Identifying wind turbine blade defects using YOLOv5 and YOLOv8 deep learning algorithms

Abstract

Wind power generation solves global energy supply and demand problems and promotes energy conservation, environmental protection and sustainable development. Accurate detection and identification of wind turbine blade defects is essential for the maintenance and upkeep of wind power systems. In this paper, four versions of You Only Look Once (YOLO) object recognition algorithm (YOLOv5n, YOLOv5s, YOLOv8n, YOLOv8s) are evaluated for wind turbine blade defect detection. The construction of wind turbine blade defect dataset is completed using image enhancement by flipping, rotating, colour changing, and noise injection, and a target detection model is trained using the YOLO model variant, which is able to quickly and accurately detect the defects on the blades, including soiling, oil leakage, erosion, and gelcoat peeling. The experimental results show that YOLOv5n, YOLOv5s, YOLOv8n, and YOLOv8s were trained for about 12.87, 11.32, 14.01, and 31.29 hours, the best mAP@0.5:0.95 is 0.83, 0.86, 0.93 and 0.95 respectively. Utilizing YOLOv5n, YOLOv5s, YOLOv8n, and YOLOv5s trained models to detect the same picture of wind power defects are about 77ms, 179ms, 139ms, and 261ms, respectively. In this study, YOLOv8n achieves the optimal combination of average accuracy and speed, which should be considered for wind blade defect detection and can improve the safety and reliability of wind power generation systems.