Abstract: In order to meet the changing processing requirements of products, bearings need to work in different working conditions. Aiming at the problems of the performance limitation of bearing fault diagnosis model due to data distribution difference caused by variation of working conditions, to the same time, there was currency a lack of methads to achieve balanced sampling of fault categories in the absence of labels, 
a bearing fault diagnosis method based on model-based sampling (MBS) and domain adversarial neural network (DANN) was proposed. Firstly, model-based sampling (MBS) was used to take the classification probability distribution of the output of the pre-training model as the sampling basis, which overcame the difficulty of realizing the classification equilibrium sampling under the target condition of unlabeled samples. Then, combining with domain adversarial neural network (DANN), a network structure was designed to transfer features from target to source conditions. Finally, a high-precision bearing fault diagnosis model under variable working conditions was realized based on the fault simulation experimental data, and the results were compared with those of several control methods,to verify the effectiveness and superiority of the method in variable condition bearing fault diagnosis. The research results show that the average diagnostic accuracy of the proposed method is 98.41% in the simulation experiment, which is more than 10% higher than that of the random sampling method; which can reveal the importance of class equalization sampling for unlabeled samples, and also verify the effectiveness and superiority of the proposed method in bearing fault diagnosis under variable conditions.
Key words:  fixed thickness roller bearing; unlabeled sampling method; classification equilibrium sampling; feature transfer learning; modelbased sampling(MBS); domain adversarial neural network (DANN)