Abstract 

Fault diagnosis of rotating machinery remains a critical challenge due to the need for precise localization of fault-related patterns in time-series signals, which are often influenced by variable speed conditions. Moreover, the limited interpretability of deep learning models restricts their deployment in practical industrial applications. To address these issues, this study proposes the Multi-Scale Signal Transformer (MSSiT), a novel deep learning architecture that adaptively captures multi-resolution features using a multi-scale self-attention mechanism. In addition, a Signal Processing-Based Attention Interpretation (SPAI) method is developed to enhance model interpretability by analyzing attention weight in both time and frequency domains through signal processing techniques.

The proposed framework is validated through two case studies involving various and time-varying speed conditions. Experimental results demonstrate that MSSiT outperforms existing fault diagnosis methods under variable speed conditions. The SPAI results reveal that multi-scale attention mechanism focuses on high-frequency components for fault-related features and low-frequency components for speed-dependent trends at each scale. Furthermore, interpretation results are quantitatively validated, confirming that attention weight reliably highlights physically meaningful features. These findings demonstrate the effectiveness and robustness of the proposed framework and its potential applicability in real-world fault diagnosis under complex operational conditions.