Optimizing UAV Altitude for Computer Vision-Based Visible Fault Detection in Solar Farms

Kiong, Oliver Han (2025) Optimizing UAV Altitude for Computer Vision-Based Visible Fault Detection in Solar Farms. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text OLIVER KIONG HAN_FULL TEXT.pdf Restricted to Registered users only Download (52MB)

Abstract

The maintenance and operational efficiency of large-scale solar farms critically depend on timely and accurate fault detection. This research investigates the optimisation of UAV inspection altitude for visible fault detection, with a specific focus on bird droppings on solar panels, a common and impactful contaminant. The objectives of the study were to determine the maximum UAV altitudes at which detection accuracy remains above recall thresholds of 0.85, 0.80, and 0.75, and to quantify the resulting reductions in flight distance and time when inspecting a 10 MW solar farm. To overcome the limitations of scarce real-world training data, a hybrid dataset comprising both real aerial imagery and photorealistic synthetic images generated in Blender was developed. Camera parameters in Blenderwere carefully matched to theDJI Phantom 4 Pro V2.0 to ensure realism, and manual annotation of both datasets was performed to maintain high labelling accuracy. Multiple YOLO-based computer vision models, including YOLOv8, YOLOv11, and YOLOv12 variants, were trained and evaluated across varying UAV inspection altitudes. Piecewise linear interpolation was employed to determine optimal altitudes corresponding to target recall thresholds. The findings demonstrate that optimised altitudes significantly improve operational efficiency. For instance, at an altitude of approximately 16 metres, inspections can be completed within a single 30-minute flight, while maintaining recall above 0.80. Higher altitudes such as 21.78 metres further reduce inspection time to under 20 minutes, albeit with a slight trade-off in detection accuracy. Models trained with hybrid datasets provided inconsistent improvements but still slightly outperformed those relying solely on real data, particularly at lower altitudes, highlighting the potential benefits of simulation-to-real transfer learning in enhancing defect detection capabilities. Overall, this study confirms that careful selection of UAV flight altitude, coupled with the integration of synthetic data to supplement real-world training datasets, effectively balances detection performance with operational efficiency. The findings contribute valuable insights for scaling UAV-based solar farm inspections and provide a foundation for future advancements in automated aerial monitoring of renewable energy assets.