基于SWT与改进卷积神经网络的轴承故障诊断  

Bearing fault diagnosis based on SWT and improved convolutional neural network

作　　者：龚俊 张月义 陈思戢 刘靖楠 GONG Jun;ZHANG Yueyi;CHEN Siji;LIU Jingnan(School of Economics and Management,China Jiliang University,Hangzhou 310018,China)

机构地区：[1]中国计量大学经济与管理学院,浙江杭州310018

出　　处：《现代电子技术》2024年第6期68-74,共7页Modern Electronics Technique

基　　金：国家社会科学基金项目(18BJY033)。

摘　　要：针对传统轴承故障诊断依赖专家经验且存在时频特征提取效果不佳,导致故障诊断效率和精度较低的问题,提出一种基于同步压缩小波变换(SWT)与改进卷积神经网络(CNN)的轴承故障诊断模型(SICNN)。首先,将一维的非平稳轴承振动信号通过SWT转换为高频率表达的二维时频图像,作为卷积神经网络的输入;然后,引入SRM对提取的特征进行风格池化与融合,调整卷积通道合适的特征权重,提高重要特征的关注度进而提高网络的表征能力;最后,通过Softmax层输出故障诊断结果。为了验证所提出的模型性能,使用凯斯西储大学采集的轴承数据集开展实验。结果表明,该模型故障诊断准确率可达到99.88%,与其他传统方法相比,具有良好的可行性和收敛性能,实践层面应用价值较高。In allusion to the issue of traditional bearing fault diagnosis relying on expert experience and poor time-frequency feature extraction,resulting in low efficiency and accuracy,a bearing fault diagnosis model(SICNN) based on synchronous squeezed wavelet transform(SWT) and improved convolutional neural network(CNN) is proposed.The one-dimensional non-stationary vibration signal is converted into a high-frequency two-dimensional time-frequency map through SWT,which is used as the input of the convolutional neural network.The SRM module is introduced to pool and fuse the extracted features,adjust the appropriate feature weights of the convolutional channel,and improve the network′s representation ability.The fault diagnosis results are output by means of the Softmax layer.In order to verify the performance of the proposed model,experiments were conducted by means of the Case Western Reserve University bearing dataset.The results show that the fault diagnosis accuracy of the model was 99.88%.Compared with other methods,it has good feasibility and convergence performance,and has high practical application value.

关 键 词：故障诊断 滚动轴承 同步压缩小波变换 卷积神经网络 通道注意力模块 注意力机制 

分 类 号：TN624-34[电子电信—电路与系统] TH133.33[机械工程—机械制造及自动化]