Abstract
Fault Tree Analysis (FTA) is an indispensable tool in high-stakes industries like nuclear power for conducting thorough risk assessments. However, the development of fault trees for Nuclear Power Plants (NPPs) is often marred by the necessity of interdisciplinary and intricate knowledge, posing a significant hurdle for non-experts. This necessity for specialized knowledge limits the wider adoption of FTA across various sectors. In response to these challenges, this study introduces the Nuclear Large Language Model Fault Tree Generator (NuLLM-FTG), an innovative solution aims at streamlining and augmenting the FTA process. Central to our approach is the implementation of large language models (LLMs) supervised fine-tuning (SFT) techniques. Specifically, we have developed a uniquely textual data structure, meticulously crafted to encapsulate the distinct characteristics of fault trees. To enhance the precision of our methodology, a comprehensive dataset containing several thousand examples is constructed for SFT. NuLLM-FTG’s performance is subject to a thorough evaluation. This process involved unraveling the “black box” nature of the model, allowing for an in-depth examination of performance enhancements, particularly in terms of horizontal conversational pattern alignment and vertical fault tree knowledge evolution. The practicality and effectiveness of NuLLM-FTG are corroborated through online experiments involving real operators, coupled with evaluations conducted by domain experts. Additionally, the applicability of our method in real-world scenarios is demonstrated through its integration with the Risk Spectrum (Version 14.0), thereby confirming its effectiveness and robustness. Crucially, our findings indicate that the finely-tuned NuLLM-FTG attains a level of performance comparable to experienced fault tree professionals across various metrics, including professionalism, completeness, and satisfaction. Notably, under specific conditions, our model outperformed GPT-4, and the utilization of an English-language corpus within our model proved to be more effective. Ultimately, our proposed method facilitates novices’ involvement in FTA.
