Fault Detection and Identification of Dynamic Systems using Multiple Feedforward Neural Networks

Abstract Combining feedforward neural network (FNN) and multiple model adaptive estimator (MMAE), a new approach for fault detection and identification (FDI) of nonlinear systems as well as linear systems is proposed in this paper. Instead of Kalman filters, a bank of FNNs is used in the MMAE which are trained for the normal operation and possible fault situations. In order to overcome the drawbacks of the traditional BP training algorithm for FNN, singular value decomposition is used for the selection of hidden neurons, and then a new fast learning algorithm for training FNN by using a variable time-varying forgetting factor technique and U-D factorization baaed extended Kalman filter (EKF) is proposed. The new approach is then used for FDI of nonlinear systems as well as linear systems. The effectiveness of the method proposed is demonstrated by two simulation examples.