Aiming at the problem of low recognition accuracy of the training model in the case of less data samples in deep belief network (DBN),which led to the classification recognition rate is not ideal, the system performance needs to be improved. Research the equivalent model of DBN, analyze the problem of poor recognition rate in the case of small samples ;Then propose a weight expansion algorithm to enlarge the matching space between the sample and weight ,so that the decision is more conducive to correct classification, which improves the accuracy of image classification under the condition of small sample size; And the algorithm is proved that can promote the performance of the system by using the detection and estimation theory, the test of different sample banks is further proof of this completion. Finally, the method is applied to small sample faulted insulator recognition.