Redes neurais artificiais aplicadas no monitoramento de ativos em subestações de energia elétrica

Abstract

With the constant growth of the electrical system, the Brazilian electrical system, efficiency in the method of performing maintenance in these environments becomes increasingly important. A thermographic inspection is one of these alternatives, and it is effective for predicting failures in power substations equipment. However, the operators’ safety and the need for a high degree of experience make the activation very expensive. The central idea of the present work is to propose automation of the thermographic inspection process of electric power substations using algorithms of convolutional neural networks and random forest to detect disconnecting switches. One of the most efficient neural networks nowadays is YOLO (You Only Look Once), this technique linked to the application of image segmentation through the random forest algorithm can automatically find the hot spot of given equipment. Using a database of images of switch-disconnectors from an electric power substation provided by COPEL from the research project PD-02866-0528/2020 for training, the technology can satisfactorily identify the objects and thus help the energy concessionaires in this process. The points to be improved are in the backgrounds and lighting effects in the images, which end up causing some confusion on the network.