ST ‐ FDGNN : An adaptive spatio‐temporal graph neural network for fault diagnosis in chemical processes

Abstract

The high‐stakes nature of chemical production drives the demand for efficient and precise fault diagnosis techniques. However, chemical process data exhibit complex temporal evolution patterns alongside spatial correlation characteristics, making it challenging to simultaneously capture time‐varying dynamic features and spatial dependencies among variables, which consequently compromises the diagnostic accuracy. This manuscript proposes an adaptive spatio‐temporal graph neural network for fault diagnosis in chemical processes (ST‐FDGNN). The framework introduces a data‐driven adaptive graph structure learning mechanism, eliminating the reliance on predefined graph structures and enabling automatic discovery of latent variable association. By integrating temporal convolutional networks (TCN) and graph neural networks (GNN) deeply, the model achieves collaborative learning of complex spatio‐temporal features among multiple process variables. Additionally, an improved spatial pooling strategy is incorporated to enhance the efficiency of feature extraction. ST‐FDGNN is applied to the Tennessee Eastman process and the continuous stirred tank reactor system, and compared with several advanced fault diagnosis methods. The results demonstrate that the proposed model outperforms existing approaches across multiple statistical metrics, exhibiting superior diagnostic accuracy and generalization capability.