An Intelligent Hierarchical Framework for Efficient Fault Detection and Diagnosis in Nuclear Power Plants

The increasing demand for usable power and the pressure to reduce carbon dioxide emissions have increased interest in fossil fuel alternatives. Specifically, nuclear power has received more attention from energy agencies globally, resulting in positive growth for the sector. The need for improved safety systems has increased with the expansion of nuclear power plants (NPP). Traditional fault detection and diagnosis (FDD) methods require high upfront and operational costs. Integrating Machine learning (ML) strategies can present a robust and equally effective alternative while minimizing the necessary time and money. This paper presents a novel framework for fault detection in NPPs. Unlike existing FDD methods that usually rely on single-model designs, we propose a hierarchical framework using a combination of multi and single-class classifiers. For data-driven FDD, one primary consideration is handling noisy scenarios in NPP. We design an algorithm that integrates deep learning multi- and single-class classifiers to improve fault diagnosis robustness, especially under noisy sensor readings. We evaluate our framework across various models and explore the need for a hierarchical approach under noisy and clean data. Our deep learning solution produces comparable results when no noise is present and significantly improves performance as noise is added to the system.