Abstract: According to the Takens embedding theorem, the nonlinear time series combined with system output is converted into discrete dynamic system. 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 for the unknown disturbance. The prediction of time series is achieved by the observation of system states. So a one-step prediction method is proposed. Using the so-called moving horizon optimization method, one-step prediction is extended to N-steps prediction. The boundedness of prediction error is proved. The fault is predicted by the prediction error through density function estimation and hypothesis test. The simulation of the structure fault prediction of a fighter F-16 proved the efficiency of the model.