Virtual Reality Simulation of Autonomous Solar Plants Inspections with Unmanned Aerial Systems

This paper presents the development of a virtual reality simulation environment for Unmanned Aerial Systems (UAS) solar plant inspection. The objective of this work is to provide a tool to test autonomous inspection and computer vision algorithms and generate realistic synthetic data for deep learning. These techniques demand realistic synthetic data, which can be made available by high-quality graphics engines, such as the ones used for game development. In this work, Unreal Engine 4 is used to host the virtual solar plant. The solar panels were modeled using Blender and Photoshop. Microsoft's AirSim plugin is used to simulate the UAS motion, together with the ArduPilot Software-In-The-Loop flight controller. The environment was evaluated through a virtual autonomous inspection of a plant with 9200 panels, where a georeferencing algorithm was used to locate the defective solar panel in a raster plant layout, based on the pixel position of the defects in the aerial images. The virtual inspection resulted on more than 1000 images and the localization of the defective panels in the layout plant using the georeferencing algorithm had an error of 0.34 meters on the North axis and 0.26 meters on the East axis, which is acceptable for large solar plants with sparse modules' arrangement.