基于改进YOLO11n的地铁车辆轮对踏面损伤检测

Detection of Wheelset Tread Defects in Metro Vehicles Based on Improved YOLO11n

摘要:

针对地铁车辆轮对踏面易出现擦伤、剥离及腐蚀等损伤且图像存在特征冗余大、背景干扰强及多尺度检测困难等问题，提出一种基于改进YOLO11n的地铁车辆轮对踏面损伤检测方法。在主干网络后段引入ScConv模块抑制空间与通道冗余，在特征融合层嵌入GCSA模块增强多尺度特征表达，并在回归阶段采用LWIoUv3损失函数优化定位性能，从而提升复杂场景下的检测能力。以长沙地铁6号线车辆段采集的1438张轮对踏面图像为实验对象，相较基线YOLO11n，所提方法的Precision、F1-score、mAP@0.5和mAP@0.5-0.95分别提升5.99%、2.95%、1.65%和1.32%。结果表明，所提方法能够在保持轻量级规模和较好实时性的同时提升地铁轮对踏面损伤检测精度，可为轨旁智能检测系统提供技术参考。

Abstract:

Surface defects such as scratches, spalling, and corrosion are prone to occur on the wheelset treads of metro vehicles. In addition, tread images are characterized by considerable feature redundancy, strong background interference, and difficulty in multi-scale detection. To address these problems, an improved YOLO11n-based method is proposed for wheelset tread defect detection in metro vehicles. A Spatial and Channel Reconstruction Convolution (ScConv) module is introduced into the later stage of the backbone network to suppress spatial and channel redundancy. A Global Channel-Spatial Attention (GCSA) module is embedded in the feature fusion layer to enhance multi-scale feature representation. The LWIoUv3 loss function is adopted in the regression stage to improve localization performance, thereby enhancing detection capability in complex scenarios. Experiments were conducted on 1,438 wheelset tread images collected from the Changsha Metro Line 6 depot. Compared with the baseline YOLO11n, the proposed method improves Precision, F1-score, mAP@0.5, and mAP@0.5–0.95 by 5.99%, 2.95%, 1.65%, and 1.32%, respectively. The results show that the proposed method improves the detection accuracy of metro vehicle wheelset tread defects while maintaining a lightweight model size and good real-time performance, and can provide technical support for trackside intelligent inspection systems.