Research On Differential Evolution Neural Network Algorithm

Artificial neural network is the complex network system by interactiveconnection of massive neurons and simulating brain'S method in parallelprocessing and the non-linearity transform It has very strongly auto-adapted．auto-organization,auto-learning capability,and has widely applied in faultdiagnosis,pattern recognition,hydrology forecast and SO on Weights training andstructure optimization are of the most important issues that affect a typical neuralnetwork learning capability,but the traditional neural network training algorithmsuch as the BP algorithm has some own inherent flaw-s:the search efficiency islimited and easy jump into the local optimal points In recent years,due to themerits of robustness and parallel computing abmHeV01ution algorithm has moreand more applied in training weights and optimize the structure of neuralnetworkBecause of its easy to use,robust,and has excellent global convergenceproperties,differential evolution(DE)algorithm has the ability to handlenon-differentiable,nonlinear and multimodal cost functions,However,as thesearch of DE algorithm is of some blindness,its local search ability and thewhole search efficiency is limited,the convergence rate during the latter part ofsearch is slow',Based on analysis of the principle and the shortcoming of DE,amodified differential evolution(MDE)algorithm w-as proposed,then the MDE w-asapplied to train the neural network and optimize the structure of neural networkfinally the MDE and neural network w-as combined and applied in fault diagnosisof analog circuits to verify the validity and practicabilityThe main COntents and research COntributions of this dissertation are asfollows:(1) As the local search ability of DE is w-eaR,and the convergence speedis slow"during the latter part ofthe search Based on analysis of search mechanismof DE,a modified DE with hybrid optimization strategy w-as proposed,whichmakes a little of individuals of the population search around the current bestindividual to enhance its convergence speed,and makes a majority of the population search on the other area to enhance its global search ability the newalgorithm show be~er convergence speed and be~er global convergencecharacteristic when dealing with more dispersed issue compared with the basicDE algorithm。(2) As the neural network is non-linear and is hard to globaloptimization,a new neural network training algorithm is proposed which basedon the modified differential evolution algorithm,then train the neural network bymeans of BP algorithm,PSO algorithm,DE algorithm and MDE algorithm andthe test result show that the MDE algorithm is prior to other algorithms(3) A new neural network structure optimization algorithm is proposedwhich based on the modified differential evolution algorithm It optimize theimplicit strata node'S number and other network parameter on method ofalgorithm nesting The simulation results show the effectiveness ofthis method(4)The neural network and DE are combined and applied in faultdiagnosis of analog circuits The simulation results show that the differentialevolution neural network can distinguish the different fault modes effectively...