引入改进YOLOv4算法的铁路轨道形变智能检测技术研究  

Research on intelligent detection technology of railroad track deformation by introducing improved YOLOv4 algorithm

作　　者：王芳 WANG Fang(China Railwany First Survey And Design Institute Group Co.,Ltd.,Lanzhouu 730000,China)

机构地区：[1]兰州铁道设计院有限公司,兰州730070

出　　处：《自动化与仪器仪表》2025年第1期143-147,共5页Automation & Instrumentation

基　　金：自然科学基金《垂直冻结技术在大埋深山岭铁路隧道中的应用研究》(18JR3RA014)。

摘　　要：随着国家城市化建设的加快,铁路运输在城市化发展中扮演着重要的角色,为了防止铁轨形变带来不必要损失。研究针对这些问题,构建了改进YOLOv4算法的铁轨形变检测模型。首先在YOLOv4中构建模型,然后在此基础上进行优化,最后利用数据集训练。将模型与卷积神经网络(Convolutional Neural Networks,CNN)、深度置信网络(Deep Belief Nets,DBN)进行误差对比时,模型方法的误差在[0.018,-0.019]之间,准确度最高。同时利用YOLOv4和MobileNet三代改进的YOLOv4模型去验证所选用模型的可靠性。其中V3-YOLOv4的F1平均值为64.12%;迭代训练中每次迭代耗时为前五十次3 min46 s,后五十次4 min16 s,在四种模型中均为最佳。这说明研究提出的模型在轨道形变智能检测中具有较高准确性的同时,还能够提高检测的效率。为铁轨形变检测提供了一种新的思路。With the accelerated urbanization of the country,railroad transportation plays an important role in the development of urbanization,in order to prevent unnecessary losses caused by rail deformation.The study addresses these problems and constructs a rail deformation detection model with improved YOLOv4 algorithm.The model is first constructed in YOLOv4,then optimized on this basis,and finally trained using the dataset.In the results,when the model is compared with Convolutional Neural Networks(CNN)and Deep Belief Nets(DBN)in terms of error,the error of the model method is between[0.018,-0.019]with the highest accuracy.YOLOv4 and MobileNet three generations of improved YOLOv4 models were also used to verify the reliability of the selected models.Among them,the mean F1 of V3-YOLOv4 is 64.12%;the elapsed time of each iteration in the iterative training is 3 min46s for the first fifty iterations and 4 min16s for the last fifty iterations,which is the best among all four models.This indicates that the model proposed in the study has high accuracy in the intelligent detection of rail deformation while improving the efficiency of detection.It provides a new way of thinking for rail deformation detection.

关 键 词：改进YOLOv4算法 铁路 轨道形变 智能检测 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]