Causality-Aware Spatiotemporal Adversarial Learning for Knowledge-Data Fault Diagnosis in Large-Scale Industrial Processes.

Abstract

Effective process monitoring and fault diagnosis (PMFD) are essential for safe and efficient operation in large-scale industrial processes. However, many existing methods still suffer from limited interpretability, insufficient exploitation of process knowledge, and difficulty in representing causal relationships in strongly coupled systems. This article addresses these challenges by proposing a causality-aware spatiotemporal adversarial learning framework for fault diagnosis (FD) that integrates both process knowledge and data. First, a two-stage procedure for constructing a causal graph is developed. In the first stage, a temporal neural network with attention is used to discover candidate spatiotemporal causal relations from historical data. In the second stage, perturbation-based validation guided by domain expertise filters spurious links and yields an interpretable causal graph consistent with the underlying process mechanisms. The resulting global causal graph is then partitioned into subgraphs associated with physical subsystems. On this basis, a causality-aware spatiotemporal adversarial model is designed to extract fault-sensitive features under subgraph constraints and to build monitoring statistics at both subsystem and plant-wide levels. Finally, an interpretable fault root-cause diagnosis strategy is introduced, which combines Shapley additive explanation (SHAP)-based variable contribution analysis with causal path inspection to identify fault sources and propagation routes. Experiments on a real hot strip mill process (HSMP) demonstrate that the proposed framework improves fault detection performance over representative existing methods while providing physically meaningful explanations that support engineering decision-making.