摘要
铁道车辆车轮故障的产生,不仅会增大列车的振动和噪声使乘坐舒适性下降,而且会加速车辆及轨道零部件的损伤,严重时还会引发事故,因此对车轮服役状态的实时监测对保证列车安全运营具有重要意义。针对现有铁道车辆车轮故障诊断方法存在自适应能力弱、准确率低等不足,提出一种基于多尺度时频图与卷积神经网络(CNN)相结合的车轮故障智能诊断方法,该方法利用车轮所在轴箱垂向振动加速度来间接识别车轮服役状态。1)首先采用形态学滤波器对车辆轴箱振动加速度信号进行滤波降噪,然后采用完全噪声辅助聚合经验模态分解(CEEMDAN)将滤波后的信号自适应地分解为若干固有模态函数(IMF),选取能量熵增量相对较大的三阶分量作为信号的主分量。2)分别求各主分量的Wigner-Ville分布(WVD),然后叠加转化为多尺度时频图。3)对经典的LeNet-5模型进行结构改进和网络参数优化,构建适合车轮故障诊断的CNN模型,来学习提取车轮在不同工况下的时频图特征,并对时频图进行分类,将特征学习提取与故障分类融为一体,一定程度上实现了端到端的车轮故障诊断。经仿真试验和现场试验验证表明:所提出的方法对于车速、故障类型和故障程度都有很好的自适应能力,故障识别准确率可达97%,且泛化能力强。因此所提方法在车辆运营状态在线监测的应用中具有一定的理论意义和工程价值。
The generation of wheel faults in railway vehicles will not only increase the vibration and noise of the train to make the ride comfortable,but also accelerate the damage of the vehicle and track parts.In the serious cases,it will also cause accidents.So the real-time monitoring of wheel service status is of great significance to ensure the safe operation of trains.Aiming at the shortcomings of the existing railway vehicle wheel fault diagnosis methods such as weak adaptive capability and low accuracy,an intelligent wheel fault diagnosis method based on multi-scale time-frequency map and convolutional neural network(CNN) was proposed,which used the vertical vibration acceleration of the axle box where the wheel was located to indirectly identify the wheel service status.(1) Firstly,a morphological filter was used to filter and reduce the noise of the vehicle axlebox vibration acceleration signal,and then the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise(CEEMDAN) was used to adaptively decompose the filtered signal into several intrinsic mode functions(IMFs).The third-order component with relatively large energy entropy increment was selected as the main component of the signal.(2) The Wigner-Ville distribution(WVD) of each principal component was found separately,and then superimposed and transformed into a multi-scale time-frequency map.(3) The classical LeNet-5 model was structurally improved,and the network parameters were optimized to build a CNN model suitable for wheel fault diagnosis to learn to extract the time-frequency diagram features of wheels under different operating conditions and classify the time-frequency diagram,which integrated feature learning and extraction with fault classification and achieved end-to-end wheel fault diagnosis to a certain extent.The simulation tests and field trials show that the proposed method has good adaptive capability for vehicle speed,fault type and fault degree,and the fault recognition accuracy can reach 97% with strong generalization capabili
作者
李大柱
牛江
梁树林
池茂儒
LI Dazhu;NIU Jiang;LIANG Shuling;CHI Maoru(State Key Laboratory of Traction Power,Southwest Jiaotong University,Chengdu 610031,China)
出处
《铁道科学与工程学报》
EI
CAS
CSCD
北大核心
2023年第3期1032-1043,共12页
Journal of Railway Science and Engineering
基金
国家自然科学基金资助项目(U21A20168)。
