Inserção automática de componentes em ambientes virtuais de treinamento para subestações de energia utilizando inteligência artificial

Abstract

The training with Virtual Reality (VR) has gained prominence in corporate and industrial environments. The execution of educational maneuvers in virtual environments reduces the risk that inexperienced employees face in hostile and highly dangerous situations, such as those found in power substations. However, modeling these environments requires specialized labor and consumes numerous hours of development. This dissertation proposes a prototype to automate the creation of Virtual Environments by identifying and inserting Core Air Reactors, present in power substations, using images captured by Unmanned Aerial Vehicles (UAVs) and YOLOv8, a Convolutional Neural Network (CNN), with previous versions and different hyperparameter configurations. The training results indicated the superiority of YOLOv8 and its most accurate configuration. With the trained model, an application integrated with Unity software was developed, capable of receiving untrained photos, identifying the number of reactors present in the image, and finding their corresponding modeled objects, inserting them into the VR scene.