摘要
在对不同工况下的滚动轴承进行故障诊断时,要收集足够多标记的故障样本是非常困难的。为此,以原始振动信号作为神经网络的输入,通过多表示动态自适应(MRDA)算法多表示对齐可迁移的特征、自适应动态的衡量边缘分布和条件分布相对重要性,从而构建了一种新的深度迁移模型,即一维多表示空洞动态自适应迁移网络(1D MRDDATN)。首先,对迁移学习数据分布进行了问题分析,对DDA进行了理论推导;然后,在一维空洞卷积基础上,创建了一维多表示空洞卷积神经网络(1D MRDCNN),并提出了MRDA算法和多表示动态自适应结构(MRDAM),形成了一维多表示空洞动态自适应迁移网络(1D MRDDATN);最后,采用美国凯斯西储大学(CWRU)的滚动轴承数据集进行了实验验证。研究结果表明:与传统的深度迁移学习方法相比,上述方法的平均诊断准确率有所提升,可达到98%以上;MRDA通过多表示对齐来完成不同工况下的跨域分类任务,自适应地捕获不同方面的信息,可以获得更好的性能。
In fault diagnosis of rolling bearing under different working conditions,abundantly labeled data was often difficult or even impossible to obtain.Therefore,using the original vibration signal as the input of the neural network,through the multiple representation dynamic adaptive(MRDA)algorithm,the multi-representation aligns the transferable features,and the adaptive dynamic measures the relative importance of the marginal distribution and the conditional distribution,a new deep transfer model was constructed,namely the one-dimensional multi-representation hole dynamic adaptive transfer network(1D MRDDATN).Firstly,transfer learning data distribution was analyzed,and DDA was theoretically deduced.Then,on the basis of one-dimensional dilated convolution,a one-dimensional multi-representation dilated convolution neural network(1D MRDCNN)was created,and the MRDA algorithm and multi-representation dynamic adaptive structure(MRDAM)were proposed to construct a one-dimensional multi-representation dilated dynamic adaptive transfer network(1D MRDDATN).Finally,the rolling bearing data set of CWRU was verified by experiments.The research results show that compared with the traditional deep transfer learning methods,the average diagnostic accuracy of the proposed method is improved and reaches more than 98%.MRDA completes cross-domain classification tasks under different working conditions through multi-representation alignment,and adaptively captures information from different aspects,which achieves better performance.
作者
朱继扬
孙虎儿
张天源
赵扬
白晓艺
ZHU Ji-yang;SUN Hu-er;ZHANG Tian-yuan;ZHAO Yang;BAI Xiao-yi(School of Mechanical Engineering,The North University of China,Taiyuan 030051,China)
出处
《机电工程》
CAS
北大核心
2023年第2期178-185,203,共9页
Journal of Mechanical & Electrical Engineering
基金
山西省自然科学基金资助项目(201801D121186)。
关键词
不同工况
一维多表示空洞动态自适应迁移网络
故障样本
深度迁移学习
多表示动态自适应算法
神经网络
一维多表示空洞卷积神经网络
different working conditions
one-dimensional multi-representation hole dynamic adaptive transfer network(1D MRDDATN)
fault sample
deep transfer learning
multi-representation dynamic adaptive(MRDA)algorithm
neural network
one-dimensional multi-representation dilated convolution neural network(1D MRDCNN)
作者简介
朱继扬(1996-),男,河南信阳人,硕士研究生,主要从事机械系统故障诊断与状态检测方面的研究。E-mail:1664546742@qq.com;通信联系人:孙虎儿,男,博士,副教授,硕士生导师。E-mail:sunhuer@163.com。
