| Abstract:To address irregular damage boundaries, unstable scale estimation, insufficient recall for small-sample categories, and limited support for maintenance review in wind turbine blade inspection images, a damage identification and mechanical repair scheme recommendation method based on edge-consistency attention is proposed. A BDS-Net with YOLOv8-seg as the backbone is constructed, integrating edge-consistency attention, class-balanced Focal Loss, and rare-class augmentation to achieve localization, classification, and instance segmentation of surface cracks, leading-edge erosion, coating peeling, lightning ablation, and delamination bulging. Chord-length scale correction is used to recover damage dimensions, and repair schemes are recommended by combining damage category, position, size, confidence, load-bearing region, material curing conditions, and review constraints. Training and testing are conducted on 5,083 images, with 642 ground-review images and 218 maintenance records used for validation. The results show that the precision, recall, mAP@0.5, and mIoU are 91.8%, 89.6%, 88.1%, and 80.2%, respectively; the AP@0.5 and recall of lightning ablation are 86.8% and 87.3%, respectively. The complete scheme achieves a consistency rate of 90.6% and a review inapplicability rate of 3.7%, providing a reference for post-inspection maintenance review. |
点此下载全文
