Interpretable visual transmission lines inspections using pseudo-prototypical part network

To guarantee the reliability of the electric energy supply, it is necessary that the transmission lines are operating without interruptions. To improve the identification of faults in the electrical power system, the unmanned aerial vehicle is used for inspection by recording photos. Based on computer vision, deep learning structures stand out for image classification have been an alternative to improve the identification of defects in transmission lines inspections. In this paper, the PseudoPrototypical Part Network (Ps-ProtoPNet) model is applied to perform the classification of missing insulators of high voltage transmission lines. To identify the position of the insulators chain and have the focus of the classification on the difference of insulators with failure, the YOLOv5 (n, s, m, l, and x), YOLOv6 (n, t, s, m, and l), YOLOv7 (std and x), and YOLOv8 (n, s, m, l, and x) are compared. The YOLOv8m is defined as the standard architecture for object detection since it has an mAP[0.5] of 0.9950 and mAP[0.5:0.95] of 0.9125. To classify the images, the Ps-ProtoPNet compares its various parts with the prototypes from all classes, and the image is classified based on the closest similarity to the prototypes class. The results show that the Ps-ProtoPNet achieves accuracy values sufficient to be applied in field inspections.