基于轻量化YOLOv8s的轨道扣件状态检测方法  

Lightweight detection method for track fastener status based on improved YOLOv8s

作　　者：武福[1] 蒋鹏民 李忠学[1] 杨喜娟[2] 吕金旺 WU Fu;JIANG Pengmin;LI Zhongxue;YANG Xijuan;LYU Jinwang(School of Mechanical Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China;School of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China)

机构地区：[1]兰州交通大学机电工程学院,兰州730070 [2]兰州交通大学电子与信息工程学院,兰州730070

出　　处：《北京交通大学学报》2024年第5期59-68,共10页JOURNAL OF BEIJING JIAOTONG UNIVERSITY

基　　金：国家自然科学基金(56062028);甘肃省教育厅产业支撑计划项目(2021CYZC-11);甘肃省教育厅创新基金项目(2022A-036)。

摘　　要：铁路基础结构设施长期遭受车辆荷载以及外在环境因素影响,沿线的轨道扣件容易产生弹条丢失、偏移、损坏等问题,严重威胁轨道线路的安全运营.针对目前依赖人工目测和采样抽检等主观性强的检修方式所导致的检测效率低、漏检率高以及在边缘设备上无法实时检测的问题,提出一种基于YOLOv8s的轻量化轨道扣件状态检测模型FTEL-YOLO,旨在提高检测准确率和实时性.首先,参考FasterNet-Block的思想设计C2f-Faster模块以减小模型参数量;然后,为解决网络轻量化导致的模型检测精度下降的问题,在空间金字塔池化(Spatial Pyramid Pooling Fast,SPPF)模块之后引入三元注意力机制(Triplet Attention),并引用EIoU作为边界框回归损失函数来提升对复杂背景下轨道扣件不同状态的特征提取能力;最后,对改进后的模型进行基于层自适应幅度的剪枝(Layer-Adaptive Magnitude-based Pruning,LAMP)操作,进一步压缩模型以减小冗余,提高其在边缘设备上的应用能力.实验结果表明:改进后的模型FTEL-YOLO检测精度仅损失0.3%,但计算量、参数量和模型大小分别下降63.1%、65.6%和66.2%,在保持准确性的同时实现了轻量化.Railway infrastructure is continuously impacted by vehicle loads and external environmental factors,causing issues such as the loss,displacement,and damage of track fasteners along railway lines.These problems pose significant threats to the safe operation of railways.To address the low detection efficiency,high omission rates,and lack of real-time detection capabilities on edge devices associated with traditional manual visual inspections and subjective sampling methods,this paper proposes a lightweight detection model for track fastener status,FTEL-YOLO,based on YOLOv8s.The model is designed to enhance detection accuracy and real-time performance.First,the C2f-Faster module,inspired by the FasterNet-Block concept,is introduced to reduce the model’s parameters.Second,to mitigate the decline in detection accuracy caused by network lightweighting,a Triplet Attention Mechanism is incorporated after the Spatial Pyramid Pooling Fast(SPPF)layer,and EIoU is utilized as the bounding box regression loss function,enhancing the model’s feature extraction capability for track fastener conditions in complex backgrounds.Finally,Layer-Adaptive Magnitude-based Pruning(LAMP)is applied to the improved model to further compress it,reducing redundancy and enhancing its deployment capability on edge devices.Experimental results demonstrate that the improved FTEL-YOLO model achieves a minimal detection accuracy loss of 0.3%,while the computation,parameters,and model size are reduced by 63.1%,65.6%,and 66.2%,respectively,achieving lightweight design without compromising accuracy.

关 键 词：深度学习 故障检测 轨道扣件 YOLOv8s 三元注意力机制 模型轻量化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] U216.3[自动化与计算机技术—计算机科学与技术]