A Spatial-Temporal Graph Convolutional Network for Small Sample Fault Diagnosis of Rotating Machinery

Graph convolutional network (GCN) provides an excellent feature extraction way for fusing condition monitored multi-channel signals, making it widely employed in fault diagnosis of rotating machinery. Some spatial-temporal networks based on GCN have been proposed with the aim of improving diagnostic performance by extracting features from both the spatial and temporal domains. The improvement in feature extraction has also enhanced the diagnostic performance under small sample scenarios. However, most of the existing studies construct spatial-temporal framework by stacking spatial and temporal modules, which could degrade the spatial-temporal correlation. To obtain more inherent fault representations and improve diagnostic performance under small sample scenarios, this paper proposed a spatial-temporal graph convolutional network. In the proposed method, each sensor channel is treated as one node to build a multi-channel interactive graph. The data at each time step is constructed as a subgraph, and every three adjacent subgraphs are connected by the same nodes to form a spatial-temporal graph. With the help of learnable position embeddings, graph convolution is conducted on spatial-temporal graphs to synchronize the extraction of spatial-temporal features. The effectiveness of the proposed method is verified through a subway train transmission system fault dataset. Comparative studies show that the proposed method possesses strong feature extraction capability, and exhibits better diagnostic performance in scenarios with small samples.