A framework for automating failure modes and effects analysis (FMEA) using large language models (LLMs) and retrieval-augmented generation (RAG)

Abstract Under the influence of various failure mechanisms – such as corrosion, erosion, fatigue, and deformation – process equipment (chemical, mechanical, and metallurgical) degrades over its service life. This degradation undermines equipment integrity and heightens the risk of failure. Although total elimination of failure risk is unattainable, robust asset management programs seek to maximize safety, reliability and availability through effective and efficient inspection and maintenance strategies. Designing and executing these programs, however, incurs significant complexity and cost. Risk-based methodologies, such as Risk-Based Inspection (RBI) and Reliability-Centered Maintenance (RCM), provide structured approaches for prioritizing resources and mitigating accident potential. Failure Modes and Effects Analysis (FMEA) remains a cornerstone of risk assessment. While FMEA can incorporate quantitative elements – such as estimating failure probabilities and consequences – it is fundamentally qualitative. Consequently, considerable time and effort are devoted to manually identifying hazards, failure modes, failure mechanisms, and failure causes, making traditional expert-driven FMEA processes time-consuming, error-prone, and inherently static. Advances in generative AI have the potential to automate aspects of the FMEA workflow. This study introduces a framework that combines open-source large language models (LLMs) with Retrieval-Augmented Generation (RAG) to extract insights from industry standards and authoritative literature for automated FMEA table generation. The paper describes the proposed architecture and presents initial evaluation results.