Acoustic-based detection of stator faults in three-phase induction motors using machine learning techniques

Three-phase induction motors are widely employed in industrial systems due to their mechanical simplicity and cost-effectiveness. However, operational stresses can induce stator faults, particularly inter-turn short circuits, which may compromise system reliability. This study proposes a diagnostic methodology based on the direct analysis of unfiltered acoustic signals acquired by five distinct microelectromechanical system (MEMS) microphones under varying mechanical load and supply voltage imbalance conditions. Features such as standard deviation, spectral centroid, zero-crossing rate (ZCR), spectral bandwidth, root mean square (RMS), and Mel Frequency Cepstral Coefficients (MFCC) were extracted to characterize the motor acoustic emissions. Supervised machine learning classifiers, including k-Nearest Neighbors (KNN), Random Forest (RF), Decision Tree (DT), Multilayer Perceptron (MLP), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost), were trained and evaluated using 5-fold cross-validation to ensure statistical reliability. After feature selection, ensemble-based models such as RF and XGBoost achieved classification accuracies up to 100%, even for incipient fault conditions (i.e., early-stage faults with minimal acoustic signature deviations). The findings demonstrate the feasibility of using acoustic-based techniques combined with machine learning for reliable stator fault diagnosis in induction motors, offering a non-invasive, scalable, and cost-effective solution for predictive maintenance strategies in industrial environments. Limitations regarding external noise sensitivity and experimental scalability are also discussed as future work directions.