Available online at ScienceDirect. Procedia Computer Science 48 (2015 ) (ICCC-2014) (ICCC-2015)

Transcription

1 Avalable onlne at ScenceDrect Proceda Computer Scence 48 (2015 ) Internatonal Conference on Intellgent Computng, Communcaton & Convergence (ICCC-2015) (ICCC-2014) Conference Organzed by Interscence Insttute of Management and Technology, Bhubaneswar, Odsha, Inda Inducton Motor Control Usng PSO-ANFIS Sakuntala Mahapatra*, Raju Danel**, Deep Narayan Dey*, Santanu Kumar Nayak *Dept. of Electroncs and Telecommuncaton Engneerng, Trdent Academy of Technology, Bhubaneswar, Odsha **Scentst-SF at IPR (Autonomous Inst. of the Department of Atomc Energy, Inda) Emal: mahapatra.sakuntala@gmal.com Abstract The speed of the Inducton motor can be adjusted to a great extent so as to provde easy control and hgh performance. In ths paper an effort s made to develop a prototype model to control the speed of an nducton motor usng PSO-ANFIS hybrd technque. We descrbe the use of Partcle Swarm Optmzaton (PSO) and ANFIS for desgnng an optmal fuzzy logc controller of an Inducton Motor. We have used two nput parameters lke speed, torque and output s frng angle. PSO-ANFIS speed controller obtans better dynamc behavor and superor performance of the Inducton motor speed control. Smlar approach can be correlated to the control of plasma column and whch can be mplemented n fuson reactor to control the plasma column. Plasma poston and shape control s very crucal for the overall performance of the fuson reactor such as tokamak. The qualty of the dscharge n the Saskatchewan TORus-Modfed (STOR-M) tokamak s strongly related to the poston of the plasma column wthn the dscharge vessel. If the plasma column approaches too near the wall, then ether mnor or complete dsrupton occurs. Consequently t s necessary to be able to control dynamcally the poston of the plasma column throughout the entre dscharge. A comparson analyss of PSO-ANFIS and Fuzzy Back Propagaton algorthm s taken n our work to control the speed of nducton motor, where the PSO-ANFIS gves better result n terms of fast computng. Keywords- Fuson reactors; Plasma confnement; PSO; ANFIS; Fuzzy Logc; Optmzaton 1.INTRODUCTION PLASMA column control s very necessary for the successful operaton of a thermonuclear fuson reactor such as a tokamak. The tokamak s a magnetc confnement system, and s currently the most promsng system for the near term development of a fuson reactor. One of the man challenges assocated wth plasma confnement s the stable The Authors. Publshed by Elsever B.V. Ths s an open access artcle under the CC BY-NC-ND lcense ( Peer-revew under responsblty of scentfc commttee of Internatonal Conference on Computer, Communcaton and Convergence (ICCC 2015) do: /j.procs

2 754 Sakuntala Mahapatra et al. / Proceda Computer Scence 48 ( 2015 ) mantenance of the plasma column near ts equlbrum poston. Ths problem has represented a great challenge to scentsts and engneers over the years. PSO-ANFIS hybrd model s mplemented for the control and predcton of nducton motor and smlar approach can be used to control plasma column. In recent years, hgh performance and hgh effcency nducton motor (IM) drves are n great demand to serve the ndustral needs for sophstcated products and servces, such as steel mlls, paper mlls, servos, machnes tools, robotcs, elevators, transportaton system etc. Ths s due to ts well-known advantages of smple constructon, relablty, ruggedness, and low cost. Although IM has been wdely used for constant speed operaton earler, t s more and more accepted n advanced varable speed applcatons. Wth the advent of vector or feld orented control whereby IM can be controlled lke a separately excted dc motor, the hgh performance control of IM drves entered nto ts new era. The man problem wth the nducton motor s dffculty of control. Because of ths, ts applcatons n the area of speed control have tradtonally been lmted. The advent of Feld Orented Control or Vector Control [2] was a major breakthrough n the area of nducton motor control. A major drawback of a lot of vector control schemes s that they requre a speed sensor. Speed sensors are usually expensve, bulky and reduce the advantage of an nducton motor drve. Also, t s not practcal to employ speed sensors n some applcatons, such as motor drves n hostle envronments and hgh speed motor drves. Because of these drawbacks researchers have put n a lot of effort n developng speed sensor less nducton motor drves [4]. In recent years, Artfcal Neural Networks (ANN) has found wdespread use n functon approxmaton [4]. It has been shown that theoretcally a three layer ANN can approxmate arbtrarly closely any nonlnear functon provded the functon s non-sngular. The man objectve of ths work s to develop a general purpose nducton motor speed estmator based on the dynamc equatons of the nducton motor. A technque s presented for estmatng the speed of an nducton motor usng PSO-ANFIS. Fuzzy theory was frst proposed and nvestgated by Prof. Zadeh n The Mamdan Fuzzy Inference System (FIS) was presented to control a steam engne and boler combnaton by lngustc rules. Fuzzy logc s expressed by means of IF-THEN rules wth the human language. In the desgn of a fuzzy logc controller, the mathematcal model s not necessary. Therefore the Fuzzy Logc Controller (FLC) s of good robustness [1]. Owng to ts easy applcaton, t has been wdely used n ndustry. However, the rules and the membershp functons of a fuzzy logc controller are based on expert experence or knowledge database. Much work has been done on the analyss of fuzzy control rules and membershp functon parameters [14]. The PSO (partcle swarm optmzaton) algorthms are used to get the optmal values and parameters of our FLC. The PSO s based on a metaphor of socal nteracton. It searches a space by adjustng the trajectores of ndvdual vectors, called partcles, as they are conceptualzed as movng as ponts n multdmensonal space. The ndvdual partcles are drawn stochastcally towards the postons of ther own prevous best performances and the best prevous performance of ther neghbors. Of these s the PSO algorthms are appled to choose membershp functons and fuzzy rules [15]. However, the expert experences or knowledge are stll necessary for the ranges of membershp functons. In ths paper, a novel strategy s proposed for desgnng the optmal fuzzy controller. PSO algorthms are appled to search globally optmal parameters of fuzzy logc. The best ranges of membershp functons, the best shapes of membershp functons and the best fuzzy nference rules are dug out at the same tme. Smulaton results are gven to show the effectveness of FLC-Swarm controller [11]. An nducton or asynchronous motor s an AC motor n whch all electromagnetc energy s transferred by nductve couplng from a prmary wndng to a secondary wndng, the two wndngs beng separated by an ar gap. In threephase nducton motors, that are nherently self-startng, energy transfer s usually from the stator to ether a wound rotor or a short-crcuted squrrel cage rotor. Three-phase cage rotor nducton motors are wdely used n ndustral drves because they are rugged, relable and economcal. Sngle-phase nducton motors are also used extensvely for smaller loads. In both nducton and synchronous motors, the AC power suppled to the motor's stator creates a magnetc feld that rotates n tme wth the AC oscllatons. Whereas a synchronous motor's rotor turns at the same rate as the stator

3 Sakuntala Mahapatra et al. / Proceda Computer Scence 48 ( 2015 ) feld, an nducton motor's rotor rotates at a slower speed than the stator feld. The nducton motor stator's magnetc feld s therefore changng or rotatng relatve to the rotor. Ths nduces an opposng current n the nducton motor's rotor, n effect the motor's secondary wndng, when the later s short-crcuted or closed through external mpedance. The rotatng magnetc flux nduces currents n the wndngs of the rotor, n a manner smlar to currents nduced n transformer s secondary wndngs. These currents n turn create magnetc felds n the rotor that react aganst the stator feld. Due to Lenz's Law, the drecton of the magnetc feld created wll be such as to oppose the change n current through the wndngs. The cause of nduced current n the rotor s the rotatng stator magnetc feld, so to oppose ths; the rotor wll start to rotate n the drecton of the rotatng stator magnetc feld. The rotor accelerates untl the magntude of nduced rotor current and torque balances the appled load. Snce rotaton at synchronous speed would result n no nduced rotor current, an nducton motor always operates slower than synchronous speed [14]. For these currents to be nduced, the speed of the physcal rotor must be lower than that of the stator's rotatng magnetc feld ( ), or the magnetc feld would not be movng relatve to the rotor conductors and no currents would be nduced. As the speed of the rotor drops below synchronous speed, the rotaton rate of the magnetc feld n the rotor ncreases, nducng more current n the wndngs and creatng more torque. The rato between the rotaton rate of the magnetc feld as seen by the rotor (slp speed) and the rotaton rate of the stator's rotatng feld s called slp. Under load, the speed drops and the slp ncreases enough to create suffcent torque to turn the load. For ths reason, nducton motors are sometmes referred to as asynchronous motors. An nducton motor can be used as an nducton, or t can be unrolled to form the lnear nducton motor whch can drectly generate lnear moton.[15,16] A. Neurofuzzy Back Propagaton Algorthm II. THE PROPOSED EXPERIMENTAL MODEL In ths paper we have mplemented a hybrd fuzzy neural network model n whch we can map for both real nput real output and fuzzy nput to fuzzy output wth use of fuzzy weghts and fuzzy bases. The back propagaton algorthm s mplemented for neural networks, to tran fuzzy systems to match desred nput-output pars. The key deas n developng ths tranng algorthm are to vew a fuzzy system as a three-layer feed forward network, and to use the chan rule to determne gradents of the output errors of the fuzzy system wth respect to ts desgn parameters. It s shown that ths tranng algorthm performs an error back propagaton procedure: hence, the fuzzy system equpped wth the back propagaton tranng algorthm s called the Fuzzy back propagaton [1, 8].

4 756 Sakuntala Mahapatra et al. / Proceda Computer Scence 48 ( 2015 ) W j (1) W j (2) +1 X1 X y(k) + d(k) - e(k) Xn Back Propagaton Algorthm Fg-1: Fuzzy back propagaton model An MLP network wth neurons (2, 3 and 1 denotes the number of neurons n the nput layer, the hdden layer, the second hdden layer and the output layer respectvely) wth the back-propagaton (BP) learnng algorthm, s depcted n Fg-1. The parameters of the neural network can be updated ether n sequental and batch mode of operaton. The output from the k th node s (1) Where f(.) s a nonlnear logstc functon. Trangular membershp functon s taken for the weghts whch are represented by (2) Smlarly Equaton (2) s computed by mn-max prncple as; The error at (M) th layer s, Where, =( / / ) Here a generalzed cost functon s used to mnmze and update weghts. (3) (5) (6) (4) (7)

5 Sakuntala Mahapatra et al. / Proceda Computer Scence 48 ( 2015 ) (8) The stepwse algorthm for back propagaton algorthm s lsted out here. The tranng phase of back propagaton nvolves four basc stages. They are as follows: () () () (v) Intalzaton of Weghts Feed Forward Back Propagaton of the errors Updatng of the weghts and bases. The ntal random weghts are selected between [-0.5, 0.5]. Each unt receves an nput sgnal and delvers t to all the hdden unts wth actvaton functon. The sgnal s then sent to each of the output unts. Fnally the output unt provdes the desred output to form the response of the net for the gven nput pattern appled to the structure. Each output unts compares ts computed actvaton functon y(k) wth ts target value d(k) to determne the assocated error for the nput pattern wth that unt. The fnal output y(k) s compared wth the desred output d(k) and the resultng error sgnal e(k) s thus produced usng the error equaton. Ths error functon s then feed back to the structure. B. Adaptve Network Based Fuzzy Inference Systems (ANFIS) Fuzzy systems present partcular problems to a developer: Rules: The IF-THEN rules have to be determned. Ths s usually done by knowledge acquston from an expert. It s a tme consumng process that s fraught wth problems. Membershp functons: A fuzzy set s fully determned by ts membershp functon. Thus proper choce of membershp functon s challengng one. The ANFIS approach learns the rules and membershp functons from data. ANFIS s an adaptve network. An adaptve network s network of nodes and drectonal lnks. Assocated wth the network s a learnng rule - for example back propagaton. It s called adaptve because some, or all, of the nodes have parameters whch affect the output of the node. These networks are learnng a relatonshp between nputs and outputs [14]. Adaptve networks cover a number of dfferent approaches but for our purposes we wll nvestgate n some detal the method proposed by Jang known as ANFIS [17]. The ANFIS archtecture s shown below. The crcular nodes represent nodes that are fxed whereas the square nodes are nodes that have parameters to be learnt.

6 758 Sakuntala Mahapatra et al. / Proceda Computer Scence 48 ( 2015 ) Layer 1 Layer 2 Layer 3 Layer 4 Layer 5 X A 1 w 1 w 1 w 1 f1 A 2 F B 1 Y w 2 w 2 w 2 f 2 B 2 A Two Rule Sugeno ANFIS has rules of the form: Fg-2: An ANFIS archtecture for a two rule Sugeno system If If x s A xs A 1 and y s B1 THEN f1 p1x q1 y r1 2 and y sb2 THEN f2 p2x q2 y r2 (9) For the tranng of the network, there s a forward pass and a backward pass. We now look at each layer n turn for the forward pass. The forward pass propagates the nput vector through the network layer by layer. In the backward pass, the error s sent back through the network n a smlar manner to back-propagaton. Layer 1 The output of each node s: O1, A ( x) for O1, B ( y) for 2 1, x 1,2 3,4 So, the O ( ) s essentally the membershp grade for x and y. The membershp functons could be anythng but for llustraton purposes we wll use the bell shaped functon gven by: 1 A x) 2 x c ( (10) 1 a a, b, c b Where are parameters to be learnt. These are the premse parameters. Layer 2 Every node n ths layer s fxed. Ths s where the t-norm s used to AND the membershp grades - for example the product: O2, w A ( x) B ( y), 1,2 (11) Layer 3 Layer 3 contans fxed nodes whch calculate the rato of the frng strengths of the rules:

7 Sakuntala Mahapatra et al. / Proceda Computer Scence 48 ( 2015 ) O w w 3, w1 w2 Layer 4 The nodes n ths layer are adaptve and perform the consequent of the rules: O w f w ( p x q y r ) 4, (13) p The parameters n ths layer (, q, r ) are to be determned and are referred to as the consequent parameters. Layer 5 There s a sngle node here that computes the overall output: w f O 5, w f w (14) Ths output s taken as the actual output of our proposed model to compare wth the desred output whch s taken from the expermental data to calculate the error. (12) C. Partcle Swarm Optmzaton (PSO) Algorthm Operaton Partcle Swarm Optmzaton [7-9] optmzes an objectve functon by undertakng a populaton based search. The populaton conssts of potental solutons, named partcles, whch are metaphor of brds n flocks. These partcles are randomly ntalzed and freely fly across the mult dmensonal search space. Durng flght, each partcle updates ts own velocty and poston based on the best experence of ts own and the entre populaton. The varous steps nvolved n Partcle Swarm Optmzaton Algorthm [8] are as follows: Step 1: The velocty and poston of all partcles are randomly set to wthn pre-defned ranges. Step 2: Velocty updatng each teraton, the veloctes of all partcles are updated accordng to, v (15) v c 1 R 1 ( p, best p ) c 2 R 2 ( g, best p ) where p and v are the poston and velocty of partcle, respectvely; p,best and g,best s the poston wth the best objectve value found so far by partcle and the entre populaton respectvely; w s a parameter controllng the dynamcs of flyng; R 1 and R 2 are random varables n the range [0,1]; c 1 and c 2 are factors controllng the related weghtng of correspondng terms. The random varables help the PSO wth the ablty of stochastc searchng.

8 760 Sakuntala Mahapatra et al. / Proceda Computer Scence 48 ( 2015 ) Step 3: Poston updatng The postons of all partcles are updated accordng to, p (16) p v After updatng, p should be checked and lmted to the allowed range. Step 4: Memory updatng Update p,best and g,best when condton s met, p g, best, best p g f f f ( p ) f ( g ) f ( p f ( g, best, best ) ) (17) Where f(x) s the objectve functon to be optmzed. Step 5: Stoppng The algorthm repeats steps 2 to 4 untl certan stoppng condtons are met, such as a pre-defned number of teratons. Once stopped, the algorthm reports the values of g best and f(g best ) as ts soluton. PSO [16] utlzes several searchng ponts and the searchng ponts gradually get close to the global optmal pont usng ts pbest and gbest. Intal postons of pbest and gbest are dfferent. However, usng thee dfferent drecton of pbest and gbest, all agents gradually get close to the global optmum.

9 Sakuntala Mahapatra et al. / Proceda Computer Scence 48 ( 2015 ) D. Flow Chart of our proposed model Start Intalze scaled value of nputs (speed & torque) Gaussan membershp functons of 1 st order Sugeno fuzzy model are generated wth two rules (A & B) Desred frng angle (d) are collected from expermental data Optmze the parameters of MF usng PSO Algorthm p,q,r value are taken randomly for 1 st teraton of ANFIS Algorthm Calculate the actual output (y) usng ANFIS Algorthm Contnue to complete all teraton Error (e) =d-y Update p, q, r value usng back propagaton algorthm Stop: error s mnmzed Fg-3: flow chart of our proposed model

10 762 Sakuntala Mahapatra et al. / Proceda Computer Scence 48 ( 2015 ) E. Algorthm of our proposed model 1. Scaled value of motor speed and torque s taken as nputs (x and y). 2. Usng two nput frst-order Sugeno Fuzzy model wth two rules (A and B) are generated. 3. Gaussan Membershp functons are plotted. 4. Desred Frng angles are taken from expermental data as d. 5. Usng ANFIS w1, w2 are found out. w1=a1(x) *B1(y) w2=a2(x)*b2(y) 6. Then p1, p2, q1, q2, r1, r2 are randomly taken for 1 st teraton. 7. Then the actual output s calculated usng ANFIS technque. IF x s AND y s THEN = x + x+ IF x s AND y s THEN = x + x+ The reasonng mechansm for ths model s: f= = + 8. Then the error s calculated from the dfference of desred and actual output. e=d-f 9. Then the p1, p2, q1, q2 are updated usng back propagaton algorthm for each teraton tll the outputs are matched. 10. At last Error convergng graph and Desred-actual output matchng graphs are plotted. III. SIMULATION RESULT Desred Vs Actual Value of Frng Angle Motor Specfcaton: AC Inducton Motor, 220volt, 1600 rpm, P=30w

11 Sakuntala Mahapatra et al. / Proceda Computer Scence 48 ( 2015 ) SL NO DESIRED F.A ACTUAL F.A

12 764 Sakuntala Mahapatra et al. / Proceda Computer Scence 48 ( 2015 ) Fg-4: plot of membershp functon usng 2 nputs 3 rules Fg-5: Comparson of Actual and Desred output plot

13 Sakuntala Mahapatra et al. / Proceda Computer Scence 48 ( 2015 ) Fg-6: Error convergence plot for ANFIS Fg-7: Error convergence plot of Fuzzy back propagaton

14 766 Sakuntala Mahapatra et al. / Proceda Computer Scence 48 ( 2015 ) Fg-8: Error convergence comparson between ANFIS and Fuzzy back propagaton Fg.9:- Workng Prototype model of nducton motor control nterfaced to PC

15 Sakuntala Mahapatra et al. / Proceda Computer Scence 48 ( 2015 ) IV. CONCLUSION In ths paper, the speed of an Inducton Motor drve s controlled by the hybrd of PSO- ANFIS algorthms. Accordng to the results of the MATLAB smulaton, the Adaptve Neuro Fuzzy (ANFIS) controller effcently s better than the tradtonal FLC. The ANFIS-PSO s the best controller whch presented satsfactory performances. The major drawback of the fuzzy controller presents an nsuffcent analytcal technque desgn (choce of the rules, the membershp functons and the scalng factors).thus we chose wth the use of the Neural Networks and Partcle Swarm Optmzaton for the optmzaton of ths controller n order to control Inducton motor speed. Fnally, the proposed controller (PSO-ANFIS) gves a very good result. The ANFIS-PSO s the best controller whch presented satsfactory performances. In the next phase, PSO-ANFIS model would be appled on a nonlnear system followed by development for tokamak plasma poston control. By mplementng ths PSO-ANFIS concept we can predct and control the plasma layer n a fuson reactor whch s our future work whch s a very challengng but necessary requrement for tokamak. Acknowledgements The authors are thankful to Board of Research n Fuson Scence & Technology (BRFST), Insttute of Plasma Research, Government of Inda, Ahmedabad for provdng the fundng to carry out the research work on Predcton and Control of Plasma layers n a fuson reactor usng Hybrd PSO-ANFIS REFERENCES [1] Sakuntala Mahapatra, Santanu K. Nayak, Samrat L. Sabat, Neuro fuzzy model for adaptve flterng of oscllatory sgnals, Elsever Scence, Measurement 30, , [2] Sakuntala Mahapatra, Samrat L. Sabat, Santanu K. Nayak, An ntellgent nstrument for trackng and adaptve flterng of oscllatory sgnals usng Hebban learnng rules, Elsever Scence, Measurement 26, , July 1999 [3] Samrat L. Sabat, Santanu K. Nayak, Sakuntala Mahapatra, WNN based ntellgent energy meter, Scence Drect, Measurement 41, , 2008 [4] T.S.Radwan, Perfect Speed Trackng of Drect Torque Controlled-Inducton Motor Drve Usng Fuzzy Logc, SMIEEE, Ryadh College of Technology, Saud Araba, IEEE, [5] D. W. Novotny and T. A. Lpo, Vector Control and Dynamcs of AC Drves, Oxford, UK, Clarendon, [6] P.Ttnen, "The next generaton motor control method, DTC drect torque control", Proceedngs of the IEEE Intl. Conf on Power Electroncs, Drves, and Energy Systems for lndustralgrowth, 1996, pp [7] Ben-Brahm. Motor Speed Identfcaton va Neural Networks.IEER Ind. Applcat. Magazne, pp , Jan/Feb [8] R. H. Nelsen. Theory of Backpropogaton Neural Network. In Internatonal Jont Conf. on Neural Networks, pp Il592, [9] C. Schauder. Adaptve Speed Identfcaton for Vector Control of Inducton Motors wthout Rotatonal Transducers.IEEE trans. Ind. Applcat., vol. 28, no. 5, pp ,Sep/Oct [10] Y. Valle, G. Venayagamoorthy, S. Mohaghegh, J. Hernandez and R.Harley, "Partcle swarm optmzaton: Basc concepts, varants and applcatons n power systems, IEEE trans. on evolutonary computaton, vol. 12, no. 2, Aprl 2008.pp [11] S. Wahsh and A. Elwer, Improved performance of Permanent Magnet synchronous motor by usng Partcle swarm optmzaton technques, n Proc.of 2007 IEEE Internatonal Conference on Robotcs, 2008, pp [12] Roht G. Kanojya, Student, Y.C.C.E, and P. M. Meshram, Assocate professor, Y.C.C.E Optmal Tunng of PI Controller for Speed Control of DC motor drve usng Partcle Swarm Optmzaton, IEEE,2012 [13] Y. Tpsuwan, Y. Chow, "Fuzzy Logc Mcrocontroller Implementaton for DC Motor Speed Control", IEEE, [14] L. Rajaj,C. Kumar and M. Vasudevan, Fuzzy and Anfs Based Soft Starter Fed Inducton Motor Drve For Hgh Performance Applcatons. Sathyabama Unversty, Inda S.K.P. Engneerng College, Inda Vestas RRB Inda Ltd., Inda, Vol. 3, No. 4, August 2008 [15] Boumedene Allaoua, Abdellah Laouf, Brahm Gasbaou, And Abdessalam Abderrahman. Neuro-Fuzzy DC Motor Speed Control Usng Partcle Swarm Optmzaton, Department of Electrcal Engneerng, Bechar Unversty, B.P 417 BECHAR (08000) Algera

16 768 Sakuntala Mahapatra et al. / Proceda Computer Scence 48 ( 2015 ) [16] Makarand S. Ballal, Hralal M. Suryawansh and Mahesh K. Mshra Detecton of Incpent Faults n Inducton Motors usng FIS, ANN and ANFIS Technques, Journal of Power Electroncs, Vol. 8, No. 2, Aprl 2008 [17] J.S.R. Jang, ANFIS: Adaptve-Network-Ba sed Fuzzy Inference System, IEEE Trans. Systems, Man, Cybernetcs, 23(5/6): , [18] Jordan E. Morell, Akra Hrose, and Hugh C. Wood, Fuzzy-Logc-Based Plasma-Poston Controller for Stor- M, IEEE Transactons on Control Systems Technology, Vol. 13, No. 2, March 2005