ABSTRACT

CNC spindle bearing failures cause ₹15–38 lakh in downtime per incident. Traditional vibration sensors trigger alerts too late, offering less than 48 hours of warning. This study finetuned GPT-4o-mini and LLaMA-3-8B on 12,000 anonymized technician voice-to-text logs to detect lubrication degradation 4.2 days earlier using informal slang cues such as "grease like peanut butter" and "squealing at startup." The models achieved 86.7% accuracy with only 6.3% false positives, enabling 28% bearing life extension through just-in-time relubrication alerts. Federated learning architecture keeps raw chat data on technician devices, while opt-in dashboards empower workers with visibility into their own predictions. Results significantly outperform VADER sentiment analysis and BERT-base baselines, demonstrating that large language models can enable proactive maintenance without surveillance overreach.

Keywords: lubrication failure, large language models, predictive maintenance, CNC spindle, technician chat, ethical AI