An Integrated ML Framework for Predictive Maintenance and Energy Optimization

Abstract

The industrial machinery provides constant stream of sensor readings, which can be effectively used for failure prediction and maintenance decisions. This research work proposes a novel approach to enable predictive maintenance and energy saving through the use of AI4I 2020 Predictive Maintenance Dataset. In the proposed model, XGBoost algorithm is used for performing machine failure prediction, along with the use of SMOTE to handle the imbalanced classes and feature engineering for revealing the relationship between sensor readings. The SHAP-based method is used for explaining the key features affecting failure prediction and prediction-based alarm system is introduced as well. Experimental evaluation on the held-out test set demonstrated 97.4% accuracy, which is better than that of baseline models. In an experiment evaluating the performance in the 48-hour deployment, the model detected all the heat dissipation failure instances that were observed, alerting about them on average 3.2 cycles before the threshold was breached. The energy-awareness policy provided 18.4% of energy savings in the period when the machine operated with degradation.