Nonlinear FaultDiagnosis basedon RBFwithSliding WindowErrorFeedback
Mingxing Jia, Xiaoping Guo · 2006
Nonlinear fault diagnosis isoneofthedifficulties in fault diagnosis field. Thepaperpresents thenonlinear fault estimator basedonRBF withsliding windowerrorfeedback fora class ofnonlinear system. Theinputofestimator is input andoutputofthesystem, andtheoutputisthefault estimate. Theneuralnetworkweight adjusting algorithm adopts sliding windowerrorfeedback, whichenforces the amountoffault information andspeeduptheconvergence. Thepaperanalyses therobustness ofalgorithm andthe windowlengthinfluence uponfaultestimate, gives the variable windowlength strategy, andqualitatively presents a methodofchoosing windowlength. Thesimulation results provethatthemethodimproves greatly theresponse speed andaccuracy infault diagnosis underthecircumstances of choosing theproperwindowlength. I.INTRODUCTION Thegrowing needs offault detection andisolation (FDI) forcontrol systems haveattracted a lotofattentions. Especially, itisveryimportant forcomplex systems. With thedevelopment offault diagnosis andneural network technology, amethodcombining observer withneural networkis used to solvenonlinear fault problems(ALC,97) ,(JIA,03), (REN,00), (MA,02). The observer isusedtocompose thewholeframe, andthe neural network diagnoses fault online asfault estimator. Thetransforming function ofRBFnetwork hidden units is radial basis function, which cansolve nonlinear problems. Inaddition, thetransformation fromhidden layers to output layers islinear. Whenthecenter ofRBF is confirmed, theweights ofnetwork canbesolved directly bylinear equation grouporrecursive computation using iteration least square algorithm. Consequently, itquickens thelearning speed greatly, avoids thelocal minimum value problem andhavegoodsample grading andfunction approximate feature (LIU,03), (LIU,01), soonline fault estimator chooses RBFnetwork mostly. RBFnetwork learning algorithms aremostly: Moolyand Darkenalgorithm(MOO,89)