Thermal Image-Based Fault Diagnosis of Induction Motors Using Deep Transfer and Ensemble Learning

Fault diagnosis of inductive motors (IM) is crucial in many industries. Infrared imaging has been popular in this regard, allowing for a non-invasive and effective capturing of fault-related thermal patterns. This paper addresses the thermal image-based IM fault diagnosis using deep transfer learning (TL) and ensemble learning. Different pretrained convolutional neural networks are leveraged: ResNet50, VGG16, MobileNetV2, EfficientNetB0, and InceptionV3. Several TL and training techniques are investigated: 1) Feature extraction. 2) Fine tuning the entire models. 3) Fine tuning coupled with hyperparameter optimization; here the learning rate (LR) using an LR scheduler. Ensemble learning is also applied on the base models trained using the aforementioned three techniques. Based on the experimental results, a test accuracy of 100% was achieved by the ResNet50 and VGG16 models trained using fine-tuning with an LR scheduler.