Hierarchical Spatiotemporal Graph Network for Fault Diagnosis of Industrial Processes

Intelligent fault diagnosis of industrial processes has received enormous attention in recent years, and deep learning-based methods have excellent performance in accurate health state detection. However, existing methods cannot fully exploit the complex relationship between different subsystems of industrial processes. To address this limit, we convert industrial data into graph structure data with sensors (as nodes) and topological connections between sensors (as edges) to represent complex interactive information. Specifically, we propose a spatiotemporal graph convolutional network with a hierarchical structure (HiSTGCN) for fault diagnosis of industrial processes. First, a local-global graph framework is constructed to explore the correlation between subsystems fully. Particularly, we propose a hierarchical graph structure with a global graph representing the correlation of sensors between subsystems and several local graphs capturing the correlation of sensors within subsystems to enrich the feature extraction. Then, we design a hierarchical spatiotemporal graph neural network to perform a local-global graph framework in both temporal and spatial dimensions. Finally, a synthesized residual health monitor module based on the principal component analysis (PCA) is designed for fault detection and location. Experiment results on an industrial simulation process dataset and a real wind farm dataset show that HiSTGCN has reliable and superior fault diagnosis performance compared to existing methods.