Anomaly diagnosis in SWRO desalination plants using a physics-informed spatiotemporal graph attention network

Reliable operation of seawater reverse osmosis (SWRO) desalination demonstration plants requires accurate anomaly detection and rapid fault localization. Existing graph-based methods rely on statistical similarities to establish sensor connections, failing to capture physical pathways and step-specific propagation delays. To enable accurate anomaly detection and interpretable fault diagnosis in SWRO plants, this study proposes the physics-informed spatiotemporal graph attention network (PI-STGAT), which extends the spatiotemporal graph attention network (STGAT) with three elements: a hybrid graph that combines domain knowledge with mutual information (MI) edges, a spatiotemporal extension with lag-aware edges for time-delayed propagation, and Shapley additive explanations (SHAP)-based sensor time-step attribution. Operational data from a 1,000 m³/day SWRO plant verified performance across energy, flow, and conductivity metrics. PI-STGAT achieved per-fold R² values of 0.946, 0.946, and 0.963, establishing a reliable detection reference and improving F1 scores by 2.7-21.7 percentage points over the MI-based configuration, particularly for multistep conductivity anomalies. Compared with five deep-learning benchmarks, including reconstruction-based autoencoders, a temporal transformer, and graph-based methods, PI-STGAT achieved the highest average F1 score of 0.953, outperforming all competitors by 13.4-29.5 percentage points. Furthermore, SHAP attribution successfully localized responsible sensors and time steps, with peaks coinciding with the raw signal deviations. By unifying physical pathways, step-specific delays, and sensor time-step attribution, the proposed framework supports rapid SWRO plant stabilization during the initial operational phase, enables precise operator intervention, and remains generalizable to other multistep industries where domain knowledge edges can be defined.