Purpose <p>Deep learning–based intelligent fault diagnosis has achieved remarkable performance in improving the reliability and safety of mechanical systems. However, its effectiveness is often limited by scarce labeled data and domain shifts across operating conditions. Unsupervised domain adaptation has therefore emerged as a promising solution to these challenges. This study proposes a global-to-local domain adaptation framework based on hybrid discrepancy measures for cross-domain fault diagnosis of rotating machinery. Methods Sliced Wasserstein discrepancy is first employed to achieve global coarse alignment by reducing the optimal transport discrepancy between domains. On top of this, a neighbor-aware propagation mechanism is introduced to generate more reliable pseudo-label information, which is further leveraged for reliability-aware subdomain alignment and confidence-weighted self-training, further reducing class-wise discrepancies and enhancing target-domain discriminability. </p> Results <p>Four open-source laboratory datasets covering cross-domain fault diagnosis tasks for both bearing and gearbox are used for evaluating the performance of the method. The results indicate that the proposed method achieves accuracy improvement over all comparison methods, with statistically reliable overall enhancement.</p> Conclusion <p>A global-to-local domain adaptation framework based on hybrid discrepancy measures and neighbor-aware pseudo-label refinement is proposed for fault diagnosis of rotating machinery. Unlike adversarial methods, the proposed framework explicitly reduces cross-domain discrepancies through transparent optimization, resulting in higher diagnostic accuracy, more stable convergence, and stronger domain alignment. By coupling distribution alignment and discriminative learning in a global-to-local manner, the proposed method delivers robust cross-domain diagnosis.</p>