铁道标准设计2025，Vol.69Issue(8)：39-46,8.DOI:10.13238/j.issn.1004-2954.202312070001
深度学习视域下的轨面伤损检测与量化分析
Detection and Quantitative Analysis of Rail Surface Damage Based on Deep Learning
摘要
Abstract
To address the problems of slow identification and low accuracy in rail damage detection caused by increased train operation density,growing transport volumes,and shorter time windows for maintenance,this study proposed a rail surface damage detection algorithm based on the deep learning YOLO v5 model,aiming to further improve the speed and accuracy of rail damage identification.Based on high-quality images collected from real railway environments,a dataset of 1 792 non-repetitive rail surface damage images in three categories was constructed through screening and labeling.The self-attention mechanism and shifted window feature of the Swin Transformer network were used to replace the C3 modules at the end of the feature extraction and feature fusion stages in the YOLO v5 network,further enhancing the network's global perception capability.A lightweight Convolutional Block Attention Module(CBAM)was integrated into the Neck stage to refine feature responses in the spatial and channel dimensions,directing model attention more effectively to the rail surface.Ablation experiments were conducted to analyze the new algorithm,and its performance was compared with the original YOLO v5s,YOLO v3,and Faster RCNN.The experimental results showed that the proposed algorithm achieved accuracy rates of 92.5％for head checks,93.7％for spalling,and 99.5％for joint damage,with an average detection speed of 19.2 ms per frame.The proposed algorithm provided higher detection accuracy and stronger feature extraction capabilities for rail surface damage detection,effectively reducing false positives and missed detections.It achieves a balance between speed and accuracy and meets the requirements for rail surface inspection in real railway environments.关键词
轨面伤损检测/深度学习/YOLO v5/Swin Transformer/CBAMKey words
rail surface damage detection/deep learning/YOLO v5/Swin Transformer/CBAM分类
交通工程引用本文复制引用
王超,韩峰,李媛,王成祥..深度学习视域下的轨面伤损检测与量化分析[J].铁道标准设计,2025,69(8):39-46,8.基金项目
国家自然科学基金项目(51568037) （51568037）
对口支援高校联合创新资金项目(LH2023018) （LH2023018）
