Explainable AI for Predictive Maintenance in Manufacturing Process Using RTDB Sensor Data

Predictive maintenance in continuous manufacturing processes is a critical challenge for smart factories. Papermaking, with its high energy demand and sensitivity to equipment failures, requires early and reliable anomaly detection methods. This paper presents a framework leveraging Real-Time Database (RTDB) sensor data from a papermaking process to predict binary alarm events before they occur. Using Random Forest models trained with lead-time shifted labels, we achieve reliable performance with an ROC AUC of 0.78 and an F1-score of 0.62 at a 5 -step horizon. Explainable AI techniques, including feature importance analysis and partial dependence plots, visualize the impact of key process variables such as steam pressure (PR001), basis weight (WT001), and moisture (MO001) on model predictions, providing interpretability for operators. We further propose a dynamic multi-level warning strategy based on optimal thresholds to provide actionable alerts. The proposed approach demonstrates both academic contributions to explainable predictive maintenance and industrial relevance for smart manufacturing in papermaking.