Abstract: A Deep Belief Network(DBN) fault identification method based on cross-entropy cost function is proposed to solve the problems of slow convergence rate and low recognition accuracy in traditional rotor system fault diagnosis. Unsupervised algorithm is used to initialize parameter space of Restricted Boltzmann Machine(RBM), cross-entropy cost function is used to transfer errors in the opposite direction and optimize parameter space, and the deep model of RBM is built after layer-by-layer stack resetting. The intelligent identification database of rotor system is established by using the existing data. MNISI hand-written digital set and rotor system fault data set are used for verification, compared with the traditional DBN, Deep Belief Network based on cross-entropy cost function can eliminate the influence of the gradient of the activation function on the updating speed of the parameter space, and improve the accuracy of the classification effectively.