Improving Robotic UAV Wind-Turbine Inspection with Digital-Twin Synthetic Data

Abstract This paper presents a robotic operating system (ROS)/Gazebo-based synthetic data generation and evaluation workflow for offshore wind-turbine inspection using UAV imagery, combining a digital-twin synthetic data generation pipeline with a controlled experimental assessment of hybrid real–synthetic training strategies for damage detection. The core perception task is structural damage detection, addressed with a You Only Look Once version 11 (YOLOv11) object detector. To mitigate the scarcity and acquisition cost of annotated offshore data, we generate synthetic inspection imagery in simulation and further expand it through a compositional copy-paste augmentation strategy that increases scene diversity and reduces context bias. We perform a controlled comparison across four training scenarios (100% real-world, 85/15, 70/30, and 100% synthetic) using k -fold cross-validation. Results show that hybrid training improves both accuracy and stability: The 70/30 configuration increases mAP50 for the ‘damage’ class from 0.754 (real-only baseline) to 0.810 and improves localization stringency on the real hold-out set. Overall, the findings validate digital-twin synthetic generation as a practical augmentation strategy and support hybrid real-synthetic training as an effective route to more robust UAV-based inspection systems for predictive maintenance.