基于tSNE多特征融合的JTC轨旁设备故障检测  

Fault detection of JTC trackside equipment based on tSNE multi-feature fusion

作　　者：武晓春[1] 郜文祥 WU Xiaochun;GAO Wenxiang(School of Automation and Electrical Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China)

机构地区：[1]兰州交通大学自动化与电气工程学院,甘肃兰州730070

出　　处：《铁道科学与工程学报》2024年第3期1244-1255,共12页Journal of Railway Science and Engineering

基　　金：中国国家铁路集团有限公司基金资助项目(N2022G012);国家自然科学基金地区项目(61661027);2023年甘肃省优秀研究生“创新之星”项目(2023CXZX-619)。

摘　　要：无绝缘轨道电路(Jointless Track Circuit,JTC)的轨旁设备在室外长期运营过程中,其可靠性会逐渐降低,进而给列车行车安全带来严重威胁。以轨道电路读取器(Track Circuit Reader,TCR)感应电压为基础,针对JTC故障诊断研究中轨旁设备故障类型复杂和故障特征提取不充分等问题,提出一种基于t分布随机邻域嵌入(t-distribution Stochastic Neighbor Embedding,tSNE)多特征融合的JTC轨旁设备故障检测模型。首先,根据不同轨旁设备故障对TCR感应电压信号的影响,分析各轨旁设备的故障特性。其次,提取TCR感应电压信号的方差、有效值、峰值因子等幅值域特征,以及排列熵、散布熵特征构成原始故障特征集。为了去除其中的冗余信息,得到具有较高判别性的融合流形特征,利用tSNE算法进行特征融合。最后输入深度残差网络(Deep Residual Network,DRN)得到故障检测混淆矩阵,实现轨旁设备故障定位。实验结果表明:tSNE算法融合后的特征在异类和同类故障样本之间分别有较大的类间间距和较小的类内间距,相比主成分分析(Principal Component Analysis, PCA)、随机相似性嵌入(Stochastic Proximity Embedding, SPE)、随机邻域嵌入(Stochastic Neighbor Embedding,SNE)算法具有更优的融合特征提取效果。此外,结合DRN可以有效识别多种轨旁设备故障,达到98.28%的故障检测准确率。通过现场信号进行实例验证,结果表明该故障检测模型能满足铁路现场对室外设备进行故障定位的实际需求。The reliability of the trackside equipment of the jointless track circuit(JTC)will gradually decrease during long-term outdoor operations,which poses a severe threat to the safety of train operations.Aiming at the problems of complex fault types and insufficient fault feature extraction of trackside equipment in JTC fault diagnosis research,a fault detection model of JTC trackside equipment based on t-distribution stochastic neighbor embedding(tSNE)multi-feature fusion was proposed.Firstly,according to the influence of different trackside equipment faults on track circuit reader(TCR)induced voltage signals,the fault characteristics of each trackside equipment were analyzed.Secondly,the amplitude domain features,such as the variance,root-mean-square,and peak factor of the TCR induced voltage signal,were extracted to form the original fault feature set with the permutation entropy and dispersion entropy.To remove the redundant information and obtain the fusion manifold features with high discrimination,the tSNE algorithm was used for feature fusion.Finally,the fault detection confusion matrix was obtained through the deep residual network(DRN)to realize the fault location of the trackside equipment.The experimental results show that the features fused by tSNE have larger inter-class distances between heterogeneous fault samples and smaller intra-class distances between homogeneous fault samples.Compared with principal component analysis(PCA),stochastic proximity embedding(SPE),and stochastic neighbor embedding(SNE)algorithms,tSNE has a better feature extraction effect.In addition,combined with DRN,it can effectively identify various trackside equipment faults,and the fault detection accuracy can reach 98.28%.The example verification results of field signals show that the proposed fault detection model can meet the actual needs of the railway field for fault location of outdoor equipment.

关 键 词：轨旁设备 幅值域 排列熵 散布熵 多特征融合 故障检测 

分 类 号：U284.2[交通运输工程—交通信息工程及控制]