Machine Learning-Based Fault Detection and Predictive Maintenance in Smart Electrical Power Systems
Abstract
Predictive maintenance is becoming more critical for intelligent asset management as any failure of the equipment can disrupt operations, lead to higher maintenance costs and affect system reliability, but they are not frequent. This paper presents a machine learning based decision support system for fault detection and maintenance prioritization based on the AI4I 2020 predictive-maintenance benchmark. Six operational predictors were examined over 10,000 observations, of which 339 were machine failures, after the removal of identifiers and leakage indicators of failure modes. Stratified cross validation and independent validation set were used to compare logistic regression, random forest, Extra Trees and HistGradientBoosting based on precision-recall area under the curve (PR-AUC) as the primary selection criterion. HistGradientBoosting has had the best PR-AUC (cross-validated) of 0.796 and was chosen for final evaluation. A validation-derived F2 threshold of 0.06 produced locked-test accuracy of 0.979, balanced accuracy of 0.951, recall of 0.922, F2 of 0.845, MCC of 0.756, PR-AUC of 0.904, and ROC-AUC of 0.984. The most important predictors were torque, air temperature, rotational speed and tool wear. Probabilities were converted to maintenance priorities due to risk-band stratification, which resulted in almost all observed failures falling into the high and critical bands. The framework illustrates how predictive analytics is made interpretable and can inform and trigger decisions for condition based maintenance in the context of intelligent management of industrial and power-system assets.