基于深度残差收缩网络的地铁车轮扁疤故障诊断  

Fault diagnosis method for metro wheel flat based on deep residual shrinkage network

作　　者：梁红琴[1] 姜进南 龙辉 陶功权[2] 卢纯 温泽峰[2] 张楷 LIANG Hongqin;JIANG Jinnan;LONG Hui;TAO Gongquan;LU Chun;WEN Zefeng;ZHANG Kai(School of Mechanical Engineering,Southwest Jiaotong University,Chengdu 610031,China;State Key Laboratory of Rail Transit Vehicle System,Southwest Jiaotong University,Chengdu 610031,China)

机构地区：[1]西南交通大学机械工程学院,四川成都610031 [2]西南交通大学轨道交通运载系统全国重点实验室,四川成都610031

出　　处：《中南大学学报(自然科学版)》2025年第3期1234-1248,共15页Journal of Central South University:Science and Technology

基　　金：国家自然科学基金资助项目(52002342,52205130);四川省自然科学基金资助项目(2022NSFSC1914);四川省科技计划资助项目(2023YFQ0091)。

摘　　要：针对地铁实际运营环境恶劣的问题,本文以轴箱振动加速度作为监测信号,基于深度残差收缩网络(DRSN),提出1种适用于强噪声背景的车轮扁疤故障严重程度辨识方法。首先,基于地铁车辆-轨道刚柔耦合动力学模型生成车轮扁疤故障数据集,并采用数据增强技术提升数据集的多样性,同时满足深度学习对数据规模的要求。其次,设计1种结构合理的深度残差收缩网络,能够自适应地提取轴箱振动加速度信号的特征,从而实现车轮扁疤故障程度的智能分类。研究结果表明:在无噪声条件下,所提方法对正常车轮及轻度、中度和重度扁疤车轮的平均诊断精度达到99.88%(标准差为0.05);同时,在不同噪声等级下,该方法的平均诊断精度仍稳定保持在95%以上。与遗传算法结合支持向量机(GA-SVM)、卷积神经网络(CNN)、宽深度卷积神经网络(WDCNN)以及深度残差网络(ResNet)相比,所提方法具有更优异的辨识能力和鲁棒性。To tackle the harsh operating conditions of metro systems,a method for identifying the severity of wheel flats under high-noise interference was proposed,utilizing a deep residual shrinkage network(DRSN)with axle box vibration acceleration as the monitoring signal.Firstly,a dataset of wheel flats was generated using a vehicle-track rigid-flexible coupled dynamics model,and data augmentation techniques were employed to enhance the dataset's diversity,meeting the data volume requirements for deep learning.Nextly,a properly structured DRSN was designed to adaptively extract features from the axle box vibration acceleration signals,enabling intelligent classification of wheel flat severity.The results show that the proposed method achieves an average diagnostic accuracy of 99.88%(with a standard deviation of 0.05)for four wheel conditions:normal wheels,mild flats,moderate flats,and severe flats,under noise-free conditions.Furthermore,the method maintains an average diagnostic accuracy of over 95% at various noise levels.Compared to approaches such as genetic algorithmsupported vector machines(GA-SVM),convolutional neural networks(CNN),wide deep convolutional neural networks(WDCNN),and deep residual networks(ResNet),the proposed method demonstrates superior identification capabilities and robustness.

关 键 词：车辆-轨道耦合动力学模型 车轮扁疤 深度残差收缩网络 轴箱振动加速度 数据增强 

分 类 号：U211.5[交通运输工程—道路与铁道工程]