Purpose <p>Based on multi-source domain adaptation technology, which effectively combines feature information with diagnostic knowledge, has become a reliable and efficient solution for diagnosing faults under variable operation conditions in the overhaul of modern machinery and equipment. However, compared with single-source domain adaptation methods, multi-source domain adaptation faces a greater challenge: it is more complex to quantify the differences between multiple domains.</p> Methods <p>In this paper, a multi-source domain two-stage joint distribution alignment (MSD-TSJDA) method based on fault diagnosis in variable operating conditions is proposed. The method achieves alignment between the source and target domains data distributions by implementing a shared three-branch feature extractor, along with several domain-specific feature extractors and classifiers, integrated into a two-stage joint distribution alignment strategy. In the first stage, maximum mean square discrepancy and local maximum mean discrepancy are used to reduce the differences between domains initially. The second stage further facilitates the acquisition of domain-invariant features by minimizing the mutual information in the potential feature space of the network. By training on source-domain data from rotating machinery operating under known conditions, the method enables fault diagnosis on target-domain data from unknown operating conditions.</p> Results and Conclusion <p>The results of the experimental study demonstrate that MSD-TSJDA attains an average accuracy of 96.81% and 98.88% on the Paderborn University (PU) and Huazhong University of Science and Technology (HUST) bearing datasets, respectively. MSD-TSJDA outperformed the comparison methods, demonstrating greater stability and accuracy.</p>