A bearing fault diagnosis method based on dual attention mechanism and multimodal feature fusion

To address the issue of incomplete and easily submerged features in single-modality methods for bearing fault diagnosis under small sample and high noise conditions, a fault diagnosis method based on dual attention and multimodal fusion is proposed. This method constructs a parallel dual-channel architecture, taking one-dimensional vibration signals and two-dimensional time-frequency images obtained by continuous wavelet transform as inputs respectively, to achieve feature complementarity between time-domain transient details and time-frequency energy distribution. During feature extraction, corresponding attention mechanisms are introduced for each channel: the image channel adopts an improved CBAM-ResNet module, which cascades feature refinement modules on the basis of CBAM to achieve progressive enhancement of fault regions; the signal channel proposes a Local Window Temporal Attention Mechanism (LWTA), which computes attention weights within multi-scale local windows to strengthen transient impact features masked by noise. To address the fusion imbalance caused by dimensional mismatch between bimodal features, a dimension alignment and adaptive weighting module is designed to achieve balanced fusion and synergistic enhancement of features. Experiments on CWRU and self-built gearbox datasets show that the proposed method achieves high accuracy and strong robustness under small sample and high noise conditions, outperforming traditional and mainstream methods.