Accurate Classification of Minor Leaks in Primary Loop Systems Using a Multi-Scale Attention Temporal Convolutional Autoencoder

Minor leaks in the reactor coolant system (RCS) can provide early warning of degradation that may progress towards safety-relevant failures in nuclear power plants. Their timely localization remains difficult because they generate weak thermal-hydraulic perturbations and labeled leak data are rarely available. Here, we present MATCA-TM, a physics-informed zero-shot framework that combines a Multi-scale Attention Temporal Convolutional Autoencoder (MATCA) with theoretical response-template matching for early leak localization. The framework learns multivariate temporal dependencies solely from normal operating data, enabling reconstruction residuals to highlight subtle anomalies caused by leaks. Physics-informed templates constructed from thermal-hydraulic response characteristics then provide a physically motivated basis for localizing leak sources. On the II + CPR1000 simulation platform, which represents a CPR1000 pressurized water reactor model, MATCA-TM achieved 65.3% localization accuracy for an equivalent leakage diameter of 6.7 × 10−3 cm and outperformed the evaluated machine learning and deep learning baselines under the same protocol. By combining data-driven representation learning with domain physics, MATCA-TM provides a physics-informed, simulation-validated approach to diagnosing incipient RCS leaks and a basis for further validation under plant-relevant operating conditions.