Aiming at drawbacks of current methods for power electronics circuits feature extraction. there is not enough accuracy and not obvious classification. The process of feature extraction is easily affected by noise. Firstly the faulty circuit information feature was analyzed and extracted by cross-wavelet transform. Initial feature matrix are obtained representing cross-wavelet spectrum. Finally principle component analysis is applied for reducing the dimension of initial feature matrix and those which redundant information are eliminated. The back propagation neural network classifiers are utilized for fault diagnosis simulation test. The results show that the fault detection accuracy is up to 98.2%.Simulation results demonstrate that the proposed method is accuracy.