Unsupervised anomaly detection for Nuclear Power Plants based on Denoising Diffusion Probabilistic Models

Abstract

The abnormal state detection in nuclear reactors constitutes a critical concern within the broader context of Nuclear Power Plants (NPPs) safety management. Deep learning techniques have exhibited exceptional performance in addressing issues pertaining to NPPs safety control. However, acquiring the large amount of labeled data required by supervised learning methodologies poses a significant challenge in practical applications. This paper addresses a key challenge in NPPs safety—abnormal state detection in nuclear reactors. Leveraging unsupervised learning due to the limited availability of labeled data, we propose an anomaly detection method using the Denoising Diffusion Probabilistic Model (DDPM) with a noise-to-noise training strategy. Comparative evaluation against AE, VAE, and GAN shows that DDPM outperforms in all metrics, demonstrating strong potential for NPPs anomaly diagnosis. Experimental results suggest that a feature count of 50 optimizes DDPM performance for NPPs anomaly detection, while the noise-to-noise training strategy improves model robustness.