铁道科学与工程学报2026，Vol.23Issue(8)：4077-4090,14.DOI:10.19713/j.cnki.43-1423/u.T20251679
TreadDet-YOLOv12:一种基于改进YOLOv12的轮对踏面损伤检测模型
TreadDet-YOLOv12:a train wheelset tread damage detection model based on improved YOLOv12
摘要
Abstract
To address the issues of low efficiency in manual inspection,insufficient edge detail perception in existing deep learning models,poor adaptability to complex scenes,and redundant parameter quantity in train wheel tread damage detection,a wheel tread damage detection model,TreadDet-YOLOv12,based on an improved YOLOv12,was proposed.First,the Edge_GAM module was designed:considering the characteristics of tread fine cracks and multi-directional damage,three scale gradient extraction kernels and four directional gradient sub-kernels were introduced on the basis of the global attention mechanism GAM to extract multi-scale directional gradient features.After lightweight convolutional encoding,these features were fused with channel attention output to enhance the fine capture ability of damage contours and subtle textures.Second,a lightweight DWC_ECA module was constructed:considering the sparsity of tread damage channel features,the efficient channel attention mechanism ECA was integrated into the depthwise separable convolution DWConv network.Through deep convolution,independent kernels were allocated to each channel to extract spatial features.ECA dynamically enhances the channel response of key damage-related texture edges,and pointwise convolution completes cross-channel fusion,balancing lightweight and feature representation capabilities.Finally,the Focal_EIoU loss function was introduced to replace the traditional CIoU loss function.The Focal mechanism was used to increase the weight of difficult samples,and EIoU was combined to precisely constrain bounding box regression,alleviating sample imbalance issues and reducing the missed detection rate of minor damages.Experimental results show that the precision,recall,mAP50,and mAP50-95 of the TreadDet-YOLOv12 model are improved by 1.7,3.7,2.3,and 1.6 percentage points,respectively,as compared to the baseline model YOLOv12n;and the floating-point operations per billion(GFLOPs)increase by 0.7.In engineering verification,the model was deployed on an industrial host computer,achieving an average detection frame rate of 43 frames per second,a damage detection rate of over 93%,a false detection rate of less than 3%,and a missed detection rate of less than 1.5%.The efficiency is improved by 3 times compared to manual inspection,and the model has been stably running for over 18 months.This model improves detection accuracy while ensuring lightweight,especially enhancing the recognition ability of fine cracks and minor damages.It is suitable for deployment on industrial host computers and meets the engineering requirements for real-time detection of train wheel tread damage,verifying the effectiveness,practicality,and adaptability of the model in industrial scenarios.关键词
踏面损伤检测/YOLOv12/Edge_GAM/DWC_ECA/Focal_EIoUKey words
tread damage detection/YOLOv12/Edge_GAM/DWC_ECA/Focal_EIoU分类
信息技术与安全科学引用本文复制引用
杨钰鑫,宋冬利,张志扬,雷蕾..TreadDet-YOLOv12:一种基于改进YOLOv12的轮对踏面损伤检测模型[J].铁道科学与工程学报,2026,23(8):4077-4090,14.基金项目
国家自然科学基金资助项目(52472368) （52472368）
四川省重大科技专项(2023ZDZX0009) （2023ZDZX0009）
