Aimed at the current situation that the amount of mechanical equipment fault is large and diverse and the fault features are extracted mainly by supervised learning, a stacked sparse autocoded depth neural network was proposed. The unsupervised learning to extract the inherent characteristics of vibration signals was realized,and it was used to fault diagnosis of rolling bearings. The spectrum envelope was used as the low level input to train network layer-by-layer to obtain the fault feature expression, and it was input into Softmax classifier to achieve fault classification. The optimization algorithm was used to fine tune the whole deep neural network to improve the classification accuracy.The rolling bearing fault diagnosis experiment results show that the proposed deep neural network can be used to realize fault diagnosis more accurately; while ensuring the accuracy, taking the frequency spectrum envelope as the low level input can improve the computational efficiency