Online Adaptive Fuzzy Logic Controller Using Genetic Algorithm and Neural Network for Networked Control Systems

1 Online Adaptive Fuzzy Logic Controller Uing Genetic Algorithm and Neural Network for Networked Control Sytem Pooya Hajebi*, Seyed Mohammad Taghi AlModarrei* *Electrical and Computer Engineering Department, Yazd Univerity, Yazd, Iran Hajebi@tu.yazduni.ac.ir, mta@yazduni.ac.ir Abtract Networked Control Sytem are ued for controlling remote plant via hared data communication network uch a Ethernet. Thee ytem have found many application in indutrial, medical and pace cience field. However there are ome drawback in thee ytem, which make them challenging to deign. One of the mot common problem in thee ytem i the tochatic time delay. Packet witching in internet bring about the randomly varying delay time and conequently make thee ytem intable. Convenient controller uch a PID and PI type controller which are jut match with a contant delay time could not be a olution for thee ytem. Fuzzy logic controller due to their none-linear characteritic which i compatible with thee ytem are potentially a wie option for their control purpoe. Fuzzy logic controller could become adaptive by mean of neural network and beneficial to deal with the varying time delay problem. Further, they do have more capabilitie to tackle packet dropout and dynamically ytem variable. Thi paper introduce a novel control method which addree the varying time delay problem effectively. Thi novel method ugget an online adaptive fuzzy logic controller which have been controlled and adapted through the neural network. Thi method take the advantage of the genetic algorithm to optimize the memberhip function for it fuzzy logic controller. Thi deigned controller i applied to an AC W ervo motor a a remote plant in order to poition control it via Ethernet. The meaurement of round-trip time (RTT) i ued to etimate the online time delay a a parameter in online adaptive fuzzy logic controller. The rule-baed table of deigned fuzzy logic controller rotate in relation to thi etimated time delay. The value of rotating i obtained from a trained neural network. Comparion of reult from imulation of different controller and their comparion indicate that thi novel deigned controller provide a better performance over the varying time delay. The propoed method follow the input eaily, depite claical method which reult in an untable ytem epecially over the large time delay a large a 6 m. Reult get even more improved when genetic algorithm i applied to fuzzy logic controller. Index Term Data Communication Network, Genetic Algorithm, Networked Control Sytem, Neural Network, Online Manucript received June 9, 1. P. Hajebi i with the Electrical and Computer Engineering Department, Yazd Univerity, Yazd, Iran (phone: 98-311-685; fax: 98-311-685; e-mail: Hajebi@tu.yazduni.ac.ir). S.M.T. AlModarrei i with the Electrical and Computer Engineering Department, Yazd Univerity, Yazd, Iran (e-mail: mta@yazduni.ac.ir). Adaptive Optimized Fuzzy Logic Controller, Rule-Table Rotation. N I. INTRODUCTION ERWORKED control ytem (NCS) are patially ditributed ytem in which the communication between enor, actuator and controller occur through a hared band-limited digital communication network [1], []. Thi multipurpoe hared network connecting, patially ditributed element, create a flexible architecture which generally reduce intallation and maintenance cot. NCS have been finding application in a broad range of area uch a mobile enor network [3], remote urgery [], haptic collaboration over the Internet [5] [7], and automated highway ytem and unmanned aerial vehicle [8], [9]. Murry et al. in [1] ha identified control over network a one of the key future direction for control. However, application of a hared network veru everal dedicated independent connection, introduce new challenge. Drop and variable delay in NCS are two major problematic iue that were addreed in [11], [1]. Packet dropout and finite level quantization make NCS untable [1]. When the delay time i le than the ampling time of NCS, reult how that the time delay ha inignificant effect on control ytem. However, delay time greater than the ampling time degrade the performance of the NCS [13]. Many controller uch a conventional PID and fuzzy logic controller are utilized to tabilize the NCS cloed loop feedback and to reduce the error. Claical Smith predictor i one of the controller which are efficient for time delay procee [13], [1]. Lai and Hu propoed an adaptive Smith predictor a a controller for NCS in [1]. Depite howing relatively a good performance, there are ome drawback in thee controller. For intance, the accuracy of the model depend on plant tranfer function etimation. Moreover; each new plant require changing the controller deign. Practically etimation of plant tranfer function i not exact. Recently Pan et al. in [15] and Dezong et al. in [16] have hown that fuzzy logic controller offer a better performance in tackling packet dropout and varying time delay, at the ame time are more compatible with nonlinear procee. W. Du and F. Du propoed a Smith predictor integrated with fuzzy adaptive PID

Actuator Plant Senor Reference - Fuzzy Logic Controller t e 1 G p () Output Delay t 1 Data Network Delay t t e Fig.. Fuzzy logic controller. Controller Fig. 1. A SISO networked control ytem tructure. controller for the NCS in [17]. However they did not meaure the network delay online. They applied fuzzy logic controller for tuning the coefficient of PID controller. Thi paper firt, ugget a fuzzy logic controller (without PID controller) to poition control of an AC W ervo motor via Ethernet. At the next tep, it propoe a novel control method which i an online adaptive fuzzy logic controller for the imilar application. Thi reearch provide the advantage of no PID controller application while offer an adaptive controller which it fuzzy logic rule are rotating during the plant control. The round-trip time (RTT) i meaured online and thi value i utilized a, t m, time delay parameter. Then, thi time delay value i mapped to an angle by mean of trained neural network. Thi neural network ha been already trained by different time delay in adaptive fuzzy logic controller. Reult verify the better performance of thi novel deign which it fuzzy logic controller rule-table rotate through a trained neural network. The fact i that in communication network time delay could exceed m (v. m, 6m). However, reult from [17] and [18] how that the repone would be degraded in thee ytem for time delay over m depite the application of deigned offline controller. Thi paper propoed method ha hown an improved repone epecially in the cae of time delay over m. Even with time delay of 6m, there i no degradation in tep repone. Genetic algorithm could be implied in order to optimizing a uitable objective function for tuning the fuzzy logic controller memberhip function. The reult indicate that thi optimized controller how better performance compare to a non optimized fuzzy logic controller. Thi paper include the following ection. In Section II, NCS, tochatic time delay and packet dropout are introduced. Section III,firt decribe, deigning a fuzzy logic controller in order to poition control of an AC W ervo motor and next introduce a novel adaptive fuzzy logic controller with a rotating rule-table by mean of trained neural network. Section IV introduce the genetic algorithm and decribe it application in optimization of memberhip function of the deigned fuzzy controller. Section V contain the related imulation and equation. Thi paper end with concluion in ection VI. II. NETWORKED CONTROL SYSTEMS Due to quantum leap in communication ytem, in recent year, it ha become more common to apply a hared communication channel uch a Ethernet or Controller Area Network (CAN) bu etc. for tranmiion of the control ignal and the meaured output. Thi method help reducing the wiring cot a well a eliminate the neceity for maintaining dedicated communication channel for each control parameter [15]. However, thi type of networked control ytem i not a perfect olution and own it variou unolved iue uch a tranmiion delay and packet dropout [1], [15] which can degrade control performance. The SISO (ingle input-ingle output) NCS tructure in the cloed loop model i hown in Fig. 1. A illutrated in thi figure, t 1 and t indicate, time-delay induced in the network tructure for the controller-to-actuator direction and the enor-to-controller direction, repectively. Baically, the induced network delay varie according to the network load, cheduling policie, number of node, and different protocol. Time-varying characteritic of thee NCS make the deign and modeling of them more complicated. The total time-delay can be categorized into three clae, baed on the part where they occur, namely, the erver node, the network channel, and the client node [], [1]. In addition, the round-trip time (RTT) meaurement i crucial a it provide of accurate delay meaurement periodically [19] [1]. RTT i defined a the total time delay in SISO NCS. Obviouly the longer ditance increae, the delay time of a network ince more node are involved and conequently reult in a larger RTT. In a claical Smith predictor deign, the value of t m i contant and uually equal to average approximation of delay time between two node in the network. The value of RTT could be applied to fuzzy logic controller for compenating of variable delay. Normally in a fuzzy logic controller rule-bae table i contant during the control proce action. In thi paper uggeted method, RTT applied by neural network mapping, generate a rotating Rule-Table. III. FUZZY LOGIC CONTROLLER USING RULES-TABLE ROTATION A it ha been already mentioned, an online adaptive fuzzy logic controller could be a olution for tackling the tochatic time delay problem in NCS. However it control with imple PID or PI controller which how limited potential, epecially in nonlinearity procee. Recently proved that fuzzy logic controller i the bet option for controlling nonlinear procee while make the ytem more robut againt the varying time delay [15] and [16]. In the following part, firt a fuzzy logic controller i deigned then a claical Smith predictor would be integrated with thi deigned fuzzy logic controller baed on our plant. Finally a novel rotating rule-table online adaptive fuzzy logic controller i decribed.

3 A. Deigning Fuzzy Logic Controller To implement a NCS controller, firt the output of plant i meaured and then it would be compared with a reference ignal. Thi comparion generate the error ignal. The error ignal and derivation of error ignal are both input for the fuzzy logic controller. Here in thi paper, the plant i an AC W ervo motor which it poition a an output i meaured with an encoder with gain 1 P/R. The coefficient of the equivalent PI controller for thi plant are K p =.1 and K i =.1, [1]. The open loop poition control i obtained from Equation (1). Equation () and (3) repreent the continuou-tate pace form of tranfer function decribed in Equation (1). In Fig. 3, even triangular memberhip function have been devoted to either, input (error and derivation of error) and output. In Fig. 3, the fuzzy linguitic variable ( NB, NM, NS, ZE, PS, PM, PB ) repreent (Negative Big, Negative Medium, Negative Small, Zero, Poitive Small, Poitive Medium and Poitive Big) repectively. In poition control, the output follow the input. Therefore, at firt they are aumed to have imilar memberhip function. However in the following ection, output memberhip would be optimized uing an objective function. Here are provided ome deign pecification, applied in thi fuzzy logic controller: 1) The inference, ued in thi deign i Mamdani-type [], ) Fuzzy logic and operator wa implemented by min method while the fuzzy logic implication i baed on the min method a well and rule are aggregated uing fuzzy max operator, 3) The fuzzy logic output ha been determined through the center of gravity method by mean of defuzzification, ) Fuzzy rule are opted baed on Table I which contain 9 rule,5) Due to 1.58 3.1 G P (1) (.1.19 1 19 78.15 8 x x 18 U () 1 y.13 18.7 x (3) 1.5 1.5 Memberhip function for Input -5 - -3 - -1 1 3 5 Memberhip function for Output -5 - -3 - -1 1 3 5 Fig. 3. Memberhip function for input and output. Reference - - G Smith () Gˆ ( )(1 e p G C t m ) t e 1 t e Fig.. A control tructure of Smith predictor. G p () Output high gain of encoder the caling factor value elected for fuzzy logic controller output i,1 -. B. Claical Smith Predictor with Fuzzy Logic Controller Claical Smith predictor i one of the controller which are efficient for time delay proce [13], [1]. Here a claical Smith predictor i deigned for comparing the reult. In thi claical Smith predictor which i hown in Fig., G C i the deigned fuzzy logic controller decribed in ection III. A. G P i the tranfer function of the plant while ĜP i the etimation of plant tranfer function. Uually t m i the approximation of total time delay from controller to plant and plant to controller. If t m i the appropriate etimation of overall time delay the performance of ytem will be reaonable. t m i aumed m in the imulation. Ĝ P, i the etimation of G P, and practically difference between thee two tranfer function reult in intabilitie and increae of the error of repone. Thi i the main problem for claical Smith predictor and online adaptive Smith predictor. In thi paper claical mith predictor with fuzzy logic controller wa aumed ideal, thu the G P and ĜP are equal in the imulation. The fuzzy rule are elected baed on Table I. C. Deigning Rule-Table Rotation of Online Adaptive Fuzzy Logic Controller Uing Neural Network RTT i etimated in network [13], [] and then thi meaurement would be applied to online fuzzy logic controller. In thi tage the deigned fuzzy logic controller in part A would be integrated with an online neural network. The meaurement of round-trip time (RTT) i applied for etimating of online time delay which in turn provide the value for rotation angle of fuzzy rule-table. A already mentioned, the controller in thi paper ha not included any PID or PI method. The nonlinear TABLE I RULE BASES FOR ERROR, ERROR DERIVATION AND OUTPUT (WITHOUT ROTATION). e de NB NB NB NB NB NM NS ZE NM NB NB NB NM NS ZE PS NS NB NB NM NS ZE PS PM ZE PS NM NS ZE PS PM PB PB PM NS ZE PS PM PB PB PB PB ZE PS PM PB PB PB PB

Online Adaptive Fuzzy Controller Reference - e de PB ZE PS PS PM PM PB PB PM NS ZE PS PS PM PM PB PS NS NS ZE PS PS PM PM ZE NM NS NS ZE PS PS PM NS NM NM NS NS ZE PS PS NM NB NM NM NS NS ZE PS NB NB NB NM NM NS NS ZE Φ t e 1 G p () Output e de PB ZE PS PS PM PM PB PB PM NS ZE PS PS PM PM PB Neural Network Input Hidden Layer Output Layer p1 IW1,1 1 b1 a1 LW,1 n1 1 b a=y n a1 = tanig (IW1,1p1b1) a = pureline (LW,1 a1b) t Fig. 5. Structure of online adaptive fuzzy logic controller uing neural network. e PS NS NS ZE PS PS PM PM ZE NM NS NS ZE PS PS PM NS NM NM NS NS ZE PS PS NM NB NM NM NS NS ZE PS NB NB NB NM NM NS NS ZE Φ fuzzy logic controller doe have the potential to control the againt the dynamically ytem variable pecially occur at the and nonlinear procee while i more robut beginning of the proce. In a control proce with no delay, the error and the derivation of error change periodically. However, thee change ugget non-linear function pattern. During the proce control and bae on the taken time, the error and derivation of error run a way on the fuzzy rule-table, which imulate a circle hape behavior. While delay caued the error and derivation of error do not have deired time value, application of thi uggeted rotation method could overcome problem of delay. Thu, thi paper ha uggeted a control method which integrated fuzzy logic controller with a neural network. Fig. 5, how the tructure of thi propoed controller. Here in thi figure, the value of RTT i mapped to an angle by neural network. The tructure of neural network ha two-layer feedforward. Firt thi neural network i trained by everal et point time delay. It mean the value of rotation for everal time delay i obtained manually Then thee value will be applied for training the neural network. The value of angle for rotating rule-table in online adaptive fuzzy logic controller, change periodically baed on the RTT value. Fuzzy rule are opted baed on Table II, but other parameter (memberhip function, fuzzy logic operator and fuzzy logic method) are imilar to data in ection III.A. Equation () how the mapping relation of the error and variation of the error in new coordinate. Matrix A, in Equation (5) i the rotation tranform matrix which rotate coordinate by TABLE II RULE BASES FOR ERROR, ERROR DERIVATION AND OUTPUT (ROTATION TABLE). e de PB ZE PS PS PM PM PB PB PM NS ZE PS PS PM PM PB PS NS NS ZE PS PS PM PM ZE NM NS NS ZE PS PS PM NS NM NM NS NS ZE PS PS NM NB NM NM NS NS ZE PS NB NB NB NM NM NS NS ZE Fig. 6. Rotating of fuzzy rule-table. the angle of φ radian. The rule-rotation tructure of fuzzy logic controller and trend of rotation i hown in Fig. 6. e e new new e A e in co A (5) co in IV. OPTIMIZATION OF FUZZY LOGIC CONTROLLER USING GENETIC ALGORITHM To improve the propoed fuzzy logic controller, genetic algorithm i ued to find the optimal memberhip function [15]. Here, firtly the genetic algorithm i explained, and then the optimization of fuzzy logic controller by mean of genetic algorithm i dicued. A. Genetic Algorithm Genetic Algorithm wa firtly introduced by John Holland and developed by him, hi tudent and colleague [3]. Thee algorithm are heuritic optimization proce inpired by natural evolution and could be ued to minimize a uitable objective function or fitne function for tuning the fuzzy logic controller parameter. It i more effective at avoiding local minima than differentiation baed method. The genetic algorithm will generally include three fundamental genetic operation of election, croover and mutation. Thee operation are ued to modify the choen olution and elect the mot appropriate offpring to ucceeding generation [3]. In a genetic algorithm, a population of tring (called chromoome), which encode candidate olution (called individual) to an optimization problem, evolve toward better olution. Traditionally, olution repreented in binary a tring of and 1. The evolution uually tart from a population of randomly generated individual and happen in generation. In each generation, the fitne of every individual in the population i evaluated, multiple individual are tochatically elected from current population (baed on their ()

5 fitne), and modified (recombined and poibly randomly mutated) to form a new population. The new population i then ued in the next iteration of the algorithm. Commonly the algorithm terminate when either a maximum number of generation ha been produced, or atifactory fitne level ha been reached for the population. The baic iteration of genetic algorithm can be ummarized a follow [3], []: 1) Genetic repreentation: encoding the variable. Genetic algorithm often encode olution a fixed length bittring (e.g. 1111, 111111, and 11). ) A method for generating the initial population: population may be generated randomly or problem pecific knowledge can be ued to contruct the chromoome with the population. 3) An evaluation function, which aign a real number to meaure the fitne of each chromoome. ) A reproduction election cheme, which i ued to elect chromoome to be expoed to genetic operation. On of the mot famou approach i roulette wheel, which elect chromoome proportional to their fitne. 5) Genetic operator: croover and mutation are two main operator in genetic algorithm. Thee operator are applied to modify the choen chromoome (parent) and elect the mot appropriate offpring to pa on to ucceed generation. The baic mechanim of genetic algorithm and croover of the parental and mutation illutrated in Fig. 7. Croover i done by electing two parent during reproduction and combining gene to produce offpring. Two elected parent are combined with ome probability (croover rate); therefore two new offpring will born. One gene or everal gene of each offpring may then change randomly (mutation) with ome probability (mutation rate). Uually the croover ha high probability (typically value are between.8 and.95) and mutation ha mall probability (typically value are between.1 and.1). 6) Termination: The cycle of genetic algorithm will be continued until the genetic algorithm reache to topping criteria. There are everal approache to terminate the genetic algorithm. A common approach i to terminate genetic algorithm when the number of generation reache to pecific value. The genetic algorithm proce may alo run jut for limited time duration. It i alo poible to terminate a genetic Evaluated Offpring Chromoome String Fitne A 111 176 B 11111 56 C 11 3 D 11 Selection (Roulette Wheel) algorithm when the objective function of the bet chromoome ha not improved in the everal generation. B. Uing Genetic Algorithm for Optimizing of Fuzzy Logic Controller Memberhip function In III. A., input and output memberhip of fuzzy logic controller are aumed imilar. Now, with taking to conider the value from the lat input memberhip function, the output memberhip function could be optimized. The hape of memberhip function in fuzzy logic controller are triangular. Therefore in the cae of triangular fuzzy et, three characteritic point (center and two width) are ued a the parameter hould be optimized. Here, the number of triangular memberhip function i even ( NB, NM, NS, ZE, PS, PM, PB ). To deign ymmetric controller for poitive and negative input pule, the memberhip function are aumed ymmetric to Y axi. Alo the center of ZE i aumed zero. Therefore there are nine point or variable for optimizing the output memberhip function; two point for NB, three point for NM, three point for NS and one point for ZE. The range of change in thee variable i between.1 and. Moreover, there are contraint in optimization which guarantee that the memberhip function are ordered according to their value (e.g. NB<NM<NS<ZE<PS<PM<PB); for intance, the center of NB mut be le than the center of NM. Thee contraint are conidered to optimize the memberhip function. The number of initial population i aumed 1. Minimizing integral of time-weighted abolute error (ITAE) i commonly referred to a good performance index in deigning PID controller. Thu the ITAE aumed a objective function. Equation (6) i how the mathematical formula of ITAE. Where t i the time and e i the different between output and reference in control proce. ITAE t e( t) dt (6) Selection of parent i bae on roulette wheel. The rate of croover which applied to parent i.8. Croover trend to make the chromoome within the population more imilar, wherea mutation trend them more diver and uually have low rate. Here, becaue of exiting contraint for memberhip function, the mutation i not applied. Alo here the genetic algorithm i terminated when 3 generation have been produced. Fitne Function Altered Offpring Offpring 1 E 111 5 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 Croover Mutation Parent 1 1 1 Offpring Parent Mutated Offpring Offpring 1 1 1 Fig. 7. Mechanim of genetic algorithm. V. SIMULATION RESULTS A cloed loop NCS unit in thi paper include thee ection: online adaptive fuzzy logic controller, neural network, plant, data communication network. In order to analye the whole unit each ection hould be analyed eparately. Thi paper ha applied the tate equation to plot the tep repone of thi NCS. At thi firt tage, tranform function of plant and controller are converted to tate equation. Since the data tranmitted over the network i digital thee equation tate need to be dicretized tate pace equation to be able to imulate the procee. Equation (7) how the dicrete tate-pace form of proce. While A, B, C and D are the continuou tate pace

Time Delay (m) purelin (n) tanig (n) 6 matrice then their equal dicrete tate pace matrice (A d, B d, C d, D d ) would be obtained from (8), (9), (1), (11). By ubtituting the (), (3) in to (7)-(11), dicrete tate pace form of plant i obtained and repreented in (1), (13). x[ k 1] Ad x[ Bdu[ (7) y[ Cd x[ Ddu[ AT A e (8) d T A Bd e d B (9) C d C (1) D d D (11).66 -.973 7.79 x[ k 1].87.795 x[ 7.99 U[ (1).35.89 1.77 y[.13 18.7 x[ k (13) ] In thee imulation a random delay time i provided to analyi the feedback of NCS. The total of command delay, t 1, and forward delay, t, generate the total time delay (RTT) which i hown in Fig. 8. After training of neural network, the biae and weight value are obtained hown in Table III. Neural network tranfer function for layer 1 (hidden layer) and layer (output layer) are tanig and purelin repectively. Equation (1) and (15) how thee two tranfer function in mathematical form, while figure of thee tranfer function are plotted in Fig. 9. tanig ( n) 1 n 1 e (1) purelin( n) n (15) The ampling time i aumed.1 econd and the model of NCS i baed on the model decribed in [5]. Here in thi paper the imulation and comparion of the tep repone are provided among three controller type: 1) Online adaptive fuzzy logic Controller, ) Claical Smith predictor with fuzzy controller and 3) Pure fuzzy logic controller. The reult are illutrated in Fig. 1. Reult how that the online adaptive fuzzy logic controller offer a better performance compared to other two controller. 7 6 5 3 1 Total Time Delay 5 1 15 5 3 Time (Second) Fig. 8. A ample of imulated varying time delay in NCS. 1.5 -.5 (a) Neural Network Tranfer Function for Layer 1-1 -3 - -1 1 3 n (b) Neural Network Tranfer Function for Layer - - -3 - -1 1 3 n Fig. 9. Neural network tranfer function; a) tanig; b) purelin. A can be een the output ignal of online adaptive fuzzy logic method doe have mall overhoot and fat repone. Therefore thi controller i recommended for networked control ytem purpoe. W. Du and F. Du have uggeted a Smith predictor integrated with adaptive fuzzy-pid controller for the NCS in [17]. However they did not meaure the network delay online. They applied a fuzzy logic controller jut for tuning the coefficient of PID controller, which mean that their uggeted controller i worked offline. They alo have deigned in [18] a RBF neural network control with Smith predictor for NCS which i worked offline a well. In communication network time delay could exceed m (v. m, 6m). The reult from [17], [18] how that repone would degrade with time delay over m. In [17], [18], it wa aumed that the maximum of burt time delay i about m while thi time delay wa applied dicretely. Our propoed method ha more improved repone epecially when the time delay i over m even with the time delay of 6m there i no degradation in tep repone. In the pite of applying thi large value of time delay continuouly, reult in Fig. 13 how that thi doe not have any more effect on the tep repone a well. A decribed in ection IV, Thi paper take the advantage of genetic algorithm method to optimize the fuzzy controller memberhip function and conequently improve the overall reult. For thi propoe, it tart with aumption of a no delay ytem while optimize the fuzzy logic controller. The applied objective function type in thi paper i ITAE. In thi cae, which i a poition control example, the goal i that the output follow the input more cloely. The cloer output follow the input the more accurate i the performance. Tuning of an online adaptive fuzzy logic controller reduce the error value even with the preence of delay time in the ytem. Hence genetic algorithm i applied for ome input pule, and a a reult, the output memberhip function become tuned. The value of objective function in each generation are hown in Fig. 1. A could be perceived, the objective function value in each generation become maller compare to their previou generation which i inherently an algorithm genetic characteritic. Another point i that objective function do not

Fitne value 7 how ignificant change in lat five generation, which indicate that, thi genetic algorithm roughly obtain it bet anwer. The bet value of ITAE in 3th generation i.8. After applying the genetic algorithm to fuzzy logic controller, tuned output memberhip function hape could be illutrated in Fig. 11. Thee memberhip function are ued for lat three type of controller. For online adaptive fuzzy logic controller, the neural network i trained baed on new optimized memberhip function. Obtained biae and weight value are hown in Table IV. Rotation value are obtained according to different delay in experiment. Then the neural network i trained by thee value. Experiment in Fig. 1 and Fig. 13 are repeated and are hown in Fig. 1 and Fig. 15. Reult of online adaptive fuzzy logic controller are better compare to the other two controller. Since the pure fuzzy logic controller i tuned in condition with no delay, change in memberhip make it untable with the preence of delay time. Moreover, Claical mith predictor a hown in Fig.15 become untable for delay time over m imilar to pure fuzzy logic controller. It i even wore when the time it i not tuned. Table V how the comparion of reult. Two main evaluation indexe for comparion of reult are the rie time (T r ) and the Percent Overhoot (P.O.). Rie time i that time taken for the output of plant to rie from 1% to 9% of it final value when imulated by tep input. The Percent Overhoot i defined a (16). For a unit tep input, where M Pt, i the firt peak value of the time repone, and f v i the final value of the repone. Normally, f v i the magnitude of the input. M.. Pt f P O v 1% (16) fv Peak time (T p ) i the time that take for output of plant to reache the peak value. The maller the P.O. and M Pt value, the better i the controller performance. The reult from Table V how that our propoed method ha no overhoot. Furthermore, when the memberhip function are optimized with genetic algorithm, the T r i more maller compare to the noneoptimized cae. Since the mith predictor controller i tuned for m delay, when optimization i applied thi controller how better reult. However, for other induced delay, it performance will decreae. Pure fuzzy logic controller tuned with no delay time and how poor performance reult compared to other two controller..95 Bet:.8 1.5 1.5 Memberhip function for Input -5 - -3 - -1 1 3 5 Memberhip function for Output NB NMNS ZE PSPM PB -5 - -3 - -1 1 3 5 Fig. 11. Optimal fuzzy logic memberhip function. VI. CONCLUSION NCS have found widely application in variou field recently. However there are ome drawback in their tructure uch a varying time delay and packet dropout, which make the control deign of thee ytem challenging. Conventional PID and fuzzy logic controller are motly deigned to addre the intability problem in NCS. Fuzzy logic controller with the great potential in tackling the nonlinear procee and making NCS more robut againt the dynamically variable parameter could be a more reaonable option for NCS. Moreover, a rotating Rule-Table fuzzy logic controller i a great olution for tochatic time delay. Thu according to above mentioned characteritic of both fuzzy logic controller and rule-table rotation, thi paper have propoed a novel controller for NCS. Thi novel deign have integrated a rotating rule-table fuzzy logic controller with a neural network to poition control a W ervo motor a a remote plant via Ethernet. Simulation reult and their comparion for three different method of controlling over thi plant verified that thi novel controller deign i more beneficial epecially over the big value of delay time a large a 6 m. The propoed method how better performance when the fuzzy logic memberhip function are optimized by mean of genetic algorithm..9.85.8.75.7.65.6.55.5.5 5 1 15 5 3 Generation Fig. 1. Objective function value for different generation of genetic algorithm. TABLE III WEIGHTS AND BIASES VALUES FOR NEURAL NETWORK..57.IW{1}.57 net net. LW{}.3566.811 1.13 net.b{1} net. b{} 1..615

8 TABLE IV WEIGHTS AND BIASES VALUES FOR NEURAL NETWORK (OPTIMIZED )..36.IW{1}.15 net net. LW{}.995 15.386 1.86 net.b{1} net. b{} 15. 9361 1.9 REFERENCES [1] J. P. Hepanha, P. Naghhtabrizi, and Y. Xu, A urvey of recent reult in networked control ytem, Proc. IEEE, vol. 95, no. 1, pp. 138-161, Jan. 7. [] F. L. Lian, J. R. Moyne, and D. M. Tilbury, Performance evaluation of control networke: Ethernet control net and device net, IEEE Control Syt. Mag., pp. 66-83, Feb. 1. [3] P. Ogren, E. Fiorelli, and N. E. Leonard, Cooperative control of mobile enor network: Adaptive gradient climbing in a ditributed environment, IEEE Tran. Automat. Contr., vol. 9, no. 8, pp. 19 13, Aug.. [] C. Meng, T. Wang, W. Chou, S. Luan, Y. Zhang, and Z. Tian, Remote urgery cae: Robot-aited teleneurourgery, in IEEE Int. Conf. Robot. and Auto. (ICRA ), Apr., vol. 1, pp. 819 83. [5] J. P. Hepanha, M. L. McLaughlin, and G. Sukhatme, Haptic collaboration over the Internet, in Proc. 5th Phantom Uer Group Workhop, Oct.. [6] K. Hikichi, H. Morino, I. Arimoto, K. Sezaki, and Y. Yauda, The evaluation of delay jitter for haptic collaboration over the Internet, in Proc. IEEE Global Telecomm. Conf. (GLOBECOM), Nov., vol., pp. 19 196. [7] S. Shirmohammadi and N. H. Woo, Evaluating decorator for haptic collaboration over the Internet, in Proc. 3rd IEEE Int. Workhop Haptic, Audio and Viual Env. Applic., Oct., pp. 15 19. [8] P. Seiler and R. Sengupta, Analyi of communication loe in vehicle control problem, in Proc. 1 Amer. Contr. Conf., Jun. 1, vol., pp. 191 196. [9] P. Seiler and P. R Sengupta, An H approach to networked control, IEEE Tran. Automat. Contr., vol. 5, no. 3, pp. 356 36, Mar. 5. [1] R. M. Murray, K. J. Atrom, S. P. Boyd, R. W. Brockett, and G. Stein, Control in an information rich world, IEEE Contr. Syt. Mag., vol. 3, no., pp. -33, Apr. 3. [11] M. G. Rivera and A. Barreiro, Analyi of networked control ytem with drop and variable delay, Automatica, vol. 3, no. 1, pp. 5-59, Dec. 7. [1] Y. Ihido, K. Takaba, and D. E. Quevedo, Stability analyi of networked control ytem ubject to packet-dropout and finite-level quantization, Sytem & Control Letter, vol. 6, no. 5, pp. 35-33, May. 11. [13] C. W. Cheng, C. L. Lai, B. C. Wang, and P. L. Hu, The time-delay effect of multiple-network ytem in NCS, (SICE7) Annual Conference, Sep. 7, pp. 99-93. [1] C. L. Lai, and P. L. Hu, Deign the remote control ytem with the time-delay etimator and the adaptive mith predictor, IEEE Tran. Ind. Inform., vol. 6, no. 1, pp. 73-8, Feb. 1. [15] I. Pan, S. Da, and A. Gupta, Tuning of an optimal fuzzy PID controller with tochatic algorithm for networked control ytem with random time delay, ISA Tranaction, vol. 5, no. 1, pp. 8-36, Jan. 11. [16] Z. Dezong, L. Chunwen, and R. Jun, Fuzzy peed control and tability analyi of a networked induction motor ytem with time delay and packet dropout, Nonlinear Analyi: Real World Application, vol. 1, no. 1, pp. 73-87, Feb. 11. [17] W. Du and F. Du, Novel mith predictor and fuzzy control for networked control ytem, Aia-Pacific Conference on Information Proceing (APCIP), July 9, pp. 75-78. [18] F. Du and W. Du, RBF neural network control and novel mith predictor for networked control ytem, Aia-Pacific Conference on Information Proceing (APCIP), July 9, pp. 67-7. [19] C. L. Lai and P. L.Hu, Deign of the adaptive mith predictor for time-varying network control ytem, IEEE International Conference on Sytem Man and Cybernetic (SMC), Oct. 1, pp. 66-73. [] C. L. Lai, P. L. Hu, and B. C. Wang, Deign of the adaptive mith predictor for time-varying network control ytem, SICE Annual Conference, Aug. 8, pp. 933-938. [1] N. Vatanki, J. P. George, C. Aubrun, E. Rondeau, and S. L. J. Jounela, Realization of networked control ytem on Ethernet with varied time delay, Control Engineering Practice, vol. 17, no., pp. 31-, Feb. 9. [] T. J. Ro, Fuzzy Logic with Engineering Application. nd ed. John Wiley & Son, Ltd.,, pp. 151-161. [3] A. Mellit and S.A. Kalogirou, Artificial intelligence technique for photovoltaic application: A review, Progre in Energy and Combution Science, vol. 3, no. 5, pp. 57 63, Oct. 8. [] C. Hick, A genetic algorithm tool for optimiing cellular or functional layout in the capital good indutry, International Journal of Production Economic, vol. 1, no., pp. 598 61, Dec. 6. [5] M. Y. Chow and Y. Tipuwan, Network-baed control ytem: A tutorial, in Proc. 7 th Annu. Confl IEEE Indl Electrol Socl, pp. 1593-16, 1. Pooya Hajebi (S 1) wa born in Ifahan, Iran, in 1981. He received the B.S. degree in electrical engineering (Electronic) and the M.S. degree in electrical engineering (Communication Sytem) both from Yazd Univerity, Yazd, Iran, in 5 and 9, repectively. Currently, He i working toward the Ph.D. degree in the Electrical and Computer Engineering Department, Yazd Univerity, Yazd, Iran. Hi reearch interet include Networked Control Sytem, Fuzzy Sytem, Neural Network, Data Communication Network, Digital Signal Proceing, Biological Signal Proceing, Digital Image Proceing, Cellular Network, Optimization, Time Delay Sytem and Real Time Sytem. Seyed Mohammad Taghi AlModarrei obtained hi B.S. degree in Electronic Engineering and M.S. degree in Communication Sytem, both from the Ifahan Univerity of Technology. He alo hold Ph.D. in Electronic (Intelligent Signal Proceing) from Univerity of Southampton (Department of Electrical and Computer Science: ECS). He work at the Department of Electrical and Computer Engineering in Yazd Univerity where he purue hi reearch interet in: (i) Networked Control Sytem (NCS) (ii) Neuro-Fuzzy Network (iii) Wirele Network.

9 (a) Total Time Delay 5 5 1 15 5 3 (b) Input (Reference) 5 1 15 5 3 (c) Online Adaptive Fuzzy Logic Controller Uing Rule-Tabale Rotation Method 5 1 15 5 3 (d) Claical Smith Predictor with Fuzzy Logic Controller, tm=. Second 5 1 15 5 3 (e) Fuzzy Logic Controller 5 5 1 15 5 3 Time (Second) Fig. 1. Simulation reult (The maximum delay time i about m); a) Time delay; b) Reference ignal; c) Poition (revolution of the motor haft) for online adaptive fuzzy logic controller uing rule-table rotation; d) Poition (revolution of the motor haft) for claical Smith predictor with fuzzy logic controller; e) Poition (revolution of the motor haft) for fuzzy logic controller. 6 (a) Total Time Delay 5 1 15 5 3 (b) Input (Reference) 5 1 15 5 3 (c) Online Adaptive Fuzzy Logic Controller Uing Rule-Tabale Rotation Method 5 1 15 5 3 (d) Claical Smith Predictor with Fuzzy Logic Controller, tm=. Second 5 1 15 5 3 Time (Second) Fig. 13. Simulation reult (The maximum delay time i about 6 m); a) Time delay; b) Reference ignal; c) Poition (revolution of the motor haft) for online adaptive fuzzy logic controller uing rule-table rotation; d) Poition (revolution of the motor haft) for claical Smith predictor with fuzzy logic controller.

1 (a) Total Time Delay 5 5 1 15 5 3 (b) Input (Reference) 5 1 15 5 3 (c) Online Adaptive Fuzzy Logic Controller Uing Rule-Tabale Rotation Method 5 1 15 5 3 (d) Claical Smith Predictor with Fuzzy Logic Controller, tm=. Second 5 5 1 15 5 3 (e) Fuzzy Logic Controller 1-1 5 1 15 5 3 Time (Second) Fig. 1. Simulation reult uing genetic algorithm (The maximum delay time i about m); a) Time delay; b) Reference ignal; c) Poition (revolution of the motor haft) for online adaptive optimized fuzzy logic controller uing rule-table rotation; d) Poition (revolution of the motor haft) for Claical Smith predictor with optimized fuzzy logic controller; e) Poition (revolution of the motor haft) for optimized fuzzy logic controller. (a) Total Time Delay 5 5 1 15 5 3 (b) Input (Reference) 5 1 15 5 3 (c) Online Adaptive Fuzzy Logic Controller Uing Rule-Tabale Rotation Method 5 1 15 5 3 (d) Claical Smith Predictor with Fuzzy Logic Controller, tm=. Second 1-1 5 1 15 5 3 (e) Fuzzy Logic Controller 1-1 5 1 15 5 3 Time (Second) Fig. 15. Simulation reult uing genetic algorithm (The maximum delay time i about 6 m); a) Time delay; b) Reference ignal; c) Poition (revolution of the motor haft) for online adaptive optimized fuzzy logic controller uing rule-table rotation; d) Poition (revolution of the motor haft) for claical Smith predictor with optimized fuzzy logic controller. e) Poition (revolution of the motor haft) for optimized fuzzy logic controller.

11 Approximate value of delay m m m 6 m TABLE V COMPARISON OF RESULTS Method Optimization T r () P.O. (%) ITAE T p () Online Adaptive Claical Smith Predictor Online Adaptive Claical Smith Predictor Online Adaptive Claical Smith Predictor Online Adaptive Claical Smith Predictor Optimized.9.6 - Non-Optimized.839.99 - Optimized.6913.93 - Non-Optimized 1.9.79 - Optimized.37.6 - Non-Optimized.598.9 - Optimized.581.37 - Non-Optimized.815..619 1.6 Optimized.339 1.83.155.56 Non-Optimized.693 -.316 1.9 Optimized.61 17.558.18 Non-Optimized.56.9.38.6 Optimized.9996 1.65 - Non-Optimized 1.166. 56 1.58.35 Optimized.156 139.16 3.779.86 Non-Optimized.3586 6.5 1.83 1 Optimized.16 51.93 9.97.8 Non-Optimized.37.3 6.363.91 Optimized 1.199.573 - Non-Optimized 1.5663.769 - Optimized.3861 1.58 11.15 1.16 Non-Optimized.39 16.6 5.9 1.3 Optimized.1616 7.96 15.78 1.1 Non-Optimized - - - -