Enhanced rolling bearing fault diagnosis using a multi-stage attention fusion network

Fault diagnosis of rolling bearings can promptly identify issues in the operation of mechanical equipment, which is crucial for the performance and safety of the equipment. In practical applications, mechanical equipment typically faces challenges such as limited sample sizes, high noise levels, and complex operating conditions, making it difficult to extract fault features from vibration signals. The introduction of the attention mechanism has significantly enhanced the feature extraction capability of neural networks, and it is increasingly being applied to the task of fault feature extraction in rolling bearings. This paper proposes a rolling bearing fault diagnosis method based on a multi-stage attention fusion network. First, the GADF method is used to convert the bearing one-dimensional time-series vibration signal into an RGB two-dimensional image. Then, global and local feature extraction is performed using the Swin Transformer and Parallelized Patch-Aware Attention, respectively. Finally, Content-Guided Attention is employed to perform weighted fusion of the two types of features. By integrating multi-stage attention mechanisms, the method improves fault recognition accuracy. Experimental results demonstrate that the proposed method outperforms other comparative models in terms of recognition accuracy, few-shot learning ability, generalization capability, and noise resistance.