Abstract: It's difficult to determine the structure of deep belief network (DBN), so raise the method that use the wolves algorithm to determine the number of DBN per layer of neurons, and K-means Clustering Accuracy Determines whether or not to increase the hidden layer method. According to the minimal reconstruction error function of wolves algorithm, it can work out the quantity of each neuron, so as to determine the initial structure of the DBN. In order to verify the effectiveness of the DBN structure, using DBN to extract the data characteristics of the clustering test, so can get the final structure of the DBN. The Iris dataset was tested in the experiment, the results show that the proposed method handles the raw data by the effective structure of the DBN obtained and improve the accuracy of clustering.