Abstract: A new type of nonlinear observerbased robust fault detection and isolation (FDI) using neural networks is presented in this paper. Firstly, a radial basis function neural networks is used to approximate the nonlinear item of the monitored system to improve the accuracy of state estimation, and the state estimation error is proved to be zero asymptotically. On the other hand, a neural network classifier is applied to identify the type and location of faults. In order to improve the robustness of fault classification, the neural network has been trained with noise injected inputs and the generalization capability an remarkably be enhanced. Therefore, this FDI strategy has good robustness against modeling error and environment disturbance. At last, this method is applied to faults detection of fighter aircraft with structure damage, and simulation results reveal that this FDI strategy is effective.