LiMS-MFormer: A Lightweight Multi-Scale and Multi-Dimensional Attention Transformer for Robust Rolling Bearing Fault Diagnosis Under Complex Conditions

Bearings are critical components in industrial machinery, and their failures can lead to equipment downtime and significant safety hazards. Traditional fault diagnosis methods rely on manually crafted features and classical classifiers, often suffering from poor robustness, weak generalization under noisy or small-sample conditions, and limited suitability for lightweight deployment. This study proposes a Lightweight Multi-Scale Multi-Dimensional Self-Attention Transformer (LiMS-MFormer)—an end-to-end lightweight fault diagnosis framework integrating multi-scale feature extraction and multi-dimensional attention. The model integrates lightweight multi-scale convolutional feature extraction, hierarchical feature fusion, and a multi-dimensional self-attention mechanism to balance feature expressiveness with computational efficiency. Specifically, the front end employs Ghost convolution and enhanced residual structures for efficient multi-scale feature extraction. The middle layers perform cross-scale concatenation and fusion to enrich contextual representations. The back end introduces a lightweight temporal-channel-spatial attention module for global modeling and focuses on key patterns. Experiments on the Paderborn University (PU) dataset and the University of Ottawa bearing vibration dataset (Ottawa dataset) show that LiMS-MFormer achieves an accuracy of 96.68% on the small-sample PU dataset while maintaining minimal parameters (0.07 M) and low computational cost (13.55 M FLOPs). Moreover, under complex noisy conditions, the proposed model demonstrates strong fault diagnosis capability. On the University of Ottawa dataset, LiMS-MFormer consistently outperforms several state-of-the-art lightweight models, exhibiting superior accuracy, robustness, and generalization in challenging diagnostic tasks.