Objective Aiming at the problem of difficulty in feature extraction and domain alignment under different operating conditions of bearings a multi-domain joint adaptive fault diagnosis method was proposed which included feature extraction networks classifiers and domain discriminators. Methods Considerations were made from the aspects of feature extraction and domain alignment. By utilizing deep separable convolutions and incorporating the idea of attention mechanisms the DRWNet feature extraction network was constructed to improve the network?? s capability to extract deep features from vibration signals. Through building multi-domain discriminators the transferability of different category samples was quantitatively evaluated and difficult-to-transfer samples were reweighted to fully align the data distributions between the source domain and the target domain thus improving the diagnostic accuracy of the model. Results Simulation experiments demonstrated that on the Case Western Reserve University bearing dataset M-DJC achieved diagnostic accuracy of over 99% on 12 transfer tasks. Compared with DANN CDAN DDAN MRAN and MRDA M-DJC showed an improvement in diagnostic accuracy ranging from 1. 84% to 7. 44% with accelerated convergence speed and enhanced stability of the model. Conclusion The M-DJC model not only reduces the noise impact in bearing vibration signals but also enhances the domain alignment capability of signal features better meeting the requirements for bearing fault diagnosis under actual operating conditions.