Abstract: This paper put forward a new extended Back-Propagation neural network based on the standard Back-Propagation neural network, which can get rid of the shortcoming of slow convergence speed and is easy to enter local maximum. A structure of a main neural network (MNN) and three decentralized neural networks (DNN) are designed. The parameter estimations are given by on-line learning of neural network. The fault can be fount out by comparing the estimators to the actual data acquired by sensors, and the examples of simulation are given.