where: - survival cost; - steady state error;- percentage overshoot; - diffe-

rence between overshoots within each region; -settling time with respect to 5% criterion; -difference between settling times within each region;

- control signal energy; - difference bet-ween regional control signals; - saturation sta-tes; - oscillation in control signal; - nor-mality of fuzzy partitioning; - consistency offuzzy partitioning.

The algorithm flow chart is presented in Fig. 7, whilethe convergence of the GA over the generation is shown inFig. 8.

4.1. Utilizing GA

et al.�

�

�

�

�

JS e EeM M M M M

t T tt T t

u Uu u U uaw Aaw

du Ddu n Nncm Ccm

T

T T

T

T

T T

T

T T

T

% % % %

R R R

R R R

� � �

� � �

� � �

Fig. 7. GA flow chart.

Fig. 8. GA convergence.


No

3

4

5

Region

l

l

l

3

4

5

Function name

Gauss

Gauss

Sigmoidal

Formula

sig = 0.1146c = 3.155sig = 0.604c = 4.78

sig = 0.671c = 6.231sig = 0.612c = 6.84

a = 5.38c = 7.514

1

1

2

2

1

1

2

2

Parameters

Table 2. MFs and their parameters.

A� ( ( )) =DO k1+exp( ( ( ) )� �a DO k c

1

JS e Ee M MM + M M M + t T t += +T T T T% % % % R R R� � �

+ t T t + u Uu + u U u + aw Aaw +� � � � � �R R RT T T T

+ du Ddu + n Nn + cm CcmT T T

Articles20


Table 3 shows the last generation (100 ) of the popu-lation, sorted in ascending order with regard to the cost.

The graphical representation of the functions is shownin Figure 9.

The calibration process exploiting GA (in example [5])makes this uneasy task achievable (see Fig. 8). In order toclearly show the differences in performance of MFs withand without calibration process a standard root mean squ-are (RMS) criterion, given by (6) was used:

(6)

where is the number of samples and is defined asfollows (7):

(7)

The data collected have been presented in Table 4,along with the percentage value of performance enhance-ment (compared with the state before calibration).

th

Fig. 9. MFs with calibrated parameters.

4.2. Calibration summary

N J

As it can easily been seen examining the comparisonresults collected in the Table 4 the GA based approachincreases the performance of the control system up to X%according to the RMS criterion.

In this section the derived controller is validated bysimulation, based on data recorded from Kartuzy WWTP.

The simulation conditions assume highly varying dis-turbances: , and , which are presented in thesame order in Figs 10-12.

The performance, error and control signal trajectoriesof the closed loop system before and after the calibrationprocess have been shown in Figs 13-18a).

Figs 13-18 indicate that the designed controllerenables the system to realize the reference trajectoryunder the heavy time varying disturbances (Figs 10-12)keeping good performance over the whole operatingrange of the plant. It can be seen (Figs 15, 15a, 18, 18a)that both the magnitudes and rates of the demandcontrol signal can be accommodated by the plantactuator system.

The control signal generated by not calibrated andcalibrated systems are illustrated in Figs 15, 15a and 18,18a, respectively. It can be seen that the control signalsare equally demanding in terms of magnitudes and rates,however as it has been already stated the calibratedsystem achieves much better tracking error (see Table 4).The simulation was carried out under worst-case plantdisturbance scenario (see Figs. 10 to 18).

As sample, the result of a daily performance, currentlyachieved by the control system, at the plant site, isillustrated in Fig. 19.

A significant improvement of operating performanceby the proposed controller can be clearly seen.

5. Simulation results

Q COD TN

VOLUME 4, N° 3 2010

Table 3. Calibrated MFs parameters.

Table 4. RMS and performance enhancement.

Region

MF

Parameters

l

MF

1

1

l

MF

2

2

l

MF

3

3

l

MF

4

4

l

MF

5

5

a = -10.932

c = 1.2052

sig = 0.371c = 1.868sig = 0.104c = 2.844

1

1

2

2

sig = 0.079c = 3.149sig = 0.427c = 5.049

1

1

2

2

sig = 0.412c = 6.311sig = 0.440c = 8.232

1

1

2

2

a = 6.578

a = 6.578

J e Ee M MM + M M M + t T t += +T T T T% % % % R R R� � �

RMS = � J1N

+ t T t + u Uu + u U u + aw Aaw� � � � � �R R RT T T T

No

1

2

MF state

before calibration

after calibration

RMS

0,0218

0,0173

Percentage performance enhancement

0%

20,51%

Articles 21


Fig. 10. Inflow to WWTP.

Fig. 12. Total nitrogen.

Fig. 14. Control error before calibration.

Fig. 11. Chemical oxygen demand.

Fig. 13. Closed loop system performance before calibration.

Fig. 15. Controller output signal before calibration.

Articles22


Fig. 15a. Zoom in of controller output - signal beforecalibration.

Fig. 16. System performance after calibration.

Fig. 17. Control error after calibration. Fig. 18. Controller output signal after calibration.

Fig. 18a. Zoom in of controller output - signal aftercalibration.

5. ConclusionsThe paper has addressed design, calibration, imple-

mentation and simulation of the intelligent PI controllerused for DO tracking in WWTP.

A classical PI controller has been used to derive mul-tiregional intelligent controller. The controller is capableof maintaining globally well-known attractive local pro-perties of the PI controller, when applied to nonlinearprocesses.

An advantageous property of the proposed controllerdesign methodology is that it can be applied withoutchanging the on plant (commonly used) hardware. Fur-thermore it allows obtaining great enhancement in thesystem performance via simple algorithm change, whichis a low cost solution.

Enhancements are possible by calibrating MFs para-meters.

The controller has been validated by simulation basedon real data recorded from the Kartuzy WWTP and ex-cellent results have been obtained, however a rigorousanalysis of the closed loop stability is under currentresearch.

Articles 23


AUTHORSTomasz Zubowicz*

Mieczysław A. Brdys

Robert Piotrowski

References

- Department of Electrical Engine-ering and Automation, Gdansk University of Technology,80-233 Gdansk, Poland. E-mail:[email protected].

- Gdansk University of Technology,Department of Electrical Engineering and Automation,80-233 Gdansk, Poland. E-mail: [email protected] of Electronic, Electrical and Computer Engi-neering, The University of Birmingham, Birmingham B152TT, UK. E-mail: [email protected].

- Department of Electrical Engine-ering and Automation, Gdansk University of Technology,80-233 Gdansk, Poland. E-mail:[email protected].* Corresponding author

[1] Bohn C., Atherton D.P., An analysis package comparingPID anti - windup strategies. , Ruhr-Univ., Bochum, 1995.

[2] Brdys M.A., Chang T., Konarczak K., “Estimation ofwastewater treatment plant state for model predictivecontrol of N-P removal at medium time scale”. In:

, Osaka, 26 -28 July 2004. (invite session).[3] Brdys M.A., Chotkowski W., Duzinkiewicz K., Konarczak

K., Piotrowski R., “Two-level dissolved oxygen controlfor activated sludge processes”. In:

, Barcelona, 21 -26 July 2002.[4] Domanski P., Brdys M.A., Tatjewski P., “Design and sta-

bility of fuzzy logic multi - regional output controllers”,, vol. 9, no. 4, 1999, pp.

883-897.[5] Haupt R.L., Haupt S.E.,

Wiley-Interscience, New Jersey, 2004.[6] Henze M., Gujer W., Mino W., Matsuo T., Wentzel M.C.,

Marais G.v.R., Scientificand Technical Report No. 3, IAWQ, London, 1995.

Dept. of Electr. Eng.

IFAC10 Symposium Large Scale Systems: Theory and Applica-tions

15 IFAC World Con-gress

Int. Appl. Math. And Comp. Sci.

Practical Genetic Algorithms.Second Edition.

Activated Sludge Model No. 2.

th

th

th th

st th

[7] O'Reilly U-M., Yu T., Riolo R., Worze B.,Springer Science, Boston,

2005.[8] Piotrowski R., Duzinkiewicz K., Brdys M.A, “Dissolved

oxygen tracking and control of blowers at fast timescale”. In:

, Osaka, 26 -28 July 2004.[9] Piotrowski R., Brdys M.A., Konarczak K., Duzinkiewicz

K., Chotkowski W., “Hierarchical dissolved oxygen con-trol for activated sludge processes”,

, vol. 16, issue 1, 2008, pp. 114-131.[10] Simba, User's guide, 2005.

http://simba.ifak.eu/simba.[11] Tanaka K., Sugeno M., “Stability analysis and designer

of fuzzy control systems”, , vol. 45,1992, pp. 135-166.

[12] Yaochu Jin,. Physica - Verlag. Springer - Verlag Company,

2002.[13] Yoo C.K., Lee H.K., Beum Lee I., “Comparison of process

identification methods and supervisory control in thefull scale wastewater treatment plant”. In:

, Barcelona, 21 -26 July 2002.

Genetic Progra-ming Theory and Practice II.

IFAC 10 Symposium Large Scale Systems:Theory and Applications

Control EngineeringPractice

Fuzzy Sets Syst.

Advanced Fuzzy Systems Design and Appli-cations

15 IFACWorld Congress

th

th

th th

st th

VOLUME 4, N° 3 2010

Fig. 19. Daily performance of the control systemat the plant site.

Articles24

Abstract:

1. IntroductionThe drive of a mobile underwater robot is indisputably

associated with various kinds of propellers. However,there is the lack of such solutions in nature. Thus, a morenatural way of moving underwater seems to be worthtaking into consideration while designing a robot whichis going to be work in the water environment. There aremore and more surveys on such kind of propulsion, car-ried out in many scientific centres around the world.H. Kim and [6] studied the motion mechanism ofreal fishes and proposed the dynamic Lagrange'sequations of a fish robot modelled as a four-link system.A similar approach is presented by L. Zang and in[12], however the team focused on developing efficientdiving mechanism, which uses pectoral fins and fuzzylogic controller. The main propulsion is also implementedas 4-section tail, driven by four servomotors. Above twopapers refers to so called carangiform swimming whereasthere are also many surveys on different forms (anguil-liform, rajiform, gymnotiform etc.) of fish movement.Some of them are presented by K.H. Low in [8]. Authorhowever considered a gymnotiform robot, which mimica Black Ghost Knifefish. He presented several similar so-lutions as well as his own prototype of such a robot. Themathematical model of such propulsion is also describedin the paper.

The main goals of the project described in this paperwere to develop a concept, design and build prototype ofa carangiform robotic fish equipped with proximity sen-sors, temperature sensor, wireless digital miniature videocamera and wireless communication system. The maximalrepresentation of a fish-like movement was, however, themost important priority for the authors. There were seve-ral assumptions to the project. Firstly, the construction

In this paper, authors present a new approach to thedesign of a mobile underwater robot inspired by a fish. Theydescribe a prototype of a self-designed and a self-mademobile underwater robot called the CyberFish, which resem-bles fish in the way it looks and behaves. In the beginning,a short consideration on fish-like swimming is presented.Then the biological inspiration for swimming robots aredescribed by means of comparison of robots parts to fishorgans. In the next section authors focus on electroniccontrol system as well as on applications written in C/C++that are used to control the robot in three different modes.

et al.

et al.

Keywords: robotic fish, underwater mobile robot, biolo-gical inspiration.

should be cheap and easy to be built without using so-phisticated materials and tools. Secondly, parts used tobuild CyberFish should be easy available. Thirdly, the ro-bot should be able to operate autonomously as well asbeing controlled computer. Taking into considerationthese assumptions authors, in cooperation with collea-gue Dominik Wojtas, have studied fish movement andhave tried to find ways of translating and transforming itinto a mechanical device. Based on the findings andconclusions of the study, the 3D CAD model of CyberFishwas created in Catia v5 system. The results of the compu-ter simulation confirmed that the kinematics of the mo-del was correct. The underwater robot was capable toswim like a fish. Therefore, the physical prototype hasbeen built. The next step of the project was robot testingin the two thousand - litre tank. Tests showed that theproposed concept is correct and CyberFish swam likea real fish. Nevertheless, it was still much work to do. Themain tasks concerned with developing electronic controlsystem and software based on an appropriate controlalgorithms, which would give the robot the ability to beoperated computer or swim autonomously.

The concept itself appeared during studies of Auto-matics and Robotics at the Faculty of Mechanical Engine-ering of Cracow University of Technology. Authors wereinterested in making original bionic robot different fromexisting rolling robots created in the likeness of arach-nids or crustaceans. The real challenge was to design andbuild underwater mobile robot. The first thought was tocreate a fish-like device. In that case the robot must notbe driven by propeller but only by means of undulatingmovement of its “body”.

It is obvious that swimming is the most convenientway of moving underwater. In general, there are two ty-pes of fish swimming methods: BCF (body and/or caudalfin propulsion) and MPF (median and/or paired fin undu-lations) [3]. In the paper authors focused on BCF-likemotion. The many kinds of fishes swim by means of wavymovement of theirs body and/or tail. The frequency andamplitude of those vibrations depend on the species(Fig. 1). Anyway, the force that pushes fish forward isa result of consecutive muscle contractions. When fish isswimming, water is pushed sideways and backwards. For-ces, which act sideways, compensate each other whereasthe force which pushes water backwards gives reactionthat enables fish to move forward. The majority of fishspecies have two types of muscles. The white muscles

via

via

2. Biological inspiration

2.1. Fish-like swimming

FISH-LIKE SWIMMING PROTOTYPE OF MOBILE UNDERWATER ROBOT

Marcin Malec, Marcin Morawski, Jerzy Zając

Received 12 ; accepted 20 May 2010.th May 2010 th


Articles 25

give fish the ability to swim very fast and turn rapidly.The red muscles are used for smooth and gentle swim-ming without excessive fatigue. The types of muscles arenamed because of theirs colour [5].

Strong muscles themselves are insufficient to performunderwater manoeuvres. The flexible and lightweightskeleton is also very important. What is important, fishskeleton does not carry whole weight of the fish becausethe buoyancy force compensates to some extent the forceof gravity. Thus, the skeleton can contain hundreds ofsmall bones and cartilages, which form with musclesa very flexible construction that helps fish to swim veryefficiently.

Another very important fact is that the density of fishbody is close to density of water. This gives the fish theability to easy move vertically underwater by means of itsswim bladder and pectoral fins. The swim bladder isa hydrostatic organ in the shape of a flexible thin-walledcontainer, which can be filled with air thus changingbuoyancy of fish. The size of the swim bladder depends onthe species. Freshwater fishes have larger swim bladdersthan seawater fishes because of the differences in thedensity of fresh water and sea water [5]. In some kinds ofgroundfish and sharks there is no such organ [4]. Sharkuses its pectoral fins to dive and emerge. Its body hashigher density than water thus is unable to maintaindepth. Shark uses dynamic lift of their pectoral fins sothey sink when they stop swimming [10].

In order to build the prototype of fishlike underwatermobile robot, authors have to be acquainted with fishanatomy and its behaviour mentioned above. The firstimportant goal was to design a mechanism which kine-matics is similar to the undulating motion of a fish's bo-dy. The mechanism consists of four segments connectedin series with the rotary kinematic pairs. The head - thebiggest segment, two tail segments of similar size andtail-fin segment. First three of them contain drives. Nextsegment is driven by a servomotor placed in a previoussegment. When the mechanism is in motion the properrotation of each segment and appropriate synchroniza-tion of movement of segments create the effect similar toswimming motion of a fish. Computer simulation showsthat such motion is really similar to fish motion. Figure 2presents top view of the CyberFish 3D model in one mo-ment of movement.

Fig. 1. BCF fish motion a) mackerel, b) trout, c) eel [2].

2.2. Robot's design

Fig. 2. Motion of the proposed mechanism, a) the firstsegment the head, b) the second segment, c) the thirdsegment, d) the fourth segment and the caudal fin, e)superimposed image of sinusoid.

Fig. 3. The transmission of the second and the thirdsegment.

Each segment is driven by micro servomotor. Thatsolution is compact, cheap, and easy to control and giveshigh torque in comparison to its size. The transmissionfrom the servo shaft to the segment axis is carried out bygear with ratio 1:1 (Fig. 3).

Another important feature of fish anatomy is swimbladder, which is essential for depth control. Such anartificial organ was implemented in the CyberFish. Itconsists of two thin-walled silicone tubes sealed at oneend. The tubes can be compressed and stretched by theadditional servomotor and the special linkingmechanism. Such a pumping mechanism can draw waterthrough a hole in the bottom of the housing. Artificialbladder's servo is also used to change angle of pectoralfins by means of levers and ties. This allows currying outup-and-down motion of the robotic fish just like sharksdo when they swim. Diving mechanism has also got thesmall additional weight, which moves forward whilepectoral fins move up, and moves backward in theopposite case. This changing slightly the centre ofgravity of the robot and allows the CyberFish to swim likea real fish. Diving mechanism is shown in the Fig. 4.

Based upon the 3D CAD model of the robot, theprototype has been built using PCV, acrylic, rubber,aluminum and stainless steel. The volume of the robotwas estimated by Catia software and used to calculatebuoyancy that allows the prototype to float in waterrather than sink. The mass of CyberFish's body is 3.5 kgand the robot's density is slightly lower than the densityof water. This solution enables the robot to change itsdepth using small changes in volume of the swim bladder.The CyberFish operating underwater is shown in Fig. 5.


Articles26

VOLUME 4, N° 3 2010

des: (1) autonomous mode, (2) manual control com-puter equipped with wireless communication module and(3) tracking submersed object recognized in camera ima-ge by image recognition algorithm. In order to fulfil thatassumption, the control system was divided into twoparts. The low level control part is based on micro control-ler electronic board embedded in the CyberFish, whereasthe high level control part is formed by an external com-puter with control software. The communication betweenthese two parts is performed by means of exchangingmessages, which are sent wireless communicationchannel. The autonomous mode is fully implemented inembedded part of the control system, thus the robot needno extra device to operate underwater. In this mode, oneof eighteen predetermined robot's activities is randomlyselected at 15 seconds intervals. Four proximity sensors,mounted on the head of the CyberFish, are turned on.Thus the robot is able to detect and avoid obstacles. Theautonomous mode can be turned on when no data is recei-ved from the computer for more than 60 seconds. If it ison and the robot receives messages, it immediatelyswitches to the manual control.

The high level part of the control system is based ona specially designed computer software which gives anoperator the ability to control the robot by clicking but-tons or pressing keys on the keyboard. An image recogni-tion algorithm is also implemented in the software. Itrecognizes red round object in the video received from therobot's onboard wireless video camera. Based upon thecoordinates of the centre of the object in the image, algo-rithm sends messages to the robot in order to maintainthe object in the middle of the frame.

The core of the robot's electronic control board is theAtmel Atmega 32 micro controller clocked by 8 MHz crys-tal. The typical application of the Atmega micro controller

via

via

3.1. Hardware

Fig. 4. Diving mechanism, a) servo, b) pectoral fin, c) pec-toral fin's axis, d) two silicon tubes, e) additional weight.

The work on the control system has been carried outsimultaneously with the building of the mechanical partof the prototype. The concept assumes three control mo-

Fig. 5. The prototype of fish-like mobile underwater robot -The CyberFish.

3. Control system


Articles 27

VOLUME 4, N° 3 2010

Fig. 6. The complete scheme of the robot's electronic control board.

which can be found in [1], was enriched with: two stabi-lized power supplies, NE555 timer to generate proximitysensors carrier wave signal, transistor keys used to modu-late proximity sensors signal, connectors used to connectthe radio communication module, servo motors, DS18B20digital thermometer, IR detectors, video camera andprogramming socket. The complete scheme is shown inFig. 6.

Several parts of the scheme need to be commented.First of all robots proximity sensors are built with fourTSOP1736 IR detectors, which need the appropriate inputsignal. The signal consists of packets of 30 pulses with thefrequency of 36 kHz with approximately 9 ms gap. The36 kHz square wave is generated by NE555 timer. Thissignal is modulated by the appropriate signal from microcontroller with the use of set of transistor keys (T1 - T8).Such a modulated signal is send by IR diode and if itreflects off the obstacle, TSOP detects it and sets a lowstate on its output.

Two sets of stabilized power sources are used to eli-minate interference caused by DC servos motors. Servosare supplied from the separate 6 V source whereas otherdevices are supplied from the 5 V source. Each source issupplied from Ni-MH 9.6 V 2700 mAh battery. ConnectorsJP12 and JP13 are used to either connect charger (pins 2and 3) or supply the system (shorting pins 1 and 2). Theminiature wireless video camera mounted in the front ofthe robot and connected to JP11 is supplied directly frombatt1 by K1 relay. This enables the operator of the robot

to turn the video camera on and off depending on theneeds. The video signal from the camera is received by thecomputer with the use of video camera receiver and a USBTV tuner. The receiver gives a composite video signal onits output, which is then transferred, to the USB TV tuner,which works as an image-capturing device.

The wireless communication module MOBOT RCRv2type A is connected to USART port by the JP5 connector.The micro controller communicates with MOBOT by asyn-chronous serial transmission, which parameters are as fol-lows: 56 kbps, 8 data bits, 1 stop bit, and no parity. Ano-ther MOBOT RCRv2 type B module is connected to PCUSB and communicates with type A module by using433 kHz radio signal.

The DS18B20 digital thermometer is connected to themicro controller by means of one-wire interface using PD2line. The temperature sensor is located near the dorsal finof the CyberFish.

The micro controller software has been written in Cusing the WinAVR development environment and avrgcccompiler. Setting WinAVR to work with the compiler wasmade with help of information presented in [7]. Compiledcode was written to the device by means of STK200 serialprogrammer and PonyProg 2000 application. The programconsists of various functions like initialisation of: timers,USART module and one-wire interface, which are calledbefore the main control loop. When the program enters

via

3.2. Software


Articles28

VOLUME 4, N° 3 2010

Fig. 7. The robot's control application window.

the main loop, function used to analyse messages is beingcalled. It compares incoming messages received from theUSART to messages stored in memory and then the appro-priate action is being performed. Moreover, message ofconfirmation is sent back to the computer. If no messageis received from USART within the 60 seconds and theautonomy flag is set, the autonomy function is beingcalled. The robot switches to the autonomous mode des-cribed previously in the paper. Sending and receivingmessages are performed in the USART interrupt handlers.PWM signals are software generated within the timerinterrupt handler. This is because Atmega32 is not equip-ped with a sufficient number of independent PWM chan-nels to control four servomotors. There is a set of variousfunctions, which are being called after receiving a messa-ge. These functions adjust the duty cycle of PWM in orderto achieve appropriate servo movement and synchro-nization.

The micro controller program also contains functionsused to handle communication with the DS18B20temperature sensor and a function to receive signals fromproximity sensors.

PC computer application has been written in C++ byDominik Wojtas using the Microsoft Visual Studio 2008development environment. OpenCV and EmguCV free com-puter vision libraries were used to build an image recogni-tion application, which is further used to tracking sub-mersed red round object by the CyberFish. The applicationitself is a single window form (Fig. 7), which containsseveral areas:

communication parameters settings area,a text box used to send and receive messages (controlcommands),a set of buttons to control the robot,image recognition parameters settings area,a frame grabbing device settings area,a video screen with the resolution of 640x480 pixelson which the underwater view (as well as detectedobject) is displayed.

The image recognition algorithm was developed withthe help of information contained in [11]. Images recei-ved from a capturing device are the algorithm's input datawhereas commands used to control the robot are theoutput data. The algorithm consists of six major steps:

capturing image,image processing,frame binary conversion using appropriate threshold,morphological operation of closing and opening to fillgaps in the image of the object,calculating coordinates of the centre of the object inthe image,coordinates analysis during fixed time intervals.

Based upon coordinates analysis of the centre of theobject in the image, appropriate commands are being sentto the CyberFish in order to maintain the object in themiddle of the frame. If the result of the analysis locatesthe object in the left side of the screen during fixed timeinterval, „turn left“ command is being sent. A similar situ-ation is observed when result of the analysis locates theobject in the right side of the screen. If the result of the

�

�

�

�

�

�

�

�

�

�

�

�

analysis locates the object in the middle of the screen, inthe so-called “dead zone”, there is no reaction from thecontrol system. Taking into account problems with main-taining undisturbed signal from the onboard video cameraat the greater depths, authors resigned from implemen-ting up-and-down control in this control mode. Therefore,tracking submersed object by the robot works only whenthe CyberFish swims just below the surface of the water.Using more expensive and sophisticated wireless videocamera should allow to implement up-and-down controlin this control mode.

After several months of work on Master of ScienceThesis in Automation and Robotics at Cracow Universityof Technology, the CyberFish has finally been made. Therobot was designed so that it could be built using thecheapest materials and widely available tools and parts.Financial constraints did not allow for the implementa-tion of sonar system or sophisticated video camera,which could operate at greater depths or at low light in-tensity. However, with the use of popular and well-knownsolutions, it was possible to build a unique underwatermobile robot, which has been tested in 2000 litres pool.Despite the difficulty of sealing, construction, manufac-turing and logistics problems and bugs in the software,the aim of the project has eventually been achieved. TheCyberFish represents an original underwater craft thatcan be drive without propeller. This solution seems to bemore efficient or even irreplaceable if the device is goingto operate in rushes or seaweed. Any further developmentwould require funding that allows to create an under-water robot performing various functions, ranging fromanalysis of water pollution, ending the stand-alone waterpenetration in the search for missing items or people.

- CracowUniversity of Technology, 31-864 Kraków, Al. Jana PawłaII 37. E-mails:[email protected],[email protected].* Corresponding author

4. Conclusions

AUTHORSMarcin Malec*, Marcin Morawski, Jerzy Zając

References[1] Baranowski R., ,

Warsaw: BTC, 2005. (in Polish)[2] Chmielewski T., , http://ryby.fishing.pl/

dodatek_3.php, October 2008. (in Polish)[3] Evans D.H., The Physiology of Fishes. ,

Boca Raton (Florida): CRC Press LCC, 1998, pp. 3-25.[4] Frey H., , Warsaw: Sport i Turys-

tyka Publ. Comp., 1990, pp. 168-174. (in Polish)[5] Jobling M., , London:

Chapman & Hill, 1995, pp. 251-297.[6] Kim H., Lee B., Kim R., „A Study on the Motion Mecha-

nism of Articulated Fish Robot”. In:

, 2007, Harbin, China, pp. 485-490.[7] Koppel R., „Programowanie procesorów w języku C”,

Mikrokontrolery AVR ATmega w praktyce

Wciąż do przodu...

Second Edition

Akwarium Słodkowodne

Environmental Biology of Fishes

Proceedings of the2007 IEEE International Conference on Mechatronics andAutomation


Articles 29

VOLUME 4, N° 3 2010


Articles30

Elektronika dla wszystkich

Mechanism andMachine Theory

Pomiary - Automatyka -Robotyka

http://en.wikipedia.org/wiki/Fish_locomotion

Praktyka analizyobrazu

Proc. of the 2007 IEEE/RSJ International Conferenceon Intelligent Robots and Systems

, no. 5, 2005, pp. 36-39.(in Polish)

[8] Low K.H., „Modelling and parametric study of modularundulating fin rays for fish robots”,

, vol. 44, 2009, pp. 615-632.[9] Malec M., Morawski M., Wojtas D., Zając J., „CyberRyba

podwodny robot mobilny”,, no 2, 2010, pp. 331-340. (in Polish)

[10] Wikipedia, the free encyclopedia, „Fish locomotion”,, April

2010.[11] Wojnar L., Kurzydłowski K.J., Szala J.,

, Cracow: Polskie Towarzystwo Stereologiczne,2002. (in Polish)

[12] Zhang L., Zhao W., Hu Y., Zhang D., Wang L., „Develop-ment and Depth Control of Biomimetic Robotic Fish”.In:

, 2007, San Diego,pp. 3560-3565.

VOLUME 4, N° 3 2010

Abstract:

1. Introduction

2. Partial discharges

Present technological solutions enable us to work outmore and more new concepts for completions of thoseproblems, which used to be difficult to solve before. Thedevelopment of bio cybernetics as well as computerscience has created favourable conditions for supportingautomation and robotics by means of computer mecha-nisms [1]. Even few years ago a direct control of a humanbrain by such devices as robots was difficult to completein terms of technical requirements [2]. Nowadays it ispossible to analyse brain waves emitted by neurons bymeans of electroencephalography. Then, after an appro-priate classification, they may be used in the process ofcontrolling [3]. The present article describes the con-cepts of the use of brain waves collected by an electro-encephalograph in order to control a mobile robot, whichwould help an automatization of the process of a mea-surement of partial discharges

Partial discharges (PD) are those phenomena, whichoccur in isolators of such electrical devices as trans-formers, energetic capacitors, electric motors, etc. Thisphenomenon is unfavourable in terms of the descriptionof an insulator, and an increasing frequency of theoccurrence of partial discharges means its considerabledegradation. The occurrence of different physical mecha-nisms, which accompany both complete and partialelectric discharges, shows a complexity of this issue.Signs of damages of insulating materials are accompa-nied with physiochemical factors. The most important ofthese factors are [4], [5], [6]:- the emission of an electromagnetic wave, from the

place where a discharge appears, which is a result ofa power-driven impulse,

- chemical conversions within the structure of insu-lation,

The article presents the concepts of a mobile system formeasurements of partial discharges controlled by brainwaves. In order to describe that, a robot, which takes mea-surements of partial discharges, has been worked out. Thedischarges may occur in an isolator of electrical devicessuch as capacitors and transformers. What is more, a con-cept of a link between the robot and a human brain isdescribed, in order to ensure a direct communication on thelevel of the human brain and the robot.

Keywords: EEG, a mobile robot, partial discharges.

.

- elastic strains on the molecular level which lead tothe emission of a sound wave,

- light flashes, which emit radiation on the level ofa visible, an infrared and an ultraviolet spectrum,

- local implosions, which cause a rise in temperatureand changes in a pressure of gas.

The occurrence of specific factors is directly connec-ted with the complexity of energetic devices, whichcontain insulators. For that reason, methods of detectionof partial discharges are suggested that are based on thedetection about the level of the occurrence of a particularphysiochemical phenomenon [7]. Nowadays the firstplace in diagnostics occupies non-invasive methods.In terms of partial discharges, the most important aremethods for the measurement of an Acoustic Emission(AE) [8, 9, 10] and an Optic Spectrum Diagnostics (OPD)[11, 12]. Figure 1 presents a modular diagram of thesystem for the measurement of partial discharges bymeans of the acoustic emission’s method.

The main elements of the structure are a piezoelectricdetector, input measuring amplifiers, a digital to ana-logue rapid converter and a system of data collection. Thepiezoelectric detector is situated close to the place wherea partial discharge appears. In terms of measurements on

Fig. 1. A modular diagram of the detection of the partialdischarges by means of the acoustic emission’s method.

A MOBILE SYSTEM FOR MEASUREMENTS OF PARTIAL

DISCHARGES CONTROLLED BY ELECTROENCEPHALOGRAPHIC WAVES

Andrzej Błachowicz, Szczepan Paszkiel

Received 20 ; accepted 26 .th January 2010 April 2010th


Articles 31

a real object, it is necessary to ‘sound’ as big as possiblesurface in order to determine areas with a high proba-bility of the occurrence of partial discharges. The place ofthe contact between the detector and the surface of theobject should be fitted, as much as it is possible, in orderto eliminate auditions from the surrounding. It is signi-ficant to have the opportunity to transfer the detectorswhat makes it easier to conduct the diagnosis. Figure 2presents a configuration of the system in terms of themeasurements of partial discharges by means of the me-thod of detection of the optical spectrum. Three opticaldetectors are situated close to the place where the phe-nomenon of partial discharges appears. Every detector isadapted to a different optical band. The bands of a visi-ble, an ultraviolet and an infrared light are enumerated.

The detectors are built from semiconductor, photo-sensitive elements which purpose is directed to specificlengths of a light wave. During the measurements, a con-siderable activity in terms of the ultraviolet waves isobserved, what directs research in this field. Because ofa very low energy level of partial discharges, which meana degradation of the insulator, very sensitive measuringtracks should be used. According to this fact, the detec-tors should be equipped with a high selectivity of theband, a high level of the reinforcement and a very goodcoefficient of a condition between the interference andthe useful signal. Because of that, highly selected mea-suring detectors should be used for work with photosen-sitive semiconductor elements. It is not necessary to userapid measuring systems because of a low dynamics ofthe signal. The digital to analogue processing in terms ofthe method of detection of partial discharges, under pre-sent consideration, does not have to exceed 100 kS/s.Then, there is a real opportunity to use a converter withan effective resolution above 16 bits.

Fig. 2. A modular diagram of the detection of partial dis-charges by means of the method of the optical spectrum’smeasurement.

3. The mobile robotThe possible measurements constitute a real technical

challenge within areas where there are big difficulties for

analysts with the access to the subject under research.A significant advantage of the mobile research unit is theopportunity to deliver the equipment to those placeswhere the staff, which operates the robot, may suffer fromhealth damages. The SQ1 robot is presented in the Figures3 and 5. It is designed and equipped with an apparatusthat enables to move it within the area of its potentialuse. The robot is fitted with such technologies as ultra-sound detectors of distance, accelerometrical detectorslocated on the robot’s legs and on the central part of itsmain body, digital cameras and laser scanner. The com-munication between an operator and the robot takes pla-ce cordlessly. The operator is able to guide the robot inthe architectural space of a distribution board of averagevoltage by making use of an original application. Mea-suring samples are taken through the detectors that areplaced on the research object. Information from the de-tectors is also transferred by means of radio waves.

The mobile measuring system consists of few modules,which enable the operator to work remotely. The mainelements of the mobile robot are a cordless moduledefined by the IEEE 802.11 standard ( –

) [13], which enables a direct communication withthe PDDetect2 controlling application; the main modulewhich consists of the measuring part; a cordless trans-mission channel ( ) which enables the communi-cation with the measuring detectors; an active underbodyand a monitoring module of work parameters of the SQ1robot. The module structure of the mobile measuring sys-tem is presented in the Figure 4.

The robot is equipped with an arm that enables theoperator to install elements of the measuring systems onthe surface of the research object. During a diagnosis,there is an analysis of the partial discharges that areemitted, together with a simultaneous consolidation ofthe knowledge of the spatial position of the detector’sinstallation in a base. In the future it will enable to docomparative research and to determine degradation’strends, which appear in an insulating system under re-search. The system of magnetic clutches, which are placedon the arm, enables to operate small modules of detectorsmore easily and to attach them to the surface of theobject.

Software of the measuring mobile system was comple-ted with support of Real Time Operation System (RTOS).

Fig. 3. The mobile robot simulated in AutoDesk Inventor2009.

Wireless FidelityWiFi

Bluetooth


Articles32

VOLUME 4, N° 3 2010

analysis and at the same time it does not causeinterferences in the work of the energy system. Regulardiagnosis of the elements of the energy system and aconsolidation of their parameters enable to gatherinformation, which are useful within the process ofplanning the service. In addition, the measuring mobilesystem memorises the track and the topology of the areaof the electro energetic station and the place of theinstallation of objects, which are diagnosed.

The PDDetect2 application, installed in the workingstation from which the process of control of the robottakes place, enables the implementation of this processby an analysis of the electroencephalographic signal. TheEEG signal is collected with a non-invasive method bymeans of active electrodes that are placed on the skin ofa person’s head under research [14]. The arrangement ofthe electrodes is specified by the 10-20 IFCN interna-tional specifications. Then, the signal is transported toan electroencephalograph, which is connected by meansof a USB 2.0 port to a working station with the PDDetect2application. Software accomplishes an appropriate clas-sification of the EEG signal in order to specify conclu-sions about the robot’s reactions on brain stimuli ina particular moment of time. Figure 7 illustrates ideasabout implementations of the communication with theuse of the EEG signal.

Fig. 6. The module of the detector: active piezoelectricprobes.

4. The control of the robot by meansof the brain

Particular modules, which constitute the robot, weredesigned according to the architecture suggested by theproducer of RTOS. The use of resources accessible in thelibrary of the operation system gives a chance for a fastimplementation of new functions, which lead to thedevelopment of the measuring mobile system.

A module of a cordless detector was implemented bythe use of a fast microcontroller with an ARM root. Itallows a registration of the signal of acoustic emissiongenerated from partial discharges with a speed to 8 MS/s.Samples registered by the detector (Figure 6) are sent tothe mobile robot. In sequence, data are gathered as a filein the fixed memory. An access to information happenswith the use of a FTP server (by means of the FTP client) orwith a PDDetect2 application, which simultaneously im-plements the process of visualization and an analysis.

The automation of the process, which eliminates thenecessity of disconnecting a live element of the electroenergy system, remarkably shortens the time of the

Fig. 5. The mobile robot – a laboratory photograph.


Articles 33

VOLUME 4, N° 3 2010

Fig. 4. The module structure of the mobile measuring system.

Fig. 7. The module structure of the system of the collectionof brain waves.

An appropriate analysis of the oscillation of the EEGsignal in terms of the ? , , rhythms has an impact onthe process of a proper control of the robot [15]. Duringthe analysis of the EEG signal a big synchronisation of the

rhythms in the attention process can be seen, what isessential from the point of view of the control of themobile robot. In the same time, a big activeness ofpyramidal cells appears and a strong synchronisation ofthem. The rhythms are seen during an intensity of theamount of the processing information by neurons in theunit of time ‘t’. In terms of the rhythms, a synchroni-sation of activity takes place, which is directly connectedwith the processing of information [16], [17]. Nonlinera-lity has a direct impact on the frequency of the rhythmsand their shapes that is connected with specific characterof stimulation, which appears then. Oscillations of the

rhythms can be observed on many electrodes withinareas of those that were placed directly above the motorcortex. During a measurement of the electroencephalo-graphic signal, it is also possible to observe a disynchro-nisation of the signal, which is seen in the difference ofthe energy of the measuring signal. The measurement ofthe disynchronisation of the signal is based on the studyof its power (1).

(1)

where is the value of an m-point of the signal in an n-repeat of the experiment and means the number ofrepeats. Then, disynchronisation may be defined as itfollows (2).

(2)

where is the level of reference (3).

(3)

– the length of the reference area,– the point of the signal.

Besides the non-invasive Brain Machine Interfacesuggested in this article, there are also invasive methods.This type of BMI was described by scientists in the UnitedStates and Europe. They use a surgical implant, which

� � �

�

�

�

�

,

,

AN

Z

lm

consists of beams of electrodes. From the technicalpoint, it is more difficult to implement and less practicalin such uses as the mobile system for the measurementsof partial discharges. That is why the non-invasive me-thod was used for the system of measurement of partialdischarges.

The mobile measuring system for research of partialcharges, which occur in the insulating structures duringthe progressive process of the degradation of the dielec-tric material in electroenergetic capacitors with the useof the EA method, will be tested during diagnostic inves-tigations in factory conditions. It will enable to specifythe scope of practical adaptations. Experimentations du-ring big levels of electromagnetic interferences, whichaccompany work of electroenergetic devices, will enableto state directions towards further transformations of thesystem what will result in its improvement. The use of themobile system of recording of the EA signals generated bypartial discharges in the insulation of the electroener-getic capacitors will allow among others to automatize ofthe measuring system which was conducted manually be-fore, to automate the process of registration, to presentthe results (automatic change of places where convertersare situated, the way of its connection with a particularobject), to separate the technical personnel effectivelyfrom the place of measurement ensuring security; a mobi-lity – the possibility for sending measuring data from theplace of measurement to the station of the operationby means of a cordless network based on the TCP/IPprotocol.

Undoubtedly, the greatest problems, which occur onthe stage of correlation between a human brain and themobile robot, are different kinds of both biological andtechnical interferences [18], [19]. Technical artefactsseem to be especially significant in the system under pre-sent analysis because the system works near big concen-trations of electroenergetic devices, which disrupt a pro-per reading of the signal [20], [21]. According to thisfact, it is necessary to isolate the control station from themobile robot’s area of work. Apart from the technical pro-blems mentioned above, a certain biological criteria arerequired from the person who controls the robot. A specialattention should be paid to people with high blood pres-sure, accelerated pulsation of heart or those who possessdifferent kind of nervous tics. Unfortunately, such peoplecannot control the mobile robot in an easy way.

The control of the mobile robot by means of brainwaves also requires a constant improvement. It is neces-sary to elaborate on population models that are imagingsof particular populations of neurons and their mutualcorrelations. By means of advanced mathematical me-thods, it is possible to simulate proper preservation of theEEG signal in specific mental states of a human being. It isthe key to success in the case of the structure and theimplementation of both Brain Computer Interface andBrain Machine Interface communication.

5. Problems on the stage of the completionof the connection between the brain andthe mobile robot

6. Conclusion


Articles34

VOLUME 4, N° 3 2010

AUTHORSAndrzej Błachowicz*, Szczepan Paszkiel

References

- Opole Uni-versity of Technology, Faculty of Electrical Engineering,Automatic Control and Computer Science, Institute ofElectrical Power Engineering, uL. Prószkowska 76, 45-758Opole. Tel. +48 600 494 771, +48 600 937 899. E-mail:[email protected].* Corresponding author

[1] Georgopoulos A.P., Langheim F.J., Leuthold A.C.,Merkle A.N., “Magnetoencephalographic signals predictmovement trajectory in space”, , vol. 25,2005, pp. 132–135.

[2] Dudek Z.T., „Interfejs BCI – próba przełamania barierypomiędzy człowiekiem a komputerem”,

, no. 7,2003. (in Polish).

[3] Tortella F., Rose J., Robles L., Moreton J., Hughes J.,Hunter J., “

”, Maryland 1997.[4] Kawada H., Honda M., Inoue T., Ameniya T., “PD auto-

matic monitor for oil-filled power transformer”,, vol. 103, no. 2, 1984, pp. 422-428.

[5] Kirtley J.L., Lavalle J.C., McCarthy D.J., “Acousticmonitoring of transformer structures”, ,Report no. 3-05.87, Vienna, 1987, pp. 1-5.

[6] Kucharska B., „Akustyczne deskryptory sygnałów emito-wanych przez wyładowania niezupełne powierzchnio-we”, ZN PO, no. 219, , vol. 44, 1996, pp. 107-124. (in Polish)

[7] Kelley N.D., Osgood R. M., Bialasewicz J.T., JakubowskiA., “Using time-frequency and wavelet analysis toaccess turbulence/rotor interactions”. In:

, Reno, Nevada, 10 –13 January, 2000.[8] Florkowska B., Zydroń P., „Analiza widmowa ulotu

elektrycznego”. In:, Poznań-Kiekrz, 2002, pp. 80-83. (in Polish)

[9] Fuhr J., Maessing M., Boss P., Tschudi D., King R.A.,“Detection and Location of Internal Defects in theInsulation of Power Transformers”, ,vol. 28, no. 6, 1993, pp. 10057-1069.

[10] Howells E., Norton E.T., “Parameters affecting thevelocity of sound in transformer oil”, , vol. 103,no. 5, 1984, pp. 1111-1115.

[11] Rabbani M., Santa Cruz D.,,

Thessaloniki, Greece 2001.[12] Skubis J., „Emisja akustyczna w badaniach izolacji

urządzeń elektroenergetycznych”, , Opole1993. (in Polish)

[13] Zmarzły D., „”, Opole University of Technology, Opole 2004.

(in Polish)[14] Venter H.S., Eloff J.H.P.,

, , RandAfrikaans University, 1992, pp. 147-154.

[15] Larrieu N., Owezarski P.: “Towards a Measurement BasedNetworking approach for Internet QoS improvement”,

, vol. 28, 2005, pp. 259-273.

Exp. Brain Res.

Przegląd Teleko-munikacyjny i Wiadomości Telekomunikacyjne

EEG Spectra Analysis of the NeroprotectiveKappa Opioids Enadoline

IEEEPAS

CIGRE Symp.

Elektryka

19 AmericanSociety of Mechanical Engineers (ASME) Wind EnergySymposium

V Ogólnopolskie Sympozjum IWN- IW2002

IEEE Trans. on EI.

IEEE PAS

The JPEG 2000 still-imagecompression standard. Materials of course in ICIP

IPPT PAN

Pomiary elektrycznych wielkości medycz-nych

“Network Security: ImportantIssues” Department of Computer Science

Computer Communications

th

th th

[16] White T., Pagurek B., Bieszczad A., “Network Modelingfor Management Applications Using Intelligent MobileAgents”, ,vol. 7, no. 3, 1999.

[17] Todorova P., “Network Control in ATM-Based LEO Satel-lite Networks”, , 22:1–4,2003, pp. 321–335.

[18] Bush S.F., Frost V.S., “A Framework for PredictiveNetwork Management of Predictive Mobile Networks”,

, vol. 7, no.2, 1999.

[19] Ibrahim T., Venin J.M., Garcia G., “Brain ComputerInterface in Multimedia Communication”,

, vol. 20, no. 1, 2003.[20] Wolański N., „

”, Warszawa2005, p. 374. (in Polish)

[21] Schumacher H.J., Ghosh S., “A fundamental frameworkfor network security”,

, 1997, pp. 305–322.

Journal of Network and Systems Management

Telecommunication Systems

Journal of Network and Systems Management

IEEE SignalProcessing Magazine

Rozwój biologiczny człowieka. Podstawyauksologii, gerontologii i promocji zdrowia

Journal of Network and ComputerApplications


Articles 35

VOLUME 4, N° 3 2010

Abstract:

1. IntroductionSome remarks on fuzzy inference systems are conside-

red in the first subsection. The second subsection inclu-des a brief review of methods of extracting of fuzzy rulesbased on fuzzy clustering and the aims of the paper.

Fuzzy inference systems are one of the most famousapplications of fuzzy logic and fuzzy sets theory. Theycan be helpful to achieve classification tasks, processsimulation and diagnosis, online decision support toolsand process control. So, the problem of generation offuzzy rules is one of more than important problems in thedevelopment of fuzzy inference systems.

There are a number of approaches to learning fuzzyrules from data based on techniques of evolutionary orneural computation, mostly aiming at optimizing para-meters of fuzzy rules. From other hand, fuzzy clusteringseems to be a very appealing method for learning fuzzyrules since there is a close and canonical connection bet-ween fuzzy clusters and fuzzy rules. The idea of derivingfuzzy classification rules from the data can be formulatedas follows: the training data set is divided into homoge-neous group and a fuzzy rule is associated to each group.

Fuzzy clustering procedures are exactly pursuing thestrategy: a fuzzy cluster is represented by the cluster cen-ter and the membership degree of a datum to the clusteris decreasing with increasing distance to the clustercenter.

So, each fuzzy rule from a fuzzy inference system canbe characterized by a typical point and membership func-

The interpretability and flexibility of fuzzy classifica-tion rules make them a popular basis for fuzzy controllers.Fuzzy control methods constitute a part of the areas ofautomation and robotics. The paper deals with the methodof extracting fuzzy classification rules based on a heuristicmethod of possibilistic clustering. The description of basicconcepts of the heuristic method of possibilistic clusteringbased on the allotment concept is provided. A general planof the D-AFC(c)-algorithm is also given. A method of cons-tructing and tuning of fuzzy rules based on clustering re-sults is proposed. An illustrative example of the method'sapplication to the Anderson's Iris data is carried out. Ananalysis of the experimental results is given and prelimi-nary conclusions are formulated.

Keywords: possibilistic clustering, fuzzy cluster, typicalpoint, tolerance threshold, fuzzy rule.

1.1. Preliminaries

tion that is decreasing with increasing distance to thetypical point.

Let us consider some methods of fuzzy rules extrac-ting from the data using fuzzy clustering algorithms.Some basic definitions must be given in the first place.

The training set contains data pairs. Each pair ismade of a -dimensional input-vector and a -dimen-sional output-vector. We assume that the number of rulesin the fuzzy inference system rule base is . So, Mamda-ni's [1] rule within the fuzzy inference system is writtenas follows:

, (1)

where and are fuzzysets that define an input and output space partitioning.

A fuzzy inference system, which is described by a setof fuzzy rules with the form (1) is the multiple inputs,multiple outputs system. Note that any fuzzy rule withthe form (1) can be presented by rules with the form ofmultiple inputs, single output:

(2)

Let be characterized by the membership function. The membership function can be triangular,

Gaussian, trapezoidal, or any other shape. In this paper,we consider trapezoidal and triangular membership func-tions.

Fuzzy classification rules can be obtained directlyfrom fuzzy clustering results. In general, a fuzzy cluste-ring algorithm aims at minimizing the objective function[2]

(3)

under the constraints

(4)

and

(5)

where is the data set, is the num-ber of fuzzy clusters in the fuzzy -parti-tion is the membership degree of object

1.2. Fuzzy clustering and fuzzy rules

n

c

m c

cl

c

c

AUTOMATIC GENERATION OF FUZZY INFERENCE SYSTEMS USING

HEURISTIC POSSIBILISTIC CLUSTERING

Dmitri A. Viattchenin

Received 10 ; accepted 26 .th February 2010 April 2010th


Articles36

to fuzzy cluster is the prototype for fuzzycluster is the distance between prototype

and object , and the parameter is the fuzzinessindex. The selection of the value of determines whetherthe cluster tend to be more crisp or fuzzy. Membershipdegrees can be calculated as following

(6)

and prototypes can be obtained from the formula

(7)

Equations (6) and (7) are necessary conditions for (3)to have a local minimum. However, the condition (5) ishard from essential positions. So, a possibilistic appro-ach to clustering was proposed in [3]. In particular, theobjective function (3) is replaced by

(8)

under the constraint of possibilistic partition

(9)

where is the number of fuzzy clusters ,in the possibilistic partition is the possibi-listic memberships which are typicality degrees,is the prototype for fuzzy cluster is the dis-tance between prototype and object , and the para-meter is the analog of the fuzziness index. Typica-lity degrees can be calculated as following

(10)

and the parameters are estimated by

(11)

where .The principal idea of extracting fuzzy classification

rules based on fuzzy clustering is the following [2]. Eachfuzzy cluster is assumed to be assigned to one class forclassification and the membership grades of the data tothe clusters determine the degree to which they can beclassified as a member of the corresponding class. So,with a fuzzy cluster that is assigned to the some class wecan associate a linguistic rule. The fuzzy cluster is pro-jected into each single dimension leading to a fuzzy seton the real numbers. From a mathematical position themembership degree of the value to the th projec-tion of the fuzzy cluster is the su-premum over the membership degrees of all vectors with

as th component to the fuzzy cluster, i.e.

(12)

c

t

t

.... .............

............

or

(13)

in the possibilistic case. An approximation of the fuzzyset by projecting only the data set and computing theconvex hull of this projected fuzzy set or approximating itby a trapezoidal or triangular membership function isused for the rules obtaining [4].

Objective function-based fuzzy clustering algorithmsare the most widespread methods in fuzzy clustering [2].Objective function-based fuzzy clustering algorithms aresensitive to initial partition selection and fuzzy rulesdepend on the selection of the fuzzy clustering method.In particular, the GG-algorithm and the GK-algorithm offuzzy clustering are recommended in [2] for fuzzy rulesgeneration. All algorithms of possibilistic clustering arealso objective functions-based algorithms.

Heuristic algorithms of clustering display low level ofa complexity. An outline for a heuristic method of pos-sibilistic clustering was presented in [5], where a basicversion of direct possibilistic clustering algorithm wasdescribed and the version of the algorithm is called the D-AFC(c)-algorithm [6].

The main goal of the paper is a detail consideration ofthe method of the rapid prototyping fuzzy inferencesystems, which was outlined in [7]. The method is basedon deriving fuzzy classification rules from the data on abasis of clustering results obtained from the D-AFC(c)-algorithm. The contents of this paper is as follows: in thesecond section basic concepts of the possibilisticclustering method based on the concept of allotmentamong fuzzy clusters are outlined and a plan of the D-AFC(c)-algorithm is given, in the third section a methodof constructing of fuzzy rules is proposed, in the fourthsection an illustrative example of deriving fuzzy rulesfrom the Anderson's Iris data are given, in the fifthsection preliminary conclusions are stated and someperspectives are outlined

The basic concepts of the heuristic method of pos-sibilistic clustering are considered in the first subsection.A plan of the direct clustering algorithm is given in thesecond subsection. The third subsection includes a re-view of methods of the data preprocessing.

The D-AFC(c)-algorithm is based on a concept of anallotment of elements of the set of classified objectsamong fuzzy -clusters. The allotment of elements of theset of objects among the fixed number of fuzzy -clus-ters can be considered as a special case of possibilisticpartition. The fact was demonstrated in [6] and [8]. Thatis why the basic version of the algorithm, which is des-cribed in [5], can be considered as a direct algorithm ofpossibilistic clustering and the algorithm was called theD-AFC(c)-algorithm [6].

Let us remind the basic concepts of the heuristic me-thod of possibilistic clustering. The concept of fuzzy

.

2. A heuristic method of possibilisticclustering

2.1. Basic concepts

��c


Articles 37

If condition

(18)

is met for all fuzzy clustersthen the family is the allotment of elements of the set

among fuzzy clustersfor some value of the tolerance threshold .

It should be noted that several allotmentscould exist for some tolerance threshold . That is whysymbol is the index of an allotment.

The condition (18) requires that every objectmust be assigned to at least one fuzzy

cluster with the membership degreehigher than zero. The condition requires thatthe number of fuzzy clusters in each allotmentmust be more than two.

Obviously, the definition of the allotment amongfuzzy clusters (18) is similar to the definition of thepossibilistic partition (9). So, the allotment among fuzzyclusters can be considered as the possibilistic partitionand fuzzy clusters in the sense of (16) are elements of thepossibilistic partition.

If condition

(19)

where and condition

(20)

are met for all fuzzy clusters of some allot-ment then the allotment isthe allotment among particularly separate fuzzy clustersand is the maximum number of elements in theintersection area of different fuzzy clusters. Obviously, if

in conditions (19) and (20) then the intersectionarea of any pair of different fuzzy cluster is an empty setand fuzzy clusters are fully separate fuzzy clusters.

The adequate allotment for some value oftolerance threshold is a family of fuzzy clusterswhich are elements of the initial allotment for thevalue of and the family of fuzzy clusters should satisfythe conditions (19) and (20).

Several adequate allotments can exist. Thus, theproblem consists in the selection of the unique adequateallotment from the set of adequate allotments,

which is the class of possible solutions ofthe concrete classification problem. The set of adequateallotments is depending on the number of fuzzy clustersin the sought allotment. So, is the set ofadequate allotments corresponding to the formulation ofa classification problem.

The selection of the unique adequate allotmentfrom the set of adequate allot-

ments must be made on the basis of evaluation of allot-ments. The criterion

(21)

where is the number of fuzzy clusters in the allotmentand is the number

�

�z

B

c

c

.........................................

..............................

..........

...........

tolerance is the basis for the concept of fuzzy -cluster.That is why definition of fuzzy tolerance must be consi-dered in the first place.

Let be the initial set of elements andsome binary fuzzy relation on with

being its membership func-tion. Fuzzy tolerance is the fuzzy binary intransitiverelation, which possesses the symmetricity property

(14)

and the reflexivity property

(15)

Let be the initial set of objects. Letbe a fuzzy tolerance on and be -level value of

. Columns or lines of the fuzzy tolerancematrix are fuzzy sets . Let be fuzzysets on , which are generated by a fuzzy tolerance .The -level fuzzy set

is fuzzy -cluster or, simply, fuzzy cluster. Soand is the mem-

bership degree of the element for some fuzzy clus-ter . Value of is the tolerancethreshold of fuzzy clusters elements.

The membership degree of the element for so-me fuzzy cluster can be definedas a

(16)

where an -level ofa fuzzy set is the support of the fuzzy cluster . So,condition is met for each fuzzy cluster

. Membership degree can be in-terpreted as a degree of typicality of an element to a fuz-zy cluster. The value of a membership function of eachelement of the fuzzy cluster in the sense of (16) is thedegree of similarity of the object to some typical objectof fuzzy cluster. Membership degree defines a possibilitydistribution function for some fuzzy cluster .The fact was demonstrated in [8] and the possibilitydistribution function is denoted by .

Let is a fuzzy tolerance on , where is the set ofelements, and is the family of fuzzy clustersfor some . The point , for which

(17)

is called a typical point of the fuzzy cluster. A fuzzy cluster can have several typical

points. That is why symbol is the index of the typicalpoint. A set of typical points of thefuzzy cluster is a kernel of the fuzzy cluster and

is a cardinality of the kernel. Obviously,if the fuzzy cluster have a unique typical point, then .

Let be a fa-mily of fuzzy clusters for some value of tolerance thres-hold which are generated by some fuzzy tole-rance on the initial set of elements .

�

� �

��

�

�

X

T

e

T

TX

X T

X X

..............................................

...................


Articles38

of elements in the support of the fuzzy cluster , can beused for evaluation of allotments.

Maximum of criterion (21) corresponds to the bestallotment of objects among fuzzy clusters. So, the clas-sification problem can be characterized formally as deter-mination of the solution satisfying

(22)

The adequate allotment is any allotmentamong fuzzy clusters in the case. Thus, the problem ofcluster analysis can be defined in general as the problemof discovering the unique allotment, resultingfrom the classification process.

Detection of fixed number of fuzzy clusters can beconsidered as the aim of classification. There is a seven-step procedure of classification:

1. Calculate -level values of the fuzzy tolerance andconstruct the sequenceof -levels;

2. Construct the initial allotmentfor every value from the sequence

3. Let ;4. Construct allotments

which satisfy conditions (19) and (20) forevery value from the sequence

5. Construct the class of possible solutions of the classi-fication problem forthe given number of fuzzy clusters and differentvalues of the tolerance threshold as follows:

for some allotment thecondition is met,

let and go to step 4;6. Calculate the value of the criterion (21) for every

allotment7. The result of classification is formed as follows:

for some unique allotment from the setthe condition (22) is met,

the allotment is the result of classification,the number of classes is suboptimal.

So, the allotment among thegiven number of fuzzy clusters and the correspondingvalue of tolerance threshold are the results of clas-sification.

Some modifications of the D-AFC(c)-algorithm areproposed in [6], [9] and [10].

Let us consider a method for the data preprocessing.The matrix of fuzzy toleranceis the matrix of initial data for the D-AFC(c)-algorithm ofpossibilistic clustering. However, the data can be presen-ted as a matrix of attributes

where the value is the value of the thattribute for th object. In the first place, the data can be

c

T

l

c

c

t

c

c

c

i

2.2. The D-AFC(c)-algorithm

if

thenelse

if

thenelse

2.3. Notes on the data preprocessing

...............................

...........................................

..........................

.......

......................................

.................................

normalized as follows:

(23)

In the second place, the data can be normalized usinga formula

(24)

So, each object can be considered as a fuzzy setand

are their membership functions.The matrix of coefficients of pair wise dissimilarity

between objects can be ob-tained after application of some distance to the matrix ofnormalized dataThe most widely used distances for fuzzy sets

in are [11]:The normalized Hamming distance:

(25)

The matrix of fuzzy tolerancecan be obtained after application of comple-

ment operation

to the matrix of fuzzy intoleranceobtained from previous operations.

�

The normalized Euclidean distance:

(26)

The squared normalized Euclidean distance:

(27)

(28)

A technique of fuzzy rules antecedents learning is pre-sented in the first subsection. A method of consequentslearning is given in the second subsection of the section.The third subsection includes a technique of fuzzy rulestuning.

In the following, we will consider that the fuzzyinference system is a multiple inputs, multiple outputssystem. The antecedent of a fuzzy rule in a fuzzy inferencesystem defines a decision region in the -dimensionalfeature space. Let us consider a fuzzy rule (1) where

is a fuzzy set associate with thefeature variable . Let be characterized by the trape-zoidal membership function , which is presentedin Figure 1.

So, the fuzzy set can be defined by four parame-ters . A triangular fuzzy set

can be considered as a particular caseof the trapezoidal fuzzy set where .

The idea of deriving fuzzy rules from fuzzy clusters isthe following [7]. We apply the D-AFC(c)-algorithm to the

�

�

3. Deriving fuzzy rules from fuzzy clusters

3.1. Antecedents learning

m

........................................................

..................................................................

........................

........................

....


Articles 39

given data and then obtain for each fuzzy clustera kernel and a support . The

value of tolerance threshold , which correspondsto the allotment , is the additionalresult of classification. We calculate the interval

of values of every attribute , forthe support . The value can be obtained asfollows

(29)

and the value can be calculated usinga formula

(30)

The parameter can be obtained as following

(31)

and the parameter can be obtained from the conditions

(32)

We calculate the value for all typical pointsof the fuzzy cluster , as

follows:

(33)

and the value can be obtained from the equation

(34)

Thus, the parameter can be calculated from theconditions

(35)

and the parameter can be obtained as following

(36)

The height of the fuzzy cluster

, must be taken into account becausethe fuzzy cluster can be a subnormal fuzzyset [12], [13].

So, condition and condition are

Fig. 1. A trapezoidal membership function for an antece-dent fuzzy set.

...........

. ...............

met for all input variables . Obviously, ifcondition is met for the fuzzy cluster andonly one typical point is presented in the fuzzy cluster,then the condition is met.

The variables are the consequents of fuz-zy rules (1), represented by the fuzzy setswith the membership functions . Fuzzy sets

can be defined on the interval of member-ships and these fuzzy sets can be presented as fol-lows: where is the tolerance thres-hold, and .

So, membership functions of fuzzy setswill be trapezoidal membership functions.

The situation is shown in Figure 2.

Fuzzy clusters can be subnormal fuzzy sets [12]. So,the case, which is presented in Figure 2, is the generalcase.

If the allotment among fully separate fuzzyclusters is obtained and all fuzzy clusters

are normal fuzzy sets thenand for each fuzzy cluster

. So, a trapezoidal membership functionof a fuzzy set in the case of normal

and fully separate fuzzy clusters is presented in Figure 3.

Thus, trapezoidal membership functions forthe fuzzy sets can be constructed ona basis of the clustering results. The empty set

can be correspond to some output variableSo, the empty fuzzy set will be cor-

respond to the output variable andis the membership function of the correspon-

ding fuzzy set .

3.2. Consequents learning

Fig. 2. A trapezoidal membership function for a consequentfuzzy set in a general case.

Fig. 3. A trapezoidal membership function for a consequentfuzzy set in a case of normal and fully separated fuzzyclusters.

.....................

......................................

.........

.............


Articles40


If the allotment among particularly separatefuzzy clusters is obtained then some non-empty fuzzy sets

will be correspond to some output variables.

The computational accuracy must be taken into ac-count in the data processing by the D-AFC(c)-algorithm.The computational accuracy can be determined by a valueof accuracy threshold . If we decrease the value, the computational accuracy increases. Membership

values the value of tolerancethreshold , and the number of typical points in eachfuzzy cluster are depending on the value ofaccuracy threshold . An allotment can be charac-terized by the value of tolerance threshold thatis increasing with decreasing accuracy, i.e., forwe have . From other hand, a fuzzy cluster

can be characterized by a kerneland the number of typical points of the fuzzy cluster isdecreasing with increasing accuracy, i.e., forwe have . So, the accuracy thresholdcan be used as a parameter for the D-AFC(c)-algorithm.The fact was demonstrated in [14]. Moreover, the accu-racy threshold can be considered as the analog of thefuzziness index in the formula (3).

Thus, the accuracy threshold can be useful for tuningof the rules. In particular, for we haveand a crisp interval is increasing. Otherwise, ifwe decrease the accuracy threshold, , the number oftypical points of the fuzzy cluster increa-ses, and a crisp interval increases.

Membership functions for consequents fuzzysets depend on the value of accuracy thres-hold . For example, if we increase the value of accuracythreshold , the crisp interval decreases. Moreover,for we have . That iswhy parameters and increases for all fuzzy sets ,i.e., for we have and .

The proposed technique for fuzzy rules tuning will beexplained by an illustrative example in the next section.

The first subsection of the section includes the resultsof the Anderson's Iris data clustering by the D-AFC(c)-algorithm. The designed fuzzy inference system is pre-sented in the second subsection and the results are com-pared with the results of other classifier systems.

The Anderson's Iris data set consists of the sepallength, sepal width, petal length, and petal width mea-sured for 150 irises [15]. The problem is to classify theplants into three subspecies on the basis of this infor-mation. The Anderson's Iris data forms the matrix of at-tributes where thesepal length is denoted - by , sepal width - by , petallength - by and petal width - by . The Iris database isthe most known database to be found in the patternrecognition literature. The method of the data prepro-cessing which was described in the third section can beused for constructing the matrix of fuzzy tolerance and

3.3. Fuzzy rules tuning

4.1. Results of the Anderson's Iris data clustering

4. An illustrative example

...........

.........

............

............

................

the matrix of fuzzy tolerance can be processed by the D-AFC(c)-algorithm. The formula (23) and the squared nor-malized Euclidean distance (27) were used for the datapreprocessing.

Four experiments were made for different values ofthe accuracy threshold . The allotment amongthree fully separated fuzzy clusters was obtained in eachexperi-ment. The results of the Anderson's Iris data setpro-cessing by the D-AFC(c)-algorithm for differentvalues of the accuracy threshold are presented in theTable 1.

By executing the D-AFC(c)-algorithm for three classes(1, 2, 3) in each experiment we obtain the following: thefirst class is formed by 50 elements all being Iris Setosa;the second class by 52 elements, 48 of them being IrisVersicolor and 4 Iris Virginica; the third class by 48 ele-ments, 46 of them being Iris Virginica and 2 Iris Versi-color. In other words, the first class corresponds to theSetosa subspecies, the second class corresponds to theVersicolor subspecies and the third class corresponds tothe Virginica subspecies. So, there are six mistakes ofclassification in each experiment.

Let us consider results of the experiment for the valueof accuracy threshold . The ninety-fifth objectis the typical point of the fuzzy cluster, which corres-ponds to the first class, the ninety-eighth object is thetypical point of the second fuzzy cluster, and theseventy-third object is the typical point of the thirdfuzzy cluster. The height of each fuzzy cluster

is equal one. So, membership functionsand for corresponding fuzzy sets and

can be constructed immedia-tely. The rule base induced by the D-AFC(c)-algorithmclustering result can be seen in Figure 4 where labels

and denote, respectively, se-pal length, sepal width, petal length, and petal width,and is the number of rule.

Note that only one typical point is presented in eachfuzzy cluster. That is why membership functions

are triangular membership func-tions. Obviously that a meaningful linguistic label can beassigned to each fuzzy set .

From other hand, linguistic labels Setosa, Versicolorand Virginica are associated with corresponding outputvariables . Note that fuzzy sets

and are empty fuzzy sets.

Table 1. Results of the Anderson's Iris data set classi-fication obtained from the D-AFC(c)-algorithm for differentvalues of the accuracy threshold.

4.2. A fuzzy inference system

............

.................

VOLUME 4, N° 3 2010

Articles 41


We show in Figure 5 a graph of the performance of thedesigned fuzzy inference system. The example ofclassification of the ninety-fifth object, which is thetypical point of the first fuzzy cluster, is presented inFigure 5.

Total area is zero in the defuzzification procedure foroutput variables Versicolor and Virginica. That is why anaverage of the range of output variables Versicolor andVirginica are used as output values and these values areequal 0.5. The values can be interpreted as uncertainmembership degrees.

The result, which is obtained from fuzzy inferencesystem, is easily interpreted. Thus, the obtained model issuitable for interpretation since the rules consequentsare the same or close to the actual class labels, such thateach rule can be taken to describe all classes.

The Anderson's Iris data were classified using theconstructed fuzzy inference system. The rules classifyfour objects incorrectly and two objects are rejected.Thus, the total number of misclassifications is 6. Eviden-tly, that the results are correlated with the results, ob-tained from the D-AFC(c)-algorithm. So, the fuzzy infe-rence system is accurate.

The application of the constructed fuzzy inference

system to the Anderson's Iris data was made in compa-rison with other approaches. Table 2 shows the results ofsome well-known classifier systems.

For example, Höppner, Klawonn, Kruse and Runkler[2] applied the simplified version of the GG-algorithm offuzzy clustering to learn a Mamdani-type fuzzy inferencesystem for classifying the Anderson's Iris data by trainingon all 150 objects. An eight-rule fuzzy system was obtai-ned. The rules classify 3 objects incorrectly and 3 morewere not classified at all. So, the total number of misclas-sifications is 6.

From other hand, the FCM-algorithm of fuzzy cluste-

Table 2. Comparison of results of different classifier sys-tems on the Anderson's Iris data set.

VOLUME 4, N° 3 2010

Fig. 5. The performance of the fuzzy inference system.

Fig. 4. The rule base induced by the clustering result.

Rule

1

2

3

SL2

SL3

SW2

SW3

PL2

PL3

PW2

PW3

SL1 SW1 PL1 PW1 SETOSA VERSICOLOR VIRGINICA

(4.274, 5, 5.83)

(4.878, 5.5, 7.056)

(5.523, 7.7, 7.907)

(2.259, 3.4, 4.437)

(1.985, 2.4, 3.437)

(2.481, 3, 3.83)

(0.9814, 1.5, 1.915)

(2.974, 3.7, 5.152)

(4.752, 6.1, 6.93)

(0.0962, 0.2, 0.6148)

(1, 1, 1.622)

(1.574, 2.3, 2.507)

(0.9642, 0.9828, 1, 1)

(0.9642, 0.9676, 1, 1)

(0.9642, 0.9673, 1, 1)

Articles42


ring was applied by Roubos and Setnes [16] to obtain aninitial Takagi-Sugeno model with singleton consequents.All 150 samples were used in the training process. So, theinitial model with three rules was constructed from clus-tering results where each rule described a class. The clas-sification accuracy of the initial model was rather dis-couraging, giving 33 misclassifications on the trai-ningdata. A multi-objective genetic algorithm-based optimi-zation approach was applied to the initial model. So, thenumber of misclassifications was reduced to 4 samples.

Ishibuchi, Nakashima and Murata [17] applied all 150samples in the training process, and derived a fuzzy clas-sifier with five rules. The resolution was 3 misclassifi-cations.

Abonyi, Roubos and Szeifert [18] proposed a data-driven method to design compact fuzzy classifierscombining a genetic algorithm, a decision-tree initializa-tion, and a similarity-driven rule reduction technique.The final fuzzy inference system had three fuzzy rules andthe number of misclassifications was 6.

A fuzzy classifier with ellipsoidal regions was propo-sed by Abe and Thawonmas [19]. They applied clusteringmethods to extract fuzzy classification rules, with onerule around cluster center, and then they tuned the slopsof their membership functions to obtain a high recogni-tion rate. Finally, they obtained a three-rule fuzzy systemwith 2 misclassifications.

The results obtained from the constructed fuzzy infe-rence system seem appropriate in comparison with thesome well-known fuzzy systems. So, the proposed me-thod of derivation of fuzzy classification rules from datacan be considered as an effective technique of the rapidprototyping fuzzy inference systems.

Some conclusions are formulated in the first subsec-tion. The second subsection deals with the perspectiveson future investigations.

Many techniques to design fuzzy inference systemsfrom data are available; they all take advantage of theproperty of fuzzy inference systems to be universal ap-proximators. This paper presents an automatic method todesign fuzzy inference system for classification heu-ristic possibilistic clustering and the method can be con-sidered as an approach to rapid prototyping of fuzzy infe-rence systems. The proposed method is simple in compa-rison with other well-known approaches. The results ob-tained with the proposed modeling approach for theAnderson's Iris data set case illustrate the effectivenessof the proposed method of designing fuzzy inferencesystems.

Notable that the fuzzy rules obtained using the D-AFC(c)-algorithm can be interpreted very simply, becausemembership functions of fuzzy sets which correspond toinput variables of fuzzy rules has natural interpretations.

Constructing a rule base from fuzzy clusters givesa first approximation for the data, which can be used asa basis for further improvements. A technique of fuzzy

via

via

5. Concluding remarks

5.1. Discussion

5.2. Perspectives

rules tuning based on varying of the accuracy threshold isproposed in the paper. However, some other approaches,such as the genetic algorithm-based approach or neuro-fuzzy techniques can be used for fuzzy rules tuning.

Note that the computational complexity of theD-AFC(c)-algorithm is higher in comparison with objecti-ve function-based fuzzy clustering algorithms. For exam-ple, approximately 700 observations is the large data setfor the D-AFC(c)-algorithm. Of course, the computationalcomplexity of the D-AFC(c)-algorithm is the subject ofspecial considerations. However, the clustering problemin cases of large data sets can be solved in the prelimi-nary way as follows: the initial data setcan be represented as a set where

and each element is the sub-set of the data set . So, the matrix of the reduced dataset can be presented as the matrix of theinterval-valued data, and the data can be processed bythe D-AFC(c)-algorithm [20]. Fuzzy rules can be extractedfrom the interval-valued data clustering results imme-diately.

From other hand, the D-AFC(c)-algorithm can be ap-plied for classification the three-way data [13] and thefuzzy data [21]. So, the proposed method of designingfuzzy inference systems can be generalized for correspon-ding cases of the training data set.

These perspectives for investigations are of great in-terest both from the theoretical point of view and fromthe practical one as well.

- Laboratory of Images Recog-nition and Processing, United Institute of InformaticsProblems of the National Academy of Sciences of Belarus,Surganov St. 6, 220012 Minsk, Belarus, e-mail:[email protected].

ACKNOWLEDGMENTS

AUTHORDmitri A. Viattchenin

References

I am grateful to Prof. Janusz Kacprzyk, Prof. Jan W. Owsinski,Prof. Frank Klawonn and Prof. Valery Starovoitov for their inter-est in the investigations and support. I would like to thank Mr.Aliaksandr Damaratski for elaborating experimental software.I also thank the anonymous referees for their valuablecomments.

[1] Mamdani E.H., Assilian S., “An experiment in linguisticsynthesis with a fuzzy logic controller”,

, vol. 7, 1975, pp. 1-13.[2] Höppner F., Klawonn F., Kruse R., Runkler T.,

, Chichester: Wiley Intersciences,1999.

[3] Krishnapuram R., Keller J.M., “A possibilistic approachto clustering”, ,vol. 1, 1993, pp. 98-110.

[4] Sugeno M., Yasukawa T., “A fuzzy-logic-based approachto qualitative modeling“,

, vol. 1, 1993, pp. 7-31.[5] Viattchenin D.A., “A new heuristic algorithm of fuzzy

InternationalJournal of Man-Machine Studies

FuzzyCluster Analysis: Methods for Classification, Data Analysisand Image Recognition

IEEE Transactions on Fuzzy Systems

IEEE Transactions on FuzzySystems

....................

VOLUME 4, N° 3 2010

Articles 43


clustering”, , vol. 33, 2004,pp. 323-340.

[6] Viattchenin D.A., “A direct algorithm of possibilisticclustering with partial supervision”,

, vol. 1,no.3, 2007, pp. 29-38.

[7] Viattchenin D.A., “Outlines for a new approach to gene-rating fuzzy classification rules through clusteringtechniques”,

, Minsk, Belarus, 2009, pp. 82-87.[8] Viattchenin D.A., “On possibilistic interpretation of

membership values in fuzzy clustering method based onthe allotment concept”,

, no. 3, 2008, pp. 85-90. (in Russian)[9] Viattchenin D.A., “Direct algorithms of fuzzy clustering

based on the transitive closure operation and their ap-plication to outliers detection”, ,no. 3, 2007, pp. 205-216. (in Russian)

[10] Viattchenin D.A., “An algorithm for detecting the prin-cipal allotment among fuzzy clusters and its applicationas a technique of reduction of analyzed features spacedimensionality”,

, vol. 33, 2009, pp. 205-217.[11] Kaufmann A.,

, New York: Academic Press, 1975.[12] Viattchenin D.A., “Kinds of fuzzy -clusters”,

, no. 4, 2008,pp. 95-101. (in Russian)

[13] Viattchenin D.A., “An outline for a heuristic approachto possibilistic clustering of the three-way data”,

, vol. 3, 2009, pp. 64-80.[14] Damaratski A., Novikau D., “On the computational accu-

racy of the heuristic method of possibilistic clustering”,

, Minsk,Belarus, 2009, pp. 78-81.

[15] Anderson E., “The irises of the Gaspe Peninsula”,, vol. 59, 1935, pp. 2-5.

[16] Roubos H., Setnes M., “Compact and transparent fuzzymodels and classifiers through iterative complexityreduction”, , vol. 9,2001, pp. 516-524.

[17] Ishibuchi H., Nakashima T., Murata T., “Three-objectivegenetic-based machine learning for linguistic rule ex-traction”, , vol. 136, 2001, pp.109-133.

[18] Abonyi J., Roubos J.A., Szeifert F., “Data-driven gene-ration of compact, accurate and linguistically soundfuzzy classifiers based on a decision-tree initializa-tion”, ,vol. 32, 2003, pp. 1-21.

[19] Abe S., Thawonmas R., “A fuzzy classifier with ellipsoi-dal regions”, , vol. 5,1997, pp. 516-524.

[20] Viattchenin D.A., Damaratski A. “Constructing ofallotment among fuzzy clusters in case of quasi-robustcluster structure of set of objects”, , no.1, 2010, pp. 46-52. (in Russian)

[21] Viattchenin D.A., “A heuristic approach to possibilisticclustering for fuzzy data”,

, vol. 32, 2008, pp. 149-163.

Control & Cybernetics

Journal of Automa-tion, Mobile Robotics and Intelligent Systems

Proc. of the 10 International Conferenceon Pattern Recognition and Information ProcessingPRIP'2009

Proceedings of the Institute ofModern Knowledge

Artificial Intelligence

Journal of Information and Organiza-tional Sciences

Introduction to the Theory of Fuzzy Sub-sets

Proceed-ings of the Institute of Modern Knowledge

Jour-nal of Uncertain Systems

Proc. of the 10 International Conference on Pattern Re-cognition and Information Processing PRIP'2009

Bulle-tin of the American Iris Society


Information Sciences

International Journal of Approximate Reasoning


Doklady BGUIR

Journal of Information andOrganizational Sciences

th

th

�

VOLUME 4, N° 3 2010

Articles44

Abstract:

1. IntroductionDaily capacity of Warsaw Underground, according to

its technical data, is up to 500,000 passengers [1]. Thatnumber of passengers does not provoke risk only undercondition, that rush will be similar during whole day.In fact, people mostly use underground during rushhours. The main hazard for passengers in underground isto be run over by train. Except train collision and derail-ment run over is the main cause of death and injuries ofpassengers. Majority of accidents can be prevented byAutomatic Train Speed Control System (SOP-2) [2]. Thatsystem has been already applied in most undergroundsystems, also in Warsaw. That solution helped preventmain cause of most death accidents. Statistical dataabout accidents in London Underground, presented inTable 1 [3], shows that only five accidents causing pas-senger deaths have occurred due to train operation innearly 150 years. In Warsaw Underground being run overby train was the only cause of death since it was openedin 1995. Risk of fall of from the platform on track can bedecrease by automatic barriers between train and plat-form. Synchronization with Automatic Train Speed Con-

Paper presents a novel idea of capacitive sensor forhuman presence detection. Due to the use of resonantcircuit significant increase of sensor's sensitivity wasachieved. In the case of human hand, value of outputsignal changes up to 421%. Moreover, sensor createspossibility of determination of the material of the detectedobject. As a result developed sensor can be applied forhuman presence detection sensors utilized in safetysystems, which are especially suitable for undergroundtransportation network.

Keywords: capacitive sensors, people safety, undergroundtransportation network.

trol System (SOP-2) is critical for such solution becausedoors on platform should be in front of train's doors, tolet passenger go out. That brings very serious danger incase of fire and was the main cause of death of 200 pas-sengers in underground in South Korea [4]. This is thereason that on second underground line barriers are notplanned.

In conclusion there is no sufficient method to preventun down by train accident. For this

paper show an example a human detection system that incooperation with Automatic Train Speed Control Systemcan increase passenger's safety in underground.

Passenger's safety can be increase by detecting hu-man presence on track and preventing to from enteringtrain to station if such incident occurred. System SOP-2can stop the train, when track sensors indicate signalthat track is occupied by other train. The idea is to com-municate with that system to stop the train in the shor-test time before train enters the station, when dangersituation has been noticed. Three types of sensors areused to detect human presence on track. Motion detec-tors based on passive infrared detector and image proces-sing method work as human detectors. Single beam opti-cal sensors are used to detect train. Third types of detec-tor are proximity sensors, which are installed on plat-forms edge to detect human's presence between train andplatform.

Proximity sensor is based on permittivity phenomena.Materials, which are found in typical environment, havedifferent permittivity level, as shown in Table 2. There issignificant difference between water and other materialpermittivity. It is more than ten times lower then inmetals and higher then in insulators. There is no othermaterial that has the same permittivity level. Sensor usesthat difference to recognize objects, which can be foundin operation area.

2. Principles of operation

passenger from being r

CAPACITIVE HUMAN PRESENCE SENSOR FOR SAFETY APPLICATIONS

Piotr Frydrych, Roman Szewczyk

Received 29 ; accepted 10 .th April 2010 May 2010th


Articles 45

Table 1. Statistic of Accidents in London Underground.

Place

NorthwoodEdgwareStratfordMoorgateHolbornCamden TownWhite City

Year

1945194619531975198020032004

Fatality

3 killedMinor injures12 killed43 killedMinor injures6 injuredNo injured and killed

Type

CollisionHit in the and of tunnelCollisionHit in the and of tunnelCollisionDerailmentDerailment

Cause

Driver failureDeath of the driverSignalization failureUnknownOperator's failureWrong steering system configurationWrong steering system configuration

Table 2. Permittivity for different materials.

Capacitance is directly proportional to permittivity,thus it is possible to detect an object by measuring chan-ges of it. Therefore sensor should be a kind of capacitor.Capacitive sensors are very common in industry, and ot-hers applications. The novel is the form of capacitor,which is described on figure1 has the ability to recogni-ze material of which the detected object is made.

Mostly capacitive sensors have small detecting area.In this application it would be necessary to use hundredsof those devices in entire hazard area. Most of knownsensors are cylindrical, which brings problems withmounting it in the floor, or in the edge of the platform.Only capacitive sensors, developed in GM factory toprotect workers from robots, were in the form of wire [5].Ability to recognize metal from insulators is achieved bysome of sensors [6]. There are also systems, which canrecognize man from insulators, but only in case, that heis grounded and there are no other grounded objects inoperation area [7]. Those conditions cannot be achievedin under-ground. Developed sensor is flat, long andadjustable to curve of the ramp. It minimizes the cost andoptimises detecting area. Flat device is easy to install anddo not disturb architecture of the station.

The purpose was to maximize the sensitivity to chan-ges of permittivity and minimize thickness. In classicalflat capacitor two sheets are in front of each other. Thatconfiguration would be not advantageous in this case,because it indicates more than three plies of board, ortape to construct sensor and it reduced electric fieldoutside the capacitor. Changes in capacitance of capa-citor are observed only under condition that object willbe in electric field produced by capacitor what result offollowing equation is:

where is given as

(2)

In equation (2), is electric potential and isdifferential track.

When outside of the capacitor, integral givenby equation (2) is equal to zero. Therefore electric fieldshould be put outside the capacitor, because object can-not be between capacitors sheets. Technical solution ofsuch specialized capacitor is shown in Figure 1.

(1)

U

E dl

E = 0

In classical flat capacitor major part of electric streamis between sheets. When sheet's area is equal to zero andwidth is big, most of electric field lines come outside thecapacitor. There can be find-detected object (dottedarea). To prove that, fundamental Maxwell's equation canbe applied:

On closed track integral of electric field is equal tozero (doted line), thus there have to be electric field notonly between sheets, but also outside. In other caseintegral would not be equal to zero.

Capacitance is proportional to permittivity, but alsoother factors have influence on it. In detecting areainsulated metal can be found, and grounded objects, liketrain body. Sensor has to be resistant to that disturbance.Case with insulated metal can be transformed as it isshown in Figure 2. In that case two series connectedcapacitors are created. Resultant capacitance is:

Equation (4) shows that resultant capacitance issmaller then capacitance of each capacitor. In fact thosetwo capacitors can have bigger capacitance than emptysensing capacitor, because of big sheets area. In conclu-sion small increase of capacitance can be observed.

(3)

(4)

Fig. 1. Principles of operation of sensing capacitor.

Fig. 2. Analyse of the case, where insulated metal is pre-sent nearly the sensor.


Articles46

VOLUME 4, N° 3 2010

Material

AirPolyethyleneSiliconRubberPaperWaterMetals

Permittivity

1,00052,2511,6873,580>1000

�r

U

QC �

� ��d

dlEU0

� ��L

dlE 0

21

21

CC

CCC

��

�

C1

C1

C1

Insulated metal

C2

C2

C2

changes of capacitance. Too high quality factor can incre-ase noise level, and cause poor resistance to disturbances.

increases when capacitance differs to capacitan-ce of empty capacitor. Higher resistance decreases qualitylevel according to equation:

(5)

where is supply voltage, - inductance, - resistanceof circuit and - current pulsation.

Dependences of calculated for different resis-tances are presented in Figure 5. Due to use of the reso-nance circuit, significant changes of the output signalfrom the sensor are achieved.

Main part of sensor is consisted of two metal sheets,which are on the same plane. Areas of the sheets and

U

U L RT

U

U

UR

LC

LC

LC

LC

Fig. 4. Electrical connections of the sensor ( - outputsignal).

Fig. 5. Dependence of the calculated sensor output signalas a function of changes of its capacitance, calculated

for different values of resistance .

4. Experimental set-up and results

Configuration of two capacitors may be considered,when grounded metal is present nearly to capacitivesensor. That case is presented in Figure 3. Most of streamfrom sensing capacitor sheets will be absorbed by groun-ded metal. Only minority of stream will be transferredbetween sensing capacitor sheets, therefore decrease ofcapacitance of circuit will be observed. Insulator inneighbourhood of capacitor increases capacitance pro-portionally to the material permittivity, which is muchhigher for water, then other materials.

Tests with automatic capacitive bridge have proventheoretical conclusions in every case of object. Forinsulated metal small change of capacitance was obser-ved. In case of grounded metal output signal was lowerthan for empty capacitor. Insulators were increasing ca-pacitance. Results of these tests are presented in Table 3.Capacitance was measured for 1 kHz and 10 kHz bridgesupply frequency.

Fig. 3. Analyse of the case where grounded metal is presen-ted nearly the sensor.

Table 3. Capacitance of sensing capacitor for differentmaterials.

3. Developed capacitive sensor for safetyapplicationsChange of capacitance is not a typical electric signal,

thus it has to be transform to current or electric voltage.To increase signal, electrical resonance is used. Reso-nance can be achieved when capacitor is empty; it meansthat there is no object in detection area. When resonanceoccurres voltage drop on elements, according toequation (5), is the lowest ( ). Even small change ofcapacitance cause significant decrease of voltage drop,which can be observed in Figure 5. Sensors circuit isshown in Figure 4.

Frequency of supply current, resistance and induc-tance has to be matched to achieved resonance of capa-citance of empty capacitor. Significant is also quality fac-tor of circuit. It can increases sensibility of circuit to

LCULC


Articles 47

VOLUME 4, N° 3 2010

Object

Empty capacitorHandInsulated metalGrounded metalWood

C (pF) 1kHz

15,7-16,2292,0 - 390,033,7 - 34,115,9 - 16,026,7 - 26,8

C (pF) 10kHz

16,3229 - 237

33,915,322,1

2

2

2

1

1

��

��

��

��

�

CLR

CLU

U LC

��

��

ravine between them have to be optimised. Sensor wasmade on printed board. As a result an influence of dif-ferences of shape and ravine during mounting and mea-suring was decreased. Such differences may occur in thecase of metal leaf. Moreover, there was insulation oncopper ply to prevent break down.

Model of the sensor shown in Figure 6 has been testedin laboratory, for different types of materials, and confi-gurations. Testing circuit was shown in Figure 7. Accu-racy of resistance, inductance and capacitance has notinfluence on detecting ability, thus current frequency isadjustable and it is only needed to adjust resonance forempty capacitor.

Fig. 6. Model of the sensing capacitor.

Fig. 7. Schematic diagram of the experimental set-up.

For configuration shown in Figure 7 resonance fre-quency was 928 Hz for 32 V supply and 860 Hz for 100 V.During experiment the reaction of four different types ofobject was tested: insulated and grounded metal, forwood and human hand. In case of human hand signals fordifferent distance were measured.

Driving properties and results of the tests were shownin Table 4. Resonance frequency changes due to supplyvoltage. This is because of internal capacitance and resis-tance of generator.

Changes of output due to object material are very sig-nificant. They are 17% higher from wood for hand in dis-tance up to 10 mm and 421% higher for distance 0 mm incase of 32 V supply. Results confirm that sensor is able todetect living tissue. Relation of signal to distance forhuman hand is shown in Figure 8. Shape of function doesnot change with supply value, thus smaller supply vol-tage can be used.

Proximity sensor described in this paper is able toachieve the requirements. On the other hand, ability todetect in distance is not sufficient. Probably it can beimproved by increasing frequency, which can straightenelectric field propagation in detection area. Good featu-re, which was observed by experiment, is fact, that detec-ting range is not related to supply voltage. It can decreaseprobability of break down, and electric shock risk.

Fig. 8. Output signal as a function of distance of humanhand, measured for different power supply voltage.

4. Conclusion


Articles48

VOLUME 4, N° 3 2010

Table 4. Experimental results.

Object

EmptyHandHand distanceInsulated metalGrounded metalWood

32 V / 928 Hz

U [V] range 1 V

0,3400,573

1-5 mm0,3010,2640,345

LC

100 V / 860 Hz

U [V] range 10 V

0,951,20

1-5 mm1,040,941,05

LC

0,4045-10 mm

1,115-10 mm

1,8000 mm

2,140 mm

Experimental results indicate, that it is possible todevelop a simple sensor, which not only detect objects,but also recognizes material. This kind of sensor is a goodalternative for sensing mat, which uses stress to detectan object. As compared with safety mats [8], used inindustry to detect human presence in danger area, deve-loped sensor is thinner, more adjustable and does not de-pend on stress, which can provoke uncertainty in opera-tion. To detect presence of man's feet on the floor highdetection range is not needed. Therefore described sen-sor can be used in many cases to detect human presencein hazardous area.

- Institute of Metrology and BiomedicalEngineering, Warsaw University of Technology, św.A. Boboli 8, 02-525 Warszawa, Poland, tel. 0-22-234-8471, e-mail: [email protected].

- Institute of Metrology and BiomedicalEngineering, Warsaw University of Technology, św.A. Boboli 8, 02-525 Warszawa, Poland. IndustrialResearch Institute for Automation and Measurements,Al. Jerozolimskie 202, 02-486 Warszawa, Poland, tel.0-22-8740171, e-mail: [email protected].* Corresponding author

AUTHORSPiotr Frydrych*

Roman Szewczyk

References[1] Internet site www.metro.waw.pl.[2] Instruction SOP-2 DTR-94/SOP-

2/1, Katowice 2002. (in Polish)[3] Internet site en.wikipedia.org/wiki/List_of_London_

Underground_accidents.[4]

, 19 Feb, 2003. www.independent.co.uk.[5] „Protecting Workers from robots”, ,

March 1984, pp. 85-86.[6] “A Combined Inductive-Capacitive Proximity Sensor

and Its Application to Seat Occupancy Sensing”. In:

, 5 -7 May 2009.[7] George B., Zangl H., Bretterklieber T., “A Warning

System for Chainsaw Personal Safety based on Capa-citive Sensing”. In: , 26 -29October 2008, pp. 419-422.

[8] www.safetymat.net.

Urządzenia nadawcze

120 killed on South Korea underground as 'arsonist'attacks train

American Machinist

International Instruments and Measurement TechnologyConference Singapore

IEEE Sensors Conference

th

th th

th th


Articles 49

VOLUME 4, N° 3 2010

Abstract:

1. IntroductionPaper industries require several drives for paper pro-

cessing. In the past, a single mechanical line shaft wasused for all drives. Nowadays independent drives areused. Electronic synchronization has to be ensured forquality of produced paper rolls. Several control strategieshave been suggested based on robust control or elec-tronic emulation of mechanical line shaft [1]. Since thedecentralized PI control method can be applied easilyand is widely known, it has an important place in controlapplications, where many industrial web transport sys-tems have used this type of controllers [2]. But this me-thod is insensitive to parameter changes. A nonlinear de-coupled control is designed for multi-motors multi mo-tors system. At the first, an ideal feedback linearizationcontrol system is adopted in order to decouple the ten-sions and velocity of the web winding system is presen-ted in [3]centralized and decentralized fixed order Hcontroller results with model based feed-forward for mul-ti motors systems which provide improved the tensionand velocity regulation is presented in [4]. In this workthe design of sliding-mode (SMC) to control a multi mo-tors system are proposed in order to improve the perfor-mances of the control system, which are coupled mecha-nically, and Synthesis of the robust control and their ap-plication to synchronize the five sequences and to main-tain a constant mechanical tension between the rollers ofthe system [6]. The advantage of an SMC is its robustnessand ability to handle the non-linear behaviour of the

Continuous processes in the plastics, textile paper andother industries, require several drives working in synchro-nism. The aim of this paper is to control speed of the multimotors system, and to maintain a constant mechanicaltension between the rollers of the system. Several control-lers are considerer, including Proportional-Integral (PI)and sliding-mode control (SMC). Since the PI control me-thod can be applied easily and is widely known, it has animportant place in control applications. But this method isinsensitive to parameter changes. The advantage of an SMCis its robustness and ability to handle the non-linear beha-viour of the system, and is indicated in comparison withtraditional proportional-integral (P1) control scheme. The-oretical analysis and simulation results are provided toevaluate the consistency and performances of this controltechnique (SMC).

Keywords: Multi-Motors systems, sliding mode control,proportional-integral (P1) control.

�

system.

In the mechanical part, the motor M1 carries out un-reeling, M3 drives the fabric by friction and M5 is used tocarry out winding, each one of the motors M2 and M4 dri-ves two rollers gears “to grip” the band (Fig.1). Eachone of M2 and M4 could be replaced by two motors, whicheach one would drive a roller of the stages of pinchingoff. The elements of control of pressure between the rol-lers are not represented and not even considered in thestudy. The stage of pinching off can make it possible toisolate two zones and to create a buffer zone. [6,7].

(1)

Where is the coefficient of dispersion and is givenby:

The model of the multi-motors system and in parti-cular the model of the mechanical coupling are developedand presented in Section II. Section III shows the deve-lopment of sliding mode controller’s design for windingsystem. The proposed structure of the studied propulsionsystem is given in the section IV. Simulation results usingMATLAB SIMULINK of different studied cases. Finally, theconclusions are drawn in Section V.

(2)

2. System Model's

via

�

SLIDING MODE SPEED CONTROL FOR MULTI-MOTORS SYSTEM

Bousmaha Bouchiba, Abdeldjebar Hazzab, Hachemi Glaoui, Fellah Med-Karim, Ismaïl Khalil Bousserhane

Received 16 ; accepted 1 .th March 2010 June 2010st


� �

� �

� �

��

�

��

�

�

��

��

�

��

�

�

��

��

�

��

�

�

�

��

�

� �

��

��

�

��

�

�

��

�

�

�

��

�

� �

��

��

lrc

qrdsdrqs

r

mr

qr

r

rdrreqs

r

rmqr

drredr

r

rds

r

rmdr

qsqr

r

rmrdr

r

m

qsr

r

msdses

s

qs

dsrqr

r

mdr

r

rm

qsesdsr

r

ms

s

ds

TJ

P

J

fii

JL

LP

dt

d

L

Ri

L

RL

dt

d

L

Ri

L

RL

dt

d

VL

RL

L

L

iRL

LRiL

Ldt

di

VL

L

L

RL

iLiRL

LR

Ldt

di

......

.

1

1

2

2

2

2

2

��

��

��

��

��

��

��

�

�

rs

m

LL

L2

1��

Articles50

Fig.1. Five motors web transport system.

Fig. 2. Web tension on the roll.

The tension model in web transport systems is basedon Hooke’s law, Coulomb’s law, [8], [9] mass conserva-tion law and the laws of motion for each rotating roll.

The tension of an elastic web is function of the webstrain

(3)

Where is the Young modulus, is the web section, isthe web length under stress and is the nominal weblength (when no stress is applied).

The study of a web tension on a roll can be consideredas a problem of friction between solids, see [8] and [9].On The roll, the web tension is constant on a stickingzone of arc length and varies on a sliding zone of arclength (cf. Fig. 2, where is the linear velocity ofthe roll ). The web tension between the first contactpoint of a roll and the first contact point of the followingroll is given by:

if (1.51)if (1.52)if (1.53)

Where is the friction coefficient, and .The tension change occurs on the sliding zone. The webvelocity is equal to the roll velocity on the stickingzone. (1.54)

A. Hooke’s law

B. Coulomb’s law

T

E S LL

ag V tk

L = a + g + L

�

0

k

t

( )

μ

C. Mass conservation law

D. Tension model between two consecutive rolls.

E. Roll velocity calculation

F. State space representation

Consider an element of web of lengthWith a weight density , under an unidirectional stress.The cross section is supposed to be constant. Accordingto the mass conservation law, the mass of the web re-mains constant between the state without stress and thestate with stress

(4)

The equation of continuity, cf. [8], applied to the webgives:

(5)

By integrating on the variable from to (cf. Fig.2), taking into account (4), and using the fact that

, we obtain

Therefore:

(6)

This equation can be simplified by using the approxi-mation

(7)

And using Hook’s law, we get:

(8)

.

where is the web length between roll and roll ,is the tension on the web between roll and roll ,is the linear velocity of the web on roll , is the

rotational speed of roll , is the radius of roll , isthe Young modulus and is the web section.

The law of motion can be obtained with a torque ba-lance:

(9)

Where , is the rotational speed of rollis the motor torque (if the roll is driven) and is thefriction torque.

The nonlinear state-space model is composed of (10)for the different web spans and of (11) for the differentrolls. Under the assumption that isvarying only slowly, which is the case for thin webs,can be chosen as a state variable in (11), leading to the

L = L

L

L kT k kV k k

k R k

V R k CemC

V

0

-1

(1+ )

-1

� .�

�

�

xa + g

<< L

k

k

ES

J R k

0

= 2, 3, 4, 5

-1

= /

( = 2, 3, 4, 5)

t

k

k

k

k

k k k k

f

k k

k


0

0

L

LLESEST

��

)(),( 1 ttx �� )(

1 )(),( axettx �� )(),( 2 ttx ��

tLxga ��

axgaxa ��

00 SLSL ��

�

�1

1

0

( )0

V

t x

� ��

� �

.111 2

2

1

1

2 ��

�

�

� �

�

VVL

dt

d

.)1(1

)1(22

1

2

21

2 ��

��

�� VVdt

dL

11 21 �� and

)21)(1(1

)1(21

1

2

2 ��

��

.)2(

)(

1

1111

kkk

kkkkk

k

VVT

VTVVESdt

dTL

��

��

�

��

fkkkkkk CCemTTR

dt

Jd��

� )(

)(1

Articles 51

Fig. 3. Block diagram for each motor with SMC control.

For the indirect field-oriented control (IFOC) tuningparameters we need two surfaces and the first for the

regulator and the second for regulator respectivelywhere: [10,11]

(20)

(21)

The derivate of can be given as:

From equation (1) and (27 ) we can obtain:

The virtual voltage controller is given by:

(22)

The voltage discontinuous control is defined as:

(23)

According to Lyapunov stability criteria [10], our spe-ed loop system's stable if: by means that ispositive constant.

The equivalent control is given as:

(24)

The derivate of can be given as:

From equation (1) and (34) we can obtain:

S Si i

S

V

V

S S K

V

1 2

1

1 1 1

ds qs

ds

qsn

dseq

<0

following linear model:

(10)

where

, (11)

(12)

A Sliding Mode Controller (SMC) is a Variable StructureController (VSC). Basically, a VSC includes several diffe-rent continuous functions that can map plant state toa control surface, whereas switching among differentfunctions is determined by plant state represented by aswitching function [8], [9].

To control the speed of the induction machine, thesliding surface is defined as follows:

(13)

The derivative of the sliding surface can be given as:

(14)

Taking into account the mechanical equation of theinduction motor defined in the system of equations (1),the derivative of sliding surface becomes

(15)

The current control is given by:

(16)

To avoid the chattering phenomenon produced by thefunction we use the Saturation function in the

discontinuous control defined as follow:

(17)

Where is the boundary layer thickness.The discontinuous control action can be given as:

(18)

: Positive constant.

(19)

The Fig. 3 shows the SMC control strategy scheme foreach induction motor

3. Design of sliding mode speedand current controllers

Sign Sat


XtCY )(�

BUXtAXEm )( ��

X = T T T T J T t J J J J T tT [ ( ) ( ) ]2 3 4 5 1 1 2 2 3 3 4 4 5 5� � � � �

Y = T T T T VT [ ]2 3 4 5 1

U = u u u u u[ ]1 2 3 4 5

�

�

�

�

m

m

m

m

m

m

w

w

w

w

w

w

S

S

�

�

�

�

*

*� �

� ok?

� �mw �

� �

�� rm

cqs

r

drmm C

J

Pw

J

fi

JL

LPwS .

² ** �

��

qsqseqqs iii �*

��

�

��

�

�

��

� �

�

�

��

� �

�

1;

1;

��

��

� Sif

SSign

Sif

S

Ssat

kiqs

� �� ssatki iqs

n

qs ��

�

� �

� � rmc

m

drm

rqseq C

J

Pw

J

fw

LP

JLi *

*2�

�

dsds iiS �� *

1

qsqs iiS �� *2

dsds iiS �� *

1

� � �

qsqs iiS �� *

2

� � �

).

1.

.

.

..

1(

*

2

2*

1

ds

s

r

rs

rm

qssds

r

mrs

s

ds

VLLL

RL

iwiL

LRR

LiS

��

�

�

�

� �

��

� �

��

.*

dsndseqds VVV �

� �111 �ssatkVdsn ��

).

..

...

1(.

*

2

2*.

r

rs

rmqss

ds

r

mrS

s

dssdseq

LL

RLiw

iL

LRR

LiLV

��

��

��

�

� �

��

� �

��

Articles52

The voltage controller is given by:

(25)

The equivalent control actions defined as:

(26)

The voltage discontinuous control is defined as:

(27)

For the same reason condition of , are positivesconstant.

The winding system we modelled is simulated usingMATLAB/SIMULINK software and the simulation is carriedout on 10s.

To evaluate system performance we carried outnumerical simulations under the following conditions:Start with the linear velocity of the web of 5m / s.

The motor M1 has the role of Unwinder a roll radius R1(R1 = 2.25 m).

The motors M2, M3, M4 are the role is to pinch the tape.The motor M5 has the role of winding a roll of radius R5.From the Fig (4-6), we can say that: the effect of the

disturbance is neglected in the case of the SMC control-ler. It appears clearly that the classical control with PIcontroller is easy to apply. However the control with sli-ding mode controllers offers better performances in bothof the overshoot control and the tracking error.

As shown in Figs (4-6). An improvement of the linearspeed is registered, and has follows the reference speedfor both PI controller and SMC control, but in case of PIcontroller, the overshoot in linear speed of Unwinder is25%. Figs (4-6) show that with the SMC controller the

V

V

V

K K

qs

qseq

dsn

1 2

4. Simulation results

Fig. 4. The linear speed of unwinder M1.

system follows the reference speed after 1 sec, in allmotors, however, in the PI controller the system followsafter 2 sec. Fig (9) shows the phase plot of the slidingsurface(s) of SMC control.

As shown in Fig (4-6). An improvement of the linearspeed is registered, and has follows the reference speedfor both PI controller and SMC control, but in case of PIcontroller, the overshoot in linear speed of Unwinder is25%. Figs (4-6) show that with the SMC controller thesystem follows the reference speed after 1 sec, in allmotors, however, in the PI controller the system followsafter 2 sec. Fig (9) shows the phase plot of the slidingsurface(s) of SMC control.

Fig. 5. The linear speed of motors M2, M3 and M4.

Fig. 6. The linear speed of winder M5.

Fig. 7. The phase plot of the sliding surface(s).


)1

....

1()(

*2

*

2

qs

s

mr

rs

mqs

r

mrs

s

dssqsqs

VL

wLL

Li

L

LRR

LiwiiS

��

�

�

��

� �

��

� �

�

��

qsnqseqqs VVV �*

)..

1

..(.

*2

lim

*

mr

rs

mqs

r

mrs

s

dssqssqseq

wLL

Li

L

LRR

L

iwiLV

��

�

�

� �

��

� �

�

� �

� �222 �ssatkVqsn ��

Articles 53

5. ConclusionThe objective of this paper consists in developing

a model of a winding system constituted of five motorsthat is coupled mechanically by a strap whose tension isadjustable and to develop the methods of analysis andsynthesis of the commands robust and their applicationto synchronize the five sequences and to maintain a con-stant mechanical tension between the rollers of thesystem.

The simulations results show the efficiency of the SMCcontroller technique, however the strategy of SMC Con-troller brings good performances, and she is more effi-cient than the classical PI controller.

- An assistant teacher at Univer-sity of Bechar, Bechar, Algeria From 2007 right now he'spreparing his PhD degree in multi machine system con-trol. E-mail: bouchiba_bousmaha@yahoo .fr.

- Professor of electrical engineeringat University of Bechar, Bechar, Algeria.

- M.Sc. From 2009 right now he'spreparing his PhD thesis in multi machine system control.

- Professor of electrical engi-neering at University of Bechar, Bechar, Algeria.

- Preceived the Master degree in industrialelectronics from the University of Quebec in Trois-Rivie-res, Trois-Rivieres, Canada, in 1990, and a Ph.D. degree inelectrical engineering from Rensselaer Polytechnic Insti-tute, Troy NY, USA in 1993. He is professor in electricaland computer engineering at University of Quebec inTrois-Rivieres where he is director of the Research groupon industrial electronics. His research interests includethe macroscopic energetic Representation, multi-drivescontrol and the rolling unrolling system control.* Corresponding author

AUTHORSBouchiba Bousmaha*

Hazzab Abdeldjabar

Hachemi Glaoui

Ismail Khalil Bousserhane

Pierre Sicard

References[1] Thiffault Ch., Pierre Sicard P., Alain Bouscayrol A.,

Tension Control Loop Using a Linear Actuator Based onthe Energetiv Macroscopic Representation CCECE 2004-CCGEI 2004, Niagara Falls, May 2004

[2] Xu Y., , Wang D., Zhang Q., “Modeling and RobustControl of Web Winding System with Sinusoidal TensionDisturbance”. In: Proceedings of the 2006 IEEEInternational Conference on Mechatronics andAutomation, 25th - 28th June, 2006, Luoyang, China,pp. 1958-1963.

[3] Abjadi N.R., Soltani J., Askari J., Navid R.A., Jafar .S,Javad A., ”Nonlinear Sliding-Mode control of a Multi-Motors web winding system without tension sensor”.In: 2008 IEEE international conference on industrialtechnology - ICIT 2008, pp. 1-6.) .

[4] Knittel D., Vedrines M.H., Pagilla D., Prabhakar “Robust H? Fixed Order Control Strategies for Large ScaleWeb Winding Systems”. In: Proceedings of the 2006IEEE International Symposium on Intelligent Control,4th-6th Oct. 2006, Munich, Germany, DOJ10.1109/CACSD-CCA-ISIC.2006.4776941

[5] Koc H., “Modelisation et commande robuste d'un

system d'entrainement de bande flexible” , Ph.D. thesis,Universite Louis Pasteur (Strasbourg I University),2000.

[6] Glaoui H., Fuzzy sliding mode control MIMO for systemmulti motors , Master thesis, University of Bechar,Algeria 2008.

[7] Jung J., Nam K., “A Dynamic Decoupling ControlScheme for High-Speed Operation of InductionMotors”, IEEE Trans. on Ind. Elect., vol. 46/01, 1999.

[8] Mezouar A., Fellah M.K, , Hadjeri S., ”Adaptive SlidingMode Observer for Induction Motor Using Two-TimeScale Approach”, Electric Power System Research, vol.77, issues 5-6, 2007, pp. 1323-1336.

[9] Mohanty K.B., “Sensorless sliding mode control ofinduction motor drives”, TENCON-2008, IEEE Region 10Conference, Hyderabad.

[10] Zhiwen M., Zheng T., Lin F., You X., “A New Sliding-Mode Current Controller for Field Oriented ControlledInduction Motor Drives”. In: IEEE Int. Conf. IAS, 2005,pp. 1341-1346.

[11] Koshkouei A.J., Burnham K.J., Zinober A.S.I., ”DynamicSliding Mode Control Design”, IEEE Proc.--ControlTheory Appl., vol. 152, no. 4, 2005, pp. 392-396.


Articles54

Abstract:

1. IntroductionIn 26 on October 2007 on one of the biggest cons-

truction site in Europe in Grevenbroich - Neurath in NorthRhine-Westphalia in Germany a building accident tookplace. Three people were killed and another five workerswere injured. The accident was caused by faults in struc-tural strength computations. To avoid such terrible re-sults of designers faults in the future the force sensornetwork [6] for online force monitoring in steel rods wasdesigned and built in cooperation with Remak S.A.

The power boiler erection process requires temporarysuspending of buckstays (Fig. 1). The aim of buckstaysinstallation is to brace screens of combustion chamber.Suspending of buckstays is performed by means of strandjacks and steel rods. The amount, location and force rangesof each rods are modelled on the basis of static structuralstrength computations. In the paper the force sensor net-work incorporated into computer and wireless communica-tion system designed to prevent overloading of the rodsresulting from force asymmetry or computational faults isintroduced and next subjected to optimization procedure.The operational data delivered by the network and recordedon the system hard drives let the authors perform thecorrelation analysis and linear regression to reduce thenumber of sensors. Thus two stage hierarchic algorithmwhich constructs the set of models for every single sensorand estimates their parameters, and then using geneticprocedure minimizes a certain loss function to automatethe sensor network optimization process is introduced.As a result, such investigation could significantly reducecost of the whole system.

Keywords: sensor network optimization, mathematicalmodeling.

Fig. 1. Support construction of power boiler with tempo-rary suspended buckstays.

The network was applied during temporary suspen-ding of backstays procedure on the block G in Neurath.In the first section the architecture of the measuringsystem incorporated into force sensor network is introdu-ced. In second and third sections the historical, opera-tional data are investigated using correlation analysis,linear regression and parametric optimization of a cer-tain loss function to reduce the number of requiredsensor units.

2. Force sensor network measuring systemThe force measuring system (Fig. 2) consists of

24 calibrated measuring units (Fig. 3) which make thesensor network. Each of them is based on a measurementfoot designed so as to move part of a known force axiallyto tensometer sensor. The voltage across tensometerbridge is measured in the measuring transducer. The firstelement of that transducer is measuring amplifier withanalog-digital converter with Sigma-Delta modulationperforming measurements with an accuracy of 10 bits.Then, the result of measurement is filtered and scaledusing the microprocessor unit. Scaling is based on thedata from the calibration procedures [1]. The measuredforce is compared with the threshold values derived fromthe maximum load for the sensor (150%) and the valuesof the maximum acceptable load for the rod. The measu-rement data are available from SLAVE module by means ofcommunication protocol with the physical layer interfacebased on RS485. The MASTER unit is working in the pul-ling mode. Each of the 24 measuring units in the networkreplies with the information about the current value ofmeasured force and short-term exceeded limits.

MASTER module provides further aggregated datathe industrial radio modem operating in the 869 MHzband. This enables the remote measurement by the sys-tem installed on mobile cross-bar which is raised toa height of 160 m. Visualization software is installed onthe operating unit and it has been made in technologyJavaSE, where a number of procedures is responsible for:the processes of receiving data, visualization, alarmsdetection, recording and transmission of the historicaldata to the Web server [7], generating reports and ana-lysis of historical data trends. The use of the Java plat-form [2] provides user-friendly tool what is an importantadvantage in daily operation of the system.

Java platform is characterized by an open architec-ture. It is a set of standards used by many software com-panies, which guarantees extensive support for this tech-nology in the future. Detected alarms and warnings areindicated acoustically by an independent microprocessorsystem.

via

MODELLING AND OPTIMIZATION OF THE FORCE SENSOR NETWORK

Grzegorz Bialic, Marcin Zmarzły, Rafał Stanisławski

Received 13 ; accepted 26 .th January 2010 April 2010th


Articles 55

buckstays

Fig. 2. Block diagram of the system.

Fig. 3. Measurement unit.

Data visualized online allow continuous monitoringand regulation of stress in rods. Fig. 4 presents a frag-ment of the force decomposition characteristics whichdescribes the suspending process. The technology usedin the assembly process allows to correct the load only oncurrently installed buckstay. Suspending of consecutivebuckstay prevents the further load revision at a higherlevels.

The number of measurement units has the most signi-ficant impact on the cost of the whole force indicationsystem. Therefore it appeared very important the optimi-zation of the network by reducing a certain number ofsensors. Reduced sensors have to be replaced by theirmodels based on other measurements. In the next sec-tions the correlation analysis, modelling using linear re-gression and minimization of a certain objective functionin the sense of relative mean square error rate are used inthis purpose.

Fig. 4. Historical force values for every single strand.

3. Correlation analysis

4. Two stage network modellingand optimization

The correlation analysis specifies the correlation coef-ficients among all sensors. Then one can determine whe-ther the particular sensor can be directly replaced withthe another one. Table 1 shows the correlation coeffi-cients between the sample sensor (S ) and remainingsensors.

It is easy to observe that the correlation coefficientsare too low to replace the sensor S by the other one. Si-milar results (on the average) have been obtained for re-maining sensors. On the other hand, the results of corre-lation analysis indicate that signal for the particular sen-sors are strongly dependent on each other. Thereforeusing more advanced algorithm to reduce the number ofsensors one can obtain satisfactory results.

In this section two stage algorithm (consists of mo-delling stage and next optimization stage) which mini-

� 2

2


Articles56

VOLUME 4, N° 3 2010

Table 1. The correlation among the measurements for sensor S against the others.2

sensor nr

sensor nr

sensor nr

�

�

�

10.667

100.629

180.612

30.580

110.523

19-0.24

40.661

120.669

200.676

50.679

130.631

210.071

60.672

140.042

220.6534

70.639

15-0.17

230.567

80.674

160.652

24-0.02

890.109

170.677

250.641

Assuming above presented methodology one canobtain the set of models for each element of the network.Every model imitates a certain element of the networkwith some accuracy. Certainly only the model which pro-duce the lowest value of MSE could be selected to replacethe sensor in the network. This in turn means that thereplaced before sensor cannot be further used to modelremaining elements of the system. As a result, the opti-mization of the sensor selection process has to be per-formed. The selection procedure should follow the guide-lines: replace as many sensors as possible, simultaneo-usly keeping the lowest MSE rate of the model. Thus,there was required to develop the tool for optimizationthe sensor selection process. The amount of the sensorswhich are supposed to be reduced was defined as theinput of the routine. Finally the algorithm returns theindices of the sensors which should be replaced by themodel on the basis of minimization of the following lossfunction

(7)

where and is the number of reducedsensors. The complexity of the above defined loss func-tion (7) (large number of local minimums) decided thatthe genetic algorithm (GA) was chosen for finding theoptimal solution. GA parameters were as follows:

population size: 20,maximum generation: 100,selection function: stochastic uniform,cross-over function: Heuristic with cross-over proba-bility 0.9 and mutation probability: 0.1.

The stochastic nature of the genetic algorithm causedthat the optimization was started from different initialconditions. Then form the set of the solutions the mostoptimal result was selected.

The properties of models used for sensor replacementfor the number of reduced elements are shown inTable 2.

Table 2. The properties of reduced sensors models.

modelled sensor MSE [%]

1 4 17 0.9954 4.744 7 13 0.9959 3.7213 10 16 0.9950 4.5820 17 22 0.9968 4.38

In the Table 3 values of these models parameters werecollected.

Table 3. Values of the models parameters.

modelled sensor

1 -15.6591 0.5202 0.67004 3.6157 0.8150 0.30213 -21.0073 0.3382 0.833120 -2.1117 0.4892 0.4652

B. Loss function minimization

n

p p p n

n

k L

a a a

={ , …, }

=4

1

0

n

k l

�

�

�

�

�

mizes the number of sensors was employed. The hierar-chical block diagram of this method is presented in theFig. 5.

First stage involves modelling of element by meansof a linear combination of two others sensors. The equa-tion of model for the element can be written in follo-wing form

(1)

where: , , are the unknown model coefficientsand , are the outpus of the , sensors in -thsample ( ). The model equation for theunit can be easily presented in linear regression form

(2)

where estimates of unknown model parameters

and (3)

or in alternative vector/matrix form

(4)

where is the model output signal vector of the form

and

where is the number of measurements. Thus unknownparameters of the model can be estimated by means ofthe classical least squares method [3, 4]

(5)

Using this method for each elementthe set of models was developed and constructed foreach pair , where and .Among all these models for the element best fittedmodel in terms of relative mean square error value (rMSE)should be chosen. Relative MSE is calculated by means ofthe following equation

(6)

Fig. 5. Block diagram of the two stage algorithm whichminimizes the number of sensors.

A. Sensors modelling and estimationS

S

a a ay y S S i

k,l S

N

S p

y y k, l l k l, k pS

p

p

pkp

lp

ik

il

k l

p

p

l k

p

0

=1,…,24

( = 1,..,24)

( = 1 ,…, 24) � �


Articles 57

VOLUME 4, N° 3 2010

]...[ 21

p

N

ppp yyyy � ]...[ ,,

2

,

1,

lk

N

lklk

lk ��

The assumption that results in the MSE ratebelow 5% and then keeps total functionality of the consi-dered system. In the Fig. 6 current value of sensor vs.defined by the model illustrates pretty good performanceof the methodology, which has been introduced in thepaper.

In the Fig. 7 the impact of the reduced sensors num-ber n on the quality of estimation was illustrated. Thusthe values of the mean square error ( ), and the maxi-mum value of the MSE (max(MSE)) as a function of n wereplotted in this figure. The maximum value of the MSE re-flects the estimation error for the sensors with the worstfit model.

Certainly, the number of reduced elements dependson the assumed error rate. The larger acceptable error,the more elements can be reduced. For example, for rela-tive MSE < 7%, 6 sensors may be replaced by its models.In order to preserve the functionality of the designedsystem the error rate of every single modeled sensorshould be kept in range of 10%. From this point of viewthe reduction of 6 measurement units seems to be themost desired.

n

S

S

J/n

=4

4

4

Fig. 6. Plots of actual vs. estimated outputs of the modelfor the sensor .

Fig. 7. Dependence of MSE and max(MSE) on the reducedsensors number.

5. ConclusionApplication of force sensor network during suspen-

ding of buckstays has improved the safety of the wholeerection procedure. The electronic measurements in realtime and recorded data let engineers symmetrize theforce and verify values modelled before. Using presentedin the paper two stage algorithm based on correlationanalysis, linear regression and parametric minimizationof a certain loss function, gives them the opportunity toreduce the number of measurement units in the networkwhich significantly diminishes the costs of the system. Itwas presented in the paper that some sensors could bereplaced with computationally modelled data and thenerror (in the mean square sense) of the whole force indi-cation system leaves in reasonable range. Moreover, it isworth emphasizing, that the presented above solutiondiscovers new fields for applying modern measurementsystems and microprocessor, computer and informationtechnology.

- Department of Electrical Engineering,Automatic Control and Computer Science, Opole, Univer-sity of Technology, 31 K. Sosnkowskiego St., PL45-272,Opole, Poland. Tel. +48 77 4006208; fax: +48 774006338.E-mail: [email protected].

- Department of Electrical Engineering,Automatic Control and Computer Science, Opole, Univer-sity of Technology, 31 K. Sosnkowskiego St., PL45-272,Opole, Poland. Tel. +48 77 4006208; fax: +48 774006338.E-mail: [email protected].

- Department of Electrical Engineer-ing, Automatic Control and Computer Science, Opole Uni-versity of Technology, 31 K. Sosnkowskiego St., PL45-272,Opole, Poland. Tel. +48 77 4006208; fax: +48 774006338.E-mail: [email protected].* Corresponding author

AUTHORSGrzegorz Bialic*

Marcin Zmarzły

Rafał Stanisławski

References[1] Bialic G., Zmarzły M., Szmechta M., Stanisławski

R.,“Proces kalibracji głowic pomiarowych dla systemupomiaru sił wykorzystywanego podczas montażu kotłaenergetycznego 1100 MW”,

, no. 2, 2009, pp. 111-113. (in Polish)[2] Bialic G., Zmarzły M., Stanisławski R., “Design of the

scales for the power boiler fuel feeding system based onthe process identification”. In:

, Międzyzdroje,19 -21 August 2009.

[3] , Java Platform Enterprise Edition,Specification v5.0, 2005.

[4] Ljung L., , Prentice-Hall, Engle-wood Cliffs, NJ, 1987.

[5] Söderström T., Stoica P., ,Prentice-Hall, Englewood Cliffs, NJ, 1989.

[6], Patent application no.

P386137, Patent Office RP, Warszawa 2008. (in Polish)[7]

Pomiary - Automatyka -Kontrola

14 IEEE IFACInternational Conference on Methods and Models inAutomation and Robotics. MMAR'2009

java.sun.com

System Identification

System Identification

Sposób pomiaru siły obciążenia cięgna oraz głowica dopomiaru obciążenia cięgna

Sposób kontroli poziomu i urządzenie do kontroli pozio-

th

th st


Articles58

VOLUME 4, N° 3 2010

1 2 3 4 5 6 7 8

MSE (J/n) 2,72 3,03 3,76 4,35 4,49 5,41 7,98 9,98

max(MSE) 2,72 3,34 4,61 4,74 5,13 6,74 12,38 24,14

0,00

5,00

10,00

15,00

20,00

25,00

30,00

MS

Erate

[%]

n

mu elementów wiszących dużych konstrukcji

Pomiary - Automatyka- Kontrola

Pomiary - Automatyka -Kontrola

5 International Conference New Elec-trical and Electronic Technologies and Their IndustrialImplementation

, Patentapplication no. P.389559, Patent Office RP, Warszawa2009. (in Polish)

[8] Stanisławski R., Bialic G., Zmarzły M., „Identyfikacjawłasności dynamicznych układu przygotowania paliwakotła energetycznego BP-1150”,

, 2009 vol. 2, pp. 108-110. (in Polish)[9] Zmarzły M., Bialic G., Stanisławski R., Rogala A., „Sys-

tem pomiaru sił dla procesu wciągania bandaży na kons-trukcję kotłów energetycznych”.

, 2009 vol. 2, pp. 114-116. (in Polish)[10] Zmarzły M., Szmechta M., “The efficiency and reliability

analysis of a telemetric event driven data transmissionover GPRS. In:

, Zakopane, 12 -15 June 2007.

th

th th


Articles 59

VOLUME 4, N° 3 2010

� The robots will rock them all

We find robots used in the industry assembling cars, electronic devices or in the surgery room as common. We accepthumanoid robots as movie heroes like R2D2 or C3P from “Star Wars”. But nobody could imagine robots as rock stars! Alas! OnFebruary 19 in Seattle`s Key Arena five huge industrial robots equipped with five large LED screens debuted on the stageaccompanying Bon Jovi`s “The Circle Tour”.

The concept of moving screens is a RoboScreen patented technology developed by Andy Flessar who is the founder andpresident of Robotic Arts of Las Vegas in Nevada. He started to work on robots in mid 1990s and in 2006 completed a roboticprogramming, design and operation programme at the ABB in Auburn Hills in Minnesota. Previously he constructed outdoorpanels placed on building walls though then he developed the idea of a graphic screen connected with a robotic arm.Flessas provided a software to animate the movements of the ABB robots: first the desired movement is established thenRobot Animator channels the code into ABB`s robot controller and the robots replicate the movement on the stage.

“We are able to take the ABB robots out of the factory and turn them into rock stars through the power of the ABB IRC5controller and its ability to concept the precise movement established in Robot Animator” said Flessas.

The unique concept and physical presence of the IRB 7600 robot attracted Jon Bon Jovi and the tour directors. Theywere overwhelmed by elegant choreography and the visionary application of the installation choreographed precisely withthe music. “The collaboration with Robotic Arts and Bon Jovi is certainly one of the most unique applications we have beeninvolved” said Joe Campbell, vice president of sales and marketing, ABB Robotics, North America.

Five ABB7600 huge industrial six-axis robots used in “The Circle Tour” use inverse kinematics to create robotic motions.Each robot has a 2x3 metres large and 320 Kg heavy LED video panel attached to the articulated robot`s arm.

All five robots can create one large screen or split into five separate screens, each one of 24 individual sub-panelsarranged in a six column by four-row grid. They are positioned toward the back of the stage and are an integral part of theconcerts displaying approximately 85% real-time video footage of the show from multiple cameras set up on the stage andin the audience and pre-programmed 3-D computer animations. Their movements are coordinated at 30 frames per sec usingthe time code to synchronize the real time to the beat of the music. During the show the robots are operated by Gordon“Gordo” Hyndford.

th

TM

Focus on new


IN THE SPOTLIGHTby Urszula Wiączek - report from Paris

VOLUME 4, N° 3 2010

In the spotlight60

The panels are covered by plexi glass to protect their surface, when they are positioned horizontally creating steps anda catwalk on which Jon Bon Jovi is tapping 4 meters above the stage.

“Industrial robots being a part of a major concert tour are unprecedented. They provide a big show element to theperformance and help present the complete entertainment experience that is synonymous with the Bon Jovi brand” saidFlesssas.

Tait Towers, the tour production company in Lititz, in Pennsylvania, did the robots design and the system that allowsthe RoboScreens to fit seamlessly into the touring platform.“The Circle World Tour” will last two years.

More information on: http://watch.discoverchannel.ca/clip307254#clip307254


?61In the spotlight


July

August

September

October

November

9 – 11 ICCDA 2010

1 – 3 ICEIE 2010

8 – 10 ABSRC 2010

9 – 12 ABSRC 2010

10 – 12 ICMET 2010

17 – 19 ICEIT 2010

3 – 5 ICSTE 2010

17 – 22 ESF-EMBO Symposium

18 – 19 ICINA 2010

26 – 29 MobiCPS 2010

12 – 14 ICSTE 2010

16 – 18 ICMCE 2010

16 – 18 ICCEE 2010

– IEEE International Conference on Computer Design and Applications,Qinhuangdao, China.http://www.iccda.org/

– IEEE International Conference on Electronics and Information Engineering,Kyoto Japan.http://www.iceie.org

– Advances in Business-Related Scientific Research Conference, Olbia, Sardinia,Italy.http://www.absrc.org/

– 11 Country Robotic Conference, Karpacz, Poland.http://kkr11.iiar.pwr.wroc.pl/

– International Conference on Mechanical and Electrical Technology, Singapore,Singapore.http://www.icmet.ac.cn/

– IEEE International Conference on Educational and Information Technology,Chongqing, China.http://www.iceit.org/

– 2 IEEE International Conference on Software Technology and Engineering,San Juan, Puerto Rico.http://www.icste.org/

– Functional Neurobiology in Minibrains: From Flies to Robots and BackAgain, Sant Feliu de Guixols, Spain.http://www.esf.org/conferences/10324

– IEEE International Conference on Information, Networking and Automation,Kunming, Yunnan, China.http://www.icina.org/

– 1 IEEE International Workshop on Mobile Cyber-Physical Systems, Xi'an,China.http://www.cpschina.org/mobicps

– 2 International Conference on the Roles of the Humanities and Social Sciencesin Engineering, Pulau Pinang, Malaysia.http://publicweb.unimap.edu.my/~icohse2010

– International Conference on Measurement and Control Engineering 2010,Chengdu, China.http://www.icmce.org/

– 3 IEEE International Conference on Computer and Electrical Engineering,Chengdu, China.http://www.iccee.org/

th

nd

st

nd

rd

EVENTSSUMMER-AUTUMN 2010

VOLUME 4, N° 3 2010

Events62

Documents

JAMRIS 2010 Vol 4 No 3