Abstract

not-yet-known not-yet-known not-yet-known unknown For the current visual detection methods of wind turbine blade defects, their detection models are usually excessively large, making them difficult to achieve a balance between model accuracy and inference speed. To tackle this problem, this paper introduces a lightweight wind turbine blade defect detection network, GCB-YOLO, which attempts to maintain a high detection accuracy and simultaneously achieve a rapid detection speed. At the beginning, a GhostNet network is employed to replace a portion of the YOLOv5s backbone network responsible for feature extraction. This replacement serves to reduce the network’s parameter size and computational load, thereby achieving compression of the feature extraction network. Subsequently, a CA (Channel Attention) mechanism is incorporated into the backbone network, which enhances its ability to focus on small-sized defects. Finally, the neck network PANet is substituted with a Bifpn network, bolstering its ability to discern small-sized defects. A series of validation experiments were conducted using an image dataset gathered from real wind farms. The result showed that the GCB-YOLO exhibited a reduction of 46.2% of model parameter number compared to that of the YOLOv5s. The improved model only has a 7.5MB volume. Hence, in GPU computation mode, the image detection speed reached 115.3 frames per second. More importantly, the proposed method achieved an mAP@0.5 of 94.72%, simplifying the deployment on edge computing devices and simultaneously meeting the real-time defect detection requirement with a sustained high level of detection accuracy.