Vision-Driven Fault Detection and Diagnostic in Nuclear Power Plants Using Large Language Models
Conference
Tonday Rodriguez, JC, Sadekeen, D, Rahman, MA et al. (2025). Vision-Driven Fault Detection and Diagnostic in Nuclear Power Plants Using Large Language Models
. 1468-1477. 10.13182/NPICHMIT25-46669
Tonday Rodriguez, JC, Sadekeen, D, Rahman, MA et al. (2025). Vision-Driven Fault Detection and Diagnostic in Nuclear Power Plants Using Large Language Models
. 1468-1477. 10.13182/NPICHMIT25-46669
Nuclear Power Plants (NPPs) have grown in recent years due to the emerging requirement for renewable and environmentally safe energy generation. At the same time, NPPs are expensive to maintain due to the inherent complexity of operation and design, especially in the case of plant failures. Existing safety mechanisms in NPP fault detection and diagnosis (FDD) rely on rule-based or artificial intelligence techniques. However, rule-based methods require a deep understanding of the particular system and are not transferable between multiple power plants. Furthermore, artificial intelligence-based methods rely on existing architectures trained on curated numerical datasets that do not provide much interpretability of the fault or support for operators. This work aims to mitigate this by integrating modern language processing models, namely Large Language Models (LLMs) and Vision Language Models (VLMs), for FDD in nuclear systems. This work leverages VLM's ability to analyze graphical data, producing a description-based FDD architecture, which provides detailed logs of NPP system states and categorizes them into fault or benign conditions. Our proposed solution uses multiple LLMs to describe and classify the nuclear reactor states, showcasing accuracy equivalent to existing models and including additional interpretability not present in existing solutions.
