Most fault diagnosis methods heavily rely on large-scale labeled data and are difficult to generalize to industrial scenarios characterized by small samples, strong noise, and complex structural relationships. In addition, traditional graph-based approaches are limited to pairwise relationships and cannot effectively model higher-order correlations among multi-source samples, while redundant and irrelevant features further degrade diagnostic performance under small-sample conditions. To address these challenges, this paper proposes a novel hypergraph attention-based fault diagnosis framework that leverages redundancy elimination and a few-shot learning strategy. First, vibration signal samples are transformed into hypergraphs to capture higher-order structural relationships among samples. A hypergraph attention neural network is designed to learn hyperedge importance and adaptively enhance discriminative feature aggregation. Second, a redundancy elimination mechanism based on structured feature decoupling and mutual information minimization is introduced to suppress redundant representations and improve feature compactness. Third, a few-shot learning strategy integrating prototype-based metric learning and episodic training is developed to enable robust fault diagnosis under extremely limited labeled samples. Extensive experiments on the rotating machinery fault datasets demonstrate that the proposed method achieves superior diagnostic accuracy, stability, and generalization performance compared with existing deep learning and graph-based approaches, especially in small-sample and noisy environments.
Zhou et al. (Mon,) studied this question.