Neural network based fault detection and identification for fighter control surface failure

As a representative complex system, the aircraft modeled very difficultly and imprecisely. This makes the model-based fault detection methods degenerated. In this dissertation, the nonlinear time series, which is constructed by output variables of aircraft, is converted into discrete dynamic system, and then a novel series prediction method is achieved by the adaptive observation of system states. An online adaptive RBFNN is used to fit the nonlinearity of system and to compensate the unknown disturbance. Thereby a one-step-ahead prediction method is proposed. By using probability density estimation and hypothesis testing for the observation error, the fault is detected directly. Finally, a rule-table is established for fault identification. The results of simulation prove the method´s efficiency.