首页|基于子空间域对抗判别网络的不同型号滚动轴承剩余寿命预测

基于子空间域对抗判别网络的不同型号滚动轴承剩余寿命预测

扫码查看
原文链接
  • NETL
  • NSTL
  • 万方数据
针对不同型号滚动轴承因结构尺寸、运行工况等差异导致轴承退化数据分布和特征尺度不一致,引起剩余寿命预测精度下降的问题,提出基于子空间域对抗判别网络的不同型号滚动轴承剩余寿命预测方法.首先,通过高效通道注意力机制提升特征提取器各通道中重要特征的权重,自适应获取不同型号滚动轴承的深层性能退化特征,并以此预训练标签预测器;然后,在对抗判别网络框架上将域判别器与特征提取器对抗训练,最小化源域和目标域在表征子空间上的正交基距离,利用表征子空间正交基不受特征缩放影响的性质克服特征尺度变化过大引起的回归性能下降问题,实现不同型号滚动轴承间的域自适应;最后,利用训练好的特征提取器提取待预测轴承退化特征,输入标签预测器得到剩余寿命.在PRONOSTIA、XJTU-SY和自测数据集上进行了验证,实验结果表明所提方法能充分学习源域特征分布信息,有效克服不同型号下的特征尺度差异,相比其他域自适应方法效果提升20％至40％.
Remaining life prediction of different types of rolling bearings based on subspace domain adversarial discrimination network
A residual life prediction method for different types of rolling bearings is proposed based on the subspace domain adversarial discriminant network(SDADN)to address the issue of inconsistent distribution and characteristic scales of bearing degradation data caused by differences in structural dimensions,operating conditions,and other factors,leading to a decrease in life prediction accuracy.Firstly,the feature extractor can adaptively obtain deep degradation features for different types of rolling bearings by using an efficient channel attention mechanism to enhance the weight of important features in each channel and is used to train the label predictor.Then,in the asymmetric feature mapping framework,the domain discriminator and feature extractor are adversarially trained to minimize the orthogonal basis distance between the source and target domains in the representation subspace.By utilizing the property that the orthogonal basis in the representation subspace is not affected by feature scaling,the regression performance degradation caused by excessive feature scale changes is reduced,and domain adaptation among different types of rolling bearings is achieved.Finally,the trained feature extractor is used to extract the degradation features of the bearing,and the remaining lifespan is obtained by inputting them into the label predictor.The proposed method was validated on PRONOSTIA,XJTU-SY,and self-test datasets,and the experimental results showed that it can fully learn the distribution information of source domain features,effectively overcome the feature scale differences under different models,and improve the performance by 20％to 40％compared to other domain adaptive methods.

rolling bearingsremaining useful life predictionadversarial discriminative domain adaptationtime convolutional network

陈仁祥、张雁峰、徐向阳、张鹏博、杨宝军

展开 >

重庆交通大学交通工程应用机器人重庆市工程试验室 重庆 400074

重庆智能机器人研究院 重庆 400714

滚动轴承 剩余寿命预测 对抗判别域自适应 时间卷积网络

国家自然科学基金重庆市教委科学技术研究计划重庆市自然科学基金创新发展联合基金重庆市研究生联合培养基地项目重庆市专业学位研究生教学案例库项目重庆交通大学研究生科研创新项目

51975079KJZD-M202200701CSTB2023NSCQ-LZX0127JDLHPYJD2021007JDALK20220072023S0123

2024

10.19650/j.cnki.cjsi.J2312306

仪器仪表学报
中国仪器仪表学会

仪器仪表学报

CSTPCD北大核心
影响因子：2.372
ISSN：0254-3087
年,卷(期)：2024.45(3)
  • 15