University of Cincinnati, Ohio, United States
Keywords: 3D Scanning, Aircraft Inspection, Path Planning, Reinforcement LearningAircraft maintenance such as inspection, replacing films, surface repairs etc., often needs 3D models of aircraft, but many aircraft do not have a 3D model or have an inaccurate 3D model due to hardware deviation. Current reverse engineering technology requires weeks time to manually setup and scan an entire aircraft, which is a process of high cost, low efficiency. Robotic scanning is a much faster and automated process, creating more accurate models with much less human labor. The key challenge is the problem of coverage path planning for autonomous scanning, which can be solved by using a machine learning technique. A 2-stage solution approach for automating aircraft scanning and inspection is proposed. The solution employs a UAV equipping a low-cost and lightweight RGB-D camera for acquiring a coarse 3D point cloud model of aircraft, which is used to generate an optimal path for detailed scanning by using Monte Carlo Tree Search Algorithm (a reinforcement learning technique). The solution was validated in a software of dynamics simulation, and the MCTS algorithm can learn a suboptimal path for detailed scanning in short time, and the path is also benefit to aircraft visual inspection.