Autonomous Inspections of Power Towers with an UAV

Abstract

Electrical grid maintenance is a complex and time consuming task. In this study, we test a way to perform electrical grid tower inspection by using camera images taken from an autonomous UAV. The images are segmented with a type of convolutional neural network called U-Net [9]. The results of the segmentation process is used to generate the movements of the UAV around the tower. The training of a U-Net model requires a large amount of labelled images. In order to reduce the time and financial costs of the generation of a large data-set of labelled images of physical towers, we develop a physics-based simulation environment that models the UAV dynamics and graphically reproduces electric towers in multiples environmental conditions. We extract labelled images for U-Net models training from the simulator. We perform multiple training, test conditions with different amount of natural world and simulated images and we evaluate which training condition generates the most effective U-Net model for the natural world image segmentation task. The contribution of the study is to detail the characteristics of the training condition that allows to maximise the U-Net performances with the minimum amount of physical world images in the training set. With the best performing U-Net, we create a post-processing analysis on the result of the segmentation to extract the required pieces of information to properly move the UAV.