基于GIS空间分析及深度学习的调车场安全识别系统  

Safety identification and recognition system of shunting yard based on the spatial analysis of GIS and deep learning

作　　者：郑伟皓 周星宇 李红梅 吴虹坪[1,2] 唐智慧 ZHENG Wei-hao;ZHOU Xing-yu;LI Hong-mei;WU Hong-ping;TANG Zhi-hui(College of Transportation&Logistics,Southwest Jiaotong University,Chengdu 610031,China;National Engineering Laboratory of Integrated Transportation Big Data Application Technology,Southwest Jiaotong University,Chengdu 610031,China)

机构地区：[1]西南交通大学交通运输与物流学院,成都610031 [2]西南交通大学综合交通大数据应用技术国家工程实验室,成都610031

出　　处：《安全与环境学报》2020年第2期423-432,共10页Journal of Safety and Environment

摘　　要：为实现调车场的智能化安全管理,首先基于调车作业的常见事故,分析了调车场系统中的人-机-环事故致因要素。利用GIS对调车场各空间组成要素进行建模,以获得调车场的电子地图与各组成要素的属性数据库。在此基础上,利用Arc GIS空间分析引擎对调车场对象的属性数据进行操纵分析,得出电路分路情况、道岔尖轨间隙特征、人员作业热点、货物空间分布等特征图层;并利用卷积神经网络(CNN)作为融合工具,以专家判定的危险区域为学习标签,大规模自动化提取各特征图层的特征,以达到自主学习专家知识的目的;最后,利用Ecd转化工具将训练的CNN识别引擎转化为栅格分类器,供GIS系统进行安全分类评估时调用,最终达到调车场自动安全分析的目的。经仿真数据验证,结果表明:在CNN的训练过程中,Dropout机制可有效控制过拟合现象,经训练的CNN在测试集的识别正确率为73. 81%;GIS系统能有效调用CNN,实现自动融合电子地图中的对象属性数据,完成调车场安全评估及展示。The given paper intends to make an analysis of the causes of the accidents associated with the human-machine-environment in the shunting yard system so as to realize its intelligent safety management based on the common accidents in the operation process in such a human-machine interactive special environment.For such a purpose,we would like to obtain an electronic distributive map of the shunting yard and the attribute database of each spatial element document through the GIS(short for the geographic information system)software to model the spatial elements of the shunting yard.Then,on the basis above,it would be possible for us to manipulate and analyze the attributive data of the objects in the yard through the tool of the spatial analysis engine in the Arc GIS system.Next,it is also possible for us to gain the feature layers,such as the situation of the rail circuit shunts,the particular features of the gap of the turnout junction,the hotspot map of the personnel work and spatial distribution of the cargoes through the steps mentioned above.And,consequently,the paper can determine the convolution neural network(CNN)as the fusion tool and the dangerous areas taken by the experts as the labels of learning.Thus,the purpose of learning of the expert knowledge can automatically be gained by extracting the features of each layer in a large scale.In addition,this paper has also managed to convert the recognition engine of the trained CNN into a raster classifier by using the conversion tool known as the raster classifier in case to use the GIS system for the safety classification assessment.And,finally,we have managed to achieve the purpose of the automatic safety analysis of the shunting yard actuality.The simulation results show that the dropout mechanism can effectively control the over-fitting phenomenon during the training process of CNN,and the correct rate of recognition of the trained CNN in the test set can only reach 73.81%.Nevertheless,the GIS system can effectively call the CNN system to integrate the da

关 键 词：安全工程 交通地理信息系统 深度学习 数据融合 

分 类 号：X912.9[环境科学与工程—安全科学]