摘要
传统的轨道检测方法需要事先对图像进行定位和分割等预处理操作,而定位和分割操作的误差又会直接干扰到后续的分类识别,多环节误差叠加,使得识别准确率低。同时,传统检测方法还需要理想的背景环境,当背景环境或结构类型发生改变时,其算法不再适用,不具备良好的鲁棒性。因此,提出一种基于深度残差网络的轨道结构病害检测方法,该方法不需要对原始图像进行预处理,同时深度残差网络以其更深的层数和更复杂的网络结构可以高效提取出各类轨道结构图像的特征并进行分类识别。以某客货共线线路隧道的钢轨踏面、钢轨扣件和支承块图像建立数据库,通过迁移学习的方式在数据库上训练网络模型,实现对钢轨、扣件及支承块三种轨道结构的病害识别,识别准确率高达98.51%。在此基础上,从识别准确率、损失函数值等方面对深度残差网络在轨道结构病害识别中的应用效果进行对比、分析,验证方法的有效性。
The traditional method of track detection requires pre-processing of images such as image location and segmentation,and the errors of the location and segmentation will directly interfere with the subsequent classification recognition.The superposition of the errors resulted from different pre-processing steps causes low recognition accuracy.At the same time,the traditional detection method also needs an ideal background environment.When the background environment or structure type changes,the algorithm is no longer applicable nor does it show good robustness.Therefore,this paper presented a method for detecting track structure diseases based on deep residual network(ResNet),which does not require any preprocessing of the original image.Meanwhile,the ResNet can efficiently extract,classify and recognize the features of various track structure images with its deeper layers and more complex network structures.A database was established based on the images of rail tread,rail fastener and supporting block of a tunnel with a certain passenger and cargo line.The network model was trained on the database through transfer learning to realize the disease identification of three track structures of rail,fastener and supporting block,with recognition accuracy up to 98.51%.On this basis,the application effect of the deep residual network in the identification of track structure disease was analyzed and evaluated from the recognition accuracy and loss function value,to verify the validity of the method.
作者
侯博文
杨晓
高亮
肖宏
马帅
HOU Bowen;YANG Xiao;GAO Liang;XIAO Hong;MA Shuai(School of Civil Engineering,Beijing Jiaotong University,Beijing 100044,China;Beijing Key Laboratory of Track Engineering,Beijing 100044,China)
出处
《铁道学报》
EI
CAS
CSCD
北大核心
2020年第8期100-106,共7页
Journal of the China Railway Society
基金
国家重点研发计划(2016YFB1200402)
中央高校基本科研业务费(2018JBM042)
北京市科技计划(Z161100001016004)
京沪高速铁路股份有限公司科研计划(京沪科研-2016-10)
中国铁路总公司科技研究开发计划(2016G009-B)。
关键词
轨道结构
病害识别
深度残差网络
迁移学习
track structure
disease identification
deep residual network
transfer learning
