Inspection of Electric Power Distribution Systems with SCRDet

Abstract

Faults in power distribution systems are among the factors that most affect the quality and continuity of the electric power supply, leading to higher operational costs and user dissatisfaction. Given this problem, this research applies deep learning techniques to develop an application capable of detecting utility poles in images, their angles, and the presence of fuse cutouts and their status (open or closed). The result is a system that can be used, among other applications, to detect or prevent faults in power distribution systems and as an asset inventory tool. We hypothesized that it is possible to use methods that perform well in the Dataset for Object Detection in Aerial images (DOTA) to accomplish this result, so, through the review of state of the art in object detection, we choose the SCRDet network to implement. Since there was no publicly available dataset with images of poles and fuse cutouts, we built it from scratch. Through the training and test, it was possible to evaluate the results and make adjustments, reaching a satisfactory result that has proven the viability of such application.