Abstract: To overcome the difficulty in feature extraction in bearing fault diagnosis, large amounts of noises in data and inability of models trained on individual working conditions to achieve effective fault diagnosis under complex ones, a bearing fault diagnosis method with variable working conditions based on Improved Convolutional Sparse Auto Encoder (ICSAE) is proposed in this paper. Firstly, sparsity constraint conditions are added to convolutional self-coding to improve the model’s capacity of effective feature extraction, and the reconstruction errors of input signals and reconstructed signals are constructed through a combination of maximum mean difference (MMD) and mean square error (MSE) to improve the model’s generalization ability and anti-noise ability. Then, the cross-domain diagnosis performance is effectively improved by domain adaptive method and MMD loss to reduce the difference of feature distribution between the two domains. To verify the effectiveness of this method, CWRU and JNU data sets are employed for rolling bearings under varying working conditions. The experimental results show the diagnostic accuracy of CWRU data set and JNU data set reaches 99.81% and 98.32% respectively, and the proposed model effectively performs the fault diagnosis of rolling bearings under complex working conditions.