ResidualAdaptive Algorithm AppliedinIntelligent Real-time Calculation Ofcurrent RMS Value DuringResistanceSpot Welding

Chengliang Liu,Tong Jiao

2005-01-01

Abstract:cn Abstract-To solve thelarge residual problems, whichmay occurduring feed-forward neural network weight training, a comprehensive residual adaptive algorithm isproposed togive abetter stability compared tostandard Levenberg-Marquardt (L-M)algorithm andhaslesscomputational complexity than classical Newtonmethod. Thecomparison withstandard L-M algorithm checks thebetter performance ofthisalgorithm. Thenthewell-trained neural network isembedded into aDSP controller toperform real-time calculation ofcurrent RMS valueduringresistance spotwelding. Experimental result showsthevalidity oftheresidual adaptive algorithm andthe feasibility ofanintelligent current measuring method. I.INTRODUCTION Asanapplicable andveryeffective non-linear predictive method, backpropagation (BP)network calculation is adopted inawidevariety offields such assignal processing, function approximation, financial forecasting andpattern recognition. In1986Rinehart originated theerror back propagation algorithm (1), andsubsequently itisextensively employed intraining theBPnetwork. However, EBPhas theproblems ofslowconvergence andlocal optimum. And hence, someimproved network training algorithms based on EBP areintroduced (2-9), andinwhichtheL-M (Levenberg-Marquardt) algorithm (5)isconsidered tobea relatively successful onealthough italso suffers alotfrom computational complexity andconvergence uncertainty. In this paper, acomprehensive algorithm, residual adaptive algorithm, isputforward to resolve thenetwork misconvergence problem thattheL-Malgorithm cannot settle alone. A nonlinear function approximation result indicates that this algorithm possesses thefeatures offast convergence andeasycomputation. Andthewell-trained neural network paves thewayforanovel real-time current RMS(Root MeanSquare) value calculating method during resistance spotwelding. Experimental result showsthe validity andreliabity oftheresidual adaptive algorithm and theintelligent measuring system.