基于SimAM注意力机制的轴承故障迁移诊断模型

扫码查看

原文链接

NETL
NSTL
万方数据
维普

中文摘要：针对轴承故障在跨工况迁移诊断时,其域不变特征难以提取,易出现模型过拟合这一问题,提出了一种基于无参数注意力模块(SimAM)的轴承故障迁移诊断方法.首先,以一维卷积神经网络作为基本框架,利用自适应批量归一化(AdaBN)对各输出层进行了归一化处理,经两层卷积层和两层池化层后,对输出特征进行了随机节点失活操作;然后,利用改进后的参数化修正线性单元(PReLU)激活函数自适应提取负值输入权值系数,分别以交叉熵损失函数监督训练有标签的源域数据,以均方对数误差(MSLE)作为损失函数训练无标签的目标数据;最后,利用自制实验台数据和凯斯西储轴承公开数据对模型进行了验证,分别以不同的单一工况作为源域,其余工况作为目标域进行了迁移诊断任务研究.研究结果表明:基于SimAM的轴承故障迁移诊断方具有较好的域不变特征提取的性能,且所提特征具有较好的聚类效果;自制实验台中的平均迁移精度在89.1%以上,最高均值可达97.85%,CWRU数据集中的平均迁移精度达98.68%.该成果可为后续轴承故障由实验向工业现场的迁移诊断奠定基础.

外文标题：Rearing fault transfer diagnosis model based on SimAM attention mechanism

外文摘要：Aiming at the problem that domain invariant features are difficult to extract and model overfitting is easy to occur in bearing fault migration diagnosis during cross-working condition migration diagnosis,a bearing fault migration diagnosis method based on a simple parameter-free attention module(SimAM)was proposed.Firstly,the one-dimensional convolutional neural network was used as the basic framework,and adaptive batch normalization(AdaBN)was used to normalize each output layer.After two convolutional layers and two pooling layers,the output features were deactivated by random nodes.Then,the improved parametric rectified linear unit(PReLU)activation function was used to adaptively extract the negative input weight coefficient,and the cross-entropy loss function was used to monitor the trained labeled source domain data and the mean squared logarithmic error(MSLE)was used as the loss function to train the unlabeled target data.Finally,the model was verified by the self-made experimental bench data and the open data of Case Western Reserve bearing.Different single working conditions were taken as the source domain,and the other working conditions were taken as the target domain to carry out the migration diagnosis task.The experimental results show that the proposed method has good domain invariant feature extraction performance,and the proposed features have good clustering effect.The average migration accuracy of the self-made experimental bench was above 89.1%,the highest mean was up to 97.85%,and the average migration accuracy of CWRU dataset was up to 98.68%.The results can lay a foundation for subsequent bearing fault transfer diagnosis from experimental data to industrial sites.

外文关键词：

bearing fault diagnosistransfer learningsimple parameter-free attention module(SimAM)adaptive batch normalization(AdaBN)parametric rectified linear unit(PReLU)mean squared logarithmic error(MSLE)convolutional neural network

作者：

包从望、朱广勇、邹旺、郭灏

展开 >

作者单位：

六盘水师范学院矿业与机械工程学院,贵州六盘水 553000

中国矿业大学机电工程学院,江苏徐州 221116

关键词：

轴承故障诊断迁移学习无参数注意力机制自适应批量归一化参数化修正线性单元均方对数误差卷积神经网络

基金：

贵州省教育厅项目六盘水市科技计划六盘水市科技计划六盘水师范学院项目

项目编号：

黔教合KY字[2020]11752020-2022-PT-0252020-2019-05-12LPSSYylzy2205

出版年：

2024

DOI：

10.3969/j.issn.1001-4551.2024.05.012

机电工程

浙江大学　浙江省机电集团有限公司

机电工程

CSTPCD北大核心

影响因子：0.785

ISSN：1001-4551

年,卷(期)：2024.41(5)

参考文献量18