Research on the controller of an arc welding process based on a PID neural network Research on the controller of an arc welding process based on a PID neural network

Research on the controller of an arc welding process based on a PID neural network

  • 期刊名字:控制理论与应用(英文版)
  • 文件大小:805kb
  • 论文作者:Kuanfang HE,Shisheng HUANG
  • 作者单位:College of Mechanical Engineering
  • 更新时间:2020-11-11
  • 下载次数:
论文简介

J Control TheoryAppl 2008 6(3) 327-329DOI 10.1007/11768-008-6189-9Research on the controller of an arc welding processbased on a PID neural networkKuanfang HE, Shisheng HUANG(College of Mechanical Engineering, South China University of Technology, Guangzhou Guangdong 510640, China)Abstract: A controller based on a PID neural network (PIDNN) is proposed for an arc welding power source whoseoutput characteristic in responding to a given value is quickly and intelligently controlled in the welding process. The newmethod syncretizes the PID control strategy and neural network to control the welding process intelligently, so it has themerit of PID control rules and the trait of better information disposal ability of the neural network. The results of simulationshow that the controller has the properties of quick response, low overshoot, quick convergence and good stable accuracy,which meet the requirements for control of the welding process.Keywords: Welding process; Characeristic of output; PID neural network; Contoller1 Introductiondigital control technology in the arc welding field. Some in-At present, arce welding power sources contolled by mi- tllualized control algorithms have been adopted for thecroprocessor are used widely in the field of arc welding [1]. arc welding power source, which in tum has good quality,Arc welding is a complex process [2]. The output character- high credibility of the U-I and a dynamic charateristic; itistic of welding is a controlled object that is nonlinear, time-also has the fine effect of output quality. A negative feed-changing and uncertain, and owns time-changing, nonlinearback control system of two parameters which have arc weld-traits, many disturbed factors and so on, so it is difcult to ing voltage and current is shown in Fig.1.establish an accurate mathematical model. Because of the| AcH AC/DC H Filter, H DC/AC H ACDC H Arccomplexity of the arc welding process, the PID controllerAH circuit r H circuit H circuit H circuit H loaddoes not meet the requirement of the welding process whichhas been widely used recently. As artificial intelligence isI Driving circuitDetect and testapplied to welding, the neural network algorithm is a repre-CircuitiIcuLsentative method which can replace the mathematical model了of the controlled object completely. The neural network al-LDIA[ Analoggorithm has better robustness than that of adaptive control个| multi -switchin the systemic identifcation of moderm control theory inthe control process of uncertain and nonlinear objects. How-MircorcssoHAD ], (Algorithm)ever, the traditional neural network has the disadvantage ofhaving a slow rate of network-learning convergence; it isFig. 1 Sketch of welding power controlled by microprocessor.very difficult to obtain the anticipated purpose because ofIn the controller, feedback signals of the arc welding volt-the random itilization weights of conections in the net- age and curent are changed into digital signals by the ADworks, so the traditional neural network algorithm cannotcircuit of detecting and testing, and sent into the micropro-meet the quick response and real-time requirements of thecessor. The interrupt program runs and executes to save theare welding process. A new algorithm using the PID neu- data on the fedback are welding voltage and curent sepa-ral network, which syncretizes PID control rule with neuralrately, and then deals with the data through transfering thenetwork superiority, can achieve quick convergence shorten memorized data at real time. When the adusing offset isthe learning time and advance the algorithm rate of the neu-wrapped by outside interference, the CPU of the micropro-ral network. The papers [3~5] introduced some succesfulcessor can judge the property and magnitude of the errorexamples of PID neural network methods applied to projectby synthesizing former memorized data. Then it orders thecontrol which obtained fine effects, but the related controlcorresponding algorithm to calculate the new control valuesmethod of the PID neural network has never been applied towhich are changed into analog signals to drive the main cir-the welding field. In this paper, an arc welding process con-cuit of the arc welding power source according to the giventoller based on PID neural network is studied to guarantee values. The above mentioned process acomplishes one ad-the quality of the arc welding process control by the meritsjustment in a sampling cycle. It needs continual circulationof PID and neural networks.of adjustment to bring the output close to the given values.The dynamic responding rate and controlled quality are de-2 Arc welding systemtermined by choice and the achievement of the algorithmsThe advanced control algorithm applied to a welding in the中国煤化工,a new type of con-power source is achieved along with the development of tolleralgorithm to quicklyYHCNMHGReceived 24 October 2006; revised 21 June 2007.This work was supported by National Nature Science Foundation of China No.50575074).328K. He et al.1J Control TheoryAppl 2008 6(3) 327-329achieve output adjustment and itelligent control of the arc and differential coficients of the hidden layers are shownwelding power source is developed in this paper.in formulas (4)~(6). Expression of output is shown in for-mula (7),3 PID neural network controller of the arcu山= u1(k),(4)welding process以=u2(k-1) + u2(k),(5)3.1 Structure of the PID neural network controller ofu的=us(k) -u3(k-1),(6)the arc welding process1,u;(k)> 1,The arc welding process controller based on the PID neu-ral network is a three layer feed forward network which in-x};(k)= { u(k),-1≤u(k)≤ 1,(7)cludes an input layer, a hidden layer and an output layer, andu;(k)<-1.whose structure is shown in Fig.2.The subscript j is specified as j = 1, 2, 3. The output layercontains a neuron which completes the summation of net-works, whose input and output expressions are shown informulas (8) and (9) separately.u(k)=它wlzx;(k),(8)=/w(1,u";(k)> 1,x"(k)=<"(),-1≤u()≤1,(9)1,u(k)<-1. .Welding powergwj are the weights of the connections between the hid-den and output layers, whose initialization values are" Feedback value ofAre loads-chosen according to the characteristic of the PID con-l voltage and current y JFig. 2 Sketch of welding process controller based on PID neural network.trol algorithm u"(k) = Kp + Ki 2 e(i) + Ko(e(k) -The input layer has two neurons corresponding to the e(k - 1)). Performance function is specifed as E =feedback arc welding voltage or curent and the given val-ues r of voltage or current. The hidden layer has three neural1 S ()-()2 = 1之e(k)2 which is tbe rule andcells corresponding respectively to the proportion and thepurpose of learning. The weights of the connections are ad-integral and differential cofficients of the PID model. Thejusted according to the grads method, where the iterativeoutput layer completes the synthesis of PID-NN rule con-equation is shown in formula (10),trol. The control values U are generated by the controller,8Ewhich are changed into analog signals by a D/A circuit, arew(m+ 1) = w(m) -η'Ow”10)sent to the main circuit of the arc welding power source, andexecuted quickly to adjust output values. The control rule is where η is the length of a leaming step. The change of theachieved by a feed forward algorithm, and the parameters weights of the connections between the hidden and outputof PID-NN are adjusted by a back propagation algorithm.layers is shown in formula (11). The change of the weights3.2 Learning algorithm of the PID neural networkof the connections between the input and hidden layers isThe input layer of PID neural networks has two neurons shown in formula (12) withoy_ y(k+1)- y(k)corresponding to feedback values and given values respec-(k)-0(k-1)tively, which is shown in Fig.2. The input ui and output xi8E 8EOy 8v 8x" u”of the input layer are shown in formulas (1) and (2). The可yo’ ax" Fu" q;= -e(k)"r(k), (1))u"subscript i,j are specifed asi= 1,2,j = 1,2,3, respec-tively.aE 8y 8u ax'u1=(k), u2= y(k),(1)8wij8y 8v 8x' Ou'x(k)-x)(k-1)(12)(1,uj(k)> 1,Wij-=--e(k)Fnjsgn;'" u(1)-u(k-1)x;(k).zx(k)={uj(k),-1≤u(k)≤1,(2)uj(k)< -1.The hidden layer has three neurons corresponding respec-4 Simulations and analysistively to the proportion, and the integral and differential co-This paper takes an IGBT inverter TIG welding powerefficients, whose expressions are all shown in formula (3).source for the simulation example. The simulation modelof the welding power source is established by SIMULINKu;(k)=亡Wjz:(k). .(3) of MATTiule of the electrical=1power s中国煤化工(5 The snulatononWij is the weights of the connections between the input and model|YCNMHGudesthemaincirhidden layers, whose initialization values are chosen ac-cuit, theubsysem. The con-cording to the characteristic of the PID control algorithm trol subsystem is composed of the PID neural network con-e = r- y. The expressions of proportion and the integral troller. The structure of the neural network is 2- 3 -1, and theK. HE et al.1J Control Theory Appl 2008 6(3) 327 -329329input value is x(1) = [y(k),r]. The arc welding current of00 rthe DC output of the welding power source is controlled by80 Fthe controller based on the proposed PID neural network.60 FIn the PID neural network, the learning step length p is en-40-acted by p = 0.2, the initialization weights of the connec-tions between the input and hidden layers are specifed as20 Fw11(0)= w21(0)= 1, w23(0)= w13(0)= -1, w12(0) =400 wwwwwM.wwwwwMNW0.1, w22(0) = -0.1, and the initialization weights of theconnections between the hidden and output layers are spec-ified by Kp, K, Kp which are 0.2, 0.25,0.2.The arc welding current response curve of this simulationsystem from idling to the load controlled by the PID neuralnetwork after five learning steps of input values is shownin Fig.3. As shown in Fig.3, it just takes 2 ms for the arc300210welding current to go from 0 A to 400 A.t/msFig. 5 Output current response while load changing.00 -5 Conclusions00 FAn arc welding process controller based on a PID neural00network, which makes use of PID control rules and the meritof neural networks, can achieve quick real-time and stablecontrol of arc welding. The results of simulation show thatthe controller based on a PID neural network has the prop-02412erties of quick resporise, low overshoot, quick convergenceand good stable accuracy, which provide the arc welding1/ msprocess with an effective control method.Fig. 3 Output curent afer five learning steps.The arc welding current response curve of this simulation Referencessystem from idling to the load controlled by the PID neural [1] K Wu, s. Huang. Research on the digital control of arc welding powernetwork after ten learning steps of input values is shown insource[J]. Design and Exploitation, 2005, 2(4): 72 -75.Fig.4.[2] S. Huang, J. Xue. A discussion of the inelligeot control and novelpower sources for arc welding[小J. Engineering Science, 2000, 2(4): ;[3] H. Shu, Y. Pi. The analysis of PID neurons and PID neuralnetworks[CW/Proceedings of Chinese Control Conference. Ningbo:Ningbo University, 1998: 607 - 613.00 t[4] H. Shu, Y. PL. PID neural networks for time delay systems[J].Computers and Chemical Engineering, 2000, 24(2): 859 - 886.[5] H. Shu. PID Neural Networks Algorithm for Control Applicarion[M].60 12Bejing: Publishing House of National Defence Industry. 2006.t/ ms6] G. Li. Inelligent Control with MATLAB Achievemen[M. Bejing:Publishing House of Electronics Inodusty, 2005.Fig. 4 Output curent after ten learning steps.As shown in Fig.4, it just takes 1ms for the arce weldingKuanfang HE rceived the M.S. degree in Me-current to go from 0 A to 400 A.chanical Design and Theory from College of Me-chanical Engineering,Hunan University of Sci-system when the original load suddenly changes from 0.20ence and Technology in 2006. Now he stud-to 0.102 is shown in Fig.5.ies for Ph.D. degree in Mechanical Engineer-ing, South China University of Technology. HisAs shown in Fig.5, when the load changes suddenly incurrent research area includes itelligent controltransient time, the arc welding current peak of 480A instan-国of digital power and welding process. E-mail:taneously appears in the cut, but it just takes 0.5ms to returnhe.kuanfang@ mail.scut.edu.cn.to the previous state. It fully proves that the arce weldingcontroller of the PID neural network has a speedy dynamicShisheng HUANG reeived the B.S. degree fromresponse characteristic and quick convergence rate.Huanan University of technology in 1964. He isAs shown in Fgs.3~s5, the results of simulation show thatcurrently a professor in College of Mechanical En-the controller based on the PID neural network has proper-中国煤化Iincludes new typee arrties of quick response, low overshoot, quick convergencenelligent control.and good stable accuracy, which can meet the requirementsHCNMHGfor real-time control of the arc welding process.

论文截图
版权:如无特殊注明,文章转载自网络,侵权请联系cnmhg168#163.com删除!文件均为网友上传,仅供研究和学习使用,务必24小时内删除。