A causal-aware spatiotemporal network with prediction-reconstruction for interpretable anomaly monitoring in complex industrial processes

Abstract

Anomaly monitoring in complex industrial processes is challenged by high-dimensional, nonstationary sensor data exhibiting intricate inter-variable couplings. Existing deep learning approaches, whether based on flat sequence models or graph-structured representations, construct variable relationships from statistical correlations, conflating genuine process dependencies with spurious confounding effects and limiting detection reliability and fault interpretability. We propose the Causal-Aware Spatiotemporal Network with Prediction-Reconstruction (CaST-PR), trained solely on unlabeled normal operating data. A Sequential Causal Graph Inference (SCGI) module learns directed inter-variable dependencies via a gated predictive causal inference mechanism with permutation-based edge validation. The SCGI's gated adjacency output guides a spatiotemporal encoder integrating time-frequency feature extraction with residual graph attention; the permutation-refined causal graph enables root-cause tracing by overlaying per-variable anomaly contributions to isolate initiating fault variables from downstream consequences, supporting actionable maintenance decisions within enterprise information workflows. Anomalies are detected via variational residual reconstruction, yielding a statistically grounded anomaly score. Validated on a benchmark dataset and two real-world industrial cases, CaST-PR achieves near-zero false alarm rates with high fault detection rates, consistently outperforming seven baselines. Our code is publicly available at https://github.com/wuyouxiaozhu/Code.