基于CWT-IDenseNet的滚动轴承故障诊断方法  

Fault diagnosis method for rolling bearings based on CWT-IDenseNet

作　　者：贾广飞[1] 梁汉文 杨金秋 武哲 韩雨欣 JIA Guangfei;LIANG Hanwen;YANG Jinqiu;WU Zhe;HAN Yuxin(School of Mechanical Engineering,Hebei University of Science and Technology,Shijiazhuang,Hebei 050018,China;Research Center of Resources Utilization and Environmental Conservation,Hebei University,Baoding,Hebei 071002,China)

机构地区：[1]河北科技大学机械工程学院,河北石家庄050018 [2]河北大学资源利用与环境保护研究中心,河北保定071002

出　　处：《河北科技大学学报》2025年第2期129-140,共12页Journal of Hebei University of Science and Technology

基　　金：国家自然科学基金(52206224);中央引导地方科技发展资金项目(226Z1906G);河北省教育厅科学研究项目(CXY2024038)。

摘　　要：针对一维信号所含信息不全面和DenseNet网络在变工况下存在过拟合等问题,提出了基于连续小波变换时频图像和改进密集连接卷积网络(improved DenseNet,IDenseNet)的滚动轴承故障诊断方法CWT-IDenseNet。首先,将一维振动信号通过CWT转为二维时频图像;其次,对DenseNet网络进行改进,将DenseNet第1个卷积块中的ReLU激活函数替换为Swish激活函数(Swish激活函数更平滑);同时,在网络中引入基于风格的卷积神经网络重校准模块(style-based recalibration module,SRM)和空间与通道注意力机制模块(convolutional block attention module,CBAM),SRM关注特征通道权重,CBAM则从通道和空间2个维度增强特征表达能力,进而得到IDenseNet;最后,将二维时频图像输入到IDenseNet模型中进行特征提取和故障诊断,通过模型的Softmax层输出故障诊断结果。结果表明,所提方法在恒定工况及变工况下的平均故障识别准确率均达到97.80%,且在迁移学习模型中,平均故障识别准确率达到了99.44%。CWT-IDenseNet方法可以有效提高模型的泛化能力,在恒定工况及变工况下具有显著优势,对提高滚动轴承故障诊断的准确率和可靠性具有参考价值。Aiming at the problems of incomplete information contained in one-dimensional signals and overfitting of the DenseNet under variable working conditions,a rolling bearing fault diagnosis method based on continuous wavelet transform(CWT)time-frequency images and an improved densely connected convolutional network(IDenseNet)was proposed.Firstly,the one-dimensional vibration signal was converted into two-dimensional time-frequency images by CWT.Then,the DenseNet network was turned into IDenseNet,the ReLU activation function in the first convolutional block of DenseNet was replaced by the Swish activation function(which was smoother),and the style-based recalibration module(SRM)and the convolutional block attention module(CBAM)were introduced into the DenseNet network.The SRM focused on the weight of feature channels,while CBAM enhanced the feature expression ability from the two dimensions of channel and space.Finally,the two-dimensional time-frequency image was input into the IDenseNet model for feature extraction and fault diagnosis,and the fault diagnosis results were output through the Softmax layer of the model.The results show that the average fault recognition accuracy of the proposed method reaches 97.80%under constant and variable conditions,and the average fault recognition accuracy reaches 99.44%in the transfer learning model.The CWT-IDenseNet method can effectively improve the generalization ability of the model,which has significant advantages under constant and variable conditions,providing reference for improving the accuracy and reliability of rolling bearing fault diagnosis.

关 键 词：机械动力学与振动 滚动轴承故障诊断 连续小波变换 密集连接卷积网络 注意力机制 

分 类 号：TH133.33[机械工程—机械制造及自动化]