Objective A bearing fault diagnosis method based on multiscale subdomain adaptation is proposed to address the significant differences in the distribution of vibration signal features of bearings under varying operating conditions and the potential inadequacy of a single receptive field in effectively capturing multi-scale fault features. Methods First local maximum mean discrepancy LMMD was employed to quantify the distribution differences of bearing fault samples from the same category within the feature space under different operational conditions. This step was crucial for extracting domain-invariant features which ensured the accurate capture of fault characteristics regardless of the varying operational states. Second a multi-scale dilated convolutional network architecture was constructed during the feature extraction stage. This design utilized multiple receptive fields to secure the optimal feature extraction scope for each operating condition thereby preventing the omission of critical fault information due to a limited receptive field size. Subsequently a squeeze-and-excitation SE attention mechanism was introduced to further enhance the model?? s capability to discern sub-class fault features across all conditions. This mechanism assigned varying weights of attention to features from different scales enabling the model to focus more intensively on features closely linked to fault diagnosis while suppressing redundant information. This process strengthened the consistency and effectiveness of the sub-class fault feature representation under various operating conditions. Results Finally an empirical analysis was conducted on a public dataset. The experimental results demonstrated that the proposed method achieved an average diagnostic accuracy of approximately 98% across all operating conditions. Conclusion These findings confirm the effectiveness and superiority of the proposed method. Compared with other existing diagnostic techniques the proposed method exhibits higher diagnostic performance and greater reliability.