A novel rolling bearing fault diagnosis method based on time-series fusion transformer with interpretability analysis

Rolling bearing fault diagnosis enhances equipment reliability, reduces maintenance costs, and enables effective non-destructive testing (NDT). However, current research often emphasizes model design and performance optimization, overlooking the long-term dependencies of fault signals and the need for interpretability. This study proposes a rolling bearing fault diagnosis model utilizing a time-series fusion transformer with interpretability analysis. The model introduces multi-scale feature adaptive fusion to automatically capture and integrate features across different scales, enhancing the global pattern detection in time-series data. A dynamic patch auto-encoder module transforms feature embeddings into a low-dimensional space to better retain local information. The model’s design, particularly the decoding layer of the time-series Transformer, is optimized with a multi-head self-attentive mechanism, and multi-dimensional attention weights visualization methods are employed to clarify the fault feature extraction process. Quantitative visualizations throughout training improve interpretability and insight into learning dynamics. Experimental results indicate that this model surpasses state-of-the-art approaches on benchmark datasets, proving its generalizability and robustness in diverse testing scenarios.