摘要
转辙机是铁路系统中控制道岔转换的关键信号设备,准确诊断其工作状态对列车的行车安全和可靠至关重要。鉴于现场数据常缺乏标签且故障样本少,提出一种基于自监督表示学习的转辙机故障诊断方法。该方法首先将转辙机动作曲线转换为图像形式,随后通过自编码器提取数据的潜在特征。然后,基于同故障类型转辙机数据具有内在相似性,设计了一种表示学习模型,该模型通过比较批次数据间的相似关系来监督训练过程,实现高相似性数据的表示分布聚合,从而挖掘数据的潜在分类结构。此外,得益于对比学习的泛化优势,该方法能够利用不同类型转辙机的数据来增强模型训练效果。最终,通过聚类挖掘表示特征中的故障模式,并训练下游分类网络分类表示特征,从而实现不依赖人工标记数据的故障诊断模型。研究结果表明,相较于传统自编码器,对比表示学习模型可以更有效地区分不同故障类型的转辙机监测数据。在无标签训练模型的情况下,采用ZDJ9型转辙机现场数据训练模型后,在测试数据集上故障诊断准确率为99.63%,融合ZDJ9型和ZYJ7型转辙机现场数据训练模型后,故障诊断准确率提升到99.88%,比传统无监督学习模型准确率提升了8个百分点,并与监督学习模型性能相当。该方法结合了无监督模型不依赖人工标记数据和监督模型故障诊断准确率高的优势,为铁路现场转辙机故障诊断提供了一种新的可行方案。
Switch machines are key signaling devices in railway turnout systems,essential for the control of track switching,where accurate diagnosis of their operational status is crucial for the safety and reliability of train operation.This study proposed a novel self-supervised fault diagnosis method of switch machines based on contrastive learning of representations,addressing the challenges associated with the lack of labeled data and the rarity of fault samples in field data.This approach initially transforms the action curves of switch machines into an image format,allowing better extraction of latent features using an autoencoder.A representation learning model is then devised,leveraging the inherent similarities among data of switch machines with the same fault type.This model supervises the training process by comparing similarities across batches of data,thereby clustering representations with high similarity and uncovering the latent structural classification of the data.Furthermore,the generalization advantage of contrastive learning allows for using data from various types of switch machines to enhance the training effect.Ultimately,a downstream classification network is trained to classify these features,facilitating a fault diagnosis model independent of manually labeled data.Experimental findings suggest that,compared to traditional autoencoders,the contrastive representation learning model more effectively differentiates between different fault types in switch machine monitoring data.When trained on unlabeled field data from ZDJ9-type switch machines,the model achieves a fault diagnosis accuracy of 99.63%on the test dataset;this accuracy increases to 99.88%when data from both ZDJ9-type and ZYJ7-type switch machines are combined for training.This represents an eight percentage point improvement over traditional unsupervised learning models and is comparable to the performance of supervised models.By merging the benefits of unsupervised models not relying on manual labels and the high fault diagnosis accuracy of supervised models,this method presents a new feasible solution for fault diagnosis of switch machines in the railway field.
作者
郑启明
王小敏
江磊
ZHENG Qiming;WANG Xiaomin;JIANG Lei(School of Information Science and Technology,Southwest Jiaotong University,Chengdu 611756,China;Sichuan Provincial Engineering Research Center of Train Operation Control Technology,Chengdu 611756,China)
出处
《铁道科学与工程学报》
北大核心
2025年第1期404-415,共12页
Journal of Railway Science and Engineering
基金
中国国家铁路集团有限公司科技研发计划(P2021G053,N2022G010)
敏捷智能计算四川省重点实验室开放式基金资助项目。
关键词
转辙机
故障诊断
自监督
对比学习
表示学习
switch machine
fault diagnosis
self-supervised
contrastive learning
representations learning
作者简介
通信作者:王小敏(1974-),男,江西萍乡人,教授,博士,从事轨道交通智能运维研究,E-mail:xmwang@swjtu.edu.cn。
