关键词: 桥梁工程；螺栓病害识别；YOLO11；小波变换；特征融合

Abstract: Existing bridge bolt disease detection models based on deep learning suffer from high computational complexity and slow inference speed, making it difficult to meet the real-time monitoring needs of edge devices. To address this issue, a lightweight bridge bolt defect detection model based on an improved YOLO11n was proposed, and its improvement mainly consisted of frequency-domain feature enhancement module (FFEM) and cross-scale feature fusion network (CFFN). FEEM combined wavelet convolution and attention mechanism, and the model could extract the detailed features of bolts more accurately, enhance the perception ability of global information, and improve the identification accuracy of bolt diseases. CFFN optimized the feature fusion process at different scales, improving inference speed. Moreover, a lightweight downsampling module ADown was introduced in the backbone network to reduce the number of model parameters while retaining more useful feature information. Shape IoU considering the shape and scale of the target box was used as the bounding box loss function to improve the regression accuracy of bolt bounding boxes. The model was trained and tested on a dataset of 5 050 self-made bolt images covering three typical types of defective bolts such as rust, looseness and detachment, as well as normal bolts. Experimental results show that, compared with YOLO11n, the improved model has reduced the number of parameters and FLOPs by 47% and 34%, respectively. The accuracy and mAP50 have reached 94% and 91.7%, respectively, surpassing YOLO11n by 1 % and 0.8 %. The proposed model not only achieves high precision, but also has excellent lightweight characteristics, making it suitable for deployment on edge devices such as smartphones.

Key words: bridge engineering; bolt defect identification; YOLO11; wavelet transform; feature fusion