A Quality-Related Fault Detection Method Based on Deep Spatial–Temporal Networks and Canonical Variable Analysis for Industrial Processes

Modern manufacturing processes typically comprise multiple subprocesses, and the temporal-spatial features of process variables are difficult to extract, which bring significant challenges for traditional quality-related fault diagnosis. To solve this problem, we propose a fault detection model driven by graph attention networks integrated gated recurrent unit-canonical variable analysis (GATRU-CVA). First, the knowledge of domain experts and historical data are used to construct a subblock knowledge graph. Next, a spatial feature extractor for the graph attention network (GAT) is constructed for global variables. In addition, the subblock knowledge graph is used to divide the global spatial features into subblocks, and the corresponding temporal feature extractor is constructed. Then, taking the process dynamic characteristic into account, CVA is used to model the process based on the spatial-temporal features, upon which the corresponding statistics are calculated. The threshold is determined by the kernel density estimation (KDE) method. Finally, actual production data of the hot strip mill process (HSMP) are used to validate the proposed model. The results show that the proposed method has a 97% correct monitoring rate (CMR) for HSMP, which is better than the other comparative fault detection methods.