Generating Troubleshooting Trees With FMEA Using Large Language Models (LLM)
Lasitha Vidyaratne, Huijuan Shao, Tsubasa Watanabe, Ahmed Farahat, Chetan Gupta · 2025
Troubleshooting trees are essential tools for industrial diagnostics, providing a structured framework for fault detection and resolution in complex systems. Construction of a new troubleshooting tree requires a systematic approach that integrates domain expertise, beginning with a comprehensive analysis of system architecture, operational parameters, and documented failure modes. However, the traditional development of troubleshooting trees is a labor-intensive process that relies heavily on expert knowledge, a large number of highly technical design documents, and iterative refinement. To address these challenges, this paper proposes an automated approach that leverages Large Language Models (LLMs) to generate troubleshooting trees using Failure Mode and Effects Analysis (FMEA) documentation as the primary input. The proposed method employs an LLM-based framework to extract and organize failure-related information from FMEA records, translating it into structured troubleshooting logic. Additionally, the framework introduces the possibility of including domain knowledge using a human-in-the-loop verification process, which also ensures traceability by linking each generated node to its source and enhances explainability through expert feedback consolidation. Experiments conducted on multiple FMEA documents demonstrate that the framework achieves an average coverage rate of 96.9%, with troubleshooting node precision exceeding 96% and a hallucination rate below 4% in most cases. These results demonstrate the effectiveness of LLM-based pipelines in automating structured fault analysis, offering a scalable and efficient alternative to traditional manual troubleshooting tree development.