Investigation of wind turbine blade defects using hybrid YOLOv8 and vision transformer

Abstract

Ongoing supervision of wind turbine blades and their frameworks is an essential element of Structural Health Monitoring (SHM) systems. Unmanned Aerial Vehicles (UAV) have become increasingly significant for automated inspections in contemporary studies because of the impracticality of constant human monitoring. The use of computer vision in conjunction with machine learning is essential one for identifying the damages from images captured by UAVs. The work herein presents a hybrid two-stage pipeline that integrates YOLOv8 with CNN (DarkNet53) and lightweight Vision Transformer models (Swin-Tiny and DeiT-Small) to evaluate the detection performance under different architectural configurations. Two variants of YOLO-based object detection frameworks were implemented using different backbone feature extractors, including the CNN(DarkNet53), the Swin-Tiny transformer, and the DeiT-Small transformer. This investigation evaluates the model’s capacity to distinguish among five essential types of defects in wind turbine blades: paint peeling, leading-edge erosion, open damage, missing vortex generator, and non-open cracks. The framework is assessed using both imbalanced and balanced datasets, with performance comparisons conducted for quantitative criteria including accuracy, mAP, F1-score, recall, and FPS. Qualitative measurements are derived from the visual evaluation of predictions as bounding boxes. The YOLOv8 + DeiT-Small configuration trained on a balanced dataset has achieved excellent performance among the evaluated configurations. Dataset balancing plays a critical role in maximizing model performance, thus yielding comparably better accuracy, improved mAP, and smoother convergence, particularly when combined with computationally efficient backbone architectures.