AI powered Failure Mode and Effect Analysis in safety critical industry

This paper presents the first systematic examination of Large Language Model (LLM) capabilities for automating the development of Failure Mode and Effects Analysis (FMEA) utilizing architectural diagrams as input. Although prior research has examined LLMs for FMEA tasks, our methodology incorporates innovative aspects, such as the direct analysis of architectural diagrams for component extraction, prediction of failure modes, causes, estimation of risk and a human-in-the-loop (Hu-IL) validation framework. We examine the capability of general-purpose LLMs to accurately automate the creation of FMEA by formulating a methodology that extracts components and signals from architectural diagrams, conducts automated component classification, and produces a comprehensive FMEA form sheet encompassing Severity, Occurrence, and Detectability (S/O/D) scoring. Our methodology is grounded in structured prompt engineering theory, utilizing scope bounding techniques to reduce hallucination while preserving extraction accuracy. Assessment against expert-validated ground truth (over 12 years of functional safety experience) across several automotive system diagrams indicates a 92% accuracy rate for signal extraction and component categorization, with S/O/D scoring obtaining an accuracy range of 70-90%. The results demonstrate substantial potential to reduce manual FMEA development processes. Key limitations include sensitivity to diagram complexity and quality, as inadequately designed diagrams markedly affect output precision along with the inability of LLMs to create new detection measures reliably. Our Hu-IL validation process mitigates these limitations while preserving the advantages of automation. This study provides baseline performance indicators for LLM-based FMEA automation and demonstrates significant potential in transforming traditional FMEA workflows in safety critical industries.