JAMRIS 2010 Vol 4 No 3 by JAMRIS

JOURNAL of AUTOMATION, MOBILE ROBOTICS & INTELLIGENT SYSTEMS

Editor-in-Chief Janusz Kacprzyk

Executive Editor: Anna Ładan aladan@piap.pl

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

Associate Editors: Mariusz Andrzejczak (PIAP, Poland) Katarzyna Rzeplińska-Rykała (PIAP, Poland)

Co-Editors: Dimitar Filev (Research & Advanced Engineering, Ford Motor Company, USA)

Kaoru Hirota

Webmaster: Tomasz Kobyliński tkobylinski@piap.pl

(Interdisciplinary Graduate School of Science and Engineering, Tokyo Institute of Technology, Japan)

Witold Pedrycz

Proofreading: Urszula Wiączek

(ECERF, University of Alberta, Canada)

Roman Szewczyk (PIAP, Warsaw University of Technology, Poland)

Editorial Office: Industrial Research Institute for Automation and Measurements PIAP Al. Jerozolimskie 202, 02-486 Warsaw, POLAND Tel. +48-22-8740109, office@jamris.org

Editorial Board: Chairman: Janusz Kacprzyk (Polish Academy of Sciences; PIAP, Poland) Plamen Angelov (Lancaster University, UK) Zenn Bien (Korea Advanced Institute of Science and Technology, Korea) Adam Borkowski (Polish Academy of Sciences, Poland) Wolfgang Borutzky (Fachhochschule Bonn-Rhein-Sieg, Germany) Oscar Castillo (Tijuana Institute of Technology, Mexico) Chin Chen Chang (Feng Chia University, Taiwan) Jorge Manuel Miranda Dias (University of Coimbra, Portugal) Bogdan Gabryś (Bournemouth University, UK) Jan Jabłkowski (PIAP, Poland) Stanisław Kaczanowski (PIAP, Poland) Tadeusz Kaczorek (Warsaw University of Technology, Poland) Marian P. Kaźmierkowski (Warsaw University of Technology, Poland) Józef Korbicz (University of Zielona Góra, Poland) Krzysztof Kozłowski (Poznań University of Technology, Poland) Eckart Kramer (Fachhochschule Eberswalde, Germany) Andrew Kusiak (University of Iowa, USA) Mark Last (Ben–Gurion University of the Negev, Israel) Anthony Maciejewski (Colorado State University, USA) Krzysztof Malinowski (Warsaw University of Technology, Poland)

Andrzej Masłowski (PIAP, Poland) Tadeusz Missala (PIAP, Poland) Fazel Naghdy (University of Wollongong, Australia) Zbigniew Nahorski (Polish Academy of Science, Poland) Antoni Niederliński (Silesian University of Technology, Poland) Witold Pedrycz (University of Alberta, Canada) Duc Truong Pham (Cardiff University, UK) Lech Polkowski (Polish-Japanese Institute of Information Technology, Poland) Alain Pruski (University of Metz, France) Leszek Rutkowski (Częstochowa University of Technology, Poland) Klaus Schilling (Julius-Maximilians-University Würzburg, Germany) Ryszard Tadeusiewicz (AGH University of Science and Technology in Kraków, Poland)

Stanisław Tarasiewicz (University of Laval, Canada) Piotr Tatjewski (Warsaw University of Technology, Poland) Władysław Torbicz (Polish Academy of Sciences, Poland) Leszek Trybus (Rzeszów University of Technology, Poland) René Wamkeue (University of Québec, Canada) Janusz Zalewski (Florida Gulf Coast University, USA) Marek Zaremba (University of Québec, Canada) Teresa Zielińska (Warsaw University of Technology, Poland)

Publisher: Industrial Research Institute for Automation and Measurements PIAP

If in doubt about the proper edition of contributions, please contact the Executive Editor. Articles are reviewed, 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.

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

CONTENTS REGULAR PAPER 60

A comparative study of incremental algorithms for computing the inverse kinematics of redundant articulated systems A. Moussaoui, R. Otmani, A. Pruski 10

A novel flexible micro assembly system: implementation and performance analysis D. Um, D. Ryu, B. Dong, D. Foor, K. Hawkins 16

Intelligent PI controller and its application to dissolved oxygen tracking problem T. Zubowicz, M.A. Brdys, R. Piotrowski 25

Fish-like swimming prototype of mobile underwater robot M. Malec, M. Morawski, J. Zając 31

A mobile system for measurements of partial discharges controlled by electroencephalographic waves A. Błachowicz, S. Paszkiel 36

Automatic generation of fuzzy inference systems using heuristic possibilistic clustering D.A. Viattchenin 45

Capacitive human presence sensor for safety applications P. Frydrych, R. Szewczyk 50

Sliding mode speed control for multi-motors system B. Bouchiba, A. Hazzab, H. Glaoui, F. Med-Karim, I.K. Bousserhane 55

Modelling and optimization of the force sensor network G. Bialic, M. Zmarzły, R. Stanisławski

DEPARTMENTS IN THE SPOTLIGHT 62

EVENTS

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

NÂ° 3

2010

A COMPARATIVE STUDY OF INCREMENTAL ALGORITHMS FOR COMPUTING THE INVERSE KINEMATICS OF REDUNDANT ARTICULATED SYSTEMS Received 9th April 2010; accepted 26th May 2010.

Abdelak Moussaoui, Rafaa Otmani, Alain Pruski

Abstract: The field of control of redundant articulated mechanical systems requires the use of algorithms to compute the inverse kinematics. The fields of animation, virtual reality or game in particular, are very interested in these algorithms. We propose in this paper a comparison between several algorithms of incremental type. The considered application concerns the accessibility evaluation of an environment used by a handicapped person (an apartment, a house, an institutionâ&#x20AC;Ś). The physical disability involves a particular characteristic of the human articulated structure that gives rise to constraints that must be taken into account in computing the inverse kinematics. Keywords: inverse kinematics, accessibility, physical disability.

1. Introduction The field of robotics has been the first that was interested in the problem of inverse kinematics. The articulated systems have long been made up of six degrees of freedom and their controls were based on linearization methods, which gave solutions consistent with expectations. Currently the humanoid robots with high redundancy require more computing time. On the other hand, the animation of avatars requires the use of the same types of computing. We note that the type of algorithm is highly dependent on the type of application envisaged. In this paper we contribute to propose algorithms for computing the inverse kinematics with constraints related to an application in the field of accessibility assessment. We analyse the relevance of each in a specific context.

2. Background The comparative study that we present in this paper concerns algorithms to determine the inverse kinematics of redundant articulated systems with very high number degrees of freedom. These systems are highly non-linear and require special methods of resolution. The principle is to calculate the value of joint variables such as: Q=f-1[(X]) With Q the joint variables and [X] the constraint vector. Many methods are proposed in the literature that we can classify into three categories: - Analytical methods;

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 computing time and/or the accuracy obtained. A redundant system has no single solution and other criteria are frequently used to converge toward a particular solution. The analytical methods are used when the number of variables 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 of solving the inverse kinematics specific for obtaining a solution quickly. By cons, it is linked to this particular structure and cannot extend to other structures. Linearization methods are commonly used when the problem is complex and non-linear. The method is to approach the solution by successive increments considering that the system is linear around the operating point. The problem is solved by inverting a Jacobian matrix that is usually singular implying to calculate its inverse by the pseudo inverse techniques whose computational cost is important. [3] The optimization methods are the most interesting and often used when the number of variables is important if we wish to obtain solutions meeting certain criteria. The principle involves formulating the problem as a cost function minimisation problem. Many approaches have been developed which include the gradient descent with special adaptations [4], [5] and genetic algorithms [6]. These algorithms are effective but can often lead to local minima. They can be very fast especially using BFGS-type methods, which approximate calculation of the inverse matrix. The work presented in this paper concerns algorithms that are part of the class of optimization methods.

3. Context 3.1. Introduction The work presented in this article is made in the context of a specific application that is to test the accessibility manipulation (Fig. 1) of a living place for a disabled person. For this purpose the articulated system consists of a humanoid-type structure that moves with a mobile base that models a walker or a wheelchair. The displacement device is important since we must consider the physical placement of the articulated structure with its mobile base in the environment. This application requires taking into account several parameters: Articles

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

The number of degrees of freedom taking into account the human trunk, one arm and the mobility system which is 24 degrees of freedom, as will see later; In essence, the human articulated system has limitations that we consider; We want verify only if a solution exists. We do not take into account other characteristics such as comfort or energy expended in computing the solution. We believe that if a solution exists then the person can reach the considered point; The accessibility evaluation is conducted in statics. We do not consider the motion type (e.g. the wheelchair non holonomy is not taken into account). Only the admissible geometric placement, which is to say without intersection with the physical environment, is considered; The required computing accuracy is not very important since we can consider that the compliance of the human body compensates for errors.

N° 3

2010

his right hand. We believe that mobility is achieved by a rectangular base corresponding to a wheelchair. The motion device is modelled using three degrees of freedom: a rotation and two translations. The frame position of the root structure is situated at the height of the waist of a seated person and the area swept by the wheelchair is a rectangle. Figure 2 shows the kinematics chain of the global user with his mobile base. The mathematical model is established from the multiplication of Denavit-Hartenberg matrices [8] whose prototype is given below: écos Qi - cosa i sin Qi sin a i sin Qi ai cos Qi ù ê sin Q cosa i cos Qi - sin a i cos Qi ai sin Qi úú i DH i = ê ê 0 sina i cosa i di ú ê ú 0 0 1 û ë 0 (1) And for n joints: DH 0,n = DH 0 ...DH i ...DH n

(2)

For any vector V in the frame n in homogeneous coordinates, we will have in the R0 world frame: VR0 = DH 0,nVRn

(3)

Two of the three algorithms that we analyse, work in an iterative manner that is to say that the joints are modified one after the other in order to verify if the endeffector approaches the goal. This sequential aspect of the algorithm allows to speeds up the computation of the direct model without having to reconsider all the joints. When we change one variable, only the variable corresponding matrix is modified. We can write DH0, n by considering two matrices. Fig. 1. Accessibility of the environment with a wheelchair. 3.2.

DH 0,n = DH 0,i -1 * DH i,n

(4)

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

DH 0,q+n 1 = DH 0,qi -1 * DH iq+1 * ( DH i ) q-1 * DH i,qn = = DH 0q,+i 1 * DH iq++11,n

(5)

DH 0,q+n 1 = DH 0,qi * ( DH i ) -q1 * DH iq+1 * DH iq+1n = (6)

= DH 0q,+i -11 * DH iq,n+1

With q for the iteration computing. This method requires only three matrix multiplications instead of n. The inverse DH-1 is directly given by the following expression: Fig. 2. The used kinematic joint structure. The incremental algorithms that we study in this article require knowledge of the direct kinematics model. The system is described by a biomechanical model of the human being with limits in the range of joint movement. These are taken into account in the algorithm of the inverse kinematics, as we will detail later. The model we use is that proposed by [7] from which we extracted a model with 21 degrees of freedom from waist toward the tip of 4

Articles

écosQi ê ê- cosai sinQi DH -1 = ê i êsinai sinQi ê ëê0

sinQi

cosai cosQi

sinai

- sinai cosQi

cosai

- ai

ù ú - d i sinai ú ú - d i cosai ú ú 1 ûú

(7) The two equations (5) and (6) are used, the first when the variables are changed sequentially from 0 to n (from

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

root to tip) and the second when the variables are changed from n to 0 (root to tip). The proposed algorithms use the two equations. 3.3. The algorithms We propose in this section three algorithms whose performance depends on usage. All these algorithms are of incremental type with optimising a cost function, which corresponds to the error between the point to be reached, and the current point. This point is defined by the position of the hand and the orientation of the two vectors X and Z (Fig. 3). The error is the sum of the differences between the projections of current vectors X21 and Z21 and the desired vector in world frame. Z21 X21

Fig. 3. Definition of vectors used in defining orientation. Thus we determine: e = w0*(Current Position - Desired Position) + w1* (Current Orientation - Desired Orientation) And generally:

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

(8) (9)

With f([Q]) the instantaneous position and [X] the desired position and orientation. 3.3.1. Algorithm Cyclic Coordinate Descent (CCD) This algorithm has been developed over many years and the following paper [9] details the fundamental principles. It is based on an incremental computing of the variables from the system end toward the root by minimizing the error between the desired point (and direction) and the current point (and direction). Some drawbacks are well known as an inhomogeneous variables adaptation in the chain joints and a slow convergence near the target. Where no limits imposed on variables, then the algorithm converges without local minima. It is possible to impose limits on the joints as we do in our application, in that case local minima may occur. 1. Initialise randomly the joint variables Qi 2. Do For each variable Qi from i=n to 1 Find the optimum value of Qi that minimise the error e 4. While Stop Conditions not verified The joint limits are taken into account by checking if the optimum value of Qi is included in the admissible limits. Otherwise, the calculated value Qi is not taken into account and we move to the next variable. In Phase 2 of the algorithm, it is necessary to compute the direct model in particular to obtain the error between the goal and that current point corresponding to the cost function e. In computing the loop, only one variable is changed thereby applying equation (6).

N° 3

2010

3.3.2. Incremental approximation algorithm (IAA) The principle of this algorithm is to modify each value of variables Qi from root to tip in order to minimize the magnitude of the cost function e. Unlike the CCD algorithm, the value of the increment Inc applied to each variable is not calculated but imposed. The resulted value of Q is preserved if it is within the allowed range. The increment Inc is calculated for each joint i as : Inc(i)=(Max(i)-Min(i))*IncrementRate

(10)

with Max(i) and Min(i) the minimum and maximum limits of the joint i. IncrementRate can adjust the speed of the algorithm convergence. The parameters Inc(i) is very important in two aspects of its sign and amplitude that contribute to the speed of convergence. In the methods of gradient descent as Newton-Raphson, gradient matrices and the inverse of the Hessian fulfil these roles. The optimization of these values can accelerate the convergence. In our case we modify the basic algorithm by storing the sign of Inc. For each variable i we use the same sign at the next iteration. Convergence is rapid initially and then the variation becomes smaller with the proximity of the solution. We propose a modification of the algorithm by adjusting the step of the increment Inc(i) depending on the magnitude of the cost function De in a non-linearly manner as in equation (11). Other adaptation functions could be applied: if (De=0) then IncrementRate=IncrementRate/2

(11)

A linear adjustment does not improve the speed of convergence. If the increment Inc(i) is sufficiently large, the distance change is rapidly becoming zero around the solution. We use the cancellation of De to decrease the value of the increment Inc(i). This algorithm is adaptable to any articulated structure. In this case, it is equation (5), which is used to calculate the direct model. 1. Initialise randomly the joint variables Qi 2. Do 2.1. Define the increment Inc (i) 2.2. Do for each variable Qi 2.2.1. Qi=Qi+ Inc (i) Compute the distance between Current Solution and Goal such as e=f([Q])-[X] if (De) <0) then keep Qi Else Qi =Qi 2* Inc (i) Compute e =f([Q])-[X] if (De <0) then keep Qi Else Qi=Qi + Inc (i) (keep the original value) 3. While Stop Conditions not verified 3.3.3. The Random algorithm Approximation Algorithm (RAA) Unlike the other two algorithms, this one treats all variables simultaneously. In a range of values Inc(i) defined by equation (10), we choose randomly increments that are added to the current values of variables. If the Articles

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

choice lowers the cost function e then the setof new values is maintained if not it is rejected. The procedure is repeated until the stop conditions are satisfied. In this case, it is equation (2), which is used to determine the model because all variables are modified during each iteration. The definition of the range increments Inc(i) is performed as previously by a non-linearly adjustment of IncrementRate (11) by varying the cost function. 1. Initialise randomly the joint variables Qi 2. Do 2.1. Define the increment Inc (i) 2.2. For all variable Qi Do Qi=Qi + Rand(Inc (i)) (Rand(X) returns a random value in the range X) 2.3. Compute the distance between Current Solution and Goal such as De=f([Q])-[X] 2.4. if (De) <0) then keep Qi Else keep the original value of Qi 3. While Stop Conditions not verified When a set of values reduces the cost function then the same set of increment Inc(i) is applied at the next iteration. This accelerates the convergence.

35000

Comparative analysis of the algorithms based on their application.

3.4.1. Introduction The algorithms presented above have different behaviours depending on the type of use and constraints related to the application. The performance study of each algorithm is performed on a PC of Pentium 4 CPU running at 3.4 GHz. The algorithms stop conditions consists of two elements: - A minimum value of the error e (cost function) equal to 1 with w0 = 1 and w1 = 5000 from equation (8) - A maximum number of iterations equal to 1000.

2010

bad choice of joint variables values in initialisation during algorithms phase 1. We give our results in a success rate on a set of 10,000 trials. We consider that the algorithm has found a solution if the error is less than a minimum allowed value e. In Figure 4 we see that the algorithms converge quickly during the first iterations and take longer then. This overall behaviour remains the same but according to the initial value of variables and the goal to achieve, the algorithms speed may vary. Our application requires the accessibility verification in many the environment points requiring for each target point in space, to calculate the inverse kinematics. The comparison is performed on a statistical study on a large number of trials (10,000). We can compare the number of iterations between the CCD algorithm and IAA since the same number of matrices products is performed at each iteration, it is to say 3 * n (where n is the number of dof). By cons, RAA algorithm requires n multiplications per iteration, which is lower than the other two algorithms. A larger number of iterations is necessary in order to find the solution and the computing time will be longer. We consider that an iteration is achieved when all variables of the articulated structure has been treated.

30000

25000 CCD

20000 Erro r

3.4

N° 3

IAA RAA

15000

10000

5000

0 1

93 116 139 162 185 208 231 254 277 300 323 346 369 392 Nb Increments

Fig. 4. Global behaviour of algorithms. The RAA and IAA algorithms require imposing a rate value IncrementRate we placed in 0.015. This value, defined empirically, lead to the best results. It is linked to the size of the components of the articulated structure, which are defined in appendix. The three algorithms show no local minimum when the joint variables are not restricted. In our case we limit the amplitude of the joints based on physical abilities of the person. Thus in some cases it is possible that the algorithms do not find a solution even if it exist. This is due to

3.4.2. Comparative analysis of the computing of the inverse kinematics without mobility. We propose to decompose the analysis such as we do not consider the mobility at the first time, without the degree of freedom in rotation Z0 and without the two degrees of freedom in translation X and Y. The system has 21 degrees of freedom corresponding to a human joint structure from the waist to a hand. The target choice is achieved as follows. The joint variables values are initialised randomly

Table 1. Comparison between the described algorithms. CCD

Algorithms

IAA

RAA

Average iteration number

Average Computing time in ms

Average iteration number

Average Computing time in ms

Average iteration number

Average Computing time in ms

Position w0=1 and w1=0

2.7

131

6.9

Position and orientation w0=1 and w1=5000

211

60.9

146

46.0

4770

301

Articles

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

2010

It is to check whether there is a set of variables Q such as f([Q]) may reach the point [X] whereas the mobility space f(x,y,z). The position to achieve does not longer correspond to a point but a surface. Thus the cost function becomes

that allows to determine associated position and orientation. It is these values that we target to reach. Table 1 shows the comparative elements between the algorithms. We note that the CCD algorithm is the faster when the position is imposed. When we impose both the position and the orientation, the algorithm IAA is slightly more 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 initial values given in phase 1. For that reason all solutions are given with 100% success.

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

(13)

We can write that: e = w0*(Current Position - Desired Surface) + w1* (Current Orientation - Desired Orientation) (14) Factor w1 of the cost function remains unchanged but the other part requires the distance computation between a point and a surface. This is easy to calculate. The CCD algorithm requires a point to reach and not a surface to find the optimal variables value at each iteration. Thus this method is not applicable to the CCD algorithm. We note that the inverse kinematics computing is significantly faster than for the fixed base. The reason comes from the number of potential points to reach is more important since we have to compute the distance from one point to a surface. The algorithm must compute, for each value of the mobile base orientation, the related configuration polygon. In order to accelerate the computing, we have pre compute these polygons. Adding the mobile base accelerates the computing.

3.4.3. Comparative analysis for the computing of the inverse kinematics of articulated structure based on mobile. In this case, it is necessary to take into account both the additional degrees of freedom and the bulk of the mobile base. We assume that the mobile base moves along the floor, which allows us to define the mobility area by a polygon E that we call, envelop polygon. Some obstacles create exclusion areas in which the base cannot moves. All possible positions inside the Envelop polygon is called configuration polygon C. The shape of the polygon C is related to the orientation Z0 of the mobile. It corresponds to the Minkowski difference of which the reader will find details in [10]. The configuration polygon is determined by scanning the envelope polygon for any value of the Z0 variable since it is dependent on the orientation of the mobile (Fig. 5). Exclusion obstacle polygone

3.4.4. Comparative analysis for the computation of the inverse kinematics of articulated structure on fixed base with small variations In our application, which is to verify the existence of a solution in the case of accessibility, we test a set of points in the environment. We must check for each environment point if a solution exists. The environment points are achieved by taking each target point as situated in the neighbourhood of the previous one. Thus, we do not achieve a random joint variables initialisation as defined in the algorithms. We maintain the joint variables values as previously. We initialise the joint variables randomly only at the beginning of the process of computing when the first application of the algorithm. To model this application we perform the computation of a set of target points situated on a circle consisting of 200 points. As we saw earlier, the application of the method that takes into account the mobility can not be applied to the CCD algorithm where the shaded area in Table 3. We see that the various algorithms require only a few iterations to achieve the goal where the weak computing time.

The configuration polygon C

The mobile base The envelop Polygon E Fig. 5. Different polygons definition. Now, the problem is to find a solution to the following problem f([Q])=[X]- f(x,y,z)

N째 3

(12)

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

RAA

Average iteration number

Average Computing time in ms

Average iteration number

Average Computing time in ms

Position w0=1 and w1=0

11.3

0.7

1.6

Position and orientation w0=1 and w1=5000

35.9

15.7

2028

125

Articles

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N째 3

2010

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

Reference on mobile base

Fixed reference

Algorithms

IAA

CCD Average iteration number

Average Computing time in ms

Average iteration number

Average Computing time in ms

Average iteration number

Average Computing time in ms

Position w0=1 and w1=0

1.4

0.3

1.8

0.7

2.1

Position and orientation w0=1 and w1=5000

8.9

2.2

6.2

2.2

369

17.1

Position w0=1 and w1=0

0.7

0.5

5.2

Position and orientation w0=1 and w1=5000

4.2

1.7

132

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

4. Conclusion In this article, we compare three algorithms for computing the joint variable values in the case of the inverse kinematics of an articulated mechanical structure. The context of the use of these algorithms consists to check the accessibility of living environments for persons with disabilities. This application requires taking into account the constraints that have been detailed. The computation time is a key criterion that led the work conducted in this area. The CCD algorithm, well known in the field of animated avatars, seems particularly helpful when the base is fixed and when the target point is defined by a position. When we introduce the orientation constraint on the target point then the algorithm IAA we propose is slightly faster for a higher accuracy. The difference in results is not really significant and does not lead to a clear choice of best algorithm to use. When we consider that the mechanical structure is based on a mobile base so we get a greater difference in the results. The proposed algorithm takes into account the area swept by the mobile base is not adaptable to the CCD algorithm. It requires knowledge of a goal point for compute the joint variable values. Especially for the calculation of scalar and vector products. The proposed algorithm considers the point to reach is materialized by a surface and the number of potential points to reach is more important which has the effect of reducing the execution time and the number of iterations. We have tried to take the nearest point from the polygon solution to the articulated system. No significant improvements are noted and for the position and orientation constraint the results are lower in terms of iteration number. 8

RAA

Articles

AUTHORS Abdelak Moussaoui, Rafaa Otmani, Alain Pruski* Lasc, ISEA, 7 rue Marconi, University of Metz, France, tel. + 33 3 87 31 52 81. E-mails: {abdelhak.moussaoui, otmani, alain.pruski}@univmetz.fr. * Corresponding author

Appendix Table 4. The Denavit-Hartenberg table of the considered kinematic chain. Q D a a 0 PI/2 0 PI/2 0 1 PI/2 0 PI/2 0 2 PI/2 0 PI/2 0 3 PI/2 L1 PI/2 0 4 PI/2 0 PI/2 0 5 PI/2 0 PI/2 0 6 PI/2 L2 PI/2 0 7 PI/2 0 PI/2 0 8 PI/2 0 PI/2 0 9 PI/2 L3 PI/2 0 10 PI/2 0 PI/2 0 11 PI/2 0 PI/2 0 12 -PI/2 L4 PI/2 L5 13 0 0 PI/2 0 14 0 0 - PI/2 L6 15 0 0 PI/2 0 16 PI/2 0 PI/2 0 17 0 L7 - PI/2 0 18 0 0 PI/2 0 19 PI/2 L8 PI/2 0 20 PI/2 0 PI/2 0 21 0 0 0 0 L1=10; L2=10; L3=10; L4=5; L5=10; L6=10; L7=30; L8=30;

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 4,

N° 3

2010

References [1]

[2]

[3]

[4]

[5]

[6]

[7]

[8]

[9]

[10]

Tolani D., Goswami A., Badler N., “Real-Time Inverse Kinematics Techniques for Anthropomorphic Limbs”, Graphical Models, vol. 62, 2000, pp. 353-388. Kallmann M., “Analytical inverse kinematics with body posture control”, Computer animation and virtual worlds, 2008, 19, pp. 79-91. Baerlocher P., Inverse kinematics techniques for the interactive posture control of articulated figures. Federal school of Lausanne Swaziland, PhD Thesis, 2001. Mukundan R., A fast Inverse Kinematics Solution for an n-link Joint Chain, 5th Int. Conf. on Information Technology and Applications, 2008, pp. 349-354. Muller-Cajar R., Mukundan R., “Triangulation: A new algorithm for Inverse Kinematics”, Proc. Image and Vision Computing, New Zealand, 2007, pp. 181-186. Abdel-Malek K., Yu W., Yang J., Nebel K., “A mathematical method for ergonomic-based design placement”, Int. Journal of Industrial Ergonomics, 2004, pp. 375-394. Yang J., Pitarch E.P., ”Digital Human Modeling and Virtual Reality for FCS”, Technical Report n° VSR-04.02; the University of Iowa, Contract/PR NO.DAAE07-03-DL003/0001, 2004. Denavit J., Hartenberg R.S., “A Kinematic Notation for Lower Pair Mechanisms Based on Matrices”, Journal of Applied Mechanics, vol . 77, 1955, pp. 215-221. Wang L.T., Chen C.C., “A combined Optimization Method for Solving the Inverse Kinematics Problem of Mechanical Manipulators”, IEEE Transactions of Robotics and Automation, vol. 7, no. 4, 1991, pp. 489-499. Lozano-Perez T., “Spatial planning: a configuration Space Approach”., IEEE Trans. on Computer, vol. C 32, n°2, 1983, pp. 108-120.

Articles

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 3

2010

A NOVEL FLEXIBLE MICRO ASSEMBLY SYSTEM: IMPLEMENTATION AND PERFORMANCE ANALYSIS Received 8th April 2010; accepted 11th May 2010.

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

Abstract: Demands for micro/nano products and assembly systems have been raised significantly to meet the ever complex technical needs for modern society. In this paper, we share the experiences and results of the study on the flexible micro assembly workcell focused primarily on a novel system implementation and system performance analysis. For flexible and autonomous assembly operations, we investigated a novel model based 3D depth measurement technology for faster and cost effective means to promote autonomous micro assembly systems in various industries. Micro parts, by its nature, are known of their shapes in advance for the majority of micro applications. We take advantage of the previously known shape of micro parts and hence apply a model based approach for a faster and cost effective localization and 3D depth measure of randomly loaded micro-parts on the workcell. The proposed 3D depth measuring method is based on the pattern recognition and multi-focus technique enabling it to extract only information useful for micro parts assembly for faster recognition. For demonstration purpose, silicon based oxide gears are fabricated by bulk micromachining and are used to study performance indices and to prove the usefulness of the proposed micro 3D depth measurement technology.

cially in the implementation and performance analysis. The term “flexible” is used for multiple model assembly loaded randomly in position and orientation on the workstation. For demonstration purpose of the proposed micro-assembly system, we fabricated silicon dioxide based micro gears and assembly base via bulk micromachining technology. To avoid crystallographic edge formation [4], several options are tried out for anisotropic wet chemical etch of silicon parts with the photolithographic masks in Figure 1. The wafers used were 100 mm thick p-type doped Si with a Miller indices crystal orientation of <100>. The SiO2 layer was grown using a wet oxidation process followed by Photolithography and BOE (Buffered Oxide Etch), and several parts release processes.

Keywords: flexible manufacturing, micro assembly, micro 3D vision.

1. Introduction Demands for micro/nano products and assembly systems have been raised significantly to meet the ever complex technical needs for modern society. Among many is the MEMS (MicroElectroMechanicalSystem) technology that has demonstrated, in nearly every sector, miniaturization of mechanical parts or systems [1]. However, although many studies have shown significant advances in the micro/nano manufacturing technology during the last decades, a full blown solution with flexible manufacturing capability in mind still falls short of industrial implementation in terms of mass production. For example, Fatikow et al. developed a flexible micro-robot for complex microsystems assembly, though the vision system became much complex for easy implementation in industry [2]. In [3], Aoyama proposed a in-situ micro robots for flexible micro assembly, but the lack of means for realtime operations may hinder immediate implementations in industry. In this paper, we share the experiences and results of the study on the flexible micro assembly workcell espe10

Articles

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

Journal of Automation, Mobile Robotics & Intelligent Systems

In order to minimize the crystallographic edge effect of the bulk micromachining, boron is diffused on the silicon dioxide surface to slow down the sharp edge formation [5]. The boron doping process allowed the silicon to be removed while strengthening the resistance of the SiO2 parts to the KOH etch. When used in conjunction with polyimide coating, the developed processes produced a part with good surface quality which matched the initial CAD model. Figure 2 shows high quality micro-gear parts produced by combined methods of boron doping and polyimide coating.

VOLUME 4,

NÂ° 3

2010

1. Micro precision robotic station The assembly station used in the assembly workcell implementation is a precision micro robotic system by National Aperture with the uni-directional repeatability of 2 Îźm. Figure 3 is the picture of the 3 Degrees of Freedom (DOF) robotic system integrated in one assembly with a magnetic grippers.

Fig. 2. SiO2 gears in various sizes. Fig. 3. 3-Axes Robotic wafer platform with a micro gripper. In order to assemble fabricated SiO2 gears, a novel model based 3D depth measurement technology for micro parts is introduced in this paper. Unlike chemically released parts and assembled integrity by MEMS technology, flexible micro assembly is still a daunting task due to difficulties in parts visualization and assembly autonomy. Among many 3D visualization technologies, the most popular for micro scale parts is the confocal mapping [6]. The processing time of confocal mapping to obtain 10 to 100 photos for each pixel and expensive device value, though, hinders commercialization for various micro application industries. To overcome such a barrier, we investigate a novel model based 3D depth measurement technology for faster and cost effective means to promote micro assembly technology in various application fields. Micro parts, by its nature, are known of their shapes in advance for the majority of micro assembly applications. We take advantage of the previously known shapes of micro parts and hence apply a model based approach for a faster and cost effective 3D depth measure. The proposed 3D depth measuring method is based on pattern recognition and multi-focus technique enabling it to extract only information useful for micro parts assembly such as location, size, and height of a micro part. In addition, a functioning micro assembly system is developed and used to prove the usefulness of the proposed 3D depth measurement technology.

2. Assembly workcell integration In this section, we present an integration methodology of a flexible micro gear assembly workcell with magnetic grippers, a micro precision robot, and a model based 3D depth measure system. A complete feedback loop between visual sensing and positioning/grasping of a micro part is the enabling technology of a flexible micro assembly workcell. As the name implies, micro parts can be loaded in the workcell randomly, and the vision system and the grippers will find their way to locate and grasp the micro parts. In our case, we use the micro gears and latches fabricated by bulk micro-machining for demonstration.

The base station that holds the wafer has two-axis robotic module for x-y motion (Figure 3). The third axis is assembled in a tilted angle for grasping and up-down motion in one assembly. The magnetic grippers open and close the tip to grip and release a micro gear down to the size of 200 Îźm. Filted design of the 3rd axis in conjunction with the lengthy grippers allows not only precision pick and place operations, but maximizes the limited work space under the microscope. 2. Model based location and 3D depth measurement system The key component to close the loop of the flexible micro assembly is the visual feedback system, with the micro 3D visualization capability. 3D measuring of micro gears is accomplished through four consecutive steps. The first stage involves the extraction of the contours so that the candidate regions of gears can be addressed. The second process entails sorting out the actual gear region among the candidate regions by comparing the extracted contours with the known information of gears. After this process, the contour that conforms the known gear shape remains, and all other contours are ignored. In the third stage, the height of each contour, which corresponds to an actual gear region, is measured with 10 different focused images. Finally, the simplified height map is developed by associating the gear regions with the obtained height information. Details of each process are discussed in the following sections. A. Contour detection and gear recognition Although the pixel-based recognition method is dominant in object recognition technology, we use the contour-based vector tracing approach for localization and 3D depth measuring of micro parts. With the pixelbased approach, an error of one pixel does not significantly affect the final result of the height map. On the other hand, one missed contour in the contour based approach causes a fatal error, though the result is signifiArticles

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

cantly faster. Thus, the contouring method is the most important process in this step, and thus, it was consequently executed with three different binary images in an effort to obtain more reliable results. First, the contour is extracted from a binary image. The focus of the image does not significantly affect the binary image, but the binary process is sensitive to the light condition. The micro assembly process is, therefore, performed in a well-controlled environment, and all light conditions are set as predefined adequate values. For even distribution of light intensity and to minimize the disturbance by ambient lights, infrared camera, filter, and fiber optic cables are used for the vision assembly (See Figure 4).

N° 3

2010

The magnitude and direction of this vector are denoted as (2)

(3) The partial derivatives in the above equations denote the direction and rate of change of the grayscale of each pixel. There are many ways to quantify the gradient values, such as Robel, Prewitt [8], Sobel, Laplacian, etc. The Laplacian operator is commonly used as a second order operator, represented by (5) In order to prevent the change in x, y directions from cancelling each other, Nayar and Nakagawa [9] proposed a modified version of the Laplacian operator with a discrete approximation as

(6) They proposed the sum-modified-Laplacian (SML) [10], as a focus measure at a point (x,y). It was denoted as for

Fig. 4. Vision system and robotic platform assembly. By analyzing the center of the recognized gear in planar coordinates, the wafer holding station moves the wafer to the exact location of the geometric center so that the gripper properly approaches the corresponding gear. The radius of the detected gear is calculated from the contour, and the gripper opens the proper amount to pick up the appropriate gear. However, the height of the gear still remains unknown. Without the height information, the gripper could be stuck by crashing into the wafer, or the gripper could miss the gear. B. Simplified height map measurement method using Sum-Modified-Laplacian Conformal mapping technique is based on measuring the focused and defocused area to create a 3D depth image. That is, due to the limited depth-of-focus of optical lenses, an object located out of the focal plane cannot be seen clearly in the image [6]. This implies measuring blurriness 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 brightness function, I(x,y), its gradient at position (x,y) can be represented by a vector: (1) 12

Articles

(7)

Where N is the measuring window size around the point (x,y). The parameter T is a distinct threshold value. In this research, the SML value of each pixel in a contour is calculated for height, and one representative height value for each contour region is defined by averaging SML values in that region. Although SML method works well with the micro parts, we experienced much delays on the 3D depth measure of the fabricated gears, which, in fact, becomes the bottleneck toward the near realtime assembly operation [11]. In order to overcome the latency of the SML method, we propose a model based position and height measure technique, namely “Simplified Height Map (SHM)” measurement method to facilitate the gear identification process using the object’s geometry information. In this method, we assume that the height of each gear are known a priori and stored in the database. In addition, for demonstration purpose, we used gears with the equivalent height but without excluding overlapping possibilities. To measure the height of each region, 10 pictures with different focal planes are taken from the same view. Figures 5 (a)-(d) show four examples of different focal planes. Finally, a simplified height map is composed of the recognized gear regions with several levels of height, as shown in Figure 6. Figure 6a) shows the extracted contours from an actual microscopic image in 3D format. The detected contours are compared with the known information of the gears. The size and shape of the gears are well-defined during the fabrication process, and the information of gears used in assembly is stored in a database before the

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

assembly process. Contours which don’t appear to be gears are ignored, and only matched contours with the database remain as regions of interest for height measurement. Figure 6b) shows the recognized gears. (a) focus on wafer

(b) focus on lower gear

(d) focus on above gears

N° 3

2010

10–100 micrometer for precision operations. Therefore the arms of the lever would magnify the deflection up to 5 times to grip various sizes of micro-gears.

Fig. 7. Magnetic gripper assembly.

Fig. 5. Height measure using conformal mapping technique. (a) without model

(b) with model

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 reliability problems over time as shown in Figure 8a). The amount of displacement reduces significantly, thus frequent replacement was inevitable. In addition, unexpected secondary displacement occurred. For instance, when the tip of the EAP strip begins from rest with no potential applied at the starting position marker, the EAP displaces to the primary displacement with the potential applied (Figure 8b)). (a)

Fig. 6. SHM with a model based gear recognition. Using the extracted gear location and height information by SHM method, our robotic grippers in conjunction with the x-y precision platform demonstrated successful pick and place operations, though current success ratio of the pick and place is fairly low to prove the usefulness of the proposed technology (See Table 1). With the tilted angle design, we achieved up to 0.5 μm z-axis accuracy of the gripping motion. The accuracy of close and open operation of the grippers, however, is not consistent overtime. In addition, although the height measure of the micro object is near realtime, processing the assembly of the micro gears on the axial shaft is yet performed with no realtime visual feedback, reducing the assembly success ratio to some extent. 3. Precision grippers A precision pick and place grippers is designed and rapid prototyped as shown in Figure 7. The lengthy tweezer-type manipulator serves two purposes. First, it enables precision manipulation in the 1.5 cm of clearance between the assembly wafer and the microscope lens and lighting assembly. Secondly, the actuator opens and closes the grippers with the clearance on the order of

(b)

Fig. 8. EAP (Electro Active Polymer) actuator performance analysis. While the potential remains on and constant, the EAP then begins secondary displacement in the opposite diArticles

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

rection of the primary displacement (in the direction of the negative terminal). As a result, the actuator is changed to an electromagnetic driver for pick and place operations. Piezo-electric grippers demonstarted consistent performance with better controllability in grasp control.

N째 3

2010

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

3. Assembly experiments The complete system has been implemented with a precision 3 axis robot, magnetic grippers and the 3D micro vision system (see Figure 9a)). For fully autonomous micro assembly operation, integration software has been developed to close the assembly loop between the vision sensor and the assembly grippers (see Figure 9b)). (a)

Fig. 11. Magnetic grippers holing a SiO2 gear.

(b) Fig. 12. Gears on the assembly workcell. Table 1. Assembly time & root cause analysis.

Fig. 9. Assembly workcell (a) & integration software (b). A complete assembly operation starts with recognizing randomly placed micro gears in the first quadrant of the assembly wafer followed by placing them at the corresponding location at the 4th quadrant. A complete assembly cycle includes gear location identification, gear height measure, robotic gripper control for pick and place operations (Figure 10). One example of the gear grasping via the magnetic grippers is shown in Figure 11. Figure 12

Fig. 10. Assembly sequence. 14

Articles

Search for gears (x,y) Identify gear height (z) Grip gear Move gear Place gear

Aveg. (min)

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

2.5 1.4 0.75 0.5 1.2

56 25 12 15 18

91 N/A 78 97 45

In Table 1, the results of assembly time and root cause analysis for assembly performance study are shared. The success ratio of the gear identification for x,y,z coordinates was not being analyzed due to the absence of the exact location information of each gear. The SHM method introduced in the paper improved the gear height identification speed up to 4 times faster than the conventional

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

SML method. As shown in the table, gear height identification stage is no longer the bottleneck of the assembly process due to the improved process speed by the SHM method in time analysis. In terms of success ratio, the gear placement stage is the root cause of the low overall success ratio. The overall assembly success ratio was about 77% and has yet to be improved for usefulness in industry. In order to improve the success ratio, realtime visual feedback has to be implemented for pick and place operations.

[1]

[2]

[3]

[4]

[5] [6]

[7]

[8]

[9]

[10]

[11]

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

2010

References

4. Conclusion The technical challenge tackled in this research was to develop an autonomous and flexible micro parts assembly workcell. Among several components that constitute a flexible assembly workcell, the localization and 3D depth measure of micro parts were the most challenging tasks. To provide a low-cost and fast gear recognition method, we proposed the simplified height map generation for visualizing micro parts in 3D. The proposed method was designed to utilize the known models of micro parts such as size and shape, thus model based, so that it increases the reliability of the object identification and localization in near realtime fashion. Precision robotic platform and magnetic gripper have been put together with the micro 3D depth measure system. Finally, micro gear assembly experiments were performed to prove the usefulness of the proposed system. The overall assembly success ratio was about 77% and has yet to be improved for usefulness in industry. The bottleneck process is turned out to be the gear identification stage in time analysis. However, the gear placement stage is the root cause of the low overall success ratio. In order to improve the success ratio, realtime visual feedback has to be implemented for pick and place operations.

N° 3

Woods D., “The fabrication of silicon microsystems”, Engineering Science and Education Journal, vol. 9 (129), 2000, pp. 129-136. Fatikow S., Seyfried J., Fahlbusch S., Buerkle A., Schmoeckel F., “A Flexible Microrobot-Based Microassembly Station”, Journal of Intelligent and Robotic Systems, vol. 27, no. 1-2, Jan. 2000, pp. 135-169. Aoyama H., Fuchiwaki O., “Flexible Micro-Processing by Multiple Micro Robots in SEM”, In: Proc. of IEEE International Conference on Robotics & Automation, Seoul, Korea, May 2001, pp.3429-3434. McGregor M. T., Mahlke H. A., Dozier S.M., Asiabanpour B., Um D., “Producing micro scale silicon dioxide gears by bulk micro machining process” , Transactions of the NAMRI/SME, vol. 37, 2009. Madou M.J., Fundamentals of microfabrication: the science of miniaturization. 2. CRC Press, 2002. Kim T., Kim T., Lee S., Gweon D., “Optimum conditions for high-quality 3D reconstruction in confocal scanning microscopy”. In: Proc. SPIE, vol. 6090, Feb. 2006. Zlotnik A., Ben-Yaish S., Zalevsky Z., “Extending the depth of focus for enhanced three-dimensional imaging and profilometry: an overview”, Applied Optics, vol. 48, no. 34, Oct. 2009, pp. 105-112. Prewitt J.M., Object enhancement and extraction in Picture Processing and Psychopictoris, Edited by: Lipkin B.S., Rosenfeld A. New York: New York: Academic, 1970, pp. 75-149. Nayar S.K., Nakagawa Y., “Shape from focus”, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 16, no. 8, Aug. 1994, pp. 824-831. Zhao H., Lia Q., Fenga H., “Multi-focus color image fusion in the HSI space using the sum-modifiedlaplacian and a coarse edge map”, Image and Vision Computing, vol. 26, no. 9, Sep. 2008, pp. 1285-1295. Um D., Asiabanpour B., Jimenez J., “A Flexible Micro Manufacturing System for Micro Parts Assembly via Micro Visual Sensing and EAP based Grasping”, Journal of Advanced Manufacturing System, vol. 8, no. 2, Dec. 2009.

AUTHORS Dugan Um*, Dongseok Ryu - Texas A&M University – Corpus Christi, Science and Technology Suite 222, 6300 Ocean Drive, Unit 5797, Corpus Christi, Texas 78412. Tel. 361-825-5849, Fax 361-825-3056. E-mail: Dugan.Um@tamucc.edu. Bo Dong - The College of Williams & Mary, Williamsburg, VA. Dave Foor - Austin Community College, Austin, TX. Kenneth Hawkins - San Marcos High School, San Marcos, TX. * Corresponding author

Articles

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

NÂ° 3

2010

INTELLIGENT PI CONTROLLER AND ITS APPLICATION TO DISSOLVED OXYGEN TRACKING PROBLEM Received 23rd September 2009; accepted 27th April 2010.

Tomasz Zubowicz, MieczysĹ&#x201A;aw A. Brdys, Robert Piotrowski

Abstract: The paper addresses design, calibration, implementation and simulation of the intelligent PI controller used for dissolved oxygen (DO) tracking at wastewater treatment plant (WWTP). The calibration process presented in this paper utilizes both engineering and scientific methods. Verification of the control system design method was obtained via simulation experiments. Keywords: aeration process, artificial intelligence, control systems, dissolved oxygen, tracking problem, fuzzy logic controller, genetic algorithms, intelligent control, Takagi Sugeno Kang method, TSK, soft switching, wastewater treatment.

1. Introduction Control algorithms for the WWTP have been investigated intensively, particularly for DO control. The DO dynamics is nonlinear and of high dimension. Dissolved oxygen is the most important control parameter for biological processes in WWTP. Increasing performance of the DO control system is a basic action leading to improve the effectiveness and efficiency of the regarded system. It becomes more interesting when it can be done without interference into the system structure and only by simple manipulation of the control algorithm. The DO tracking problem is one of the most complex and still fundamental issue of biological WWTP involving activated sludge technology due to its influence on dynamics of biochemical processes. Its complexity derives from strong no linearity and time dependence of the process variables. In this paper a very effective controller design approach is presented, which uses both classical and artificial intelligence methods to find a solution dissolved oxygentracking problem. The paper is organized as follows. The problem statement is described in Section 2. Section 3 presents the PI controller design. The controller calibration process is described in Section 4. The engineering and scientific

Fig. 1. Scheme of the biological WWTP. 16

Articles

methods are used. Application of the controller to Kartuzy WWTP is described in Section 5 and the simulation results are presented. Section 6 concludes the paper.

2. Problem statement The previous papers [Brdys, et. al., 2002; Piotrowski, et. al., 2004; Piotrowski, et. al., 2008] propose a two level controller to track prescribed dissolved oxygen trajectory. The upper level control unit prescribes trajectories of desired airflows to be delivered into the aerobic biological reactor zones. The lower level controller forces the aeration system to follow these set point trajectories. A non-linear model predictive control algorithm is applied to design this controller unit. The goal of this paper is to provide the design methodology of intelligent multiregional controller [Domanski, et al., 1999], with enhanced regional controllers, to enable highly nonlinear system, in this case aeration system, to work in a whole control space with the same high quality performance under heavy disturbances. It also shows how to utilize optimization method to calibrate the resulting controller parameters. Activated sludge wastewater treatment plant (ASWWTP) is a very complex and demanding structure for closed-loop control due to its internal processes nonlinearities, inconsistency, and time and parameter variability. The scheme of the biological WWTP is presented in Fig. 1. The advanced biological processes with nutrient removal is accomplished in the activated sludge reactor designed and operated according the University of Cape Town (UCT) process. The first zone where the phosphorus is released is anaerobic. The second zone where the denitrification process is conducted is anoxic. The internal recirculation 2 of mixed liquor originates from the anoxic zone. The returned activated sludge from the bottom of the clarifiers and the internal recirculation 1 from the end of the aerobic zone (containing nitrates) are directed to the anoxic zone.

Journal of Automation, Mobile Robotics & Intelligent Systems

The last part of the reactor (aerobic) is aerated by a diffused aeration system. This zone is divided into four compartments of various intensity of aeration. The biologically treated wastewater and biomass (activated sludge) are separated in two parallel horizontal (rectangular) secondary clarifiers. In order to ensure a high level of phosphorus removal, iron sulphate (PIX) is added to the aerobic zone to precipitate most of the remaining soluble phosphorus (simultaneous precipitation). There is also the opportunity to precipitate phosphorus in the grit chamber (pre-precipitation). The excess biological sludge is stored in a thickener, then dewatered in two centrifuges and finally chemically stabilized in lime. It is expected that the sludge will be disposed of for agricultural applications. For the purpose of the research the activated sludge model type ASM2d [Henze, et al., 1995] was used. This selection was a good compromise between accuracy and speed in predicting dynamics of the biochemical processes [Brdys, et al., 2004]. The ASM2d type model derives from the activated sludge model family (ASM), which also consists of the types ASM1, ASM2, ASM3, ASM3 Bio-P. The differences between each model are generally situated in their mathematical representation accuracy of biochemical processes. The ASM family evolved according to the stated order with ASM2d model placed in between ASM2 and ASM3. The WWTP model was designed using commercial package Simba [Simba, 2005]. After the quality and quantity parameter identification the model calibration process was applied. For the purpose of calibration the genetic algorithms (GA) were used. Final step, the validation process verified and confirmed the usefulness of the constructed model.

3. Controller design This section addresses directly the DO tracking problem. In [Yoo, et al., 2002] for the purpose of solving the DO tracking problem the PI controller with on-line parameter adaptation was proposed. First of the mechanisms was invoked to enable in the controller ability to reject disturbances and the second to carry out control of the nonlinear process with the same performance in whole control area. In this paper the more effective (regarding both controller design methodology and computational requirements) method was proposed. For the purpose of disturbance rejection the classical PI regulator was applied regionally. In order to enhance the regional fixed parameter controllers to work as a one in whole control space the control signal switching system was utilized. Due to decreasing of performance during the hard switching techniques, the artificial intelligence (AI)Takagi-Sugeno-Kangs (TSK) soft switching method is used (Fig. 2). This also has allowed to eliminate troubling parameter adaptation procedure presented in [Yoo, et al., 2002]. For the purpose of the design process the following assumptions have been made: the aeration system is regarded as ideal thus the generated and reference airflows fulfill the expression:

VOLUME 4,

N° 3

2010

ref Qair (k ) º Qair (k) ; plant being under control is a single aeration chamber.

Fig. 2. Scheme of the control system. The intelligent controller (Fig. 2) design process can be easily divided into four main stages: stage I gathering information; stage II region indication; stage III local controller design; stage IV fuzzy logic applying; stage V calibration (optional). Stage I: The first step that needs to be taken is to represent knowledge of the process in the form of static characteristic of DO as a function of airflow Qair, generated by the aeration system. As mentioned before (see Section 2) the considered process is under heavy influence of time varying disturbances: wastewater inflow (Q), chemical oxygen demand (COD), total nitrogen (TN) and total phosphorus (TP). The model based sensitivity analysis per-formed to acquire the steady state characteristic of the process resulted in obtaining the whole family of DO(Qair) curves. In Fig. 3, the mean DO(Qair) characteristic (determined via the studies) and its boundaries have been shown. Stage II: As determined in Stage I the relationship between DO and Qair is highly nonlinear (Fig 3). In the second stage the mean steady state characteristic is piecewise linearized (Fig. 4) for the purpose of nonlinearity approximation thus divided into affine, linear spaces (region indication: li , for i = 1,..,5). This method has many advantages such as: simplicity, ability to indicate linearly representative regions of process, associated error and above all in some cases can be done graphically (mainly while regarding SISO systems). This last feature enables the designer to decide if the balance between accuracy and complexity is maintained at proper level. Piecewise linearization also can be done via an optimization algorithm with properly chosen cost function, but the first approach represents, in authors opinion, a good balance between simplicity and effectiveness and thus recommended for this and similar cases.

Articles

17 1

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 3

2010

Incorporating stated nonlinearities into PI model results in many advantages. The magnitude and rate limiter enables controller to generate control signal meeting the actuator system static and dynamic limitations. Both influence the actuators lifespan and also time needed between maintenance by reducing stress generated in the system. The negative influence of saturation states entailed mostly by the magnitude limiter is reduced by the anti wind-up filter [Bohn and Atherton, 1995] utilized in the controller. The PI controller (Fig. 5) can be described as: u(k) = g(k)(0.5 × sat(up(k))+0.5) p

up(k) = 0.5 × å sat(Duu(i)) Fig. 3. Static characteristic of the aeration process and its boundaries.

i=1

uu(k) = Kp e(k)+Kp Ki å ui(j) j=1

(1)

ui(k) = e(k) - Kaw(up(k) - 0.5 × sat(up(k)) - 0.5) where: e(k) - error; u(k) - control signal; ui(k) - signal in summation line; upre(k) - signal before magnitude limitation; Kp - proportional gain; Ki - summation gain [1/s]; Kaw - anti wind-up filter gain; g - output scaling factor. In the system (1) the error signal e(k) is defined as: e(k) =

1 (DOref (k) - DO(k)) DOmax

(2)

where: DO(k) - measured dissolved oxygen value; DOref (k) - reference value of DO; DOmax - input scaling factor. Fig. 4. Mean and piecewise linear static characteristic of the aeration process. Stage III: Indicating the regions with lack or small enough error, which can be treated as linear, allows looking for also linear regional controller. PI controller is used for a local (regional) controller (Fig. 5). Additionally magnitude and rate limiter of the generated control signal have been added along with the anti wind-up filter controlling the performance of the magnitude limiter. As a result the nonlinear PI controller is used.

Fig. 5. Local discrete nonlinear PI controller.

Articles

The saturation function is as follows: sat(.) = sign(.)min{ï.ï, 1}

(3)

The calibration of the local controllers parameters can simply be done using well-known engineering approach both via on plant experiments or model-based simulations. Parameters for the five local PI controllers (see Fig. 5) chosen arbitrary to balance the performance in each region have been presented in Table 1.

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 3

2010

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

Controller I

Controller II

Controller III

Controller IV

Controller V

Region

Ts [s] Ki, i [1/s] Kp, i Kaw, i

30 0.04 16.83 9.67

30 0.02 10.54 9.6

30 0.01 6.44 9.79

30 0.03 14.98 15.88

30 0.02 11.29 16.22

Stage IV In this step the TSK fuzzy logic mechanism (4) is being 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:

Resulting fuzzy partition is presented in Fig. 6.

å w (DO(k)) × Q i

ref air

Q (k) =

the value of membership of i-th MF should be equal to 0 when it reaches the neighboring MFs core; created fuzzy partition should be normal and consistent.

ref air, i

(k)

i=1

å w (DO(k)) i

(4)

i=1

ref

where: Qair (k) - multiregional controller output; ref Qair, i (k) - output value of i-th regional controller; p number of regions in this case p = 5. The weights wi(DO) is on-line tuned according to the process state, which in this case is denoted by the DO concentration. The possible weight values are fuzzy sets of a priori defined membership functions (MFs). Each MF is directly correlated with corresponding regional controller; hence the number of regions determines the number of MFs. In this case the Sigmoidal and Gauss conditional functions were utilized [Yaochu, 2002]. Defining the MF simply means to chose its type and shape. It should be noted that the choice of the MFs is crucial for acquiring stability and good performance of the closed loop system, however it can be done while having basic knowledge of the process. It is also useful to know the shape of the assessed process static characteristic, which in this case is already achieved (see Stage I). The procedure for choosing MFs is as follows: each regional controller should be given a corresponding MF; the core of i-th MF should overlap with the part of i-th region that possesses relative error of approximation less then 10% (which is an arbitrary chosen value);

Fig. 6. Initial MFs. The formulas for each MF and parameters in the numerical form have been presented in Table 2. Stage V Calibration of the MFs parameters is the final stage of the controller design process. This stage can be omitted with no loss of functionality of the designed control system, however it certainly improves the overall system performance. As a tool for that purpose the GA was chosen and applied. The whole calibration process is described in details in Section 4 due to its rather extensive character.

Table 2. MFs and their parameters. No

Region

Function name

Sigmoidal

Gauss

Formula mA(DO(k)) =

Parameters

1 1+exp(-a(DO(k) - c) exp

mA(DO(k)) = 1; exp

a = -10 c = 1.483

-(DO(k) - c1)2 s12

; left - most curve

-(DO(k) - c2)2 s22

; right - most curve

whenever c1 < c2

sig1 = 0.422 c1 = 1.98 sig2 = 0.145 c2 = 2.85

Articles

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 3

2010

Table 2. MFs and their parameters. No

Region

Function name

Formula

Gauss

sig1 = 0.1146 c1 = 3.155 sig2 = 0.604 c2 = 4.78

Gauss

sig1 = 0.671 c1 = 6.231 sig2 = 0.612 c2 = 6.84

Sigmoidal

mA(DO(k)) =

1 1+exp(-a(DO(k) - c)

4. MFs parameters calibration 4.1. Utilizing GA In order to calibrate parameters of the MF, the GA was used due to their large number, task computational difficulty and lack of knowledge about system model derivatives. The main features of the utilized algorithm are as follows [O'Reilly, et al., 2002]: genes in a form of real numbers (16 genes see Table 2); initial population contains one fixed chromosome with genes equal to the parameters presented in Table 2; algorithm uses operands: mating (crossing and blending) and mutation both with gene value and duplicity control; elitism; generation counter as algorithm stop condition. The cost function used to evaluate each chromosome is given by the formula: JS = eTEe+M%T MM% +DM%T DMDM% + tRTTR tR + + DtTRDTR DtR + uTUu + DuTDUDu + awTAaw + + duTDdu + nTNn + cmTCcm

(5) Fig. 7. GA flow chart.

where: JS - survival cost; eTEe - steady state error; M%T M% - percentage overshoot; DM%T DMDM% - difference between overshoots within each region; tTRTR tR settling time with respect to 5% criterion; DtTRDTR DtR difference between settling times within each region; uTUu - control signal energy; DuTDUDu - difference between regional control signals; awTAaw - saturation states; duTDdu - oscillation in control signal; nTNn - normality of fuzzy partitioning; cmTCcm - consistency of fuzzy partitioning. The algorithm flow chart is presented in Fig. 7, while the convergence of the GA over the generation is shown in Fig. 8.

Fig. 8. GA convergence. 20

Articles

Parameters

a = 5.38 c = 7.514

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

NÂ° 3

2010

Table 3. Calibrated MFs parameters. Region

MF1

MF2

MF3

MF4

MF5

sig1 = 0.371 c1 = 1.868 sig2 = 0.104 c2 = 2.844

sig1 = 0.079 c1 = 3.149 sig2 = 0.427 c2 = 5.049

sig1 = 0.412 c1 = 6.311 sig2 = 0.440 c2 = 8.232

Parameters

a = -10.932 c = 1.2052

Table 3 shows the last generation (100th) of the population, sorted in ascending order with regard to the cost. The graphical representation of the functions is shown in Figure 9.

a = 6.578 a = 6.578

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

5. Simulation results In this section the derived controller is validated by simulation, based on data recorded from Kartuzy WWTP. The simulation conditions assume highly varying disturbances: Q, COD and TN, which are presented in the same order in Figs 10-12. The performance, error and control signal trajectories of the closed loop system before and after the calibration process have been shown in Figs 13-18a).

Fig. 9. MFs with calibrated parameters. 4.2. Calibration summary The calibration process exploiting GA (in example [5]) makes this uneasy task achievable (see Fig. 8). In order to clearly show the differences in performance of MFs with and without calibration process a standard root mean square (RMS) criterion, given by (6) was used: RMS =

1 Ă&#x2014;J N

(6)

where N is the number of samples and J is defined as follows (7): J = eTEe+M%T MM% +DM%T DMDM% + tRTTR tR + + DtTRDTR DtR + uTUu + DuTDUDu + awTAaw

(7)

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

Figs 13-18 indicate that the designed controller enables the system to realize the reference trajectory under the heavy time varying disturbances (Figs 10-12) keeping good performance over the whole operating range of the plant. It can be seen (Figs 15, 15a, 18, 18a) that both the magnitudes and rates of the demand control signal can be accommodated by the plant actuator system. The control signal generated by not calibrated and calibrated systems are illustrated in Figs 15, 15a and 18, 18a, respectively. It can be seen that the control signals are equally demanding in terms of magnitudes and rates, however as it has been already stated the calibrated system achieves much better tracking error (see Table 4). The simulation was carried out under worst-case plant disturbance scenario (see Figs. 10 to 18). As sample, the result of a daily performance, currently achieved by the control system, at the plant site, is illustrated in Fig. 19. A significant improvement of operating performance by the proposed controller can be clearly seen.

Table 4. RMS and performance enhancement. No

MF state

RMS

Percentage performance enhancement

before calibration

0,0218

after calibration

0,0173

20,51%

Articles

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N째 3

2010

Fig. 10. Inflow to WWTP.

Fig. 11. Chemical oxygen demand.

Fig. 12. Total nitrogen.

Fig. 13. Closed loop system performance before calibration.

Fig. 14. Control error before calibration.

Fig. 15. Controller output signal before calibration.

Articles

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N째 3

Fig. 15a. Zoom in of controller output - signal before calibration.

Fig. 16. System performance after calibration.

Fig. 17. Control error after calibration.

Fig. 18. Controller output signal after calibration.

2010

5. Conclusions

Fig. 18a. Zoom in of controller output - signal after calibration.

The paper has addressed design, calibration, implementation and simulation of the intelligent PI controller used for DO tracking in WWTP. A classical PI controller has been used to derive multiregional intelligent controller. The controller is capable of maintaining globally well-known attractive local properties of the PI controller, when applied to nonlinear processes. An advantageous property of the proposed controller design methodology is that it can be applied without changing the on plant (commonly used) hardware. Furthermore it allows obtaining great enhancement in the system performance via simple algorithm change, which is a low cost solution. Enhancements are possible by calibrating MFs parameters. The controller has been validated by simulation based on real data recorded from the Kartuzy WWTP and excellent results have been obtained, however a rigorous analysis of the closed loop stability is under current research.

Articles

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 3

2010

Fig. 19. Daily performance of the control systemat the plant site. AUTHORS Tomasz Zubowicz* - Department of Electrical Engineering and Automation, Gdansk University of Technology, 80-233 Gdansk, Poland. E-mail: tomasz.zubowicz@gmail.com. Mieczysław A. Brdys - Gdansk University of Technology, Department of Electrical Engineering and Automation, 80-233 Gdansk, Poland. E-mail: m.brdys@ely.pg.gda.pl. Department of Electronic, Electrical and Computer Engineering, The University of Birmingham, Birmingham B15 2TT, UK. E-mail: m.brdys@bham.ac.uk. Robert Piotrowski - Department of Electrical Engineering and Automation, Gdansk University of Technology, 80-233 Gdansk, Poland. E-mail: r.piotrowski@ely.pg.gda.pl. * Corresponding author

[7]

[8]

[9]

[10] [11]

[12]

References [1]

[2]

[3]

[4]

[5] [6]

Bohn C., Atherton D.P., An analysis package comparing PID anti - windup strategies. Dept. of Electr. Eng., RuhrUniv., Bochum, 1995. Brdys M.A., Chang T., Konarczak K., “Estimation of wastewater treatment plant state for model predictive control of N-P removal at medium time scale”. In: IFAC 10th Symposium Large Scale Systems: Theory and Applications, Osaka, 26th-28th July 2004. (invite session). Brdys M.A., Chotkowski W., Duzinkiewicz K., Konarczak K., Piotrowski R., “Two-level dissolved oxygen control for activated sludge processes”. In: 15th IFAC World Congress, Barcelona, 21st-26th July 2002. Domanski P., Brdys M.A., Tatjewski P., “Design and stability of fuzzy logic multi - regional output controllers”, Int. Appl. Math. And Comp. Sci., vol. 9, no. 4, 1999, pp. 883-897. Haupt R.L., Haupt S.E., Practical Genetic Algorithms. Second Edition. Wiley-Interscience, New Jersey, 2004. Henze M., Gujer W., Mino W., Matsuo T., Wentzel M.C., Marais G.v.R., Activated Sludge Model No. 2. Scientific and Technical Report No. 3, IAWQ, London, 1995.

Articles

[13]

O'Reilly U-M., Yu T., Riolo R., Worze B., Genetic Programing Theory and Practice II. Springer Science, Boston, 2005. Piotrowski R., Duzinkiewicz K., Brdys M.A, “Dissolved oxygen tracking and control of blowers at fast time scale”. In: IFAC 10th Symposium Large Scale Systems: Theory and Applications, Osaka, 26th-28th July 2004. Piotrowski R., Brdys M.A., Konarczak K., Duzinkiewicz K., Chotkowski W., “Hierarchical dissolved oxygen control for activated sludge processes”, Control Engineering Practice, vol. 16, issue 1, 2008, pp. 114-131. Simba, User's guide, 2005. http://simba.ifak.eu/simba. Tanaka K., Sugeno M., “Stability analysis and designer of fuzzy control systems”, Fuzzy Sets Syst., vol. 45, 1992, pp. 135-166. Yaochu Jin, Advanced Fuzzy Systems Design and Applications. Physica - Verlag. Springer - Verlag Company, 2002. Yoo C.K., Lee H.K., Beum Lee I., “Comparison of process identification methods and supervisory control in the full scale wastewater treatment plant”. In: 15th IFAC World Congress, Barcelona, 21st-26th July 2002.

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 3

2010

FISH-LIKE SWIMMING PROTOTYPE OF MOBILE UNDERWATER ROBOT Received 12th May 2010; accepted 20th May 2010.

Marcin Malec, Marcin Morawski, Jerzy Zając

Abstract: In this paper, authors present a new approach to the design of a mobile underwater robot inspired by a fish. They describe a prototype of a self-designed and a self-made mobile underwater robot called the CyberFish, which resembles 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 are described by means of comparison of robots parts to fish organs. In the next section authors focus on electronic control system as well as on applications written in C/C++ that are used to control the robot in three different modes. Keywords: robotic fish, underwater mobile robot, biological inspiration.

1. Introduction The 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 more natural way of moving underwater seems to be worth taking into consideration while designing a robot which is going to be work in the water environment. There are more and more surveys on such kind of propulsion, carried out in many scientific centres around the world. H. Kim and et al. [6] studied the motion mechanism of real fishes and proposed the dynamic Lagrange's equations of a fish robot modelled as a four-link system. A similar approach is presented by L. Zang and et al. in [12], however the team focused on developing efficient diving mechanism, which uses pectoral fins and fuzzy logic controller. The main propulsion is also implemented as 4-section tail, driven by four servomotors. Above two papers refers to so called carangiform swimming whereas there are also many surveys on different forms (anguilliform, rajiform, gymnotiform etc.) of fish movement. Some of them are presented by K.H. Low in [8]. Author however considered a gymnotiform robot, which mimic a Black Ghost Knifefish. He presented several similar solutions as well as his own prototype of such a robot. The mathematical model of such propulsion is also described in the paper. The main goals of the project described in this paper were to develop a concept, design and build prototype of a carangiform robotic fish equipped with proximity sensors, temperature sensor, wireless digital miniature video camera and wireless communication system. The maximal representation of a fish-like movement was, however, the most important priority for the authors. There were several assumptions to the project. Firstly, the construction

should be cheap and easy to be built without using sophisticated materials and tools. Secondly, parts used to build CyberFish should be easy available. Thirdly, the robot should be able to operate autonomously as well as being controlled via computer. Taking into consideration these assumptions authors, in cooperation with colleague Dominik Wojtas, have studied fish movement and have tried to find ways of translating and transforming it into a mechanical device. Based on the findings and conclusions of the study, the 3D CAD model of CyberFish was created in Catia v5 system. The results of the computer simulation confirmed that the kinematics of the model was correct. The underwater robot was capable to swim like a fish. Therefore, the physical prototype has been built. The next step of the project was robot testing in the two thousand - litre tank. Tests showed that the proposed concept is correct and CyberFish swam like a real fish. Nevertheless, it was still much work to do. The main tasks concerned with developing electronic control system and software based on an appropriate control algorithms, which would give the robot the ability to be operated via computer or swim autonomously.

2. Biological inspiration The concept itself appeared during studies of Automatics and Robotics at the Faculty of Mechanical Engineering of Cracow University of Technology. Authors were interested in making original bionic robot different from existing rolling robots created in the likeness of arachnids or crustaceans. The real challenge was to design and build underwater mobile robot. The first thought was to create a fish-like device. In that case the robot must not be driven by propeller but only by means of undulating movement of its “body”. 2.1. Fish-like swimming It is obvious that swimming is the most convenient way of moving underwater. In general, there are two types of fish swimming methods: BCF (body and/or caudal fin propulsion) and MPF (median and/or paired fin undulations) [3]. In the paper authors focused on BCF-like motion. The many kinds of fishes swim by means of wavy movement of theirs body and/or tail. The frequency and amplitude of those vibrations depend on the species (Fig. 1). Anyway, the force that pushes fish forward is a result of consecutive muscle contractions. When fish is swimming, water is pushed sideways and backwards. Forces, which act sideways, compensate each other whereas the force which pushes water backwards gives reaction that enables fish to move forward. The majority of fish species have two types of muscles. The white muscles Articles

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 4,

N째 3

2010

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

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

Fig. 1. BCF fish motion a) mackerel, b) trout, c) eel [2]. Strong muscles themselves are insufficient to perform underwater manoeuvres. The flexible and lightweight skeleton is also very important. What is important, fish skeleton does not carry whole weight of the fish because the buoyancy force compensates to some extent the force of gravity. Thus, the skeleton can contain hundreds of small bones and cartilages, which form with muscles a very flexible construction that helps fish to swim very efficiently. Another very important fact is that the density of fish body is close to density of water. This gives the fish the ability to easy move vertically underwater by means of its swim bladder and pectoral fins. The swim bladder is a hydrostatic organ in the shape of a flexible thin-walled container, which can be filled with air thus changing buoyancy of fish. The size of the swim bladder depends on the species. Freshwater fishes have larger swim bladders than seawater fishes because of the differences in the density of fresh water and sea water [5]. In some kinds of groundfish and sharks there is no such organ [4]. Shark uses its pectoral fins to dive and emerge. Its body has higher density than water thus is unable to maintain depth. Shark uses dynamic lift of their pectoral fins so they sink when they stop swimming [10]. 2.2. Robot's design In order to build the prototype of fishlike underwater mobile robot, authors have to be acquainted with fish anatomy and its behaviour mentioned above. The first important goal was to design a mechanism which kinematics is similar to the undulating motion of a fish's body. The mechanism consists of four segments connected in series with the rotary kinematic pairs. The head - the biggest segment, two tail segments of similar size and tail-fin segment. First three of them contain drives. Next segment is driven by a servomotor placed in a previous segment. When the mechanism is in motion the proper rotation of each segment and appropriate synchronization of movement of segments create the effect similar to swimming motion of a fish. Computer simulation shows that such motion is really similar to fish motion. Figure 2 presents top view of the CyberFish 3D model in one moment of movement. 26

Articles

Fig. 3. The transmission of the second and the third segment. Each segment is driven by micro servomotor. That solution is compact, cheap, and easy to control and gives high torque in comparison to its size. The transmission from the servo shaft to the segment axis is carried out by gear with ratio 1:1 (Fig. 3). Another important feature of fish anatomy is swim bladder, which is essential for depth control. Such an artificial organ was implemented in the CyberFish. It consists of two thin-walled silicone tubes sealed at one end. The tubes can be compressed and stretched by the additional servomotor and the special linking mechanism. Such a pumping mechanism can draw water through a hole in the bottom of the housing. Artificial bladder's servo is also used to change angle of pectoral fins by means of levers and ties. This allows currying out up-and-down motion of the robotic fish just like sharks do when they swim. Diving mechanism has also got the small additional weight, which moves forward while pectoral fins move up, and moves backward in the opposite case. This changing slightly the centre of gravity of the robot and allows the CyberFish to swim like a real fish. Diving mechanism is shown in the Fig. 4. Based upon the 3D CAD model of the robot, the prototype has been built using PCV, acrylic, rubber, aluminum and stainless steel. The volume of the robot was estimated by Catia software and used to calculate buoyancy that allows the prototype to float in water rather than sink. The mass of CyberFish's body is 3.5 kg and the robot's density is slightly lower than the density of water. This solution enables the robot to change its depth using small changes in volume of the swim bladder. The CyberFish operating underwater is shown in Fig. 5.

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

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

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

3. Control system The work on the control system has been carried out simultaneously with the building of the mechanical part of the prototype. The concept assumes three control mo-

N째 3

2010

des: (1) autonomous mode, (2) manual control via computer equipped with wireless communication module and (3) tracking submersed object recognized in camera image by image recognition algorithm. In order to fulfil that assumption, the control system was divided into two parts. The low level control part is based on micro controller electronic board embedded in the CyberFish, whereas the high level control part is formed by an external computer with control software. The communication between these two parts is performed by means of exchanging messages, which are sent via wireless communication channel. The autonomous mode is fully implemented in embedded part of the control system, thus the robot need no extra device to operate underwater. In this mode, one of eighteen predetermined robot's activities is randomly selected 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. The autonomous mode can be turned on when no data is received from the computer for more than 60 seconds. If it is on and the robot receives messages, it immediately switches to the manual control. The high level part of the control system is based on a specially designed computer software which gives an operator the ability to control the robot by clicking buttons or pressing keys on the keyboard. An image recognition algorithm is also implemented in the software. It recognizes red round object in the video received from the robot's onboard wireless video camera. Based upon the coordinates of the centre of the object in the image, algorithm sends messages to the robot in order to maintain the object in the middle of the frame. 3.1. Hardware The core of the robot's electronic control board is the Atmel Atmega 32 micro controller clocked by 8 MHz crystal. The typical application of the Atmega micro controller

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

Journal of Automation, Mobile Robotics & Intelligent Systems

which can be found in [1], was enriched with: two stabilized power supplies, NE555 timer to generate proximity sensors carrier wave signal, transistor keys used to modulate proximity sensors signal, connectors used to connect the radio communication module, servo motors, DS18B20 digital thermometer, IR detectors, video camera and programming socket. The complete scheme is shown in Fig. 6. Several parts of the scheme need to be commented. First of all robots proximity sensors are built with four TSOP1736 IR detectors, which need the appropriate input signal. The signal consists of packets of 30 pulses with the frequency of 36 kHz with approximately 9 ms gap. The 36 kHz square wave is generated by NE555 timer. This signal is modulated by the appropriate signal from micro controller with the use of set of transistor keys (T1 - T8). Such a modulated signal is send by IR diode and if it reflects off the obstacle, TSOP detects it and sets a low state on its output. Two sets of stabilized power sources are used to eliminate interference caused by DC servos motors. Servos are supplied from the separate 6 V source whereas other devices are supplied from the 5 V source. Each source is supplied from Ni-MH 9.6 V 2700 mAh battery. Connectors JP12 and JP13 are used to either connect charger (pins 2 and 3) or supply the system (shorting pins 1 and 2). The miniature wireless video camera mounted in the front of the robot and connected to JP11 is supplied directly from batt1 by K1 relay. This enables the operator of the robot

Fig. 7. The robot's control application window. 28

Articles

VOLUME 4,

N째 3

2010

to turn the video camera on and off depending on the needs. The video signal from the camera is received by the computer with the use of video camera receiver and a USB TV tuner. The receiver gives a composite video signal on its output, which is then transferred, to the USB TV tuner, which works as an image-capturing device. The wireless communication module MOBOT RCRv2 type A is connected to USART port by the JP5 connector. The micro controller communicates with MOBOT by asynchronous serial transmission, which parameters are as follows: 56 kbps, 8 data bits, 1 stop bit, and no parity. Another MOBOT RCRv2 type B module is connected to PC via USB and communicates with type A module by using 433 kHz radio signal. The DS18B20 digital thermometer is connected to the micro controller by means of one-wire interface using PD2 line. The temperature sensor is located near the dorsal fin of the CyberFish. 3.2. Software The micro controller software has been written in C using the WinAVR development environment and avrgcc compiler. Setting WinAVR to work with the compiler was made with help of information presented in [7]. Compiled code was written to the device by means of STK200 serial programmer and PonyProg 2000 application. The program consists of various functions like initialisation of: timers, USART module and one-wire interface, which are called before the main control loop. When the program enters

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

the main loop, function used to analyse messages is being called. It compares incoming messages received from the USART to messages stored in memory and then the appropriate action is being performed. Moreover, message of confirmation is sent back to the computer. If no message is received from USART within the 60 seconds and the autonomy flag is set, the autonomy function is being called. The robot switches to the autonomous mode described previously in the paper. Sending and receiving messages are performed in the USART interrupt handlers. PWM signals are software generated within the timer interrupt handler. This is because Atmega32 is not equipped with a sufficient number of independent PWM channels to control four servomotors. There is a set of various functions, which are being called after receiving a message. These functions adjust the duty cycle of PWM in order to achieve appropriate servo movement and synchronization. The micro controller program also contains functions used to handle communication with the DS18B20 temperature sensor and a function to receive signals from proximity sensors. PC computer application has been written in C++ by Dominik Wojtas using the Microsoft Visual Studio 2008 development environment. OpenCV and EmguCV free computer vision libraries were used to build an image recognition application, which is further used to tracking submersed red round object by the CyberFish. The application itself is a single window form (Fig. 7), which contains several areas: communication parameters settings area, a text box used to send and receive messages (control commands), 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 pixels on which the underwater view (as well as detected object) is displayed. The image recognition algorithm was developed with the help of information contained in [11]. Images received from a capturing device are the algorithm's input data whereas commands used to control the robot are the output 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 fill gaps in the image of the object, calculating coordinates of the centre of the object in the image, coordinates analysis during fixed time intervals.

2010

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

4. Conclusions After several months of work on Master of Science Thesis in Automation and Robotics at Cracow University of Technology, the CyberFish has finally been made. The robot was designed so that it could be built using the cheapest materials and widely available tools and parts. Financial constraints did not allow for the implementation of sonar system or sophisticated video camera, which could operate at greater depths or at low light intensity. However, with the use of popular and well-known solutions, it was possible to build a unique underwater mobile robot, which has been tested in 2000 litres pool. Despite the difficulty of sealing, construction, manufacturing and logistics problems and bugs in the software, the aim of the project has eventually been achieved. The CyberFish represents an original underwater craft that can be drive without propeller. This solution seems to be more efficient or even irreplaceable if the device is going to operate in rushes or seaweed. Any further development would require funding that allows to create an underwater robot performing various functions, ranging from analysis of water pollution, ending the stand-alone water penetration in the search for missing items or people.

AUTHORS Marcin Malec*, Marcin Morawski, Jerzy Zając - Cracow University of Technology, 31-864 Kraków, Al. Jana Pawła II 37. E-mails: marcin.morawski.m@gmail.com, zajac@mech.pk.edu.pl. * Corresponding author

References [1] [2] [3] [4]

Based upon coordinates analysis of the centre of the object in the image, appropriate commands are being sent to the CyberFish in order to maintain the object in the middle of the frame. If the result of the analysis locates the object in the left side of the screen during fixed time interval, „turn left“ command is being sent. A similar situation is observed when result of the analysis locates the object in the right side of the screen. If the result of the

N° 3

[5] [6]

[7]

Baranowski R., Mikrokontrolery AVR ATmega w praktyce, Warsaw: BTC, 2005. (in Polish) Chmielewski T., Wciąż do przodu..., http://ryby.fishing.pl/ dodatek_3.php, October 2008. (in Polish) Evans D.H., The Physiology of Fishes. Second Edition, Boca Raton (Florida): CRC Press LCC, 1998, pp. 3-25. Frey H., Akwarium Słodkowodne, Warsaw: Sport i Turystyka Publ. Comp., 1990, pp. 168-174. (in Polish) Jobling M., Environmental Biology of Fishes, London: Chapman & Hill, 1995, pp. 251-297. Kim H., Lee B., Kim R., „A Study on the Motion Mechanism of Articulated Fish Robot”. In: Proceedings of the 2007 IEEE International Conference on Mechatronics and Automation, 2007, Harbin, China, pp. 485-490. Koppel R., „Programowanie procesorów w języku C”, Articles

Journal of Automation, Mobile Robotics & Intelligent Systems

[8]

[9]

[10]

[11]

[12]

Elektronika dla wszystkich, no. 5, 2005, pp. 36-39. (in Polish) Low K.H., „Modelling and parametric study of modular undulating fin rays for fish robots”, Mechanism and Machine Theory, vol. 44, 2009, pp. 615-632. Malec M., Morawski M., Wojtas D., Zając J., „CyberRyba podwodny robot mobilny”, Pomiary - Automatyka Robotyka, no 2, 2010, pp. 331-340. (in Polish) Wikipedia, the free encyclopedia, „Fish locomotion”, http://en.wikipedia.org/wiki/Fish_locomotion, April 2010. Wojnar L., Kurzydłowski K.J., Szala J., Praktyka analizy obrazu, Cracow: Polskie Towarzystwo Stereologiczne, 2002. (in Polish) Zhang L., Zhao W., Hu Y., Zhang D., Wang L., „Development and Depth Control of Biomimetic Robotic Fish”. In: Proc. of the 2007 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2007, San Diego, pp. 3560-3565.

Articles

VOLUME 4,

N° 3

2010

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 3

2010

A MOBILE SYSTEM FOR MEASUREMENTS OF PARTIAL DISCHARGES CONTROLLED BY ELECTROENCEPHALOGRAPHIC WAVES Received 20th January 2010; accepted 26th April 2010.

Andrzej Błachowicz, Szczepan Paszkiel

Abstract: The article presents the concepts of a mobile system for measurements of partial discharges controlled by brain waves. In order to describe that, a robot, which takes measurements of partial discharges, has been worked out. The discharges may occur in an isolator of electrical devices such as capacitors and transformers. What is more, a concept of a link between the robot and a human brain is described, in order to ensure a direct communication on the level of the human brain and the robot. Keywords: EEG, a mobile robot, partial discharges.

1. Introduction Present technological solutions enable us to work out more and more new concepts for completions of those problems, which used to be difficult to solve before. The development of bio cybernetics as well as computer science has created favourable conditions for supporting automation and robotics by means of computer mechanisms [1]. Even few years ago a direct control of a human brain by such devices as robots was difficult to complete in terms of technical requirements [2]. Nowadays it is possible to analyse brain waves emitted by neurons by means of electroencephalography. Then, after an appropriate classification, they may be used in the process of controlling [3]. The present article describes the concepts of the use of brain waves collected by an electroencephalograph in order to control a mobile robot, which would help an automatization of the process of a measurement of partial discharges.

elastic strains on the molecular level which lead to the emission of a sound wave, light flashes, which emit radiation on the level of a visible, an infrared and an ultraviolet spectrum, local implosions, which cause a rise in temperature and changes in a pressure of gas.

The occurrence of specific factors is directly connected with the complexity of energetic devices, which contain insulators. For that reason, methods of detection of partial discharges are suggested that are based on the detection about the level of the occurrence of a particular physiochemical phenomenon [7]. Nowadays the first place in diagnostics occupies non-invasive methods. In terms of partial discharges, the most important are methods 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 the system for the measurement of partial discharges by means of the acoustic emission’s method.

2. Partial discharges Partial discharges (PD) are those phenomena, which occur in isolators of such electrical devices as transformers, energetic capacitors, electric motors, etc. This phenomenon is unfavourable in terms of the description of an insulator, and an increasing frequency of the occurrence of partial discharges means its considerable degradation. The occurrence of different physical mechanisms, which accompany both complete and partial electric discharges, shows a complexity of this issue. Signs of damages of insulating materials are accompanied with physiochemical factors. The most important of these factors are [4], [5], [6]: - the emission of an electromagnetic wave, from the place where a discharge appears, which is a result of a power-driven impulse, - chemical conversions within the structure of insulation,

Fig. 1. A modular diagram of the detection of the partial discharges by means of the acoustic emission’s method. The main elements of the structure are a piezoelectric detector, input measuring amplifiers, a digital to analogue rapid converter and a system of data collection. The piezoelectric detector is situated close to the place where a partial discharge appears. In terms of measurements on Articles

Journal of Automation, Mobile Robotics & Intelligent Systems

a real object, it is necessary to ‘sound’ as big as possible surface in order to determine areas with a high probability of the occurrence of partial discharges. The place of the contact between the detector and the surface of the object should be fitted, as much as it is possible, in order to eliminate auditions from the surrounding. It is significant to have the opportunity to transfer the detectors what makes it easier to conduct the diagnosis. Figure 2 presents a configuration of the system in terms of the measurements of partial discharges by means of the method of detection of the optical spectrum. Three optical detectors are situated close to the place where the phenomenon of partial discharges appears. Every detector is adapted to a different optical band. The bands of a visible, an ultraviolet and an infrared light are enumerated.

VOLUME 4,

N° 3

2010

analysts with the access to the subject under research. A significant advantage of the mobile research unit is the opportunity to deliver the equipment to those places where the staff, which operates the robot, may suffer from health damages. The SQ1 robot is presented in the Figures 3 and 5. It is designed and equipped with an apparatus that enables to move it within the area of its potential use. The robot is fitted with such technologies as ultrasound detectors of distance, accelerometrical detectors located on the robot’s legs and on the central part of its main body, digital cameras and laser scanner. The communication between an operator and the robot takes place cordlessly. The operator is able to guide the robot in the architectural space of a distribution board of average voltage by making use of an original application. Measuring samples are taken through the detectors that are placed on the research object. Information from the detectors is also transferred by means of radio waves.

Fig. 3. The mobile robot simulated in AutoDesk Inventor 2009. Fig. 2. A modular diagram of the detection of partial discharges by means of the method of the optical spectrum’s measurement. The detectors are built from semiconductor, photosensitive elements which purpose is directed to specific lengths of a light wave. During the measurements, a considerable activity in terms of the ultraviolet waves is observed, what directs research in this field. Because of a very low energy level of partial discharges, which mean a degradation of the insulator, very sensitive measuring tracks should be used. According to this fact, the detectors should be equipped with a high selectivity of the band, a high level of the reinforcement and a very good coefficient of a condition between the interference and the useful signal. Because of that, highly selected measuring detectors should be used for work with photosensitive semiconductor elements. It is not necessary to use rapid measuring systems because of a low dynamics of the signal. The digital to analogue processing in terms of the method of detection of partial discharges, under present consideration, does not have to exceed 100 kS/s. Then, there is a real opportunity to use a converter with an effective resolution above 16 bits.

3. The mobile robot The possible measurements constitute a real technical challenge within areas where there are big difficulties for 32

Articles

The mobile measuring system consists of few modules, which enable the operator to work remotely. The main elements of the mobile robot are a cordless module defined by the IEEE 802.11 standard (Wireless Fidelity – WiFi) [13], which enables a direct communication with the PDDetect2 controlling application; the main module which consists of the measuring part; a cordless transmission channel (Bluetooth) which enables the communication with the measuring detectors; an active underbody and a monitoring module of work parameters of the SQ1 robot. The module structure of the mobile measuring system is presented in the Figure 4. The robot is equipped with an arm that enables the operator to install elements of the measuring systems on the surface of the research object. During a diagnosis, there is an analysis of the partial discharges that are emitted, together with a simultaneous consolidation of the knowledge of the spatial position of the detector’s installation in a base. In the future it will enable to do comparative research and to determine degradation’s trends, which appear in an insulating system under research. The system of magnetic clutches, which are placed on the arm, enables to operate small modules of detectors more easily and to attach them to the surface of the object. Software of the measuring mobile system was completed with support of Real Time Operation System (RTOS).

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 4,

N° 3

2010

Fig. 4. The module structure of the mobile measuring system. Particular modules, which constitute the robot, were designed according to the architecture suggested by the producer of RTOS. The use of resources accessible in the library of the operation system gives a chance for a fast implementation of new functions, which lead to the development of the measuring mobile system. A module of a cordless detector was implemented by the use of a fast microcontroller with an ARM root. It allows a registration of the signal of acoustic emission generated from partial discharges with a speed to 8 MS/s. Samples registered by the detector (Figure 6) are sent to the mobile robot. In sequence, data are gathered as a file in the fixed memory. An access to information happens with the use of a FTP server (by means of the FTP client) or with a PDDetect2 application, which simultaneously implements the process of visualization and an analysis.

analysis and at the same time it does not cause interferences in the work of the energy system. Regular diagnosis of the elements of the energy system and a consolidation of their parameters enable to gather information, which are useful within the process of planning the service. In addition, the measuring mobile system memorises the track and the topology of the area of the electro energetic station and the place of the installation of objects, which are diagnosed.

Fig. 6. The module of the detector: active piezoelectric probes.

4. The control of the robot by means of the brain

Fig. 5. The mobile robot – a laboratory photograph. The automation of the process, which eliminates the necessity of disconnecting a live element of the electro energy system, remarkably shortens the time of the

The PDDetect2 application, installed in the working station from which the process of control of the robot takes place, enables the implementation of this process by an analysis of the electroencephalographic signal. The EEG signal is collected with a non-invasive method by means of active electrodes that are placed on the skin of a person’s head under research [14]. The arrangement of the electrodes is specified by the 10-20 IFCN international specifications. Then, the signal is transported to an electroencephalograph, which is connected by means of a USB 2.0 port to a working station with the PDDetect2 application. Software accomplishes an appropriate classification of the EEG signal in order to specify conclusions about the robot’s reactions on brain stimuli in a particular moment of time. Figure 7 illustrates ideas about implementations of the communication with the use of the EEG signal.

Articles

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

Fig. 7. The module structure of the system of the collection of brain waves. An appropriate analysis of the oscillation of the EEG signal in terms of the ?a, b, g rhythms has an impact on the process of a proper control of the robot [15]. During the analysis of the EEG signal a big synchronisation of the b rhythms in the attention process can be seen, what is essential from the point of view of the control of the mobile robot. In the same time, a big activeness of pyramidal cells appears and a strong synchronisation of them. The g rhythms are seen during an intensity of the amount of the processing information by neurons in the unit of time ‘t’. In terms of the g rhythms, a synchronisation of activity takes place, which is directly connected with the processing of information [16], [17]. Nonlinerality has a direct impact on the frequency of the rhythms and their shapes that is connected with specific character of stimulation, which appears then. Oscillations of the g rhythms can be observed on many electrodes within areas of those that were placed directly above the motor cortex. During a measurement of the electroencephalographic signal, it is also possible to observe a disynchronisation of the signal, which is seen in the difference of the energy of the measuring signal. The measurement of the disynchronisation of the signal is based on the study of its power (1). ,

(1)

where A is the value of an m-point of the signal in an nrepeat of the experiment and N means the number of repeats. Then, disynchronisation may be defined as it follows (2). ,

(2)

(3) l – the length of the reference area, m – the point of the signal. Besides the non-invasive Brain Machine Interface suggested in this article, there are also invasive methods. This type of BMI was described by scientists in the United States and Europe. They use a surgical implant, which Articles

2010

consists of beams of electrodes. From the technical point, it is more difficult to implement and less practical in such uses as the mobile system for the measurements of partial discharges. That is why the non-invasive method was used for the system of measurement of partial discharges. The mobile measuring system for research of partial charges, which occur in the insulating structures during the progressive process of the degradation of the dielectric material in electroenergetic capacitors with the use of the EA method, will be tested during diagnostic investigations in factory conditions. It will enable to specify the scope of practical adaptations. Experimentations during big levels of electromagnetic interferences, which accompany work of electroenergetic devices, will enable to state directions towards further transformations of the system what will result in its improvement. The use of the mobile system of recording of the EA signals generated by partial discharges in the insulation of the electroenergetic capacitors will allow among others to automatize of the measuring system which was conducted manually before, to automate the process of registration, to present the results (automatic change of places where converters are situated, the way of its connection with a particular object), to separate the technical personnel effectively from the place of measurement ensuring security; a mobility – the possibility for sending measuring data from the place of measurement to the station of the operation by means of a cordless network based on the TCP/IP protocol.

5. Problems on the stage of the completion of the connection between the brain and the mobile robot Undoubtedly, the greatest problems, which occur on the stage of correlation between a human brain and the mobile robot, are different kinds of both biological and technical interferences [18], [19]. Technical artefacts seem to be especially significant in the system under present analysis because the system works near big concentrations of electroenergetic devices, which disrupt a proper reading of the signal [20], [21]. According to this fact, it is necessary to isolate the control station from the mobile robot’s area of work. Apart from the technical problems mentioned above, a certain biological criteria are required from the person who controls the robot. A special attention should be paid to people with high blood pressure, accelerated pulsation of heart or those who possess different kind of nervous tics. Unfortunately, such people cannot control the mobile robot in an easy way.

6. Conclusion

where Z is the level of reference (3).

N° 3

The control of the mobile robot by means of brain waves also requires a constant improvement. It is necessary to elaborate on population models that are imagings of particular populations of neurons and their mutual correlations. By means of advanced mathematical methods, it is possible to simulate proper preservation of the EEG signal in specific mental states of a human being. It is the key to success in the case of the structure and the implementation of both Brain Computer Interface and Brain Machine Interface communication.

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

AUTHORS Andrzej Błachowicz*, Szczepan Paszkiel - Opole University of Technology, Faculty of Electrical Engineering, Automatic Control and Computer Science, Institute of Electrical Power Engineering, uL. Prószkowska 76, 45-758 Opole. Tel. +48 600 494 771, +48 600 937 899. E-mail: andrew_blachowicz@poczta.fm. * Corresponding author

[16]

[17]

[18]

References [1]

[2]

[3]

[4]

[5]

[6]

[7]

[8]

[9]

[10]

[11]

[12]

[13]

[14]

[15]

Georgopoulos A.P., Langheim F.J., Leuthold A.C., Merkle A.N., “Magnetoencephalographic signals predict movement trajectory in space”, Exp. Brain Res., vol. 25, 2005, pp. 132–135. Dudek Z.T., „Interfejs BCI – próba przełamania bariery pomiędzy człowiekiem a komputerem”, Przegląd Telekomunikacyjny i Wiadomości Telekomunikacyjne, no. 7, 2003. (in Polish). Tortella F., Rose J., Robles L., Moreton J., Hughes J., Hunter J., “EEG Spectra Analysis of the Neroprotective Kappa Opioids Enadoline”, Maryland 1997. Kawada H., Honda M., Inoue T., Ameniya T., “PD automatic monitor for oil-filled power transformer”, IEEE PAS, vol. 103, no. 2, 1984, pp. 422-428. Kirtley J.L., Lavalle J.C., McCarthy D.J., “Acoustic monitoring of transformer structures”, CIGRE Symp., Report no. 3-05.87, Vienna, 1987, pp. 1-5. Kucharska B., „Akustyczne deskryptory sygnałów emitowanych przez wyładowania niezupełne powierzchniowe”, ZN PO, no. 219, Elektryka, vol. 44, 1996, pp. 107124. (in Polish) Kelley N.D., Osgood R. M., Bialasewicz J.T., Jakubowski A., “Using time-frequency and wavelet analysis to access turbulence/rotor interactions”. In: 19th American Society of Mechanical Engineers (ASME) Wind Energy Symposium, Reno, Nevada, 10th–13th January, 2000. Florkowska B., Zydroń P., „Analiza widmowa ulotu elektrycznego”. In: V Ogólnopolskie Sympozjum IWN- IW 2002, Poznań-Kiekrz, 2002, pp. 80-83. (in Polish) Fuhr J., Maessing M., Boss P., Tschudi D., King R.A., “Detection and Location of Internal Defects in the Insulation of Power Transformers”, IEEE Trans. on EI., vol. 28, no. 6, 1993, pp. 10057-1069. Howells E., Norton E.T., “Parameters affecting the velocity of sound in transformer oil”, IEEE PAS, vol. 103, no. 5, 1984, pp. 1111-1115. Rabbani M., Santa Cruz D., The JPEG 2000 still-image compression standard. Materials of course in ICIP, Thessaloniki, Greece 2001. Skubis J., „Emisja akustyczna w badaniach izolacji urządzeń elektroenergetycznych”, IPPT PAN, Opole 1993. (in Polish) Zmarzły D., „Pomiary elektrycznych wielkości medycznych”, Opole University of Technology, Opole 2004. (in Polish) Venter H.S., Eloff J.H.P., “Network Security: Important Issues”, Department of Computer Science, Rand Afrikaans University, 1992, pp. 147-154. Larrieu N., Owezarski P.: “Towards a Measurement Based Networking approach for Internet QoS improvement”, Computer Communications, vol. 28, 2005, pp. 259-273.

[19]

[20]

[21]

N° 3

2010

White T., Pagurek B., Bieszczad A., “Network Modeling for Management Applications Using Intelligent Mobile Agents”, Journal of Network and Systems Management, vol. 7, no. 3, 1999. Todorova P., “Network Control in ATM-Based LEO Satellite Networks”, Telecommunication Systems, 22:1–4, 2003, pp. 321–335. Bush S.F., Frost V.S., “A Framework for Predictive Network Management of Predictive Mobile Networks”, Journal of Network and Systems Management, vol. 7, no. 2, 1999. Ibrahim T., Venin J.M., Garcia G., “Brain Computer Interface in Multimedia Communication”, IEEE Signal Processing Magazine, vol. 20, no. 1, 2003. Wolański N., „Rozwój biologiczny człowieka. Podstawy auksologii, gerontologii i promocji zdrowia”, Warszawa 2005, p. 374. (in Polish) Schumacher H.J., Ghosh S., “A fundamental framework for network security”, Journal of Network and Computer Applications, 1997, pp. 305–322.

Articles

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N째 3

2010

AUTOMATIC GENERATION OF FUZZY INFERENCE SYSTEMS USING HEURISTIC POSSIBILISTIC CLUSTERING Received 10th February 2010; accepted 26th April 2010.

Dmitri A. Viattchenin

Abstract: The interpretability and flexibility of fuzzy classification rules make them a popular basis for fuzzy controllers. Fuzzy control methods constitute a part of the areas of automation and robotics. The paper deals with the method of extracting fuzzy classification rules based on a heuristic method of possibilistic clustering. The description of basic concepts of the heuristic method of possibilistic clustering based on the allotment concept is provided. A general plan of the D-AFC(c)-algorithm is also given. A method of constructing and tuning of fuzzy rules based on clustering results is proposed. An illustrative example of the method's application to the Anderson's Iris data is carried out. An analysis of the experimental results is given and preliminary conclusions are formulated. Keywords: possibilistic clustering, fuzzy cluster, typical point, tolerance threshold, fuzzy rule.

1. Introduction Some remarks on fuzzy inference systems are considered in the first subsection. The second subsection includes a brief review of methods of extracting of fuzzy rules based on fuzzy clustering and the aims of the paper. 1.1. Preliminaries Fuzzy inference systems are one of the most famous applications of fuzzy logic and fuzzy sets theory. They can be helpful to achieve classification tasks, process simulation and diagnosis, online decision support tools and process control. So, the problem of generation of fuzzy rules is one of more than important problems in the development of fuzzy inference systems. There are a number of approaches to learning fuzzy rules from data based on techniques of evolutionary or neural computation, mostly aiming at optimizing parameters of fuzzy rules. From other hand, fuzzy clustering seems to be a very appealing method for learning fuzzy rules since there is a close and canonical connection between fuzzy clusters and fuzzy rules. The idea of deriving fuzzy classification rules from the data can be formulated as follows: the training data set is divided into homogeneous group and a fuzzy rule is associated to each group. Fuzzy clustering procedures are exactly pursuing the strategy: a fuzzy cluster is represented by the cluster center and the membership degree of a datum to the cluster is decreasing with increasing distance to the cluster center. So, each fuzzy rule from a fuzzy inference system can be characterized by a typical point and membership func36

Articles

tion that is decreasing with increasing distance to the typical point. 1.2. Fuzzy clustering and fuzzy rules Let us consider some methods of fuzzy rules extracting from the data using fuzzy clustering algorithms. Some basic definitions must be given in the first place. The training set contains n data pairs. Each pair is made of a m-dimensional input-vector and a c-dimensional output-vector. We assume that the number of rules in the fuzzy inference system rule base is c. So, Mamdani's [1] rule l within the fuzzy inference system is written as follows: ,

(1)

where and are fuzzy sets that define an input and output space partitioning. A fuzzy inference system, which is described by a set of fuzzy rules with the form (1) is the multiple inputs, multiple outputs system. Note that any fuzzy rule with the form (1) can be presented by c rules with the form of multiple 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 functions. Fuzzy classification rules can be obtained directly from fuzzy clustering results. In general, a fuzzy clustering algorithm aims at minimizing the objective function [2] (3) under the constraints (4) and (5) where is the data set, c is the number of fuzzy clusters in the fuzzy c-partition is the membership degree of object

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

to fuzzy cluster is the prototype for fuzzy cluster is the distance between prototype and object , and the parameter is the fuzziness index. The selection of the value of determines whether the cluster tend to be more crisp or fuzzy. Membership degrees 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) is hard from essential positions. So, a possibilistic approach to clustering was proposed in [3]. In particular, the objective function (3) is replaced by (8) under the constraint of possibilistic partition (9) where c is the number of fuzzy clusters....,............. in the possibilistic partition is the possibilistic memberships which are typicality degrees,............ is the prototype for fuzzy cluster is the distance between prototype and object , and the parameter is the analog of the fuzziness index. Typicality degrees can be calculated as following (10) and the parameters

are estimated by

N째 3

2010

(13) in the possibilistic case. An approximation of the fuzzy set by projecting only the data set and computing the convex hull of this projected fuzzy set or approximating it by a trapezoidal or triangular membership function is used for the rules obtaining [4]. Objective function-based fuzzy clustering algorithms are the most widespread methods in fuzzy clustering [2]. Objective function-based fuzzy clustering algorithms are sensitive to initial partition selection and fuzzy rules depend on the selection of the fuzzy clustering method. In particular, the GG-algorithm and the GK-algorithm of fuzzy clustering are recommended in [2] for fuzzy rules generation. All algorithms of possibilistic clustering are also objective functions-based algorithms. Heuristic algorithms of clustering display low level of a complexity. An outline for a heuristic method of possibilistic clustering was presented in [5], where a basic version of direct possibilistic clustering algorithm was described and the version of the algorithm is called the DAFC(c)-algorithm [6]. The main goal of the paper is a detail consideration of the method of the rapid prototyping fuzzy inference systems, which was outlined in [7]. The method is based on deriving fuzzy classification rules from the data on a basis of clustering results obtained from the D-AFC(c)algorithm. The contents of this paper is as follows: in the second section basic concepts of the possibilistic clustering method based on the concept of allotment among fuzzy clusters are outlined and a plan of the DAFC(c)-algorithm is given, in the third section a method of constructing of fuzzy rules is proposed, in the fourth section an illustrative example of deriving fuzzy rules from the Anderson's Iris data are given, in the fifth section preliminary conclusions are stated and some perspectives are outlined.

(11)

2. A heuristic method of possibilistic clustering where . The principal idea of extracting fuzzy classification rules based on fuzzy clustering is the following [2]. Each fuzzy cluster is assumed to be assigned to one class for classification and the membership grades of the data to the clusters determine the degree to which they can be classified as a member of the corresponding class. So, with a fuzzy cluster that is assigned to the some class we can associate a linguistic rule. The fuzzy cluster is projected into each single dimension leading to a fuzzy set on the real numbers. From a mathematical position the membership degree of the value to the tth projection of the fuzzy cluster is the supremum over the membership degrees of all vectors with as tth component to the fuzzy cluster, i.e. (12)

The basic concepts of the heuristic method of possibilistic clustering are considered in the first subsection. A plan of the direct clustering algorithm is given in the second subsection. The third subsection includes a review of methods of the data preprocessing. 2.1. Basic concepts The D-AFC(c)-algorithm is based on a concept of an allotment of elements of the set of classified objects among fuzzy a-clusters. The allotment of elements of the set of objects among the fixed number c of fuzzy a-clusters can be considered as a special case of possibilistic partition. The fact was demonstrated in [6] and [8]. That is why the basic version of the algorithm, which is described in [5], can be considered as a direct algorithm of possibilistic clustering and the algorithm was called the D-AFC(c)-algorithm [6]. Let us remind the basic concepts of the heuristic method of possibilistic clustering. The concept of fuzzy Articles

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

tolerance is the basis for the concept of fuzzy a-cluster. That is why definition of fuzzy tolerance must be considered in the first place. Let be the initial set of elements and some binary fuzzy relation on X with being its membership function. Fuzzy tolerance is the fuzzy binary intransitive relation, which possesses the symmetricity property (14) and the reflexivity property (15) Let be the initial set of objects. Let T be a fuzzy tolerance on X and a be a-level value of . Columns or lines of the fuzzy tolerance matrix are fuzzy sets . Let be fuzzy sets on X, which are generated by a fuzzy tolerance T. The a-level fuzzy set.............................................. is fuzzy a-cluster or, simply, fuzzy cluster. So and is the membership degree of the element for some fuzzy cluster . Value of a is the tolerance threshold of fuzzy clusters elements. The membership degree of the element for some fuzzy cluster can be defined as a

N째 3

2010

If condition (18) is met for all fuzzy clusters......................................... then the family is the allotment of elements of the set among fuzzy clusters.............................. for some value of the tolerance threshold a. It should be noted that several allotments.......... could exist for some tolerance threshold a. That is why symbol z is the index of an allotment. The condition (18) requires that every object must be assigned to at least one fuzzy cluster with the membership degree higher than zero. The condition requires that the number of fuzzy clusters in each allotment........... must be more than two. Obviously, the definition of the allotment among fuzzy clusters (18) is similar to the definition of the possibilistic partition (9). So, the allotment among fuzzy clusters can be considered as the possibilistic partition and fuzzy clusters in the sense of (16) are elements of the possibilistic partition. If condition (19) where

and condition (20)

(16) where an a-level of a fuzzy set is the support of the fuzzy cluster . So, condition is met for each fuzzy cluster . Membership degree can be interpreted as a degree of typicality of an element to a fuzzy cluster. The value of a membership function of each element of the fuzzy cluster in the sense of (16) is the degree of similarity of the object to some typical object of fuzzy cluster. Membership degree defines a possibility distribution function for some fuzzy cluster . The fact was demonstrated in [8] and the possibility distribution function is denoted by . Let T is a fuzzy tolerance on X, where X is the set of elements, and is the family of fuzzy clusters for 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 e is the index of the typical point. A set of typical points of the fuzzy 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 family of fuzzy clusters for some value of tolerance threshold which are generated by some fuzzy tolerance T on the initial set of elements .

Articles

are met for all fuzzy clusters of some allotment then the allotment is the allotment among particularly separate fuzzy clusters and is the maximum number of elements in the intersection area of different fuzzy clusters. Obviously, if in conditions (19) and (20) then the intersection area of any pair of different fuzzy cluster is an empty set and fuzzy clusters are fully separate fuzzy clusters. The adequate allotment for some value of tolerance threshold is a family of fuzzy clusters which are elements of the initial allotment for the value of and the family of fuzzy clusters should satisfy the conditions (19) and (20). Several adequate allotments can exist. Thus, the problem consists in the selection of the unique adequate allotment from the set B of adequate allotments, which is the class of possible solutions of the concrete classification problem. The set of adequate allotments is depending on the number of fuzzy clusters c in the sought allotment. So, is the set of adequate allotments corresponding to the formulation of a classification problem. The selection of the unique adequate allotment from the set of adequate allotments must be made on the basis of evaluation of allotments. The criterion (21) where c is the number of fuzzy clusters in the allotment and is the number

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

of elements in the support of the fuzzy cluster , can be used for evaluation of allotments. Maximum of criterion (21) corresponds to the best allotment of objects among c fuzzy clusters. So, the classification problem can be characterized formally as determination of the solution satisfying

N° 3

2010

normalized as follows: (23) In the second place, the data can be normalized using a formula

(22)

(24)

The adequate allotment is any allotment among c fuzzy clusters in the case. Thus, the problem of cluster analysis can be defined in general as the problem of discovering the unique allotment, resulting from the classification process.

So, each object can be considered as a fuzzy set..... and ................................................... are their membership functions. The matrix of coefficients of pair wise dissimilarity between objects can be obtained after application of some distance to the matrix of normalized data....................................................... The most widely used distances for fuzzy sets........... in are [11]: The normalized Hamming distance:

2.2. The D-AFC(c)-algorithm Detection of fixed c number of fuzzy clusters can be considered as the aim of classification. There is a sevenstep procedure of classification: 1. Calculate -level values of the fuzzy tolerance T and construct the sequence of -levels; l 2. Construct the initial allotment............................... for every value from the sequence 3. Let ; 4. Construct allotments........................................... which satisfy conditions (19) and (20) for every value from the sequence.......................... 5. Construct the class of possible solutions of the classification problem for the given number of fuzzy clusters c and different values of the tolerance threshold as follows: the if for some allotment condition is met, then and go to step 4; else let 6. Calculate the value of the criterion (21) for every allotment 7. The result of classification is formed as follows: from the set....... if for some unique allotment the condition (22) is met, then the allotment is the result of classification, else the number c of classes is suboptimal. So, the allotment among the given number c of fuzzy clusters and the corresponding value of tolerance threshold are the results of classification. Some modifications of the D-AFC(c)-algorithm are proposed in [6], [9] and [10]. 2.3. Notes on the data preprocessing Let us consider a method for the data preprocessing. The matrix of fuzzy tolerance...................................... is the matrix of initial data for the D-AFC(c)-algorithm of possibilistic clustering. However, the data can be presented as a matrix of attributes................................. where the value is the value of the tth attribute for ith object. In the first place, the data can be

(25) The normalized Euclidean distance: (26) The squared normalized Euclidean distance: (27) The matrix of fuzzy tolerance........................ can be obtained after application of complement operation (28) to the matrix of fuzzy intolerance........................ obtained from previous operations.

3. Deriving fuzzy rules from fuzzy clusters A technique of fuzzy rules antecedents learning is presented in the first subsection. A method of consequents learning is given in the second subsection of the section. The third subsection includes a technique of fuzzy rules tuning. 3.1. Antecedents learning In the following, we will consider that the fuzzy inference system is a multiple inputs, multiple outputs system. The antecedent of a fuzzy rule in a fuzzy inference system defines a decision region in the m-dimensional feature space. Let us consider a fuzzy rule (1) where.... is a fuzzy set associate with the feature variable . Let be characterized by the trapezoidal membership function , which is presented in Figure 1. So, the fuzzy set can be defined by four parameters . A triangular fuzzy set can be considered as a particular case of the trapezoidal fuzzy set where . The idea of deriving fuzzy rules from fuzzy clusters is the following [7]. We apply the D-AFC(c)-algorithm to the Articles

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N째 3

2010

met for all input variables . Obviously, if condition is met for the fuzzy cluster and only one typical point is presented in the fuzzy cluster, then the condition is met.

Fig. 1. A trapezoidal membership function for an antecedent fuzzy set. given data and then obtain for each fuzzy cluster a kernel and a support . The value of tolerance threshold , which corresponds to the allotment , is the additional result of classification. We calculate the interval........... of values of every attribute , for the support . The value can be obtained as follows

3.2. Consequents learning The variables are the consequents of fuzzy rules (1), represented by the fuzzy sets..................... with the membership functions . Fuzzy sets can be defined on the interval of memberships and these fuzzy sets can be presented as follows: where is the tolerance threshold, and . So, membership functions of fuzzy sets will be trapezoidal membership functions. The situation is shown in Figure 2.

(29) and the value a formula

can be calculated using

(30) The parameter

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

can be obtained as following (31)

and the parameter

can be obtained from the conditions (32)

We calculate the value for all typical points of the fuzzy cluster .,............... as follows:

Fuzzy clusters can be subnormal fuzzy sets [12]. So, the case, which is presented in Figure 2, is the general case. If the allotment among fully separate fuzzy clusters is obtained and all fuzzy clusters.................... are normal fuzzy sets then.................. and for each fuzzy cluster......... . So, a trapezoidal membership function of a fuzzy set in the case of normal and fully separate fuzzy clusters is presented in Figure 3.

(33) and the value

can be obtained from the equation (34)

Thus, the parameter conditions

and the parameter

The height

the fuzzy cluster set [12], [13]. So, condition 40

Articles

can be calculated from the

(35)

Fig. 3. A trapezoidal membership function for a consequent fuzzy set in a case of normal and fully separated fuzzy clusters.

(36)

Thus, trapezoidal membership functions for the fuzzy sets can be constructed on a basis of the clustering results. The empty set............. can be correspond to some output variable So, the empty fuzzy set will be correspond to the output variable and is the membership function of the corresponding fuzzy set .

can be obtained as following

of the fuzzy cluster , must be taken into account because can be a subnormal fuzzy and condition

are

Journal of Automation, Mobile Robotics & Intelligent Systems

If the allotment among particularly separate fuzzy clusters is obtained then some non-empty fuzzy sets will be correspond to some output variables........... . 3.3. Fuzzy rules tuning The computational accuracy must be taken into account in the data processing by the D-AFC(c)-algorithm. The computational accuracy can be determined by a value of accuracy threshold . If we decrease the value , the computational accuracy increases. Membership values the value of tolerance threshold , and the number of typical points in each fuzzy cluster are depending on the value of accuracy threshold . An allotment can be characterized by the value of tolerance threshold that is increasing with decreasing accuracy, i.e., for......... we have . From other hand, a fuzzy cluster can be characterized by a kernel............ and the number of typical points of the fuzzy cluster is decreasing with increasing accuracy, i.e., for.......... we have . So, the accuracy threshold .. can be used as a parameter for the D-AFC(c)-algorithm. The fact was demonstrated in [14]. Moreover, the accuracy threshold can be considered as the analog of the fuzziness index in the formula (3). Thus, the accuracy threshold can be useful for tuning of the rules. In particular, for we have................ and a crisp interval is increasing. Otherwise, if we decrease the accuracy threshold, , the number of typical points of the fuzzy cluster increases, and a crisp interval increases. Membership functions for consequents fuzzy sets depend on the value of accuracy threshold . For example, if we increase the value of accuracy threshold , the crisp interval decreases. Moreover, for we have . That is why parameters and increases for all fuzzy sets , i.e., for we have and . The proposed technique for fuzzy rules tuning will be explained by an illustrative example in the next section.

4. An illustrative example The first subsection of the section includes the results of the Anderson's Iris data clustering by the D-AFC(c)algorithm. The designed fuzzy inference system is presented in the second subsection and the results are compared with the results of other classifier systems. 4.1. Results of the Anderson's Iris data clustering The Anderson's Iris data set consists of the sepal length, sepal width, petal length, and petal width measured for 150 irises [15]. The problem is to classify the plants into three subspecies on the basis of this information. The Anderson's Iris data forms the matrix of attributes where the sepal length is denoted - by , sepal width - by , petal length - by and petal width - by . The Iris database is the most known database to be found in the pattern recognition literature. The method of the data preprocessing which was described in the third section can be used for constructing the matrix of fuzzy tolerance and

VOLUME 4,

N째 3

2010

the matrix of fuzzy tolerance can be processed by the DAFC(c)-algorithm. The formula (23) and the squared normalized Euclidean distance (27) were used for the data preprocessing. Four experiments were made for different values of the accuracy threshold . The allotment among three fully separated fuzzy clusters was obtained in each experi-ment. The results of the Anderson's Iris data set pro-cessing by the D-AFC(c)-algorithm for different values of the accuracy threshold are presented in the Table 1. Table 1. Results of the Anderson's Iris data set classification obtained from the D-AFC(c)-algorithm for different values of the accuracy threshold.

By executing the D-AFC(c)-algorithm for three classes (1, 2, 3) in each experiment we obtain the following: the first class is formed by 50 elements all being Iris Setosa; the second class by 52 elements, 48 of them being Iris Versicolor and 4 Iris Virginica; the third class by 48 elements, 46 of them being Iris Virginica and 2 Iris Versicolor. In other words, the first class corresponds to the Setosa subspecies, the second class corresponds to the Versicolor subspecies and the third class corresponds to the Virginica subspecies. So, there are six mistakes of classification in each experiment. 4.2. A fuzzy inference system Let us consider results of the experiment for the value of accuracy threshold . The ninety-fifth object is the typical point of the fuzzy cluster, which corresponds to the first class, the ninety-eighth object is the typical point of the second fuzzy cluster, and the seventy-third object is the typical point of the third fuzzy cluster. The height of each fuzzy cluster is equal one. So, membership functions and for corresponding fuzzy sets and can be constructed immediately. The rule base induced by the D-AFC(c)-algorithm clustering result can be seen in Figure 4 where labels and denote, respectively, sepal length, sepal width, petal length, and petal width, and is the number of rule. Note that only one typical point is presented in each fuzzy cluster. That is why membership functions............ are triangular membership functions. Obviously that a meaningful linguistic label can be assigned to each fuzzy set . From other hand, linguistic labels Setosa, Versicolor and Virginica are associated with corresponding output variables . Note that fuzzy sets................. and are empty fuzzy sets. Articles

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

Rule

SL1

SW1

PL1

PW1

SETOSA

(4.274, 5, 5.83)

(2.259, 3.4, 4.437)

(0.9814, 1.5, 1.915)

(0.0962, 0.2, 0.6148)

(0.9642, 0.9828, 1, 1)

SL2

SW2

PL2

PW2

(4.878, 5.5, 7.056)

(1.985, 2.4, 3.437)

(2.974, 3.7, 5.152)

(1, 1, 1.622)

SL3

SW3

PL3

PW3

(5.523, 7.7, 7.907)

(2.481, 3, 3.83)

(4.752, 6.1, 6.93)

(1.574, 2.3, 2.507)

VERSICOLOR

NÂ° 3

2010

VIRGINICA

2 (0.9642, 0.9676, 1, 1)

3 (0.9642, 0.9673, 1, 1)

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

Fig. 5. The performance of the fuzzy inference system. We show in Figure 5 a graph of the performance of the designed fuzzy inference system. The example of classification of the ninety-fifth object, which is the typical point of the first fuzzy cluster, is presented in Figure 5. Total area is zero in the defuzzification procedure for output variables Versicolor and Virginica. That is why an average of the range of output variables Versicolor and Virginica are used as output values and these values are equal 0.5. The values can be interpreted as uncertain membership degrees. The result, which is obtained from fuzzy inference system, is easily interpreted. Thus, the obtained model is suitable for interpretation since the rules consequents are the same or close to the actual class labels, such that each rule can be taken to describe all classes. The Anderson's Iris data were classified using the constructed fuzzy inference system. The rules classify four objects incorrectly and two objects are rejected. Thus, the total number of misclassifications is 6. Evidently, that the results are correlated with the results, obtained from the D-AFC(c)-algorithm. So, the fuzzy inference system is accurate. The application of the constructed fuzzy inference 42

Articles

system to the Anderson's Iris data was made in comparison with other approaches. Table 2 shows the results of some well-known classifier systems. Table 2. Comparison of results of different classifier systems on the Anderson's Iris data set.

For example, HĂśppner, Klawonn, Kruse and Runkler [2] applied the simplified version of the GG-algorithm of fuzzy clustering to learn a Mamdani-type fuzzy inference system for classifying the Anderson's Iris data by training on all 150 objects. An eight-rule fuzzy system was obtained. The rules classify 3 objects incorrectly and 3 more were not classified at all. So, the total number of misclassifications is 6. From other hand, the FCM-algorithm of fuzzy cluste-

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

ring was applied by Roubos and Setnes [16] to obtain an initial Takagi-Sugeno model with singleton consequents. All 150 samples were used in the training process. So, the initial model with three rules was constructed from clustering results where each rule described a class. The classification accuracy of the initial model was rather discouraging, giving 33 misclassifications on the trai-ning data. A multi-objective genetic algorithm-based optimization approach was applied to the initial model. So, the number of misclassifications was reduced to 4 samples. Ishibuchi, Nakashima and Murata [17] applied all 150 samples in the training process, and derived a fuzzy classifier with five rules. The resolution was 3 misclassifications. Abonyi, Roubos and Szeifert [18] proposed a datadriven method to design compact fuzzy classifiers via combining a genetic algorithm, a decision-tree initialization, and a similarity-driven rule reduction technique. The final fuzzy inference system had three fuzzy rules and the number of misclassifications was 6. A fuzzy classifier with ellipsoidal regions was proposed by Abe and Thawonmas [19]. They applied clustering methods to extract fuzzy classification rules, with one rule around cluster center, and then they tuned the slops of their membership functions to obtain a high recognition rate. Finally, they obtained a three-rule fuzzy system with 2 misclassifications. The results obtained from the constructed fuzzy inference system seem appropriate in comparison with the some well-known fuzzy systems. So, the proposed method of derivation of fuzzy classification rules from data can be considered as an effective technique of the rapid prototyping fuzzy inference systems.

5. Concluding remarks Some conclusions are formulated in the first subsection. The second subsection deals with the perspectives on future investigations.

N° 3

2010

rules tuning based on varying of the accuracy threshold is proposed in the paper. However, some other approaches, such as the genetic algorithm-based approach or neurofuzzy techniques can be used for fuzzy rules tuning. Note that the computational complexity of the D-AFC(c)-algorithm is higher in comparison with objective function-based fuzzy clustering algorithms. For example, approximately 700 observations is the large data set for the D-AFC(c)-algorithm. Of course, the computational complexity of the D-AFC(c)-algorithm is the subject of special considerations. However, the clustering problem in cases of large data sets can be solved in the preliminary way as follows: the initial data set.................... can be represented as a set where and each element is the subset of the data set . So, the matrix of the reduced data set can be presented as the matrix of the interval-valued data, and the data can be processed by the D-AFC(c)-algorithm [20]. Fuzzy rules can be extracted from the interval-valued data clustering results immediately. From other hand, the D-AFC(c)-algorithm can be applied for classification the three-way data [13] and the fuzzy data [21]. So, the proposed method of designing fuzzy inference systems can be generalized for corresponding cases of the training data set. These perspectives for investigations are of great interest both from the theoretical point of view and from the practical one as well. ACKNOWLEDGMENTS I am grateful to Prof. Janusz Kacprzyk, Prof. Jan W. Owsinski, Prof. Frank Klawonn and Prof. Valery Starovoitov for their interest 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 valuable comments.

5.1. Discussion Many techniques to design fuzzy inference systems from data are available; they all take advantage of the property of fuzzy inference systems to be universal approximators. This paper presents an automatic method to design fuzzy inference system for classification via heuristic possibilistic clustering and the method can be considered as an approach to rapid prototyping of fuzzy inference systems. The proposed method is simple in comparison with other well-known approaches. The results obtained with the proposed modeling approach for the Anderson's Iris data set case illustrate the effectiveness of the proposed method of designing fuzzy inference systems. Notable that the fuzzy rules obtained using the DAFC(c)-algorithm can be interpreted very simply, because membership functions of fuzzy sets which correspond to input variables of fuzzy rules has natural interpretations.

AUTHOR Dmitri A. Viattchenin - Laboratory of Images Recognition and Processing, United Institute of Informatics Problems of the National Academy of Sciences of Belarus, Surganov St. 6, 220012 Minsk, Belarus, e-mail: viattchenin@mail.ru.

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

[4]

References [1]

[2]

[3]

[5]

Mamdani E.H., Assilian S., “An experiment in linguistic synthesis with a fuzzy logic controller”, International Journal of Man-Machine Studies, vol. 7, 1975, pp. 1-13. Höppner F., Klawonn F., Kruse R., Runkler T., Fuzzy Cluster Analysis: Methods for Classification, Data Analysis and Image Recognition, Chichester: Wiley Intersciences, 1999. Krishnapuram R., Keller J.M., “A possibilistic approach to clustering”, IEEE Transactions on Fuzzy Systems, vol. 1, 1993, pp. 98-110. Sugeno M., Yasukawa T., “A fuzzy-logic-based approach to qualitative modeling“, IEEE Transactions on Fuzzy Systems, vol. 1, 1993, pp. 7-31. Viattchenin D.A., “A new heuristic algorithm of fuzzy Articles

Journal of Automation, Mobile Robotics & Intelligent Systems

[6]

[7]

[8]

[9]

[10]

[11] [12]

[13]

[14]

[15] [16]

[17]

[18]

[19]

[20]

[21]

clustering”, Control & Cybernetics, vol. 33, 2004, pp. 323-340. Viattchenin D.A., “A direct algorithm of possibilistic clustering with partial supervision”, Journal of Automation, Mobile Robotics and Intelligent Systems, vol. 1, no.3, 2007, pp. 29-38. Viattchenin D.A., “Outlines for a new approach to generating fuzzy classification rules through clustering techniques”, Proc. of the 10th International Conference on Pattern Recognition and Information Processing PRIP'2009, Minsk, Belarus, 2009, pp. 82-87. Viattchenin D.A., “On possibilistic interpretation of membership values in fuzzy clustering method based on the allotment concept”, Proceedings of the Institute of Modern Knowledge, no. 3, 2008, pp. 85-90. (in Russian) Viattchenin D.A., “Direct algorithms of fuzzy clustering based on the transitive closure operation and their application to outliers detection”, Artificial Intelligence, no. 3, 2007, pp. 205-216. (in Russian) Viattchenin D.A., “An algorithm for detecting the principal allotment among fuzzy clusters and its application as a technique of reduction of analyzed features space dimensionality”, Journal of Information and Organizational Sciences, vol. 33, 2009, pp. 205-217. Kaufmann A., Introduction to the Theory of Fuzzy Subsets, New York: Academic Press, 1975. Viattchenin D.A., “Kinds of fuzzy a-clusters”, Proceedings of the Institute of Modern Knowledge, no. 4, 2008, pp. 95-101. (in Russian) Viattchenin D.A., “An outline for a heuristic approach to possibilistic clustering of the three-way data”, Journal of Uncertain Systems, vol. 3, 2009, pp. 64-80. Damaratski A., Novikau D., “On the computational accuracy of the heuristic method of possibilistic clustering”, Proc. of the 10th International Conference on Pattern Recognition and Information Processing PRIP'2009, Minsk, Belarus, 2009, pp. 78-81. Anderson E., “The irises of the Gaspe Peninsula”, Bulletin of the American Iris Society, vol. 59, 1935, pp. 2-5. Roubos H., Setnes M., “Compact and transparent fuzzy models and classifiers through iterative complexity reduction”, IEEE Transactions on Fuzzy Systems, vol. 9, 2001, pp. 516-524. Ishibuchi H., Nakashima T., Murata T., “Three-objective genetic-based machine learning for linguistic rule extraction”, Information Sciences, vol. 136, 2001, pp. 109-133. Abonyi J., Roubos J.A., Szeifert F., “Data-driven generation of compact, accurate and linguistically sound fuzzy classifiers based on a decision-tree initialization”, International Journal of Approximate Reasoning, vol. 32, 2003, pp. 1-21. Abe S., Thawonmas R., “A fuzzy classifier with ellipsoidal regions”, IEEE Transactions on Fuzzy Systems, vol. 5, 1997, pp. 516-524. Viattchenin D.A., Damaratski A. “Constructing of allotment among fuzzy clusters in case of quasi-robust cluster structure of set of objects”, Doklady BGUIR, no. 1, 2010, pp. 46-52. (in Russian) Viattchenin D.A., “A heuristic approach to possibilistic clustering for fuzzy data”, Journal of Information and Organizational Sciences, vol. 32, 2008, pp. 149-163.

Articles

VOLUME 4,

N° 3

2010

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N째 3

2010

CAPACITIVE HUMAN PRESENCE SENSOR FOR SAFETY APPLICATIONS Received 29th April 2010; accepted 10th May 2010.

Piotr Frydrych, Roman Szewczyk

Abstract: Paper presents a novel idea of capacitive sensor for human presence detection. Due to the use of resonant circuit significant increase of sensor's sensitivity was achieved. In the case of human hand, value of output signal changes up to 421%. Moreover, sensor creates possibility of determination of the material of the detected object. As a result developed sensor can be applied for human presence detection sensors utilized in safety systems, which are especially suitable for underground transportation network. Keywords: capacitive sensors, people safety, underground transportation network.

1. Introduction Daily capacity of Warsaw Underground, according to its technical data, is up to 500,000 passengers [1]. That number of passengers does not provoke risk only under condition, that rush will be similar during whole day. In fact, people mostly use underground during rush hours. The main hazard for passengers in underground is to be run over by train. Except train collision and derailment run over is the main cause of death and injuries of passengers. Majority of accidents can be prevented by Automatic Train Speed Control System (SOP-2) [2]. That system has been already applied in most underground systems, also in Warsaw. That solution helped prevent main cause of most death accidents. Statistical data about accidents in London Underground, presented in Table 1 [3], shows that only five accidents causing passenger deaths have occurred due to train operation in nearly 150 years. In Warsaw Underground being run over by train was the only cause of death since it was opened in 1995. Risk of fall of from the platform on track can be decrease by automatic barriers between train and platform. Synchronization with Automatic Train Speed Con-

trol System (SOP-2) is critical for such solution because doors on platform should be in front of train's doors, to let passenger go out. That brings very serious danger in case of fire and was the main cause of death of 200 passengers in underground in South Korea [4]. This is the reason that on second underground line barriers are not planned. In conclusion there is no sufficient method to prevent passenger from being run down by train accident. For this paper show an example a human detection system that in cooperation with Automatic Train Speed Control System can increase passenger's safety in underground.

2. Principles of operation Passenger's safety can be increase by detecting human presence on track and preventing to from entering train to station if such incident occurred. System SOP-2 can stop the train, when track sensors indicate signal that track is occupied by other train. The idea is to communicate with that system to stop the train in the shortest time before train enters the station, when danger situation has been noticed. Three types of sensors are used to detect human presence on track. Motion detectors based on passive infrared detector and image processing method work as human detectors. Single beam optical sensors are used to detect train. Third types of detector are proximity sensors, which are installed on platforms edge to detect human's presence between train and platform. Proximity sensor is based on permittivity phenomena. Materials, which are found in typical environment, have different permittivity level, as shown in Table 2. There is significant difference between water and other material permittivity. It is more than ten times lower then in metals and higher then in insulators. There is no other material that has the same permittivity level. Sensor uses that difference to recognize objects, which can be found in operation area.

Table 1. Statistic of Accidents in London Underground. Place

Year

Fatality

Type

Cause

Northwood Edgware Stratford Moorgate Holborn Camden Town White City

1945 1946 1953 1975 1980 2003 2004

3 killed Minor injures 12 killed 43 killed Minor injures 6 injured No injured and killed

Collision Hit in the and of tunnel Collision Hit in the and of tunnel Collision Derailment Derailment

Driver failure Death of the driver Signalization failure Unknown Operator's failure Wrong steering system configuration Wrong steering system configuration

Articles

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

Table 2. Permittivity for different materials. Material

Permittivity er

Air Polyethylene Silicon Rubber Paper Water Metals

1,0005 2,25 11,68 7 3,5 80 >1000

2010

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

ò E × dl = 0

(3)

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

Capacitance is directly proportional to permittivity, thus it is possible to detect an object by measuring changes of it. Therefore sensor should be a kind of capacitor. Capacitive sensors are very common in industry, and others applications. The novel is the form of capacitor, which is described on figure1 has the ability to recognize material of which the detected object is made. Mostly capacitive sensors have small detecting area. In this application it would be necessary to use hundreds of those devices in entire hazard area. Most of known sensors are cylindrical, which brings problems with mounting it in the floor, or in the edge of the platform. Only capacitive sensors, developed in GM factory to protect workers from robots, were in the form of wire [5]. Ability to recognize metal from insulators is achieved by some of sensors [6]. There are also systems, which can recognize man from insulators, but only in case, that he is grounded and there are no other grounded objects in operation area [7]. Those conditions cannot be achieved in under-ground. Developed sensor is flat, long and adjustable to curve of the ramp. It minimizes the cost and optimises detecting area. Flat device is easy to install and do not disturb architecture of the station. The purpose was to maximize the sensitivity to changes of permittivity and minimize thickness. In classical flat capacitor two sheets are in front of each other. That configuration would be not advantageous in this case, because it indicates more than three plies of board, or tape to construct sensor and it reduced electric field outside the capacitor. Changes in capacitance of capacitor are observed only under condition that object will be in electric field produced by capacitor what result of following equation is:

Q C= U

N° 3

(1)

Fig. 1. Principles of operation of sensing capacitor. Capacitance is proportional to permittivity, but also other factors have influence on it. In detecting area insulated metal can be found, and grounded objects, like train body. Sensor has to be resistant to that disturbance. Case with insulated metal can be transformed as it is shown in Figure 2. In that case two series connected capacitors are created. Resultant capacitance is:

C1 × C2 C1 + C2

(4)

Equation (4) shows that resultant capacitance is smaller then capacitance of each capacitor. In fact those two capacitors can have bigger capacitance than empty sensing capacitor, because of big sheets area. In conclusion small increase of capacitance can be observed. Insulated metal

where U is given as d

U = ò E × dl

(2)

In equation (2), E is electric potential and dl is differential track. When E = 0 outside of the capacitor, integral given by equation (2) is equal to zero. Therefore electric field should be put outside the capacitor, because object cannot be between capacitors sheets. Technical solution of such specialized capacitor is shown in Figure 1.

Fig. 2. Analyse of the case, where insulated metal is present nearly the sensor. 46

Articles

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

Configuration of two capacitors may be considered, when grounded metal is present nearly to capacitive sensor. That case is presented in Figure 3. Most of stream from sensing capacitor sheets will be absorbed by grounded metal. Only minority of stream will be transferred between sensing capacitor sheets, therefore decrease of capacitance of circuit will be observed. Insulator in neighbourhood of capacitor increases capacitance proportionally to the material permittivity, which is much higher for water, then other materials. Tests with automatic capacitive bridge have proven theoretical conclusions in every case of object. For insulated metal small change of capacitance was observed. In case of grounded metal output signal was lower than for empty capacitor. Insulators were increasing capacitance. Results of these tests are presented in Table 3. Capacitance was measured for 1 kHz and 10 kHz bridge supply frequency.

N° 3

2010

changes of capacitance. Too high quality factor can increase noise level, and cause poor resistance to disturbances. ULC increases when capacitance differs to capacitance of empty capacitor. Higher resistance decreases quality level according to equation:

U LC =

1 ö æ U × ç wL ÷ wC ø è

1 ö æ R 2 + ç wL ÷ w Cø è

(5)

where U is supply voltage, L - inductance, R - resistance of circuit and T - current pulsation. Dependences of ULC calculated for different resistances are presented in Figure 5. Due to use of the resonance circuit, significant changes of the output signal from the sensor are achieved.

Fig. 3. Analyse of the case where grounded metal is presented nearly the sensor. Table 3. Capacitance of sensing capacitor for different materials. Object Empty capacitor Hand Insulated metal Grounded metal Wood

C (pF) 1kHz

C (pF) 10kHz

15,7-16,2 292,0 - 390,0 33,7 - 34,1 15,9 - 16,0 26,7 - 26,8

16,3 229 - 237 33,9 15,3 22,1

Fig. 4. Electrical connections of the sensor (ULC - output signal).

3. Developed capacitive sensor for safety applications Change 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. Resonance can be achieved when capacitor is empty; it means that there is no object in detection area. When resonance occurres voltage drop on LC elements, according to equation (5), is the lowest (ULC). Even small change of capacitance cause significant decrease of voltage drop, which can be observed in Figure 5. Sensors circuit is shown in Figure 4. Frequency of supply current, resistance and inductance has to be matched to achieved resonance of capacitance of empty capacitor. Significant is also quality factor of circuit. It can increases sensibility of circuit to

Fig. 5. Dependence of the calculated sensor output signal ULC as a function of changes of its capacitance, calculated for different values of resistance R.

4. Experimental set-up and results Main part of sensor is consisted of two metal sheets, which are on the same plane. Areas of the sheets and Articles

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N째 3

2010

Table 4. Experimental results.

Object Empty Hand Hand distance Insulated metal Grounded metal Wood

32 V / 928 Hz

100 V / 860 Hz

ULC [V] range 1 V

ULC [V] range 10 V

0,404 5-10 mm

0,340 0,573 1-5 mm 0,301 0,264 0,345

1,800 0 mm

ravine between them have to be optimised. Sensor was made on printed board. As a result an influence of differences of shape and ravine during mounting and measuring was decreased. Such differences may occur in the case of metal leaf. Moreover, there was insulation on copper ply to prevent break down. Model of the sensor shown in Figure 6 has been tested in laboratory, for different types of materials, and configurations. Testing circuit was shown in Figure 7. Accuracy of resistance, inductance and capacitance has not influence on detecting ability, thus current frequency is adjustable and it is only needed to adjust resonance for empty capacitor.

1,11 5-10 mm

0,95 1,20 1-5 mm 1,04 0,94 1,05

2,14 0 mm

For configuration shown in Figure 7 resonance frequency was 928 Hz for 32 V supply and 860 Hz for 100 V. During experiment the reaction of four different types of object was tested: insulated and grounded metal, for wood and human hand. In case of human hand signals for different distance were measured. Driving properties and results of the tests were shown in Table 4. Resonance frequency changes due to supply voltage. This is because of internal capacitance and resistance of generator.

Fig. 6. Model of the sensing capacitor.

Fig. 8. Output signal as a function of distance of human hand, measured for different power supply voltage. Changes of output due to object material are very significant. They are 17% higher from wood for hand in distance up to 10 mm and 421% higher for distance 0 mm in case of 32 V supply. Results confirm that sensor is able to detect living tissue. Relation of signal to distance for human hand is shown in Figure 8. Shape of function does not change with supply value, thus smaller supply voltage can be used.

4. Conclusion

Fig. 7. Schematic diagram of the experimental set-up. 48

Articles

Proximity sensor described in this paper is able to achieve the requirements. On the other hand, ability to detect in distance is not sufficient. Probably it can be improved by increasing frequency, which can straighten electric field propagation in detection area. Good feature, which was observed by experiment, is fact, that detecting range is not related to supply voltage. It can decrease probability of break down, and electric shock risk.

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 4,

N° 3

2010

Experimental results indicate, that it is possible to develop a simple sensor, which not only detect objects, but also recognizes material. This kind of sensor is a good alternative for sensing mat, which uses stress to detect an object. As compared with safety mats [8], used in industry to detect human presence in danger area, developed sensor is thinner, more adjustable and does not depend on stress, which can provoke uncertainty in operation. To detect presence of man's feet on the floor high detection range is not needed. Therefore described sensor can be used in many cases to detect human presence in hazardous area.

AUTHORS Piotr Frydrych* - Institute of Metrology and Biomedical Engineering, Warsaw University of Technology, św. A. Boboli 8, 02-525 Warszawa, Poland, tel. 0-22-2348471, e-mail: piotrfrydrych@gmail.com. Roman Szewczyk - Institute of Metrology and Biomedical Engineering, Warsaw University of Technology, św. A. Boboli 8, 02-525 Warszawa, Poland. Industrial Research Institute for Automation and Measurements, Al. Jerozolimskie 202, 02-486 Warszawa, Poland, tel. 0-22-8740171, e-mail: rszewczyk@piap.pl. * Corresponding author

References [1] [2] [3] [4] [5] [6]

[7]

[8]

Internet site www.metro.waw.pl. Instruction SOP-2 Urządzenia nadawcze DTR-94/SOP2/1, Katowice 2002. (in Polish) Internet site en.wikipedia.org/wiki/List_of_London_ Underground_accidents. 120 killed on South Korea underground as 'arsonist' attacks train, 19th Feb, 2003. www.independent.co.uk. „Protecting Workers from robots”, American Machinist, March 1984, pp. 85-86. “A Combined Inductive-Capacitive Proximity Sensor and Its Application to Seat Occupancy Sensing”. In: International Instruments and Measurement Technology Conference Singapore, 5th-7th May 2009. George B., Zangl H., Bretterklieber T., “A Warning System for Chainsaw Personal Safety based on Capacitive Sensing”. In: IEEE Sensors Conference, 26th-29th October 2008, pp. 419-422. www.safetymat.net.

Articles

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 3

2010

SLIDING MODE SPEED CONTROL FOR MULTI-MOTORS SYSTEM Received 16th March 2010; accepted 1st June 2010.

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

Abstract: Continuous processes in the plastics, textile paper and other industries, require several drives working in synchronism. The aim of this paper is to control speed of the multi motors system, and to maintain a constant mechanical tension between the rollers of the system. Several controllers are considerer, including Proportional-Integral (PI) and sliding-mode control (SMC). Since the PI control method can be applied easily and is widely known, it has an important place in control applications. But this method is insensitive to parameter changes. The advantage of an SMC is its robustness and ability to handle the non-linear behaviour of the system, and is indicated in comparison with traditional proportional-integral (P1) control scheme. Theoretical analysis and simulation results are provided to evaluate the consistency and performances of this control technique (SMC). Keywords: Multi-Motors systems, sliding mode control, proportional-integral (P1) control.

1. Introduction Paper industries require several drives for paper processing. In the past, a single mechanical line shaft was used for all drives. Nowadays independent drives are used. Electronic synchronization has to be ensured for quality of produced paper rolls. Several control strategies have been suggested based on robust control or electronic emulation of mechanical line shaft [1]. Since the decentralized PI control method can be applied easily and is widely known, it has an important place in control applications, where many industrial web transport systems have used this type of controllers [2]. But this method is insensitive to parameter changes. A nonlinear decoupled control is designed for multi-motors multi motors system. At the first, an ideal feedback linearization control system is adopted in order to decouple the tensions and velocity of the web winding system is presented in [3]centralized and decentralized fixed order H¥ controller results with model based feed-forward for multi motors systems which provide improved the tension and velocity regulation is presented in [4]. In this work the design of sliding-mode (SMC) to control a multi motors system are proposed in order to improve the performances of the control system, which are coupled mechanically, and Synthesis of the robust control and their application to synchronize the five sequences and to maintain a constant mechanical tension between the rollers of the system [6]. The advantage of an SMC is its robustness and ability to handle the non-linear behaviour of the 50

Articles

system. The model of the multi-motors system and in particular the model of the mechanical coupling are developed and presented in Section II. Section III shows the development of sliding mode controller’s design for winding system. The proposed structure of the studied propulsion system is given in the section IV. Simulation results using MATLAB SIMULINK of different studied cases. Finally, the conclusions are drawn in Section V.

2. System Model's In the mechanical part, the motor M1 carries out unreeling, M3 drives the fabric by friction and M5 is used to carry out winding, each one of the motors M2 and M4 drives two rollers via gears “to grip” the band (Fig.1). Each one of M2 and M4 could be replaced by two motors, which each one would drive a roller of the stages of pinching off. The elements of control of pressure between the rollers are not represented and not even considered in the study. The stage of pinching off can make it possible to isolate two zones and to create a buffer zone. [6,7]. 2 ì æ æ ö æ Lm ö di 1 ç ds ç ï = - Rs + çç ÷÷ × Rr ÷ × ids + sLswe iqs + L ÷ ï dt s × Ls ç ç è Lr ø ø è è ï ï ö Lm × Rr L × fdr + m × fqr × wr + Vds ÷÷ ï 2 Lr Lr ï ø ï 2 æ ö æ Lm ö ï diqs 1 æç ç ÷ ï dt = s × L ç - sLswe ids - ç Rs + çç L ÷÷ × Rr ÷ × iqs - L s r è ø ï è ø è ï ö Lm L ×R ï × fdr × wr + m 2 r × fqr + Vqs ÷÷ í Lr Lr ø ï ï df L ×R R ï dr = m r × ids - r × fdr + we - wr × fdr Lr Lr ï dt ï ï dfqr = Lm × Rr × i - w - w × f - Rr × f qs e r dr qr ï dt Lr Lr ï 2 ï dwr P .Lm f P . iqs .fdr - ids .fqr - c .wr - .Tl = ï Lr .J J J ï dt ï (1) î

(

)

Where s is the coefficient of dispersion and is given by:

s = 1-

L2m Ls Lr

(2)

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 3

2010

C. Mass conservation law Consider an element of web of length L = L0 (1+e). With a weight density r, under an unidirectional stress. The cross section is supposed to be constant. According to the mass conservation law, the mass of the web remains constant between the state without stress and the state with stress

r 1 = r SL = r 0 SL 0 Þ r0 1+ e

(4)

D. Tension model between two consecutive rolls. The equation of continuity, cf. [8], applied to the web gives: Fig.1. Five motors web transport system. The tension model in web transport systems is based on Hooke’s law, Coulomb’s law, [8], [9] mass conservation law and the laws of motion for each rotating roll. A. Hooke’s law The tension T of an elastic web is function of the web strain e

T = ESe = ES

L - L0 L0

(5)

By integrating on the variable x from 0 to Lt (cf. Fig. 2), taking into account (4), and using the fact that a + g << L, we obtain

d æ L ö V V çç ÷÷ = 1 - 2 . dt è 1 + e 2 ø 1 + e 1 1 + e 2 Therefore:

(3)

Where E is the Young modulus, S is the web section, L is the web length under stress and L0 is the nominal web length (when no stress is applied). B. Coulomb’s law The study of a web tension on a roll can be considered as a problem of friction between solids, see [8] and [9]. On The roll, the web tension is constant on a sticking zone of arc length a and varies on a sliding zone of arc length g (cf. Fig. 2, where Vk(t) is the linear velocity of the roll k). The web tension between the first contact point of a roll and the first contact point of the following roll is given by:

e ( x , t ) = e 1 (t ) if x a e ( x , t ) = e 1 ( t ) e m ( x - a ) if a £ x £ a + g e ( x , t ) = e 2 (t ) if a + g £ x £ L t

¶r ¶ ( rV ) + = 0 ¶t ¶x

de 2 (1 + e 2 ) 2 = V1 - V2 (1 + e 2 ) . dt 1 + e1

-L

(6)

This equation can be simplified by using the approximation

e 1 << 1 and

e 2 << 1

(1 + e 2 ) » (1 - e1 )( 1 + 2 e 2 ) 1 + e1

(7)

And using Hook’s law, we get:

dTk @ ES (Vk - Vk -1 ) + Tk -1Vk -1 dt - Tk (2Vk -1 - Vk ) . Lk -1

(8)

k = 2, 3, 4, 5. (1.51) (1.52) (1.53)

Where μ is the friction coefficient, and Lt = a + g + L. The tension change occurs on the sliding zone. The web velocity is equal to the roll velocity on the sticking zone. (1.54)

where Lk-1 is the web length between roll k-1 and roll k, Tk is the tension on the web between roll k-1 and roll k, Vk is the linear velocity of the web on roll k, Wk is the rotational speed of roll k, Rk is the radius of roll k, E is the Young modulus and S is the web section. E. Roll velocity calculation The law of motion can be obtained with a torque balance:

d ( J k Wk ) = Rk (Tk +1 - Tk ) + Cemk + C f dt

(9)

Where Wk = Vk/Rk, is the rotational speed of roll k Cemk is the motor torque (if the roll is driven) and Cf is the friction torque.

Fig. 2. Web tension on the roll.

F. State space representation The nonlinear state-space model is composed of (10) for the different web spans and of (11) for the different rolls. Under the assumption that Jk Rk (k = 2, 3, 4, 5) is varying only slowly, which is the case for thin webs, Vk can be chosen as a state variable in (11), leading to the Articles

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 3

2010

following linear model: ·

Em X = A(t ) X + BU Y = C (t ) X

(10)

where XT = [T2 T3 T4 T5 J1T(t)W1 J2W2 J3W3 J4W4 J5T(t)W5] YT = [T2 T3 T4 T5 V1] ,

(11)

U = [u1 u2 u3 u4 u5]

(12)

3. Design of sliding mode speed and current controllers A Sliding Mode Controller (SMC) is a Variable Structure Controller (VSC). Basically, a VSC includes several different continuous functions that can map plant state to a control surface, whereas switching among different functions is determined by plant state represented by a switching function [8], [9]. To control the speed of the induction machine, the sliding surface is defined as follows:

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

S1 =i ds* -i ds

(20)

S (wm ) = wm* - wm

S 2 = i *qs - i qs

(21)

(13)

The derivative of the sliding surface can be given as:

S& (wm ) = w& m* - w& m

ok?

(14)

For the indirect field-oriented control (IFOC) tuning parameters we need two surfaces S1 and S2 the first for the ids regulator and the second for iqs regulator respectively where: [10,11]

The derivate of S1 can be given as:

S&1 =i& ds* -i& ds

From equation (1) and (27 ) we can obtain: Taking into account the mechanical equation of the induction motor defined in the system of equations (1), the derivative of sliding surface becomes

æ P ² L m f dr* f P ö S& (wm ) = w& m * -çç .i qs - c wm - C r ÷÷ J J ø è JL r

(15)

æL ö ö 1 æç S&1 =i&ds* -(R s + Rr çç m ÷÷ 2 ÷÷i ds + w s .i qs ç s .L s è è Lr ø ø L .R 1 + m r2 .f r*+ V ds) s .L s L r s .L s

The current control is given by: The virtual voltage controller Vds is given by:

iqs* = iqseq + iqs

(16)

To avoid the chattering phenomenon produced by the Sign function we use the Saturation function Sat in the discontinuous control defined as follow:

ì S ; if ï f æSö ï sat çç ÷÷ = í æSö èf ø ï Signçç ÷÷ ; if ï èf ø î

S <1 f S >1 f

(17)

Vdsn = k1 × sat (s1 f1 )

- w s .i qs -

kiqs : Positive constant.

iqseq =

JLr P Lmfdr* 2

P ö æ * fc ç w& m + wm + Cr ÷ J J ø è

(23)

According to Lyapunov stability criteria [10], our speed loop system's stable if: S1 S1<0 by means that K1 is positive constant. The equivalent control Vdseq is given as:

V dseq= s .Ls (i ds* + (18)

(22)

The voltage discontinuous control Vqsn is defined as:

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

iqsn = kiqs × sat (s (w) fw )

V ds*=V dseq+V dsn.

1 s .L s

æ æ ö ö ç R S + R r .ç L m ÷ 2 ÷.i ds çL ÷ ÷ ç è rø ø è

L m .R r * f r) s .L s L 2r

The derivate of can be given as: (19)

S&2 =i& qs* -i& qs From equation (1) and (34) we can obtain:

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

Articles

(24)

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

1 S&2 (iqs ) = i&qs* - (- w s i ds sL s

N° 3

2010

system follows the reference speed after 1 sec, in all motors, however, in the PI controller the system follows after 2 sec. Fig (9) shows the phase plot of the sliding surface(s) of SMC control.

æ æ ö ö ç R s + R r .ç L m ÷ 2 ÷.i qs - L m .f *r .wm + 1 V qs) ç ÷ ÷ ç sL s L r sL s è Lr ø ø è The voltage controller Vqs is given by:

Vqs* = Vqseq + Vqsn

(25)

The Vqseq equivalent control actions defined as: * Vqseq = s .L s .(i& qs lim + w s .i ds

1 s .L s

æ æ ö ö ç R s + Rr ç L m ÷ 2 ÷.i qs + L m f r*wm ) çL ÷ ÷ ç sL s L r è rø ø è

(26)

The voltage discontinuous control Vdsn is defined as:

(

Vqsn = k 2 × sat s2 f 2

)

(27)

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

For the same reason condition of K1, K2 are positives constant.

4. Simulation results The winding system we modelled is simulated using MATLAB/SIMULINK software and the simulation is carried out on 10s. To evaluate system performance we carried out numerical 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 controller. It appears clearly that the classical control with PI controller is easy to apply. However the control with sliding mode controllers offers better performances in both of the overshoot control and the tracking error.

Fig. 6. The linear speed of winder M5. As shown in Fig (4-6). An improvement of the linear speed is registered, and has follows the reference speed for both PI controller and SMC control, but in case of PI controller, the overshoot in linear speed of Unwinder is 25%. Figs (4-6) show that with the SMC controller the system follows the reference speed after 1 sec, in all motors, however, in the PI controller the system follows after 2 sec. Fig (9) shows the phase plot of the sliding surface(s) of SMC control.

Fig. 4. The linear speed of unwinder M1. As shown in Figs (4-6). An improvement of the linear speed is registered, and has follows the reference speed for both PI controller and SMC control, but in case of PI controller, the overshoot in linear speed of Unwinder is 25%. Figs (4-6) show that with the SMC controller the

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

Articles

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

5. Conclusion The objective of this paper consists in developing a model of a winding system constituted of five motors that is coupled mechanically by a strap whose tension is adjustable and to develop the methods of analysis and synthesis of the commands robust and their application to synchronize the five sequences and to maintain a constant mechanical tension between the rollers of the system. The simulations results show the efficiency of the SMC controller technique, however the strategy of SMC Controller brings good performances, and she is more efficient than the classical PI controller.

[6]

[7]

[8]

[9]

AUTHORS Bouchiba Bousmaha* - An assistant teacher at University of Bechar, Bechar, Algeria From 2007 right now he's preparing his PhD degree in multi machine system control. E-mail: bouchiba_bousmaha@yahoo .fr. Hazzab Abdeldjabar - Professor of electrical engineering at University of Bechar, Bechar, Algeria. Hachemi Glaoui - M.Sc. From 2009 right now he's preparing his PhD thesis in multi machine system control. Ismail Khalil Bousserhane - Professor of electrical engineering at University of Bechar, Bechar, Algeria. Pierre Sicard - Preceived the Master degree in industrial electronics from the University of Quebec in Trois-Rivieres, Trois-Rivieres, Canada, in 1990, and a Ph.D. degree in electrical engineering from Rensselaer Polytechnic Institute, Troy NY, USA in 1993. He is professor in electrical and computer engineering at University of Quebec in Trois-Rivieres where he is director of the Research group on industrial electronics. His research interests include the macroscopic energetic Representation, multi-drives control and the rolling unrolling system control. * Corresponding author

References [1]

[2]

[3]

[4]

[5] 54

Thiffault Ch., Pierre Sicard P., Alain Bouscayrol A., Tension Control Loop Using a Linear Actuator Based on the Energetiv Macroscopic Representation CCECE 2004CCGEI 2004, Niagara Falls, May 2004 Xu Y., , Wang D., Zhang Q., “Modeling and Robust Control of Web Winding System with Sinusoidal Tension Disturbance”. In: Proceedings of the 2006 IEEE International Conference on Mechatronics and Automation, 25th - 28th June, 2006, Luoyang, China, pp. 1958-1963. Abjadi N.R., Soltani J., Askari J., Navid R.A., Jafar .S, Javad A., ”Nonlinear Sliding-Mode control of a MultiMotors web winding system without tension sensor”. In: 2008 IEEE international conference on industrial technology - ICIT 2008, pp. 1-6.) . Knittel D., Vedrines M.H., Pagilla D., Prabhakar “ Robust H? Fixed Order Control Strategies for Large Scale Web Winding Systems”. In: Proceedings of the 2006 IEEE International Symposium on Intelligent Control, 4th-6th Oct. 2006, Munich, Germany, DOJ 10.1109/CACSD-CCA-ISIC.2006.4776941 Koc H., “Modelisation et commande robuste d'un

Articles

[10]

[11]

N° 3

2010

system d'entrainement de bande flexible” , Ph.D. thesis, Universite Louis Pasteur (Strasbourg I University), 2000. Glaoui H., Fuzzy sliding mode control MIMO for system multi motors , Master thesis, University of Bechar, Algeria 2008. Jung J., Nam K., “A Dynamic Decoupling Control Scheme for High-Speed Operation of Induction Motors”, IEEE Trans. on Ind. Elect., vol. 46/01, 1999. Mezouar A., Fellah M.K, , Hadjeri S., ”Adaptive Sliding Mode Observer for Induction Motor Using Two-Time Scale Approach”, Electric Power System Research, vol. 77, issues 5-6, 2007, pp. 1323-1336. Mohanty K.B., “Sensorless sliding mode control of induction motor drives”, TENCON-2008, IEEE Region 10 Conference, Hyderabad. Zhiwen M., Zheng T., Lin F., You X., “A New SlidingMode Current Controller for Field Oriented Controlled Induction Motor Drives”. In: IEEE Int. Conf. IAS, 2005, pp. 1341-1346. Koshkouei A.J., Burnham K.J., Zinober A.S.I., ”Dynamic Sliding Mode Control Design”, IEEE Proc.--Control Theory Appl., vol. 152, no. 4, 2005, pp. 392-396.

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 3

2010

MODELLING AND OPTIMIZATION OF THE FORCE SENSOR NETWORK Received 13th January 2010; accepted 26th April 2010.

Grzegorz Bialic, Marcin Zmarzły, Rafał Stanisławski

Abstract: The power boiler erection process requires temporary suspending of buckstays (Fig. 1). The aim of buckstays installation is to brace screens of combustion chamber. Suspending of buckstays is performed by means of strand jacks and steel rods. The amount, location and force ranges of each rods are modelled on the basis of static structural strength computations. In the paper the force sensor network incorporated into computer and wireless communication system designed to prevent overloading of the rods resulting from force asymmetry or computational faults is introduced and next subjected to optimization procedure. The operational data delivered by the network and recorded on the system hard drives let the authors perform the correlation analysis and linear regression to reduce the number of sensors. Thus two stage hierarchic algorithm which constructs the set of models for every single sensor and estimates their parameters, and then using genetic procedure minimizes a certain loss function to automate the sensor network optimization process is introduced. As a result, such investigation could significantly reduce cost of the whole system. Keywords: sensor network optimization, mathematical modeling.

1. Introduction In 26 on October 2007 on one of the biggest construction site in Europe in Grevenbroich - Neurath in North Rhine-Westphalia in Germany a building accident took place. Three people were killed and another five workers were injured. The accident was caused by faults in structural strength computations. To avoid such terrible results of designers faults in the future the force sensor network [6] for online force monitoring in steel rods was designed and built in cooperation with Remak S.A.

buckstays

Fig. 1. Support construction of power boiler with temporary suspended buckstays.

The network was applied during temporary suspending of backstays procedure on the block G in Neurath. In the first section the architecture of the measuring system incorporated into force sensor network is introduced. In second and third sections the historical, operational data are investigated using correlation analysis, linear regression and parametric optimization of a certain loss function to reduce the number of required sensor units.

2. Force sensor network measuring system The force measuring system (Fig. 2) consists of 24 calibrated measuring units (Fig. 3) which make the sensor network. Each of them is based on a measurement foot designed so as to move part of a known force axially to tensometer sensor. The voltage across tensometer bridge is measured in the measuring transducer. The first element of that transducer is measuring amplifier with analog-digital converter with Sigma-Delta modulation performing measurements with an accuracy of 10 bits. Then, the result of measurement is filtered and scaled using the microprocessor unit. Scaling is based on the data from the calibration procedures [1]. The measured force is compared with the threshold values derived from the maximum load for the sensor (150%) and the values of the maximum acceptable load for the rod. The measurement data are available from SLAVE module by means of communication protocol with the physical layer interface based on RS485. The MASTER unit is working in the pulling mode. Each of the 24 measuring units in the network replies with the information about the current value of measured force and short-term exceeded limits. MASTER module provides further aggregated data via the industrial radio modem operating in the 869 MHz band. This enables the remote measurement by the system installed on mobile cross-bar which is raised to a height of 160 m. Visualization software is installed on the operating unit and it has been made in technology JavaSE, where a number of procedures is responsible for: the processes of receiving data, visualization, alarms detection, recording and transmission of the historical data to the Web server [7], generating reports and analysis of historical data trends. The use of the Java platform [2] provides user-friendly tool what is an important advantage in daily operation of the system. Java platform is characterized by an open architecture. It is a set of standards used by many software companies, which guarantees extensive support for this technology in the future. Detected alarms and warnings are indicated acoustically by an independent microprocessor system. Articles

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N째 3

2010

Data visualized online allow continuous monitoring and regulation of stress in rods. Fig. 4 presents a fragment of the force decomposition characteristics which describes the suspending process. The technology used in the assembly process allows to correct the load only on currently installed buckstay. Suspending of consecutive buckstay prevents the further load revision at a higher levels. The number of measurement units has the most significant impact on the cost of the whole force indication system. Therefore it appeared very important the optimization of the network by reducing a certain number of sensors. Reduced sensors have to be replaced by their models based on other measurements. In the next sections the correlation analysis, modelling using linear regression and minimization of a certain objective function in the sense of relative mean square error rate are used in this purpose.

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

3. Correlation analysis The correlation analysis specifies the correlation coefficients among all sensors. Then one can determine whether the particular sensor can be directly replaced with the another one. Table 1 shows the correlation coefficients r between the sample sensor (S2) and remaining sensors. It is easy to observe that the correlation coefficients are too low to replace the sensor S2 by the other one. Similar results (on the average) have been obtained for remaining sensors. On the other hand, the results of correlation analysis indicate that signal for the particular sensors are strongly dependent on each other. Therefore using more advanced algorithm to reduce the number of sensors one can obtain satisfactory results.

Fig. 2. Block diagram of the system.

4. Two stage network modelling and optimization In this section two stage algorithm (consists of modelling stage and next optimization stage) which mini-

Fig. 3. Measurement unit.

Table 1. The correlation among the measurements for sensor S2 against the others.

sensor nr r

1 0.667

3 0.580

4 0.661

5 0.679

6 0.672

7 0.639

8 0.674

89 0.109

sensor nr r

10 0.629

11 0.523

12 0.669

13 0.631

14 0.042

15 -0.17

16 0.652

17 0.677

sensor nr r

18 0.612

19 -0.24

20 0.676

21 0.071

22 0.6534

23 0.567

24 -0.02

25 0.641

Articles

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

mizes the number of sensors was employed. The hierarchical block diagram of this method is presented in the Fig. 5.

Fig. 5. Block diagram of the two stage algorithm which minimizes the number of sensors. A. Sensors modelling and estimation First stage involves modelling of Sp element by means of a linear combination of two others sensors. The equation of model for the Sp element can be written in following form (1) where: a0p , akp , alp are the unknown model coefficients and yik , yil are the outpus of the Sk, Sl sensors in i-th sample (k,l=1,…,24). The model equation for the Sp unit 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

N° 3

2010

B. Loss function minimization Assuming above presented methodology one can obtain the set of models for each element of the network. Every model imitates a certain element of the network with some accuracy. Certainly only the model which produce the lowest value of MSE could be selected to replace the sensor in the network. This in turn means that the replaced before sensor cannot be further used to model remaining elements of the system. As a result, the optimization of the sensor selection process has to be performed. The selection procedure should follow the guidelines: replace as many sensors as possible, simultaneously keeping the lowest MSE rate of the model. Thus, there was required to develop the tool for optimization the sensor selection process. The amount of the sensors n which are supposed to be reduced was defined as the input of the routine. Finally the algorithm returns the indices of the sensors which should be replaced by the model on the basis of minimization of the following loss function (7) where p={p1, …, pn} and n is the number of reduced sensors. The complexity of the above defined loss function (7) (large number of local minimums) decided that the genetic algorithm (GA) was chosen for finding the optimal solution. GA parameters were as follows: population size: 20, maximum generation: 100, selection function: stochastic uniform, cross-over function: Heuristic with cross-over probability 0.9 and mutation probability: 0.1. The stochastic nature of the genetic algorithm caused that the optimization was started from different initial conditions. Then form the set of the solutions the most optimal result was selected. The properties of models used for sensor replacement for the number of reduced elements n=4 are shown in Table 2.

y p = [ y1p y2p ... y Np ] and F k ,l = [j1k ,l j 2k ,l ... j Nk ,l ]

Table 2. The properties of reduced sensors models.

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

modelled sensor

MSE [%]

1 4 13 20

4 7 10 17

17 13 16 22

0.9954 0.9959 0.9950 0.9968

4.74 3.72 4.58 4.38

(5) Using this method for each element Sp (p= 1,..,24) the set of models was developed and constructed for each pair yl, yk (k, l = 1 ,…, 24) where l ¹ k and l, k ¹ p. Among all these models for the element Sp best fitted model in terms of relative mean square error value (rMSE) should be chosen. Relative MSE is calculated by means of the following equation

(6)

In the Table 3 values of these models parameters were collected. Table 3. Values of the models parameters. modelled sensor a0

1 4 13 20

0.5202 0.8150 0.3382 0.4892

0.6700 0.302 0.8331 0.4652

-15.6591 3.6157 -21.0073 -2.1117

Articles

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

The assumption that n=4 results in the MSE rate below 5% and then keeps total functionality of the considered system. In the Fig. 6 current value of sensor S4 vs. defined by the model illustrates pretty good performance of the methodology, which has been introduced in the paper.

Fig. 6. Plots of actual vs. estimated outputs of the model for the sensor S4. In the Fig. 7 the impact of the reduced sensors number n on the quality of estimation was illustrated. Thus the values of the mean square error (J/n), and the maximum value of the MSE (max(MSE)) as a function of n were plotted in this figure. The maximum value of the MSE reflects the estimation error for the sensors with the worst fit model. 30,00

MSE rate [%]

25,00 20,00

2010

5. Conclusion Application of force sensor network during suspending of buckstays has improved the safety of the whole erection procedure. The electronic measurements in real time and recorded data let engineers symmetrize the force and verify values modelled before. Using presented in the paper two stage algorithm based on correlation analysis, linear regression and parametric minimization of a certain loss function, gives them the opportunity to reduce the number of measurement units in the network which significantly diminishes the costs of the system. It was presented in the paper that some sensors could be replaced with computationally modelled data and then error (in the mean square sense) of the whole force indication system leaves in reasonable range. Moreover, it is worth emphasizing, that the presented above solution discovers new fields for applying modern measurement systems and microprocessor, computer and information technology.

AUTHORS Grzegorz Bialic* - Department of Electrical Engineering, Automatic Control and Computer Science, Opole, University of Technology, 31 K. Sosnkowskiego St., PL45-272, Opole, Poland. Tel. +48 77 4006208; fax: +48 774006338. E-mail: bialic@sprinter.com.pl. Marcin Zmarzły - Department of Electrical Engineering, Automatic Control and Computer Science, Opole, University of Technology, 31 K. Sosnkowskiego St., PL45-272, Opole, Poland. Tel. +48 77 4006208; fax: +48 774006338. E-mail: marcin@zmarzly.com.pl. Rafał Stanisławski - Department of Electrical Engineering, Automatic Control and Computer Science, Opole University of Technology, 31 K. Sosnkowskiego St., PL45-272, Opole, Poland. Tel. +48 77 4006208; fax: +48 774006338. E-mail: r.stanislawski@po.opole.pl. * Corresponding author

15,00 10,00

References [1]

5,00 0,00

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

[2]

Fig. 7. Dependence of MSE and max(MSE) on the reduced sensors number. Certainly, the number of reduced elements depends on the assumed error rate. The larger acceptable error, the more elements can be reduced. For example, for relative MSE < 7%, 6 sensors may be replaced by its models. In order to preserve the functionality of the designed system the error rate of every single modeled sensor should be kept in range of 10%. From this point of view the reduction of 6 measurement units seems to be the most desired.

[3] [4] [5] [6]

[7] 58

N° 3

Articles

Bialic G., Zmarzły M., Szmechta M., Stanisławski R.,“Proces kalibracji głowic pomiarowych dla systemu pomiaru sił wykorzystywanego podczas montażu kotła energetycznego 1100 MW”, Pomiary - Automatyka Kontrola, no. 2, 2009, pp. 111-113. (in Polish) Bialic G., Zmarzły M., Stanisławski R., “Design of the scales for the power boiler fuel feeding system based on the process identification”. In: 14th IEEE IFAC International Conference on Methods and Models in Automation and Robotics. MMAR'2009, Międzyzdroje, 19th-21st August 2009. java.sun.com, Java Platform Enterprise Edition, Specification v5.0, 2005. Ljung L., System Identification, Prentice-Hall, Englewood Cliffs, NJ, 1987. Söderström T., Stoica P., System Identification, Prentice-Hall, Englewood Cliffs, NJ, 1989. Sposób pomiaru siły obciążenia cięgna oraz głowica do pomiaru obciążenia cięgna, Patent application no. P386137, Patent Office RP, Warszawa 2008. (in Polish) Sposób kontroli poziomu i urządzenie do kontroli pozio-

Journal of Automation, Mobile Robotics & Intelligent Systems

[8]

[9]

[10]

VOLUME 4,

N° 3

2010

mu elementów wiszących dużych konstrukcji, Patent application no. P.389559, Patent Office RP, Warszawa 2009. (in Polish) Stanisławski R., Bialic G., Zmarzły M., „Identyfikacja własności dynamicznych układu przygotowania paliwa kotła energetycznego BP-1150”, Pomiary - Automatyka - Kontrola, 2009 vol. 2, pp. 108-110. (in Polish) Zmarzły M., Bialic G., Stanisławski R., Rogala A., „System pomiaru sił dla procesu wciągania bandaży na konstrukcję kotłów energetycznych”. Pomiary - Automatyka Kontrola, 2009 vol. 2, pp. 114-116. (in Polish) Zmarzły M., Szmechta M., “The efficiency and reliability analysis of a telemetric event driven data transmission over GPRS. In: 5th International Conference New Electrical and Electronic Technologies and Their Industrial Implementation, Zakopane, 12th-15th June 2007.

Articles

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 4,

N° 3

2010

INFocus THE SPOTLIGHT on new by Urszula Wiączek - report from Paris

n 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 accept humanoid robots as movie heroes like R2D2 or C3P from “Star Wars”. But nobody could imagine robots as rock stars! Alas! On February 19th in Seattle`s Key Arena five huge industrial robots equipped with five large LED screens debuted on the stage accompanying Bon Jovi`s “The Circle Tour”. The concept of moving screens is a RoboScreenTM patented technology developed by Andy Flessar who is the founder and president of Robotic Arts of Las Vegas in Nevada. He started to work on robots in mid 1990s and in 2006 completed a robotic programming, design and operation programme at the ABB in Auburn Hills in Minnesota. Previously he constructed outdoor panels 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 then Robot 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 IRC5 controller 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. They were overwhelmed by elegant choreography and the visionary application of the installation choreographed precisely with the music. “The collaboration with Robotic Arts and Bon Jovi is certainly one of the most unique applications we have been involved” 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-panels arranged in a six column by four-row grid. They are positioned toward the back of the stage and are an integral part of the concerts displaying approximately 85% real-time video footage of the show from multiple cameras set up on the stage and in the audience and pre-programmed 3-D computer animations. Their movements are coordinated at 30 frames per sec using the time code to synchronize the real time to the beat of the music. During the show the robots are operated by Gordon “Gordo” Hyndford.

In the spotlight

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 4,

N° 3

2010

The panels are covered by plexi glass to protect their surface, when they are positioned horizontally creating steps and a 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 the performance and help present the complete entertainment experience that is synonymous with the Bon Jovi brand” said Flesssas. Tait Towers, the tour production company in Lititz, in Pennsylvania, did the robots design and the system that allows the 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

In the spotlight

61 ?

VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 3

2010

EVENTS SUMMER-AUTUMN 2010 July 9 – 11

ICCDA 2010 – IEEE International Conference on Computer Design and Applications, Qinhuangdao, China. http://www.iccda.org/

August 1–3

ICEIE 2010 – IEEE International Conference on Electronics and Information Engineering, Kyoto Japan. http://www.iceie.org

September 8 – 10

ABSRC 2010 – Advances in Business-Related Scientific Research Conference, Olbia, Sardinia, Italy. http://www.absrc.org/

9 – 12

ABSRC 2010 – 11th Country Robotic Conference, Karpacz, Poland. http://kkr11.iiar.pwr.wroc.pl/

10 – 12

ICMET 2010 – International Conference on Mechanical and Electrical Technology, Singapore, Singapore. http://www.icmet.ac.cn/

17 – 19

ICEIT 2010 – IEEE International Conference on Educational and Information Technology, Chongqing, China. http://www.iceit.org/

October 3–5

ICSTE 2010 – 2nd IEEE International Conference on Software Technology and Engineering, San Juan, Puerto Rico. http://www.icste.org/

17 – 22

ESF-EMBO Symposium – Functional Neurobiology in Minibrains: From Flies to Robots and Back Again, Sant Feliu de Guixols, Spain. http://www.esf.org/conferences/10324

18 – 19

ICINA 2010 – IEEE International Conference on Information, Networking and Automation, Kunming, Yunnan, China. http://www.icina.org/

26 – 29

MobiCPS 2010 – 1st IEEE International Workshop on Mobile Cyber-Physical Systems, Xi'an, China. http://www.cpschina.org/mobicps

November

12 – 14

ICSTE 2010 – 2nd International Conference on the Roles of the Humanities and Social Sciences in Engineering, Pulau Pinang, Malaysia. http://publicweb.unimap.edu.my/~icohse2010

16 – 18

ICMCE 2010 – International Conference on Measurement and Control Engineering 2010, Chengdu, China. http://www.icmce.org/

16 – 18

ICCEE 2010 – 3rd IEEE International Conference on Computer and Electrical Engineering, Chengdu, China. http://www.iccee.org/

Events