Abstract
Robust causal inference under noise and nonlinear dynamics is essential for industrial fault detection and root-cause diagnosis. A unified framework is presented that integrates Granger causality for structural priors, LSTM-based temporal encoding, GATv2-driven graph refinement, and FGSM adversarial training. The approach concurrently performs forecasting, detection, and diagnosis within a structure-aware pipeline. Evaluation on synthetic datasets. the Tennessee Eastman Process, and a hybrid tank system reveals substantial performance gains over Granger-only baselines in both graph reconstruction and root-cause localization. Adversarial augmentation yields consistent improvements across all metrics, with particularly strong gains in precision and noise resilience, validating the framework’s reliability for industrial diagnostic applications.
| Original language | English |
|---|---|
| Article number | 103777 |
| Journal | Journal of Process Control |
| Volume | 165 |
| DOIs | |
| Publication status | Published - Sept 2026 |
Keywords
- Adversarial training
- Causal discovery
- Fault diagnosis
- Granger causality
- Graph attention networks
- Industrial process monitoring
- Root cause analysis
Cite this
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver