Self-Supervised One-Class Prototypical Networks for Nuclear Power Plant Pipe Leak Detection

Detecting pipeline leaks in nuclear power plants is important to ensuring operational safety and minimizing economic losses. Although supervised deep learning models have shown promising performance in leak detection, they face two major challenges: the difficulty of obtaining sufficient leak data and the degradation of performance when applied to unseen environments. To address these challenges, we propose one-class prototypical networks that trains exclusively on normal state data without requiring labeled leak data. Our method leverages only normal state data from multiple domains for training and effectively adapts to new environments with minimal support data. To further enhance performance, a self-supervised anomaly detection strategy is introduced by augmenting normal data and treating them as pseudo-leak data. Experimental evaluations in multiple simulated nuclear power plant environments demonstrate that the proposed method consistently achieves high detection performance, achieving an average AUC-ROC score of 0.994 across all environments, even in unseen domains. Furthermore, explainability analysis using 1D Grad-CAM confirms that the model correctly identifies leak-relevant frequency bands, highlighting its ability to capture domain-invariant features. These results demonstrate the practicality and robustness of self-supervised one-class prototypical networks as an effective solution for real-world industrial anomaly detection tasks where fault data are unavailable.