A Method Based on Multi-Graph Convolutional Network with Graph Fusion for Rolling Bearing Fault Diagnosis

Graph data based on non-Euclidean space possesses more comprehensive information representation capabilities, which is of significant importance for fault diagnosis. Nonetheless, graph convolutional networks (GCNs) that perform well in non-Euclidean spaces have certain limitations. In this paper, we propose a multi-graph convolutional network with graph fusion called Multi-Graph Convolutional Network with Gated Recurrent Unit (MGGCN-GRU). In our method, we devise a multi-graph fusion approach that complements the single-graph features to enrich the feature information. Additionally, we design a multi-graph convolutional module that extracts features from vibration signals, including two single graphs and one fusion graph feature extraction, enabling the exploration of various relationships between signals. In the multi-graph convolution, we employ a multi receptive fields graph convolutional module as the convolutional unit, which integrates information from multi receptive fields. Furthermore, we incorporate a gated recurrent unit after the graph convolutional module to extract time-dependent features. Through experiments on a collected truck bench dataset, it is demonstrated that the proposed method exhibits excellent performance.