Convolution Neural Network in Partial Discharge Source Identification: Performance Analysis for Lab-Based Data and Electric Machine

Partial discharge (PD), caused by local electric field enhancement in insulation systems, accelerates degradation and can lead to breakdown if prolonged. Although complete mitigation of PD is not achievable due to the unavoidable manufacturing defects in the system, an early detection of PD and its source is necessary to prevent any further degradation of the electrical insulation system. In recent years, various deep-learning approaches have shown potential in classifying complex image patterns with no manual feature engineering. Therefore, this study has explored the use of deep learning techniques to classify distinct types of PD depending on their source, such as internal, surface, and corona discharges, and distinguish them from noise. To perform the classification task, convolutional neural network (CNN) models, such as AlexNet, ResNet, DenseNet, and InceptionV3, have been employed and trained utilizing phase-resolved partial discharge (PRPD) images taken from both laboratory experiments and field measurements collected from rotating electrical machinery. The models demonstrated high accuracy in identifying PD sources, highlighting CNN's potential for automating PD classification for a combination of data sources.