A Small-Sample Bearing Fault Diagnosis Method Based on Multi-Image Fusion and Multi-Scale Dynamic Residual Dual Attention Mechanism

Abstract

In recent years, fault diagnosis methods based on convolutional neural networks (CNNs) have garnered significant attention in the field of rotating bearing fault diagnosis. Addressing the challenge of extremely limited fault signal samples, this paper proposes a small-sample bearing fault diagnosis method based on multi-image fusion and a dual-attention mechanism incorporating multi-scale dynamic residuals. This method first converts the fault signal into a two-dimensional image through continuous wavelet transform and Gram angle field (GASF/GADF), thereby transforming the fault diagnosis problem into an image feature learning problem. The model extracts basic features through the initial convolutional layer and sequentially learns deep features via multi-scale dynamic residual blocks and dual attention mechanisms. Among these, the multi-scale architecture captures features across different receptive fields through parallel convolutional branches, while the dual attention mechanism performs feature recalibration in both the channel and spatial dimensions. Experimental results demonstrate that the proposed method achieves an accuracy rate of 97.47% in bearing fault diagnosis tasks, representing a significant improvement over traditional CNN models. This validates the model’s effectiveness and superiority in complex fault diagnosis scenarios.