The deep learning model increases the number of hidden layers, which makes the model have a good effect on speech recognition, image video classification and so on. However, to establish a model suitable for a specific object, a large number of data sets are required to train it for a long time to get the appropriate weights and biases. To resolve the above problems, a sub-health diagnosis method for rolling bearing was proposed based on depth autoencoder-relevance vector machine network model. Firstly, the bearing vibration signal was collected and transformed by Fourier transform and normalization. Secondly, the improved automatic encoder, named sparse edge noise reduction autoencoder, was designed, which combined the features of sparse automatic encoder and edge noise reduction automatic encoder. Then the depth autoencoder-relevance vector machine network model was designed, in which the supervised function was used to finely tune the parameters of each hidden layer, and it was trained by Relevance Vector Machine (RVM). Finally, the final classification results were obtained according to D-S (Dempster-Shafer) evidence fusion theory. The experimental results show that the proposed algorithm can effectively improve the recognition precision of the "sub-health" state of the rolling bearing and correct the error classification.