Explainable LLM-assisted fault diagnosis for smart manufacturing using multi-trace diagnostic rationales and verification

Abstract

Abstract Smart manufacturing systems increasingly rely on data-driven fault diagnosis to support condition monitoring, predictive maintenance, and operational decision-making. Although conventional machine-learning and deep-learning models can achieve high diagnostic accuracy, their predictions are often difficult for engineers to interpret, verify, or confidently use under uncertain operating conditions. This study presents CoT-FD6, an LLM-assisted diagnostic framework combining standardized sensor features, training-only analogical retrieval, multiple explicit structured diagnostic rationales, modal LLM diagnosis aggregation, independent prototype verification, and selective acceptance or human-review escalation. The framework is evaluated on controlled synthetic manufacturing faults and the NASA C-MAPSS FD001 degradation benchmark, with conventional classifiers tested on the same held-out instances. On the synthetic task, the raw LLM diagnosis achieved $$0.998 \pm 0.004$$ accuracy and macro-F1. On NASA FD001, Logistic Regression, SVM-RBF, and Random Forest achieved accuracies of 0.920, 0.920, and 0.930, respectively, whereas the raw LLM achieved $$0.618 \pm 0.018$$ . Retrieval-only classification reached 0.860 accuracy and the independent verifier 0.910. Selective governance increased accuracy to $$0.900 \pm 0.008$$ among automatically accepted NASA cases at $$0.660 \pm 0.020$$ coverage while capturing $$82.7\%$$ of raw LLM errors. Generated rationales were strongly grounded in the supplied measurements, with grounding rates of 0.982 and 0.995 on the synthetic and NASA datasets, respectively. These results position CoT-FD6 as a complementary explanation, verification, and selective-governance layer rather than a replacement for high-performing supervised fault classifiers.