Identificação automática de computadores em redes de distribuição urbana de energia elétrica

Abstract

Failures in the functioning of components in urban electricity distribution networks can causemany problems for businesses and homes, thus compromising public order. Some faults aredifficult to identify when they are not the cables. Making an identification quickly, can helpin reducing possible expenses with teams and equipment for maintenance. In order to assistin the rapid identification of possible component failures, this work makes use of machinelearning to identify aerial images captured by drones, components of the electrical network. To this end, it was necessary to expand and improve an image base provided by the ItaipuT echnological Park, label all the components contained in the base and train two Darknet-53convolutional neural networks (CNNs) in the YOLOv3 and Faster R-CNN framework in the Tensorflow framework. The assembled base had a total of 1,010 images and of these, 5,056images of seven components were labeled. The results were analyzed and compared, in whichthe Darknet-53 network presented an mean Average Precision in IoU (0.5) of 95.13% and inIoU (0.75) of 61.43%, whereas Faster R-CNN presented an mean Average Precision in IoU(0.5) of 88.72% and in IoU (0.75) of 60.71%.