Fault Diagnosis Method Based On Selective Neural Network Ensemble

Since modern industrial production systems have been increasingly larger, morecomplex and automated, once an accident takes place, it will bring huge economic lossesand casualties for the enterprise. Therefore, how to improve the fault diagnosis technologyto ensure the normal operation of the production system has become a serious problem.First of all, this paper presents the overall framework of fault diagnosis method basedon selective neural network. Then analysis the causes and the process by using selectiveneural network in fault diagnosis area, and points out two key technologies of thisdiagnostic system: one is how to build selective neural network ensemble, the other oneis how to reduce the fault data dimensions.For the problem of how to construct selective neural network ensemble, this paperproposes an improved binary particle swarm optimization method（IBPSO）, and thismethod is used to choose individual neural networks, then establishes the model ofselective neural network ensemble based on IBPSO. The results which compared withbinary particle swarm algorithm, K-means clustering algorithm and ant colony algorithmin standard data sets show that the generalization performance of the proposed model isimproved.Aiming at the problem of multiple fault features in modern industrial production, Thekernel principal component analysis method is introduced. A new fault diagnosis modelbased on KPCA and selective neural network ensemble is established by binding these twomethods. The proposed model is tested through the simulation platform of Tennessee-Eastman process and the actual case of steelworks air supply system. The outcome showsthat this algorithm has a higher correct rate, shorter diagnosis time than other neuralnetworks methods. Finally, a prototype system of fault diagnosis is developed based onselective neural network ensemble.