A novel approach for fault detection and isolation (FDI) is proposed. In order to detect the faults that reflect themselves as fault-induced frequency changes at certain time instants in the measured signal, wavelet analysis is applied to capture such changes and extract fault features on line and in real-time. An improved self-organizing feature map (SOM) neural network is then used to isolate the fault. By introducing the concept of hierarchy training and zone recognizing, the improved SOM neural network proposed in this paper has achieved higher clustering and matching-up precision compared to the conventional SOM network. Therefore, the proposed FDI scheme is more accurate.