基于特征注意匹配CYCLEGAN的高速列车轮对轴承数据均衡化方法  

Data equalization processing method for high-speed train wheelset bearings based on CYCLEGAN-FAM

作　　者：刘素艳 汪浩宁 马增强[1,2,3] 苑宗昊 LIU Suyan;WANG Haoning;MA Zengqiang;YUAN Zonghao(Hebei Province Intelligent Integration Technology and Equipment Collaborative Innovation Center,Shijiazhuang Railway University,Shijiazhuang 050043,China;School of Electrical and Electronic Engineering,Shijiazhuang Railway University,Shijiazhuang 050043,China;State Key Laboratory of Structural Mechanical Behavior and System Safety of Traffic Engineering,Shijiazhuang 050043,China;School of Transportation,Shijiazhuang Railway University,Shijiazhuang 050043,China)

机构地区：[1]石家庄铁道大学河北省交通电力网智能融合技术与装备协同创新中心,石家庄050043 [2]石家庄铁道大学电气与电子工程学院,石家庄050043 [3]省部共建交通工程结构力学行为与系统安全国家重点实验室,石家庄050043 [4]石家庄铁道大学交通运输学院,石家庄050043

出　　处：《振动与冲击》2024年第15期32-43,共12页Journal of Vibration and Shock

基　　金：国家自然科学基金(52205571,12072207);河北省自然科学基金(E2021210105)。

摘　　要：高速列车滚动轴承一旦发生故障就会停车检修,导致样本数据极度不平衡。数据集的不平衡性会对故障诊断结果的准确性和稳定性产生重要影响。针对该问题,提出一种基于特征注意匹配(feature attention matching, FAM)和循环生成对抗网络(cycle-consistent generative adversarial networks, CYCLEGAN)的轴承不平衡数据处理CYCLEGAN-FAM方法,该方法在CYCLEGAN的判别器中加入特征注意匹配模块,对从真实图像和生成图像中提取的特征进行对齐,从而提高生成样本的质量。试验表明,该方法能够生成与真实样本高度相似的生成样本,并随着不平衡数据集被逐渐平衡,故障诊断的准确率在凯斯西储大学4类和10类数据集上分别达到了99.8%和99.2%,在QPZZ-II四类和十类数据集上分别达到了99.4%和99.6%。Once rolling bearings of high-speed trains fail,they need to be stopped for maintenance to cause sample data being extremely unbalanced.The imbalance of data set can have a significant impact on the correctness and stability of fault diagnosis results.Here,aiming at this problem,a bearing imbalance data processing method called CYCLEGAN-FAM was proposed based on feature attention matching(FAM)and CYCLEGAN.This method could add a FAM module to the discriminator of CYCLEGAN to align features extracted from real images and those extracted from generated images,and improve quality of generated samples.Experiments showed that the proposed method can generate generated samples being highly similar to real samples;with imbalanced data set being gradually balanced,the correct rate of fault diagnosis can reach 99.8%and 99.2%,respectively on class 4 and class 10 data sets of CWRU,and 99.4%and 99.6%,respectively on class 4 and class 10 data sets of QPZZ-II.

关 键 词：生成对抗网络 特征注意力匹配(FAM) 不均衡数据集 故障诊断 

分 类 号：TH133.33[机械工程—机械制造及自动化]