Counter propagation（CP） arithmetic is one kind mixture arithmetic complsed supervised learning algorithm and unsupervised learning algorithm.It has good pattern recognition performance

and can solve the convergence problem of feedback neural network.This paper presents a fault diagnosis method for power transformer based on CP neural network

and builds a compound neural networks model.The training times and the number of neurons on a competition level in CP network are determined by comparing the results in different training situations.The diagnostic examples indicate the diagnostic correctness of the proposed method is higher than that of the improved three rations method.