Autonomous UAV-Based Power Line Inspection Using Lidar Point Clouds and Deep Learning for Fault Localization

Power transmission systems cannot be deemed safe and dependable without infrastructure inspection, the traditional methods of which are both perilous and nearly impossible to scale. The paper proposed a modular deep learning framework that makes use of unmanned aerial vehicles (UAVs) together with light detection and ranging (LiDAR) point cloud data to facilitate fault localization in power line infrastructure. The proposed approach unifies data collection by UAVs and preprocessing of raw LiDAR scans for further analysis, which makes use of statistical outlier removal, octree downsampling and unit ball normalization. PointNet[Formula: see text] is used for semantic segmentation, while defect classification is done by 3D-UNet. The 3D spatial features are then retrieved and displayed by heatmaps to facilitate the fault region identification. Evaluation on the Cables dataset shows exceptional segmentation results (mIoU [Formula: see text]%), along with accurate defect identification and localization with low Euclidean error. Furthermore, the models work in real-time at less than 1[Formula: see text]s, which is well below the targeted latency. Using this method, the UAV–LiDAR–deep learning combination has the potential to independently conduct assessments of infrastructure inspections in a more secure, insightful and efficient manner, thus addressing the issues the UAVs face when conducting such inspections.