Leveraging LLMs and Case-Based Reasoning for Expert-Governed Risk Knowledge Reuse

Abstract

Failure Mode and Effects Analysis (FMEA) is central to industrial risk management but remains labor-intensive, expert-dependent, and difficult to reuse across projects. Large Language Models (LLMs) can support the generation and adaptation of FMEA content, yet their application in safetycritical engineering requires semantic grounding and expert oversight to prevent hallucinated or technically implausible recommendations. This paper presents the FMEA Knowledge Reuse Framework (FMEA-KRF), which combines Case-Based Reasoning (CBR), two-phase Retrieval-Augmented Generation (RAG), and expert-governed validation. Historical FMEA data are transformed into a vector-relational case base that preserves cause-failure mode-effect relations and links them to product, process, and resource context. For new analysis tasks, similar cases are retrieved to ground LLM-generated proposals, which are reviewed and revised by domain experts before retention. An evaluation across 150 configurations and five LLMs identifies retrieval as the main quality driver. RAG-supported configurations substantially improve FMEA completeness, whereas prompting strategy has only minor influence once sufficient context is available. The best configuration achieves an SMEvalidated score of 4.46/5.00. The results further indicate that cost-efficient models such as Gemini 2.0 Flash can achieve competitive performance when grounded in a structured case base. The framework provides a lightweight blueprint for expertgoverned, LLM-augmented risk knowledge reuse that integrates with existing FMEA toolchains.