Detecção automática de riscos por vegetação em redes de distribuição na UFAM utilizando visão computacional e reconhecimento de padrões

Abstract

The efficiency and safety of the electrical network are critical elements for the operation of any institution, and UFAM is no exception. In the Amazonian context, where climatic conditions can exacerbate risks, it is essential to adopt preventive measures to ensure the integrity of the electrical network. The risk identification procedures adopted by the energy utility prove to be limited, as they rely on visual field inspections, which are constrained in terms of efficiency and scalability. Therefore, there is a clear need to automate the identification of trees near power lines to enhance the efficiency and effectiveness of risk detection and enable rapid and accurate response. After creating a local database to assess the need for tree pruning at UFAM, various pattern recognition methods from the literature are evaluated to identify the most efficient approaches for this problem. This project follows a methodological approach divided into three main stages. Initially, a dataset is collected and organized, exploring and adapting computer vision and pattern recognition techniques, including the use of convolutional neural networks. During this stage, model parameters are tuned and tests are conducted to ensure effectiveness. Next, the performance of the developed system is assessed using metrics to verify its accuracy and operational effectiveness, comparing it with widely used methods from the literature. The final stage involves estimating the distance between conductors and tree branches. Models trained with YOLOv8 and Keras achieve F1-Scores between 90% and 100% on test images and Google Maps images. For video predictions, YOLOv8 and MobileNetV2 achieve F1-Scores between 85% and 100%.