小样本下基于迁移学习的轴承状态识别方法  

Rolling Bearing State Recognition Based on Transfer Learning under Small Samples

作　　者：曹宁[1] 江志农[1] 高金吉[2] CAO Ning;JIANG Zhinong;GAO Jinji(National Defense Key Laboratory of Ministry of Education,Beijing University of Chemical Technology,Beijing 100029,China;Diagnostic and Self-healing Engineering Research Center,Beijing University of Chemical Technology,Beijing 100029,China)

机构地区：[1]北京化工大学教育部国防重点实验室,北京100029 [2]北京化工大学诊断与自愈工程研究中心,北京100029

出　　处：《噪声与振动控制》2020年第5期89-94,132,共7页Noise and Vibration Control

基　　金：装备发展部十三五预研资助项目(41404030202);GF973资助项目(BHJG2015013)。

摘　　要：传统的基于机器学习的滚动轴承状态识别方法需要满足两个前提条件,即目标数据量充足、训练数据和测试数据分布相同。然而在实际工程中,滚动轴承的工作环境非常复杂,无法满足上述条件。为了解决上述问题,提出一种基于加权混合核迁移成分分析(Weighted Mixed Kernel Transfer Component Analysis,WKTCA)的栈式自编码(Stacked Auto-Encoder,SAE)神经网络的轴承状态识别方法,用于目标数据不足时滚动轴承的状态识别。该方法引入源域数据,利用迁移成分分析(Transfer Component Analysis,TCA)理论构造加权混合核函数,将源域数据与目标域数据映射到同一特征空间进而实现迁移学习(Transfer learning,TL);进一步将特征值输入到具有分类功能的SAE神经网络进行特征自学习和轴承状态识别。对比分析不同数量的目标数据对轴承状态识别准确率的影响,实验结果表明,WKTCA算法可明显缩小目标域数据与源域数据的分布差异,并实现小样本下轴承状态的准确识别。Traditional state recognition method for rolling bearings based on machine learning needs to meet two prerequisites that the target data is sufficient and the training data and test data should have the same probability distribution.However,in engineering practice,the working environment of rolling bearings is very complicated and the above-mentioned conditions cannot be always satisfied.In order to solve the above problems,a rolling bearing state recognition method based on weighted mixed kernel transfer component analysis(WKTCA)and stacked auto-encoder(SAE)neural network is proposed,which can be used for the state recognition of rolling bearings in the case of insufficient data.This method uses transfer component analysis(TCA)theory to construct a multi-kernel function and maps the source data and target data into a same potential space to realize the transfer learning(TL).Then,the feature vectors are input into SAE neural network with classification function for feature self-learning and bearing state recognition.The effects of different amounts of target data on the accuracy of bearing state recognition are compared and analyzed.The experimental results show that the WKTCA algorithm can significantly reduce the distribution difference between the target data and the source data,and accurately identify the bearing status under small samples.

关 键 词：故障诊断 加权混合核迁移成分分析 栈式自编码神经网络 迁移学习 状态识别 滚动轴承 

分 类 号：TH133.3[机械工程—机械制造及自动化]