CoT-Driven Knowledge Distillation for Fault Diagnosis in Mineral Processing Equipment

Le résumé fourni par la source

Large language models (LLMs) have demonstrated exceptional reasoning capabilities in complex fault diagnosis tasks, but their massive parameter requirements hinder deployment in resource-constrained industrial environments. To address this challenge, this study proposes a Chain-of-Thought (CoT)-driven reasoning distillation framework for transferring logical inference capabilities from large teacher models to lightweight student models. By leveraging LLMs as teachers, structured reasoning chains containing multi-step diagnostic logic are elicited through meticulously designed prompts. Alarm signals are then used as inputs, with the reasoning chains encapsulated within ¡reasoning¿ tags serving as critical supervisory signals. The final diagnostic work orders and reasoning chains constitute a comprehensive instruction-tuning dataset. To enhance model generalization, semantic-preserving paraphrasing is employed to augment input signals. Experimental results demonstrate that the distilled student models achieve significant accuracy improvements while maintaining reasoning trace fidelity comparable to teacher models. This research establishes a feasible pathway for deploying interpretable AI diagnostic systems on industrial edge devices with limited computational resources.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.