공개 고장 데이터와 LLM을 활용한 FMEA 자동화 프레임워크: NHTSA 차량 리콜 및 불만 데이터를 기반으로

초록

Purpose: The purpose of this study is to automate the FailureModeandEffectsAnalysis (FMEA) process using the National Highway Traffic Safety Administration (NHTSA) vehicle recall and complaint data. It establishes a practical foundation for manufacturers to identify potential defects early and improve quality management. Methods: We utilized NHTSA recall data from 1979 to 2024, applied text preprocessing, Large Language Model (LLM)-based sentence embedding, and K-means clustering to group similar failure modes. Human-readable failure modes were identified using Term Frequency-Inverse Document Frequency (TF-IDF) keywords and LLM-generated labels. Risk Priority Numbers (RPNs) were quantitatively evaluated based on severity, occurrence, and detection metrics. The proposed framework was validated using recent recall data. Results: Application of the proposed methodology to the top ten manufacturing-processdefects demonstrated the potential to prevent 935 recalls from the past three years. We present a practical approach to significantly reduce recall rates and improve manufacturing quality. Conclusion: Our framework enables manufacturers to efficiently analyze and categorize data via RPN calculation, facilitating early identification of failure modes, prevention of potential defects, reduced recall response times, and improved safety

키워드