基于机器学习的高速列车抗蛇行减振器劣化状态识别方法研究  

Deterioration State Identification of Yaw Damper Based on Machine Learning for High-speed Train

作　　者：魏庆 王悦明 吕凯凯 代明睿 杨涛存 杜文然 池长欣 WEI Qing;WANG Yueming;LYU Kaikai;DAI Mingrui;YANG Taocun;DU Wenran;CHI Changxin(Postgraduate Department,China Academy of Railway Sciences,Beijing 100081,China;Locomotive&Car Research Institute,China Academy of Railway Sciences Corporation Limited,Beijing 100081,China;Institute of Computing Technologies,China Academy of Railway Sciences Corporation Limited,Beijing 100081,China;Railway Science&Technology Research&Development Center,China Academy of Railway Sciences Corporation Limited,Beijing 100081,China)

机构地区：[1]中国铁道科学研究院研究生部,北京100081 [2]中国铁道科学研究院集团有限公司机车车辆研究所,北京100081 [3]中国铁道科学研究院集团有限公司电子计算技术研究所,北京100081 [4]中国铁道科学研究院集团有限公司铁道科学技术研究发展中心,北京100081

出　　处：《铁道机车车辆》2023年第6期45-53,共9页Railway Locomotive & Car

基　　金：中国国家铁路集团有限公司科技研究开发计划系统性重大项目(P2021J005);中国铁道科学研究院集团有限公司科技研究计划重点课题(2022YJ263)。

摘　　要：为识别高速列车抗蛇行减振器服役过程中的劣化状态,首先基于运用统计选取了5种典型的组合参数,通过台架试验获取其动态频变刚度和阻尼特性;然后采用抗蛇行减振器非参数化建模方法建立了整车动力学联合仿真模型,计算得到不同工况下的车辆动力学响应;构建了机器学习分类问题,分别采用支持向量机(Support Vector Machine, SVM)和卷积神经网络(Convolutional Neural Network, CNN)的方法对减振器状态进行识别。研究结果表明,基于BP神经网络(Back Propagation Neural Network,BPNN)的非参数化模型更为准确地描述抗蛇行减振器的动态行为,建立的整车联合仿真模型计算结果与实测数据符合较好;采用SVM算法构建的机器学习模型识别效果一般,而采用CNN算法构建的机器学习模型则达到较高的识别准确度。考虑实际运用需求,将机器学习问题简化为6分类问题,信号通道数精简为4个,CNN机器学习模型仍可实现较高精度的劣化状态识别。In order to identify the deterioration state of yaw damper of high-speed trains in service,five typical dampers are selected based on statistics,and their dynamic stiffness and damping characteristics.Then,a vehicle dynamic model considering non-parametric yaw damper is established,and the vehicle dynamic responses under the condition of each degraded damper are calculated.Machine leaning methods based on SVM and CNN are used respectively to identify the state of degraded damper.The results show that the non-parametric model based on BP neural network describe the actual dynamic behavior of the damper accurately.The calculated results of vehicle dynamics model are in good agreement with the measured data.The identification accuracy of degraded damper based on the SVM model shows relatively low effect.And the recognition model constructed by CNN achieves high recognition accuracy.Considering the practical application requirements,the 15 classification problems are simplified to form 6 classification problems,including 2 damper states and 3 wheel tread states,and the number of channels is also reduced to 4,the CNN model can still achieve high-precision degradation state identification.

关 键 词：抗蛇行减振器 非参数化建模 动力学响应 劣化识别 支持向量机 卷积神经网络 

分 类 号：U279.3[机械工程—车辆工程]