Thermographic image-based diagnosis of failures in electrical motors using deep transfer learning

Abstract

Diagnosing faults in electric motors is a task of great importance for the Industrial Sector since stopping these types of equipment can cause several invaluable losses for industries. To address the problems associated with quality and quantify motor data available in the literature, an intelligent fault diagnosis model for electric motors using Deep Transfer Learning (DTL), and InfraRed Thermal (IRT) images were studied. For that end, publicly available dataset containing 11 fault conditions were balanced and fed into a Convolutional Neural Network (CNN) pre-trained with the ImageNet dataset. Then, a cropping layer is added to the network to delimit the regions of interest and finally a hyperparameter optimization is obtained using Random Search (RS). The proposal was evaluated on four CNN architectures: InceptionV3, MobileNetV2, EfficientNetSV2, and RegNetX002. The research presented good results for fault classification, reaching 98.18% accuracy for the RegNetX002 model in 3245 s. All the experiments based on deep transfer learning presented great potential adaptation in the classification of problems when applied to the diagnosis of failures of machinery using data InfraRed Thermal.