Real-Time Induction Motor Health Index Prediction in A Petrochemical Plant using Machine Learning

This paper presents real-time health prediction of induction motors (IMs) utilised in a petrochemical plant through the application of intelligent sensors and machine learning (ML) models. At present, maintenance engineers of the company implement time-based and condition-based maintenance techniques in periodically examining and diagnosing the health of IMs which results in sporadic breakdowns of IMs. Such breakdowns sometimes force the entire production process to stop for emergency maintenance resulting in a huge loss in the company's revenue. Hence, top management decides to switch the operational practice to real-time predictive maintenance instead. Intelligent sensors are installed on IMs to collect necessary information related to their working statuses. ML exploits the real-time information received from intelligent sensors to flag abnormalities of mechanical or electrical components of IMs before potential failures are reached. Four ML models are investigated to evaluate which one is the best, i.e. Artificial Neural Network (ANN), Particle Swarm Optimization (PSO), Gradient Boosting Tree (GBT) and Random Forest (RF). Standard performance metrics are used to compare the relative effectiveness among different ML models including Precision, Recall, Accuracy, F1-score, and AUC-ROC curve. The results reveal that PSO not only obtains the highest average weighted Accuracy but also can differentiate the statuses (Class 0 - Class 3) of the TM more correctly than other counterpart models.