In order to solve the problem of feature learning in traditional bearing fault diagnosis, which were necessary to master a large number of signal processing methods and diagnostic experience, a new identification method is proposed to classify the bearing fault states directly from the original data. The method was used of the original vibration data to train the stacked denoising autoencoder network. Due to the elimination of the explicit feature extraction phase of intelligent diagnosis, the artificial participation factors could be reduced and the dependence on a large number of signal processing technology and diagnosis experience was gotten rid of. The experimental results show that the proposed method has a good recognition ability and practical significance to the bearing fault identification rate of 97%, and can achieve adaptive extraction of fault feature, which enhances the intelligence of machinery fault diagnosis.