Bearing fault diagnosis via multi-receptive field backbone and dual-branch metric learning

Accurate fault diagnosis under limited labelled data is vital for industrial machinery health monitoring. To address this few-shot challenge, we propose a novel method integrating a Multi-Receptive Field Feature Backbone (MRF-Backbone), a Context Module: Dual-Attention Class-wise Metric Matching, and a Detail Module: Wasserstein Distance-Based Metric. The MRF-Backbone extracts multi-scale representations via parallel large-kernel convolutions with varied dilation rates, enhancing both local and global feature encoding. Based on this, the Context Module employs self-attention and class-wise cross-attention to model global semantic relations between queries and class prototypes. In parallel, the Detail Module estimates per-class distributions and computes Wasserstein distances to capture fine-grained statistical discrepancies between query and support samples. A weighted fusion of the two branches combines global context with local detail, enabling robust and discriminative few-shot classification. Experiments on the PU bearing dataset under standard 5-shot settings validate the effectiveness of our approach, which outperforms strong baselines and achieves competitive performance against state-of-the-art methods. These results highlight the benefit of integrating multi-scale encoding with dual-granularity metric learning for few-shot fault diagnosis.