In order to calculate the depth of deep belief network(DBN) in its applications, the reason of failure in training by using random initialization in gradient-based is analyzed in both math and biology, and then verified by the test. The theorem that the reconstruction error of restricted boltzmann machine(RBM) is related to network’s energy function is proved. After that, a method to calculate the depth by using restructure error in RBM is proposed based on the relationship between hidden layers and errors. DBN approaches human-level performance in AI tasks after the self-training. The experiment of hand writing digital recognition shows that the proposed method can improve the efficiency and lower the cost.