A comparative study of KNN, Random Forest, CNN-LSTM, and Transformer encoder for noninvasive fault detection in three-Phase induction motors via stator current analysis

Introduction: Accurate and noninvasive fault diagnosis of three-phase induction motors is essential for predictive maintenance, improved reliability, and reduced maintenance costs in industrial environments. Stator current analysis provides a practical and cost-effective approach for monitoring motor health without requiring additional sensors. This study presents a comparative evaluation of four machine learning models for fault classification using stator current signatures.

Methods: A dataset of 609 fixed-length stator current windows was generated in MATLAB/Simulink under seven operating conditions, including healthy operation, voltage imbalance (±10% and ±30% per phase), and single-phase loss. Four supervised learning models, K-Nearest Neighbors (KNN), Random Forest (RF), a hybrid CNN-LSTM, and a Transformer Encoder, were trained and evaluated using a 70%, 10%, and 20% training, validation, and testing split. Classification performance was assessed using accuracy, precision, recall, and F1-score.

Results: Random Forest and CNN-LSTM achieved 100% classification accuracy, precision, recall, and F1-score on the test dataset, while the Transformer Encoder and KNN obtained 98.4% accuracy with an F1-score of 0.98. The few misclassifications occurred primarily between adjacent voltage imbalance levels. These results demonstrate that both traditional machine learning and deep learning approaches can effectively identify motor faults from stator current signals under simulated operating conditions.

Conclusions: The proposed comparative study highlights the effectiveness of data-driven approaches for noninvasive induction motor fault diagnosis. However, the results are based on controlled, noise-free MATLAB/Simulink simulations and therefore do not fully represent real-world operating conditions. Future work will focus on validation using experimental measurements incorporating measurement noise and varying load conditions to improve robustness for real-world industrial applications.