Classifying Electrical Tree Growth Stages in XLPE Cable Insulation using Vision Transformer with Window Sequence Merging Mechanism

The presence of partial discharge (PD) activities deteriorates the high voltage cross-linked polyethylene (XLPE) cable due to void, crack, and contamination of impurities. The persistence of these discharges replicates several carbonyl channels in the form of an electrical tree. The growth rate of treeing phenomena implies a sudden and complete breakdown of in-service cable. Therefore, detecting the various stages of electrical tree growths is essential for the safety and ensuring the reliable operation of the XLPE cable over a more extended period. This study focuses on classifying the electrical tree growth stages of a 33 kV, 3-core XLPE cable insulation using a vision transformer with a window sequence merging mechanism (WSM-ViT). Initially, electrical tree growth images are collected from the experimental setup and expanded using image augmentation techniques to address the limited availability of images. Subsequently, the augmented images are used to train the WSM-ViT model, and its performance is compared with the Vision Transformer (ViT) model. The results indicate that the WSM-ViT approach exhibited enhanced scalability and achieved an impressive recognition accuracy of 99.28%, outperforming the other models. Furthermore, the proposed model demonstrated reliability and effectiveness by identifying electrical tree growth in untrained images.