Control System Failure Detection And Diagnosis Based Neural Networks

As current automatization level is improved and industry control system is increasingly complex,System failure detection and diagnosis (FDD) is increasingly important too. The research of neural networks (NN) develops quickly,which make a new way for FDD.The one main way that NN is applied to FDD is dynamic system identification with feedforward networks. To quicken convergence and improve model precision,a new algorithm is presented in this paper,which utilize construct orderliness property of self-organization feature maps (SOFM),divide system input space and adopt 1 order or 2 order local model in each subspace individually instead of a global model. The parameter of local model can be calculated by gradient descent in neighborhood with the SOFM weight together,or estimated by Least-Squared Estimation (LSE). In the algorithm,the concept is more clearly and more perspicuous. Simulation results show the algorithm has more rapid convergent speed and high fitting precision.After NN has been introduced to Principal Component Analysis (PCA),the algorithm of principal component calculation is simplified and has higher precision and finer convergent property. In FDD. PCA is mainly applied to the process that has direct redundancy. In this paper,PCA is utilized in modeling sensor sequence,and it solves the sensor FDD that has indirect redundancy,and expands the PCA application field. The algorithm has been applied in linear and nonlinear system.