A Hybrid Feature-Based Machine Learning Framework for Real-Time Fault Detection and Classification in Three-Phase Induction Motors

Abstract

Unplanned failure of three-phase induction motors is one of the leading causes of unscheduled downtime, safety hazards, and maintenance expenditure in industrial and power-system installations. This paper presents an extensive, reproducible machine learning (ML) framework for the automatic detection and classification of the three most prevalent electromechanical fault types bearing defects, broken rotor bars, and stator winding faults together with the healthy operating condition, using stator current signatures. Ten time- and frequency-domain features are extracted from simulated current waveforms generated with motor current signature analysis (MCSA)-informed fault models under randomized severity and realistic sensor noise. Four classifiers Support Vector Machine (SVM), k-Nearest Neighbors (k-NN), Random Forest (RF), and a shallow feed-forward Artificial Neural Network (ANN) are tuned via 5-fold cross-validated grid search and evaluated on a stratified 1,200-sample dataset (300 samples per class). After hyperparameter optimization, the Random Forest and proposed ANN classifiers achieve the highest test accuracy (98.33% each), followed by SVM (97.67%); k-NN trails substantially (76.67%). Paired t-tests over cross-validation folds confirm no statistically significant difference between ANN, RF, and SVM (p > 0.05), while the gap to k-NN is highly significant (p < 0.001). Beyond headline accuracy, this study contributes a feature-ablation analysis that identifies the sideband-energy ratio and harmonic ratio as indispensable features (their removal drops accuracy by up to 26.7 percentage points), a noise-robustness sweep showing Random Forest degrades most gracefully under increasing measurement noise, learning curves characterizing data efficiency, receiver operating characteristic (ROC) and precision-recall analyses per fault class, and a computational cost comparison (training time, per-sample inference latency, and parameter count) relevant to embedded deployment. The methodology, mathematical formulation of each classifier, dataset-generation procedure, and full evaluation protocol are documented to support reproducibility, benchmarking, and extension to experimentally acquired data. Keywords: machine learning; Random Forest; Predictive Maintenance; fault diagnosis; induction motor; artificial neural network; support vector machine; motor current signature analysis; condition monitoring; feature ablation.