Predictive maintenance system with a conversational Chat Interface

Introduction Predictive maintenance (PdM) anticipates equipment failure from historical maintenance data, overcoming limitations of the reactive and scheduled maintenance still common in thermal power plants. This study examines whether enterprise maintenance records alone, without dedicated sensor infrastructure, can support equipment-risk prediction, and couples the predictor with an offline conversational assistant. Methods We used 8,943 SAP Plant Maintenance records from a single thermal power station (2014–2026). After cleaning and deriving a time-to-next-defect label, 6,846 records remained, covering 388 of 447 monitored equipment items across 391 functional locations. A Random Forest classifier assigned each equipment item to a HIGH/MEDIUM/LOW risk category using a single chronological 80/20 train-test split, with time-series cross-validation for tuning. Performance was compared with a majority-class baseline using McNemar’s test and a bootstrap confidence interval. An offline Llama 3.2 assistant (via Ollama), whose responses are routed by query intent, was evaluated on a 40-question set. Results The classifier achieved 57.08% test accuracy, statistically indistinguishable from a 57.01% majority-class baseline (McNemar’s test, p = 1.00; bootstrap 95% CI, [−1.90, 2.04] percentage points). An ablation study showed that the equipment identifier contributes no measurable accuracy once functional location is included. Under the intent-routed architecture, which explicitly declines to answer rather than fabricate a response when no grounding data exist, 55% of the 40 questions produced verifiably reliable responses, up from an estimated 25–30% under the original approach. Discussion The current SAP metadata alone provides no statistically detectable improvement over predicting the majority class, so ERP metadata offers only modest, statistically unconfirmed predictive signal. The framework’s contribution is therefore primarily architectural: a retraining pipeline that correctly re-derives labels as new records arrive rather than substituting a proxy label, role-based access control, and an offline assistant. The results show that LLM grounding requires intent-aware retrieval rather than keyword-triggered lookup.