Dynamic Graph Embedding PCA to Extract Spatio–Temporal Information for Fault Detection

The complexity of coupled multivariate data in industrial settings often limits the effectiveness of principal component analysis (PCA) in revealing patterns and structures in the data. In this article, we propose a novel fault detection framework for industrial process time series data with temporal and spatial correlation. First, by applying graph theory, the framework captures the complex network structures inherent in industrial processes, enabling the discovery of hidden data associations from a topological perspective. Then, the proposed method integrates temporal and spatial correlations in the modeling process, ensuring a comprehensive and integrated analysis. Specifically, the time series data are divided into sliding window intervals, and then the graph convolution is embedded within each window. After the modeling optimization objectives are defined, the overall solution is derived. Finally, these components, which contain spatio–temporal information, are used to construct dynamic and static statistics. Experiments on a chemical dataset show that the proposed method can significantly reduce the false alarm rate and improve the fault detection rate compared with the dynamic internal PCA without considering spatial factors. In addition, by applying it to the actual hot rolling process of strip, the superiority of the method is further verified, and its practical value and robustness are highlighted.