AI-Based Fault Diagnosis of Industrial Electric Motors Using Multi-Sensor Data: A Comparative Evaluation of Random Forest, XGBoost, Support Vector Machine, and Artificial Neural Network Models

Abstract

When industrial electric motors fail unexpectedly, the resulting downtime and repair bills can be severe, which makes dependable early fault detection central to condition-based maintenance. Here, we quantitatively compare four machine learning models—Random Forest, XGBoost, Support Vector Machine (SVM), and Artificial Neural Network (ANN)—to distinguish healthy operation from several motor fault conditions using multi-sensor records (vibration, stator current, winding temperature, shaft speed, and load) drawn from industrial induction motors. Features across time and frequency domains were pulled from every sensor channel, after which the models were trained and checked via a stratified train-test split. Classifier performance was tracked through accuracy, precision, recall, F1-score, and root mean square error (RMSE); in parallel, feature-importance and sensitivity tests pinpointed the sensor metrics most influential in identifying faults. The findings, stated as values in line with trends from earlier literature, show XGBoost delivering the highest overall classification accuracy and lowest RMSE, with Random Forest finishing just behind; ANN remained competitive while capturing nonlinear fault patterns better, whereas SVM lagged in separating multiple fault classes. Feature-ranking showed that kurtosis and root-mean-square values from vibration signals, alongside harmonic components in stator current, weighed most heavily in sorting faults; by contrast, rotational speed and motor load acted mainly as baseline inputs that stabilized the models across different working regimes. These outcomes argue for using ensemble tree architectures in multi-sensor motor diagnostics, showing that pairing vibration with current data yields far better diagnostic insight than leaning on one sensor type alone.