As a key component of mechanical equipment, bearing failure will cause major safety accidents and property losses. Therefore, effective diagnosis of bearings is crucial. The success of bearing fault diagnosis relies on the following two conditions: (1) a large number of labeled fault data for model training; (2) the training data and prediction data satisfy the same probability distribution. However, some machines are not easy to obtain large amounts of labeled data. Moreover, multiple factors can impact data distribution of bearing failures. Accordingly, it is difficult to satisfy the same probability distribution for both training and prediction data. These problems limit the application of intelligent fault diagnosis methods for bearings. Therefore, a bearing fault diagnosis method based on transfer learning is proposed. The method uses the multiscale residual network to automatically extract signal features and perform fault diagnosis. Meanwhile, multi-kernel-maximum mean discrepancy is used to reduce the data distribution discrepancy between source and target domains. The effectiveness of the proposed method is verified through experiments in comparison with other models. Furthermore, to ensure robustness and high generalization capability, the model undergoes generalization error analysis. The proposed method provides a certain value for the research of bearing fault diagnosis.