This thesis investigates the feasibility of detecting dents on aircraft structures using drones. It was initiated in collaboration with Mainblades, a company specializing in aircraft inspections, and aims to integrate dent detection into its existing platform. Identifying dents for Maintenance, Repair, and Overhaul (MRO) operators is critical because their presence in aircraft structures can compromise structural integrity, negatively impact fatigue and aerodynamic performance, and dents could be accompanied by cracks due to excessive plastic deformation.

The study addresses this goal through a three-part approach. The first part focuses on identifying and adapting suitable sensors for drone-based inspection. Secondly, the thesis investigates the performance of 8tree’s dentCHECK sensor to understand the feasibility of drone-based deployment. The selected sensor, 8tree’s dentCHECK, was tested in controlled experiments using an ABB robotic arm to replicate realistic drone movements and evaluate sensor response under translational and angular disturbances. Finally, a custom machine learning algorithm is developed to classify dents from synthetic point clouds. The algorithm is designed to be drone-agnostic and capable of handling varying data quality from different sensors available on the market.

The results indicate that while structured light sensors, such as 8tree's dentCHECK, show significant promise for drone deployment, their performance is mainly limited by the drone's drift. For the vibrations, linear vibrations in the Z direction (i.e., vertical motion as the drone moves up and down) are critical, while angular vibrations for the roll and pitch were inconclusive. Nevertheless, despite not meeting Boeing's full tolerance criteria (± 0.05 mm for depth and ± 1mm for length and width) when airborne, the 8tree sensor significantly outperforms manual inspection methods in speed and measurement consistency, despite not yet meeting Boeing’s tolerance criteria when airborne. It is a good benchmark for identification purposes that can be complemented and characterized manually with handheld scanners.

At the other end, synthetic point cloud datasets incorporating mathematical and FEM-based dent geometries with Gaussian and robot-collected drone noise profiles were used to train a U-Net-based machine learning model for automatic dent segmentation. The developed machine learning model demonstrated reliable identification for dents larger than 0.5 mm, although further data diversity is needed to enhance length and width predictions. Together, these findings demonstrate that integrating dent detection into autonomous drone platforms is both feasible and promising, laying the groundwork for future refinements in drone stability and data-driven detection methods.