Rolling bearing fault diagnosis is an important technology for health monitoring and pre-maintenance of mechanical equipment, which is of great significance for improving equipment operation reliability and reducing maintenance costs. This article reviews the research progress of fault diagnosis methods for rolling bearings, with a focus on analyzing the applications, advantages, and disadvantages of traditional data-driven methods, deep learning methods, graph embedding methods, and Transformer methods in this field. In addition, further analysis was conducted on the main issues of current research, including complex network structures, insufficient information attention, difficulties in graph data processing, and challenges in long-term dependency modeling. In response to these challenges, future research should focus on designing more lightweight and efficient models, improving computational efficiency, robustness of the models, and strengthening attention and deep mining of fault features.
To address the issues of low contrast, unclear local details, and the difficulty of existing deep learning models in capturing long-range dependencies in time-frequency images generated by current time-frequency construction methods, this paper proposes a bearing fault diagnosis framework (MATE-ViT) based on multi-channel contrast-limited adaptive time-frequency enhancement (CLAHE) and vision transformer (ViT). First, an improved CLAHE algorithm is used to independently enhance the multi-channel time-frequency images, effectively improving the local contrast and detail expression of the images, thereby enhancing the recognizability of fault features. Next, a perception-driven image segmentation and rearrangement strategy is applied, dividing the enhanced image into local feature patches and rearranging them to maximize the complementarity between local features and global information, thus providing optimized input feature representations for subsequent global modeling by the transformer. Finally, the ViT model is applied to extract global features from the enhanced images, achieving accurate classification of bearing fault types. Experimental results show that the proposed MATE-ViT framework outperforms mainstream deep learning models in terms of diagnostic accuracy and robustness under various noise levels and operating conditions, offering a more effective and reliable approach for bearing fault diagnosis.