Physics-guided contrastive temporal graph learning for anomaly detection and root-cause localization in industrial...

Area of Science:

• Control Systems Engineering
• Machine Learning
• Data Science

Background:

• Industrial control systems (ICS) generate complex multivariate time series data.
• Detecting anomalies in ICS without labeled attack data remains a significant challenge.

Purpose of the Study:

• To propose a physics-guided contrastive temporal graph learning framework for anomaly detection and root-cause localization in ICS.
• To develop a method that learns robust representations of sensor data and captures temporal and cross-sensor relationships.

Main Methods:

• A self-supervised contrastive encoder is trained on normal data to learn sensor window representations.
• A physics-guided graph is constructed using plant P&ID, representing sensors and actuators as nodes and physical dependencies as edges.
• A temporal graph module processes learned embeddings to analyze time evolution and cross-sensor correlations.

Main Results:

• The proposed framework demonstrates superior anomaly detection performance compared to LSTM-AE and CNN-AE baselines on an ICS dataset.
• Robustness tests indicate effective performance under distribution shift, sensor drift, and missing channels.
• The method provides interpretable sensor rankings for root-cause analysis.

Conclusions:

• The integration of contrastive learning with physics-guided graph reasoning offers a powerful, label-free approach for ICS anomaly detection.
• This framework enhances operational safety and efficiency by enabling timely anomaly identification and root-cause localization.