Fault Diagnosis Methods Based on Spatio-Temporal Feature Fusion

Accurate fault diagnosis in industrial processes depends on the ability to model and integrate spatial dependencies among process variables and temporal dynamics of operational data. To address this challenge, this paper proposes a novel spatio-temporal fusion fault diagnosis method, SDGCN-LSTM, integrating Graph Convolutional Networks (GCN) and Long Short-Term Memory (LSTM) networks. The GCN module captures spatial relationships among process variables based on their structural connections, while the LSTM network learns temporal patterns from sequential data. After separately extracting spatial and temporal features, a two-dimensional attention mechanism is applied to adaptively enhance the most informative features. Finally, the ADaboost algorithm is employed as a classifier to perform final fault identification. Experimental results demonstrate that the proposed SDGCN-LSTM method achieves superior fault diagnosis accuracy across Three-Phase Flow Facility (TFF) and Tennessee Eastman (TE) datasets compared to baseline methods.