Fault Detection & Fault Prediction For Complex Nonlinear System

The neural network based fault detection and prediction is studied in this dissertation. The methods, which combine time series analysis and neural networks, are especially studied and applied in the model-unknown nonlinear system. The main contents of the dissertation are as follows:At first, the problem of fault fast detection for a class of uncertain nonlinear systems is addressed. The new full-order unknown input observer is adopted to estimate the state and the fault at the same time. A dead zone function is used in the adjust rule of the weight matrix of fault observer neural network, so that the robust of fault is improved under the uncertainty of system.Based on the concept named structure difference among neural networks in ensemble, a negative selection immune algorithm is designed. The algorithm can keep the structure difference among networks meanwhile reducing the train error of the network. The generalization of neural network ensemble is improved; thereby the tiny fault is predicted accurately. From the simplified mathematical model of the immune mediated by T lymphocyte, a new controller structure is constructed with the RBF neural network. The output of controller can report the fault at the expected time.Then, because of the characteristic of complex engineering systems like fighters, such as modeling difficulty, multiplicity work situations, difficulty and expensiveness for test, a new kind of fast fault predictor is designed based on an improved k-nearest neighbor method, which neither need math model of system nor need data for train and knowledge. It studies and predicts at the same time of system's running. So it overcomes the difficulty of fault data collection. This predictor has a high prediction speed, and the false alarm rate and missing alarm rate can be adjusted randomly. The method is simply and universalizable.Further the fault prediction for model-unknown nonlinear system is studied, which is based on the neural network and time series.According to the Takens embedding theorem, the nonlinear time series is converted into discrete dynamic system. The prediction of time series is achieved by the observation of system states. An autoregressive model is used to fit the linear part of series; the neural network is used to fit the nonlinear part of series and to compensate the unknown disturbance. By this method, the fault can be predicted several steps ahead.Based on the statere configuration system, an unknown input observer is presented for the prediction of time series. The approximate error of the AR model is regarded as the unknown-input of system. So an N-steps-ahead moving horizon prediction method is proposed. The fault is predicted conveniently by the prediction of unknown-input and prediction error.A subspace state space identification approach is applied in unknown nonlinear system. A linear state-space model is modeled for the system identification and prediction at the tangent space of working point. The errors of modeling and...