Integrating Large Language Models and Knowledge Graphs for System Diagnostics

Abstract

SUMMARY & CONCLUSIONSThis paper presents a novel diagnostic framework that integrates Dynamic Master Logic (DML) modeling with Large Language Models (LLMs) and Knowledge Graphs (KGs) to support fault diagnostics in complex engineering systems. By automating the extraction of the functional model of the system from textual documentation of the system and encoding it into a structured, searchable KG, the framework reduces manual modeling efforts while enabling interactive, interpretable diagnostics. The integration of LLMs allows users to engage with the system through natural language, invoking reasoning tools for fault propagation and success path analysis. The paper presents a case study of a nuclear auxiliary feedwater system with a demonstrated high model construction accuracy (>90%) and reliable model diagnostic query by the LLM agent. The approach supports both logic-driven diagnostics and contextual explanation, providing a solution for enhancing system understanding and decision-making in safety-critical domains.