Performance Evaluation of Different CNN Models for Motor Fault Detection Based on Thermal Imaging

Motor faults can lead to significant operational and financial losses. In this paper, we evaluated the performance of pre-trained Convolutional Neural Network (CNN) models, including Alexnet, VGG, and ResNet, for motor fault detection using thermal imaging. We have also tested the effect of Contrast Limited Adaptive Histogram Equalization (CLAHE) on the accuracy of these models. Our approach involves first generating a new dataset using CLAHE to enhance the original dataset, followed by training AlexNet, VGG, and ResNet models using transfer learning technique. The accuracy of the models is tested using the enhanced and original datasets as inputs. The results show that using the enhanced images significantly improves the accuracy of VGG11 and VGG13 models, but it deteriorates the performance of AlexNet. For ResNet models, the improvement is minimal or even slightly decreased. The reason could be due to insufficient training or overfitting since there are limited number of images in the dataset. Overall, ResNet models achieved the best performance with the highest accuracy of 99.62%. Thus, it can be concluded that ResNet models are better suited for detecting motor fault from thermal images. Future research could focus on optimizing the CNN models or incorporating multi-sensors fusion to improve the detection performance.