循环相关熵和一维浅卷积神经网络轴承故障诊断  

Bearing Fault Diagnosis Based on Cyclic Correntropy and One-dimensional Shallow Convolutional Neural Network

作　　者：李辉 徐伟烝 LI Hui;XU Weizheng(School of Mechanical Engineering,Tianjin University of Technology and Education,Tianjin 300222,China)

机构地区：[1]天津职业技术师范大学机械工程学院,天津300222

出　　处：《机械科学与技术》2024年第4期600-610,共11页Mechanical Science and Technology for Aerospace Engineering

基　　金：国家自然科学基金项目(51375319);天津市科技计划项目(23YFYSHZ00280)。

摘　　要：针对传统二维深度卷积神经网络结构复杂、易产生过拟合和难以有效处理低信噪比信号的问题,提出了一种基于循环相关熵和一维浅卷积神经网络的故障诊断-CCe-1D SCNN方法。该方法综合利用了一维浅卷积神经网络结构简单、计算复杂度低和循环相关熵能在低信噪比环境下有效提取故障特征的优点。首先,计算轴承故障振动信号的循环相关熵函数、循环相关熵谱密度函数和广义循环平稳度;其次,将一维归一化的广义循环平稳度作为一维浅卷积神经网络的输入层,通过一维浅卷积神经网络自动实现故障特征提取和模式分类;最后,将CCe-1D SCNN方法应用于电机轴承故障特征提取和分类,实验结果表明:CCe-1D SCNN方法在低噪声比情况下仍能保持很高的模式识别正确率,为一种自动故障特征提取和模式识别的有效方法。The traditional two-dimensional convolutional neural network(2D CNN)not only has high computational complexity and is easier to over fitting,but also is difficult to deal with low signal-to-noise ratio(SNR)signal effectively.In order to overcome the shortcomings of 2D CNN,a new fault diagnosis method based on cyclic correntropy(CCe)and one-dimensional shallow convolutional neural network(1D SCNN)is proposed.This new method(CCe-1D SCNN)makes fully use of the advantages of 1D SCNN and CCe,in which the 1D SCNN has simple structure and low computational complexity.Firstly,the cyclic correntropy function,the cyclic correntropy spectral density(CCSD)function and generalized degree of cyclostationary(DCS)of bearing fault vibration signal are calculated.Secondly,the one-dimensional normalized generalized degree of cyclostationary is used as the input layer of one-dimensional shallow convolutional neural network.The fault feature extraction and pattern classification are automatically realized by one-dimensional shallow convolutional neural network.Finally,the CCe-1D SCNN method is applied to fault feature extraction and pattern classification of motor bearing fault.The experimental results show that the CCe-1D SCNN technique can still maintain a high accuracy of pattern recognition in the case of very low signal-to-noise ratio,which is an effective method for automatic fault feature extraction and pattern recognition.

关 键 词：循环相关熵 一维浅卷积神经网络 深度学习 循环平稳信号 故障诊断 

分 类 号：TH165.3[机械工程—机械制造及自动化] TN911.72[电子电信—通信与信息系统]