A Transfer Learning Architecture to Detect Faulty Insulators in Powerlines

Regular inspection and monitoring of various critical equipment (porcelain insulators including pins or discs) available on power grid lines is required for several reasons, such as safe and uninterrupted transmission, avoiding shock accidents, and many more. In the earlier days, these inspection and maintenance activities were carried out manually by skilled experts. However, manual inspection is a challenging, dangerous and costly task. With the latest developments in the field of Artificial Intelligence and Unmanned Aerial Vehicles, several neural networks and feature-oriented research tools & techniques are introduced to identify broken components from the UAVs captured images. However, these benchmark research studies in this domain have several significant limitations, such as low generalization capability on different component types, inability to detect small-size components (pins & discs), and so on. To overcome these issues, the present research proposes a generalized model for identifying or classifying different small-size component types on the transmission lines. The proposed approach implements a deep learning pipeline involving several stages, including data preprocessing, quality assessment, augmentation, YOLOV5, and DETR for the target task. In addition, the current work also proposes a transfer learning strategy for improving classification accuracy and generalization capabilities. The performance assessment of the proposed approach and existing benchmark approaches on the porcelain insulator dataset validates the effectiveness and reliability of the proposed approach.