Enhancing nuclear power plant diagnostics: A comparative analysis of XAI-based feature selection methods for abnormal and emergency scenario detection

Abstract

This study introduces the application of explainable artificial intelligence (XAI) techniques to enhance nuclear power plant diagnostics through effective feature selection. We compared various XAI methods, including gradient-based techniques, layer-wise relevance propagation, DeepSHAP, integrated gradients, local interpretable model-agnostic explanation(LIME), and saliency maps, with traditional approaches such as principal component analysis (PCA). By applying these methods to data from an IAEA iPWR simulator, which includes 35 abnormal and emergency scenarios with 116 state variables, we demonstrated the superiority of XAI-based methods in selecting features that effectively distinguish between different plant conditions. Our approach successfully reduced the input dimensionality from 116 to 20 features while maintaining high diagnostic accuracy. XAI methods, particularly saliency map and DeepSHAP, outperformed traditional techniques by revealing distinct patterns for various abnormal situations. This reduction in dimensionality offers several benefits, including enhanced cybersecurity, improved human–machine interfaces, and increased computational efficiency. The findings have significant implications for developing more accurate, efficient, and interpretable diagnostic systems in nuclear power plants, potentially improving safety and operational effectiveness. Future work will focus on validating these methods across diverse plant designs and integrating this approach with advanced AI techniques for real-time adaptive diagnostics.