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

Area of Science:

• Industrial Process Control
• Artificial Intelligence
• Causal Inference

Background:

• Effective process monitoring and fault diagnosis (PMFD) are critical for industrial operations.
• Existing methods often lack interpretability and struggle with complex, coupled systems.
• There's a need to integrate process knowledge with data-driven approaches for robust fault diagnosis (FD).

Purpose of the Study:

• To develop a causality-aware spatiotemporal adversarial learning framework for fault diagnosis.
• To improve the interpretability and accuracy of fault detection in industrial processes.
• To integrate domain expertise with data-driven methods for enhanced PMFD.

Main Methods:

• A two-stage causal graph construction using temporal neural networks and perturbation-based validation.
• Development of a causality-aware spatiotemporal adversarial model with subgraph constraints.
• Implementation of an interpretable fault root-cause diagnosis strategy using SHAP and causal path inspection.

Main Results:

• The proposed framework demonstrated improved fault detection performance on a real hot strip mill process (HSMP).
• The method successfully identified fault sources and propagation routes with physically meaningful explanations.
• The causality-aware approach enhanced the interpretability of fault diagnosis compared to existing methods.

Conclusions:

• The causality-aware framework effectively integrates process knowledge and data for advanced PMFD.
• The approach provides interpretable explanations, supporting engineering decision-making and process safety.
• This work advances the state-of-the-art in industrial fault diagnosis for complex systems.