Effective fault diagnosis of bearings is crucial due to their susceptibility to damage and the significant impact of failures on industrial systems. This paper introduces a novel fault diagnosis method that integrates an attention mechanism with multi-scale deep transfer learning to enhance the accuracy and robustness of bearing fault detection. The detailed characteristic representation of bearing signal is extracted by time-frequency analysis method. By applying an attention mechanism, the model prioritizes critical features, improving the discrimination of fault types. Furthermore, the use of multi-scale convolutional kernels within a ResNeXt architecture facilitates effective feature extraction and adapts the model to new operational conditions through transfer learning. Experimental validation on the Case Western Reserve University bearing dataset demonstrates that our method achieves superior diagnostic performance compared to traditional approaches, confirming its potential for real-world applications in condition monitoring and fault diagnosis systems.