A dual-branch spatiotemporal dynamic residual fusion network for rolling bearing fault diagnosis

Rolling bearings are key components in rotating machinery. Their fault signals are often affected by noise interference, overlapping impulsive components, and multiscale feature coupling, which create difficulties for feature extraction and condition recognition. To improve the accuracy of rolling bearing fault diagnosis, this study proposes a dual-branch spatiotemporal dynamic residual fusion network, termed Multi-scale GAM-Transformer Dynamic Fusion Network (MS-GTDFNet). The model constructs parallel one-dimensional Convolutional Neural Network–Transformer and two-dimensional Convolutional Neural Network–Global Attention Mechanism branches to extract temporal impulsive features and spatial structural features from vibration signals, respectively. A sample-level dynamic residual fusion module is further introduced to adaptively adjust the feature contributions of the two branches for different input samples, while a residual pathway enhances complementary feature representation. In the Hanoi University of Science and Technology (HUST) 10-class task, MS-GTDFNet achieves a test accuracy of 98.70%, with macro-averaged precision, recall, and F1-score values of 98.56%, 98.47%, and 98.50%, respectively. Considering the results of multi-dataset validation, cross-load testing, and comparisons with recent methods, MS-GTDFNet maintains high diagnostic performance and good stability under different data and operating conditions; in particular, the average accuracy in the cross-load tests reaches 98.46%. The results demonstrate that the proposed model can effectively fuse temporal impulsive features and two-dimensional structural features in vibration signals, providing an effective method for rolling bearing fault diagnosis under complex operating conditions.