Intelligent fault diagnosis of rolling bearings is important for guaranteeing the safe of equipment. However， the non-sta? tionary conditions lead to the incomplete collected training datasets， which makes the data-driven model learn the limited diagnostic knowledge. This declines the testing accuracy observably. To solve this problem， a Standard Self-Learned Data Augmentation （SSDA） fault diagnosis method is proposed， which can generate disturbed samples to expand the completeness of the original data? set. The method includes two training steps： standard self-learning and data augmentation. The training process of one-dimensional convolutional neural network is regarded as the self-learned standard of model to judge disturbed samples. Based on this standard， sample parameterization and model datalization are used to generate disturbed samples. By alternately carrying out the two steps， not only the disturbed data can be generated to augment the completeness， but also the fault diagnosis model under non-stationary conditions can be obtained. In addition， by studying the sample differences with different data generating number， it is found that the randomness of distance and direction is superimposed on the randomness of the proposed method to guaranteeing the diversity of the generated samples. Experimental results show that the proposed method is effective and advantageous in diagnosing bearing fault with incomplete training data sets under non-stationary conditions.