摘要
由于轴承带标签的故障样本数量较少,且源域数据与目标域数据存在异域问题,会导致轴承诊断准确率大大下降。为此,对异源域样本条件下的轴承故障诊断问题进行了研究,提出了基于改进均衡分布适配迁移学习的轴承故障迭代诊断方法。首先,分析了滚动轴承的结构和不同部位故障的信号特征;介绍了迁移学习工作原理,基于动态的均衡因子,提出了改进均衡分布适配方法,解决了边缘分布和条件分布差异性未知导致的异源域适配难题;然后,给出了基于K近邻算法(KNN)的伪标签初步确定方法,提出了基于迁移学习和KNN算法的目标域伪标签迭代优化方法,确定了目标域样本的故障标签;最后,采用实验数据对该诊断方法的有效性进行了验证,并将其与其他两种方法进行了异域样本的故障诊断,对其诊断准确率进行了对比。研究结果表明:在凯斯西储轴承实验中,基于迁移学习、迁移成分分析(TCA)+KNN的诊断准确率均值分别为93.72%和75.52%;在西安交通大学轴承实验中,基于迁移学习、TCA+KNN的诊断准确率分别为94.80%和70.40%。上述实验结果验证了基于迁移学习的迭代诊断方法在异源域样本故障诊断中的优越性。
The number of bearing fault samples with labels is small,and there are exotic problems between the source domain data and the target domain data,which will greatly reduce the accuracy of bearing diagnosis.Therefore,the problem of bearing fault diagnosis under the condition of different source samples was studied,an iterative bearing fault diagnosis method based on improved equilibrium distribution adaptive transfer learning was proposed.Firstly,the structure of rolling bearing and the signal characteristics of faults in different parts were analyzed in detail.The working principle of transfer learning was introduced.Based on the dynamic equilibrium factor,an improved equilibrium distribution adaptation method was proposed to solve the problem of heterogeneous domain adaptation caused by the unknown difference between edge distribution and conditional distribution.Then,a pseudo-label iterative optimization method of target domain based on transfer learning and K-nearest neighbor(KNN)algorithm was proposed,and the fault labels of target domain samples were finally determined based on iterative optimization method between KNN algorithm and migration learning.Finally,the effectiveness of the diagnosis method was verified by experimental data.It was compared with other two methods to diagnose the faults of foreign samples,and the diagnostic accuracy was compared.The results show that in the bearing experiment of Case Western Reserve University(CWRU),the mean diagnostic accuracy based on transfer learning and transfer component analysis(TCA)+KNN were respectively 93.72%and 75.52%.In the bearing experiment of Xi'an Jiaotong University,the diagnostic accuracy based on transfer learning and TCA+KNN was 94.80%and 70.40%respectively.The above experimental results verify the superiority of the iterative diagnosis method based on transfer learning in the fault diagnosis of samples in different source regions.
作者
朱旭东
ZHU Xu-dong(Wuxi Institute of Art and Technology,Yixing 214200,China)
出处
《机电工程》
CAS
北大核心
2023年第3期361-369,共9页
Journal of Mechanical & Electrical Engineering
基金
中国纺织工业联合会职业教育教学改革项目(2020ZJJGLX132)
江苏省高等教育教改研究项目(2021JSJG413)。
关键词
轴承故障诊断准确率
异源域样本
改进均衡适配
迁移学习
K近邻算法
源域数据
目标域数据
accuracy of bearing diagnosis
heterologous domain samples
improve balanced fitting
transfer learning
K-nearest neighbor(KNN)algorithm
source domain data
target domain data
作者简介
朱旭东(1980-),男,江苏宜兴人,硕士,副教授,主要从事物联网应用技术、嵌入式技术、软件开发等方面的研究。E-mail:zhuxud1980@163.com。
