KG-MAFMEA: A Knowledge Graph-Driven Multi-Agent Framework for Failure Mode and Effects Analysis

Abstract

Failure mode and effects analysis (FMEA) is widely used for proactive reliability and safety analysis, but conventional FMEA knowledge is usually stored in tabular documents, making causal dependencies difficult to retrieve, reason over, and reuse. Although large language models (LLMs) provide strong natural language understanding capabilities, directly applying them to FMEA may produce unsupported or hallucinated conclusions. To address these limitations, this paper proposes KG-MAFMEA, a knowledge graph-driven multi-agent framework for FMEA reasoning. The framework constructs a failure mode knowledge graph (FMKG) from FMEA records, explicitly representing the causal relations among functions, failure modes, failure causes, failure effects, detection methods, and recommended actions. Based on the FMKG, the Reasoning Agent performs intent recognition, reasoning-anchor retrieval, and risk-aware causal traversal to support root cause analysis, effect prediction, and measure recommendation. An Evaluation Agent further assesses generated answers using intent recognition, entity grounding, faithfulness, and usefulness metrics. Experiments on an automotive design FMEA dataset show that KG-MAFMEA achieves high retrieval accuracy and traceable FMEA reasoning.