Fault Tree Analysis (FTA), a key component of Probabilistic Safety Assessment (PSA) in the nuclear industry, is traditionally a manual and labor-intensive process. Constructing fault trees requires not only the interpretation of Piping and Instrumentation Diagrams (P&IDs), but also the incorporation of diverse information related to plant design, system configuration, operating logic, and engineering knowledge. Because this process relies heavily on expert judgment, it is time-consuming, prone to human error, and often leads to inconsistencies that can affect both analytical quality and project cost.
This research introduces a novel framework that utilizes a specialized Large Language Model (a Custom GPT) to automate this critical workflow. Unlike general-purpose AI, which can be prone to factual inaccuracies, this framework's reasoning is precisely controlled by embedding domain-specific engineering knowledge and a set of explicit logical rules. This approach ensures that the generated analyses are both highly consistent and reproducible, directly addressing the limitations of less constrained AI models.
The developed framework demonstrates a full end-to-end capability, automating the entire process from the initial interpretation of P&IDs to the generation of complete FTA models. Crucially, it ensures compatibility with PSA software, AIMS-PSA, allowing its outputs to be directly integrated into existing PSA workflows. The framework's performance was rigorously validated through multiple case studies on representative nuclear systems. The outputs, including key safety metrics such as minimal cut sets and top event probabilities, were benchmarked against expert-verified analyses using the AIMS-PSA software. The results demonstrate that the framework can produce FTAs with expert-level accuracy, confirming its significant potential as a tool for cognitive automation in nuclear safety. This work paves the way for enhancing the efficiency, reliability, and overall quality of the PSA process.