Geração de trajetórias para manutenção preventiva em usinas fotovoltaicas de grande porte utilizando visão computacional e algoritmos clássicos de otimização

Abstract

The accelerated growth of solar energy in Brazil, evidenced by the expansion of centralized-generation photovoltaic plants, poses significant technical challenges related to the monitoring and maintenance of large operational areas. Traditional manual inspections in these environments demand considerable time and human resources and expose workers to safety risks. This work proposes a computer-vision-based system to automate the detection of photovoltaic module rows in aerial images and to generate coverage routes for maintenance teams. The YOLOv8m object detection architecture is trained to locate the photovoltaic tables, whose regions are treated as obstacles in the construction of a navigable mask of the corridors. Skeletonization is applied to this mask to extract the central axis of the paths, which is then converted into a weighted graph with distances expressed in pixels, given the absence of Ground Sample Distance (GSD) information in the images. The inspection route is obtained by formulating the Rural Postman Problem, in which the corridors between rows constitute the required edges and the connecting segments are used only to link the corridors and adjust the parity of the graph, using Dijkstra's algorithm to compute the shortest paths of these connections. The detection model achieved a mAP@0.5 of approximately 0.81, and the generated route covered between 96% and 100% of the detected rows in the evaluated plants, with an average coverage of 97.6% and an average travel savings of 23.4% compared with the Chinese Postman formulation. [resumo fornecido pelo autor]