Multi-scale signal transformer for rolling bearing fault diagnosis under variable conditions

Abstract Deep learning has shown promise in bearing fault diagnosis but often suffers from performance degradation under non-stationary conditions due to the difficulty in capturing time-varying fault characteristics. This study proposes a multi-scale signal transformer (MSSiT) to enhance diagnostic robustness under variable conditions. The architecture incorporates a multi-scale convolutional embedding module, leveraging multi-scale dilated convolutions and efficient channel attention to refine discriminative signal representations. Following convolution-based patch embedding for efficient tokenization, an alternating global–local scale self-attention mechanism and a multi-scale feed-forward network are employed. These modules utilize multi-scale depthwise convolutions to bolster local perception while maintaining a lightweight parameter count. Case studies on four datasets substantiate that MSSiT consistently outperforms five representative methods in cross-speed, cross-load, variable-speed tasks and cross-lateral-load tasks, demonstrating improved cross-condition diagnostic performance under the evaluated variable-condition scenarios.