Abstract: Traditional composite structure damage detection technologies suffers from operational complexity, a reliance on skilled personnel, and high detection costs. To address this issue, this paper proposes an LWDS identification method for composite structure damage based on the improved YOLOv5s model. A large selective kernel module is introduced into the original model to enhance feature extraction capability, the WPDIoU loss is adopted to optimize the bounding box regression accuracy, and the SE attention mechanism is integrated to improve the weight of key features. Experiments results on the composite structure dataset demonstrate that the proposed method achieves significant improvements in detection accuracy. Compared with the baseline model, the performance indicators of this model are significantly improved. The classification accuracy increased by 12.8%, the target detection rate by 11.7%, and the mean average precision (mAP) by 15.3%, while maintaining a high real-time processing speed. The newly proposed method can realize accurate and rapid detection of composite structure damage, and effectively avoid false detections and missed detections of damage defects.