Unsafe Behavior Detection with Adaptive Contrastive Learning in Industrial Control Systems

Unsafe behavior detection is crucial for maintaining safe and reliable operations in various industrial control systems. However, the scarcity of labeled samples for model training poses significant challenges for existing methods. Self-supervised learning, particularly contrastive learning, offers a promising solution due to its ability to learn from unlabelled data. In this paper, we present AdaTCL, a contrastive learning framework with adaptive augmentations, to detect unsafe behavior in industrial control systems. By dividing instances into task-irrelevant and informative parts and applying lossless transform functions, AdaTCL prevents ad-hoc decisions and laborious trial-and-error tuning for augmentation selection, which improves the generalization capability of contrastive learning. Our experiments demonstrate that AdaTCL significantly outperforms classic and recent baselines highlighting the potential of state-of-the-art self-supervised learning techniques for industrial control systems.