基于CEEMDAN-FastICA-MCNN的多传感信息融合轴承故障诊断  

Bearing Fault Diagnosis Based on CEEMDAN-FastICA-MCNN Multi-sensor Information Fusion

作　　者：张鑫 钟倩文 余佑民 彭乐乐 郑树彬[1] 陈谢祺 ZHANG Xin;ZHONG Qianwen;YU Youmin;PENG Lele;ZHENG Shubin;CHEN Xieqi(School of Urban Railway Transportation,Shanghai University of Engineering Science,Shanghai 201620,China;Shanghai Metro Maintenance Guarantee Co.,Ltd.,Vehicle Branch,Shanghai 200031,China)

机构地区：[1]上海工程技术大学城市轨道交通学院,上海201620 [2]上海地铁维护保障有限公司车辆分公司,上海200031

出　　处：《噪声与振动控制》2024年第4期145-152,共8页Noise and Vibration Control

基　　金：国家自然科学基金资助项目(51907117,51975347);上海市科技计划资助项目(22010501600);上海申通地铁集团资助项目(JS-KY21R008-6,JS-KY20R013-3)。

摘　　要：针对轴承振动信号易受噪声干扰、变工况及单一传感器提取特征信息不完备的问题,提出基于自适应噪声完备集成经验模态分解(Complete Ensemble Empirical Mode Decomposition of Adaptive Noise,CEEMDAN)、快速独立分量(Fast Independent Components Analysis,FastICA)降噪和多输入卷积神经网络(Multiple-input Convolutional Neural Networks,MCNN)的多传感信息融合轴承故障诊断方法。首先分别对多传感采集的振动信号划分数据集,并输入CEEMDAN得到本征模态函数(Inherent Nodal Function,IMF);随后,选择峭度大于3的IMF构造观测信号,其余IMF构造虚拟噪声信号,作为两个输入源输入FastICA,分离出特征向量;最后,设计MCNN识别故障类型。在CWRU和XJTU-SY数据集上的正确率分别为99.94%和99.64%。在信噪比为-8 dB的抗噪性能测试中,正确率分别为96.95%和98.29%;在信噪比为0 dB的抗噪性能测试中,正确率分别为99.00%和99.23%。对比实验结果表明此方法能够提取更为全面的故障特征信息,获得更高的准确率。Aiming at the problems that vibration signal of bearing is easy to be interfered by noise and variable operating conditions,and its feature information extracted by a single sensor is incomplete,a bearing fault diagnosis method based on complete ensemble empirical mode decomposition of adaptive noise(CEEMDAN),fast independent components analysis(FastICA)denoising and multiple-input convolutional neural networks(MCNN)is proposed.Firstly,the vibration signals collected by multiple sensing are divided into data sets separately and the CEEMDAN is used to obtain Inherent Modal functions(IMFs).Then,the IMFs with the kurtosis greater than 3 are selected to construct observation signal,the remaining IMFs are used to construct virtual noise signal.They are input to FastICA as two input sources to separate the feature vectors.Finally,MCNN is designed to identify fault types.The results show that the accuracy on the CWRU and XJTU-SY data sets is up to 99.94%and 99.64%.The accuracy is 96.95%and 98.29%in the anti-noise performance test with signal to noise ratio(SNR)of-8 dB.The accuracy is 99.00%and 99.23%in the anti-noise performance test with SNR of 0.The comparative experimental results show that the method can extract more comprehensive fault feature information and obtain higher accuracy of extraction.

关 键 词：故障诊断 CEEMDAN FASTICA MCNN 

分 类 号：TH133.3[机械工程—机械制造及自动化]