Fault Diagnosis in Thermal Images of Transformer and Asynchronous Motor Through Semantic Segmentation and Different Cnn Models

Abstract

Transformers are crucial power equipment that play an important role in changing voltage levels to meet consumer needs and in transmitting electricity. A fault in a transformer can cause significant economic losses and social problems. Similarly, asynchronous motors are widely used in industry, and faults in these motors can have substantial negative effects on both the economy and human life. Early detection of faults in both types of power equipment can save time and costs, as well as allow for remedial measures to prevent the failure of the entire system. Traditional fault diagnosis methods, which integrate various monitoring and measurement equipment into power systems, are not sufficient for early fault detection. Therefore, modern solutions have evolved towards more reliable and risk-free artificial intelligence (AI)-based automatic fault diagnosis methods. In our application, we aim to determine faults based on AI in thermal images of asynchronous motors and transformers in operation. Specifically, we propose a semantic segmentation application that highlights fault areas on thermal images, setting other pixels as background. This approach allows the region where the fault occurred to be taken as a reference for later fault diagnosis. As a result of semantic segmentation, the winding of the transformer and the stator region of the asynchronous motor are automatically segmented. Data augmentation techniques are then applied to these segmented images. Augmented and segmented motor and transformer images are classified using seven different Convolutional Neural Network (CNN) models. The results show that CNN models provide fault classification with accuracy reaching 100% for transformers and 96.49% for asynchronous motors.