基于改进集成多隐层小波极限学习神经网络的滚动轴承故障识别研究  

Fault identification of rolling bearing based on improved ensemble multiple hidden layers wavelet ELM network

作　　者：赵凡超[1] 戴石良 房华伟[1] 张丽敏[1] 刘伟 ZHAO Fan-chao;DAI Shi-liang;FANG Hua-wei;ZHANG Li-min;LIU Wei(China Tobacco Guangxi Industrial Company Limited,Liuzhou 535006,China;School of Civil Engineering,University of South China,Hengyang 421001,China;Hunan Nuclear Sunny Technology Engineering Company Limited,Hengyang 421001,China)

机构地区：[1]广西中烟工业有限责任公司,广西柳州535006 [2]南华大学土木工程学院,湖南衡阳421001 [3]湖南核三力技术工程有限公司,湖南衡阳421001

出　　处：《机电工程》2021年第9期1152-1159,共8页Journal of Mechanical & Electrical Engineering

基　　金：广西省工信委工业创新发展资助项目(1909014872)。

摘　　要：由于强噪声和非线性、非平稳性等特性,导致滚动轴承振动信号存在难以提取和其工况状态难以辨识的问题,对此提出了一种基于改进集成多隐层小波极限学习神经网络的滚动轴承故障识别模型。首先,使用了谱分割小波变换,将采集到的滚动轴承振动信号分解为若干本征模态分量;然后,选择了较能反映轴承运行工况特征的模态分量,并加以了重构;最后,利用了不同小波函数设计了不同的多隐层小波极限学习神经网络,并加入了卷积机制,将重构后的信号输入不同的深层网络,进行了特征学习与故障识别,利用集成方法得到了最后的滚动轴承故障识别结果。研究结果表明:提出方法的平均故障识别准确率达到99.42%,标准差仅为0.11;该方法自动特征提取能力和工况识别能力优于深度稀疏自动编码器、深度降噪自动编码器和深度信念网络等深度学习方法,适用于滚动轴承故障的自动识别。Due to the characteristics of strong noise,nonlinearity,and non-stationarity,the vibration signal of rolling bearing was difficult to extract and its working condition was difficult to identify.Therefore,a rolling bearing fault recognition model based on improved ensemble multiple hidden layers wavelet ELM network(IEMHLWEN)was proposed.Firstly,a new spectral segmentation method was proposed and the collected rolling bearing vibration signals were decomposed by spectral segmentation wavelet transform,and the decomposed components which could better reflect the characteristics of bearing conditions were selected and reconstructed.Finally,different multiple hidden layers wavelet ELM networks were designed by employing different wavelet functions,and the reconstructed signals were fed into different deep networks for automatic feature learning and fault identification.The final result was obtained by ensemble learning method.The experimental results show that the average fault identification accuracy of proposed method reaches 99.42%and the standard deviation is only 0.11.The ability of condition automatic feature extraction and automatic condition identification are better than deep learning methods such as deep sparse auto-encoder,deep de-noising auto-encoder,deep belief network and so on,and it is suitable for automatic identification of rolling bearing faults.

关 键 词：滚动轴承 集成学习 故障识别 极限学习机 小波变换 改进集成多隐层小波极限学习神经网络 

分 类 号：TH133.3[机械工程—机械制造及自动化]