RLMST: A Robust and Lightweight Multi-scale Transformer Model for Motor Fault Diagnosis

This paper presents a robust and lightweight multiscale Transformer model (RLMST) for real-time motor fault diagnosis under high-noise industrial environments. The model integrates a deconvolution-enhanced multiscale convolution module to strengthen local feature representation and suppress noise, while a novel Local-Aware Global-Enhanced (LAG) unit based on grouped linear transformations adaptively recalibrates feature channels to emphasize informative patterns. A Transformer encoder further models nonlinear dependencies among diverse features, enabling effective fusion of local and global information. With only 62.9 K parameters and 5.63 M FLOPs, RLMST achieves high computational efficiency and strong noise tolerance. Experiments on the UOEMD-VAFCVS dataset show that the model maintains superior accuracy and stability under Gaussian, Laplacian, and Pink noise, even at low signal-to-noise ratios. These results demonstrate that RLMST effectively balances accuracy, robustness, and efficiency, offering a promising solution for edge-deployable predictive maintenance in intelligent power systems.