Multi-attention collaborative temporal-spatial hypergraph attention network for machinery fault diagnosis

Graph models have emerged as an important research direction in the field of machinery fault diagnosis. However, most of these methods are unable to capture non-pairwise, high-order interactions among multiple samples, and they generally achieve satisfactory performance only when a large amount of labeled data is available. In this paper, a multi-attention collaborative temporal-spatial hypergraph attention network (MA-TSHGAN) with hypergraph label passing is proposed. First, we propose an attention-aware hypergraph construction method to establish high-order relationships among nodes, in which the attention mechanism is introduced to dynamically assign weights to hyperedges. Second, by exploiting the advantages of the attention-aware hypergraph, the semi-supervised learning problem is formulated as hypergraph information propagation, and a hypergraph label passing method (HLPM) is developed to propagate the label information of limited labeled samples through the hypergraph structure. Finally, to capture temporal-spatial features, a multi-attention collaborative hypergraph attention network is designed, in which a multi-attention collaborative hypergraph convolution layer with different attention heads is proposed to further expand the receptive field and aggregate richer neighborhood information through multi-scale attention. Experiments on three datasets confirm that the proposed method can achieve average recognition accuracies of 99.82%, 99.83% and 97.63%, respectively, with limited labeled samples, which outperforms several state-of-the-art comparison methods.