Abstract: In view of the problem of feature extraction for analog circuit fault diagnosis, model of deep belief network (DBN) is introduced in this paper. Comparison experiments on Leapfrog Filter with a group of models including DBN, single hidden layer BP and three hidden layer BP prove that the model of DBN is feasible and effective. Moreover, according to the experiment on band pass filters with methods of wavelet analysis, DBN and single hidden layer BP network, the results show that the ability of feature extraction of DBN is better and the feature of extraction is closer to the essence of data.