The fault diagnosis method based on deep learning is widely used in the fault diagnosis of key mechanical components represented by bearings. The premise of achieving ideal results is that there are enough fault samples and the training set and test set meet the same distribution requirements. However， the data distribution will change under the actual working conditions， which makes it difficult to apply the diagnostic model under the original working conditions to the new working conditions. For this reason， the domain adaptation transfer learning method is used to solve the problem of different distribution of training sets and test sets， and its key point is to achieve data distribution adaptation， that is， to measure data distribution differences and use the mea? surement results to guide model training， which can effectively improve learning efficiency and diagnostic accuracy. On this basis， this paper proposes a new domain adaptation method based on adversarial learning. The core of this method is to combine the pro? posed exponential adjustment strategy with adversarial network to make the network adapt to different data distribution in source domain and target domain more specifically in the process of fault diagnosis. The network consists of a feature extractor， a classifi? er， a global domain discriminator， and multiple local domain discriminators， and the model is optimized by using the adversarial strategy and adaptive moment estimation algorithm， and adjusted the importance of marginal distribution and conditional distribu? tion by using the exponential adaptive factor set based on the exponential adjustment strategy， so that the model could diagnose faults stably and efficiently. The proposed method is verified in bearing diagnosis cases of cross-speed， cross-load and simultaneous cross-speed load. The results show that the method in this paper is better than other domain adaptation methods in diagnosis effect and has better stability.