基于GADF与2D CNN-改进SVM的道岔故障诊断方法研究  

Turnout fault diagnosis method based on GADF and 2D CNN-improved SVM

作　　者：王彦快[1] 孟佳东[2] 张玉 杨建刚 王贵强 WANG Yankuai;MENG Jiadong;ZHANG Yu;YANG Jiangang;WANG Guiqiang(Institute of Railway Technology,Lanzhou Jiaotong University,Lanzhou 730000,China;School of Automation&Electrical Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China;Lanzhou Longneng Power Technology Co.,Ltd.,Lanzhou 730070,China;CRSC Research&Design Institute Group Co.,Ltd.,Beijing 100070,China)

机构地区：[1]兰州交通大学铁道技术学院,甘肃兰州730000 [2]兰州交通大学机电工程学院,甘肃兰州730070 [3]兰州陇能电力科技有限公司,甘肃兰州730070 [4]北京全路通信信号研究设计院集团有限公司,北京100070

出　　处：《铁道科学与工程学报》2024年第7期2944-2956,共13页Journal of Railway Science and Engineering

基　　金：甘肃省科技计划项目(21JR7RA305,23JRRA850);兰州交通大学青年科学研究基金资助项目(1200061027)。

摘　　要：针对道岔故障特征不易提取以及道岔故障诊断准确率较低的问题,提出一种格拉姆角差场(Gramian Angular Difference Fields, GADF)与二维卷积神经网络(Two Dimensional Convolutional Neural Network, 2D CNN)-改进支持向量机(Support Vector Machine, SVM)的道岔故障诊断组合方法。首先,结合现场实际应用情况,选取道岔设备正常转换与典型故障的转辙机功率曲线,建立转辙机功率曲线样本数据库;采用GADF编码将一维转辙机功率曲线信号转换为具有时间相关性的二维特征图,分别选择16×16、32×32以及64×64大小的特征图并提取图像数据。其次,在LeNet-5模型的基础上设计2D CNN网络结构,并将图像数据输入至基于2D CNN的道岔故障特征提取模型中,经多层的卷积层、池化层以及全连接层提取特征指标,建立道岔故障诊断样本数据库。最后,通过北方苍鹰优化(Northern Goshawk Optimization, NGO)算法优化SVM算法的惩罚因子与核函数方差,构建基于NGO-SVM的道岔故障诊断模型。实验结果分析表明,将转辙机功率曲线数据经GADF编码为64×64大小的特征图,并通过2D CNN模型提取道岔典型特征数据,较其他数据处理方法具有较高的故障诊断准确率,同时提高了故障诊断实时性;将建立的道岔故障诊断样本数据库输入至NGO-SVM道岔故障诊断模型,其故障诊断准确率高达97.5%,较其他故障诊断模型具有更好的故障诊断性能,为道岔故障诊断提供了一种新方法,对现场道岔设备的日常维修具有一定的指导意义。Aiming at the problem that the fault characteristics of turnout were not easy to extract and the accuracy rate of turnout fault diagnosis was low,a combination method of Gramian Angular Difference Fields(GADF)and two Dimensional Convolutional Neural Network(2D CNN)-improved Support Vector Machine(SVM)for turnout fault diagnosis was proposed.Firstly,combined with the actual application situation on site,the switch machine power curve of normal conversion and typical fault of turnout equipment was selected.The sample database of switch machine power curve was established.The GADF coding method was used to convert the one-dimensional switch machine power curve signal into a two-dimensional feature map with time correlation.The feature maps of 16×16,32×32 and 64×64 were selected respectively and the image data was extracted.Secondly,based on the LeNet-5 model,a 2D CNN network structure model was designed.The image data was input into the turnout fault feature extraction model based on 2D CNN.The feature indicators were extracted through the multi-layer convolution layer,pooling layer and full connection layer to establish the turnout fault diagnosis sample database.The experimental results show that the power curve data of the switch machine is converted into a 64×64 feature map by GADF coding,and the typical feature data of the turnout is extracted by 2D CNN model.Compared with other data processing methods,it has higher fault diagnosis accuracy and improves the real-time performance of fault diagnosis.The established turnout fault diagnosis sample database is input into the NGO-SVM turnout fault diagnosis model.The fault diagnosis accuracy is as high as 97.5%,which has better fault diagnosis performance than other fault diagnosis models.It can provide a new method for turnout fault diagnosis and has certain guiding significance for the daily maintenance of on-site turnout equipment.

关 键 词：道岔设备 故障诊断 GADF 2D CNN NGO-SVM 

分 类 号：U284.92[交通运输工程—交通信息工程及控制]