Thermal Image-Based Condition Monitoring of Induction Motors: A Comparative Study

Introduction: Reliable monitoring of induction motors is essential for maintaining process continuity and reducing unplanned downtime in industrial systems. Infrared thermography provides a non-contact approach for capturing thermal patterns associated with operating conditions and motor degradation. However, the limited availability of labeled thermal images requires classification strategies that can achieve robust performance with relatively small datasets. This study investigates machine-learning-based thermal condition monitoring by comparing a hybrid CNN–ML approach with direct CNN classification.

Methods: A dataset containing 488 original thermal images was considered, covering six operating classes associated with healthy and corrosion-related motor conditions under full load, medium load, and no load. Geometric data augmentation was applied to increase the diversity of the training data while maintaining independent validation and test samples. Three ImageNet-pretrained convolutional neural networks—EfficientNet-B0, MobileNetV2, and ResNet50—were initially investigated. Two classification approaches were then compared. In the hybrid CNN–ML approach, frozen CNN architectures were used as deep feature extractors, and the extracted representations were classified using conventional machine-learning algorithms. In the direct CNN approach, the pretrained backbone was connected to a newly trained six-class neural classification layer.

Results: EfficientNet-B0 emerged as the best-performing architecture in both approaches. In the hybrid approach, EfficientNet-B0 features combined with logistic regression achieved a test accuracy of 98.63% and a macro-F1 score of 98.55%. In the direct CNN approach, EfficientNet-B0 achieved 95.89% test accuracy and a macro-F1 score of 96.11%.

Conclusion: Both strategies demonstrated strong capability for thermal condition classification. Nevertheless, the hybrid EfficientNet-B0–logistic regression approach achieved superior accuracy and class-balanced performance, indicating that combining pretrained deep features with a conventional machine-learning classifier is particularly effective for limited thermal-image datasets.