基于数据特征识别的接触网补偿装置和中心锚结状态异常诊断方法  

A Method for Diagnosing the Abnormal State of the OCS Compensation Device and Mid-Point Anchor Based on Data Feature Recognition

作　　者：王斌 张文轩 王婧 杨志鹏 姚永明 王文昊 WANG Bin;ZHANG Wenxuan;WANG Jing;YANG Zhipeng;YAO Yongming;WANG Wenhao(Infrastructure Inspection Research Institute,China Academy of Railway Sciences Corporation Limited,Beijing 100081,China;Railway Infrastructure Inspection Center,China State Railway Group Co.,Ltd.,Beijing 100081,China)

机构地区：[1]中国铁道科学研究院集团有限公司基础设施检测研究所,北京100081 [2]中国国家铁路集团有限公司铁路基础设施检测中心,北京100081

出　　处：《中国铁道科学》2024年第6期11-21,共11页China Railway Science

基　　金：中国国家铁路集团有限公司科技研究开发计划课题(J2022G014);中国铁道科学研究院集团有限公司院基金课题(2023YJ098)。

摘　　要：常用的人工巡视、图像识别、定点监测等手段难以对接触网补偿装置和中心锚结状态进行全面、精准地评估,因此提出基于数据特征识别的状态异常诊断方法。首先,归纳补偿装置和中心锚结状态异常时的2类典型数据特征,定义描述中心锚结区域接触线高度突变的3项特征量、锚段左右2侧接触线高度差异的3项特征量以及中心锚结区域弓网接触力突变的3项特征量;然后,依托实测数据构建样本库,采用离群点检测方法检查样本可靠性,采用主成分分析法对强相关特征量进行规约,形成训练样本库;最后,采用随机森林算法训练分类模型,并通过总体准确率、假正率和真负率对分类模型进行评估。通过对该方法的特征量和分类模型及现场试用进行验证,结果表明:各特征量可分性良好,验证了归纳数据特征的可靠性和普遍性;随机森林算法在2类样本训练中均有95%以上的分类准确率;采用该方法的诊断结果现场复核准确性高,在补偿装置和中心锚结状态异常诊断中可行性较强。Common methods,such as manual inspection,image recognition,and fixed-point monitoring,struggle to comprehensively and accurately evaluate the status of the Overhead Contact System(OCS)compensation device and mid-point anchor.Therefore,a state anomaly diagnosis method based on data feature recognition is proposed.Firstly,two types of typical data features are summarized during abnormal conditions of both the compensation device and the mid-point anchor.Specifically,three feature quantities are defined to describe the Overhead Contact Line(OCL)height mutation in the mid-point anchor area,three to indicate the difference in OCL height on both sides of the anchor section,and three to describe the pantograph-catenary contact force mutation in the mid-point anchor area.Secondly,a sample library is constructed based on a large amount of detected data.Outlier diagnosis methods are employed to check the reliability of the samples,and Principal Component Analysis(PCA)is applied to reduce strongly correlated feature quantities,thereby forming a training sample library.Finally,a classification model is trained using Random Forest(RF),and the classification model is evaluated using overall accuracy,false positive rate,and true negative rate.The feature quantities,classification model,and field trial verification of the proposed method are conducted.The results indicate that each feature quantity exhibits good separability,verifying the reliability and universality of the summarized data features.RF achieves a classification accuracy of over 95%in both types of sample training.The diagnosis results obtained using the proposed method are highly accurate in field verification,demonstrating strong feasibility in diagnosing abnormal conditions of both the compensation device and the mid-point anchor.

关 键 词：故障诊断 接触网 中心锚结 补偿装置 特征建模 机器学习 

分 类 号：U225.4[交通运输工程—道路与铁道工程]