基于CNN-SN和无监督域适应的滚动轴承故障诊断
Fault Diagnosis for Rolling Bearings Based on CNN-SN and Unsupervised Domain Adaptation
陈攀 1袁逸萍 1马军岩 1樊盼盼 1田芳2
作者信息
- 1. 新疆大学 机械工程学院,乌鲁木齐 830047
- 2. 乌鲁木齐市技术创新研发与科技成果转化中心,乌鲁木齐 830047
- 折叠
摘要
针对滚动轴承在不同工况下振动数据分布差异大且难以获取所有故障标记样本,致使故障诊断模型泛化能力差的问题,提出了一种基于卷积神经网络-收缩网络(CNN-SN)和无监督域适应的变工况故障诊断方法.首先,构建领域共享的一维卷积神经网络以提取振动信号中的故障特征,同时引入软阈值学习机制构建局部特征收缩网络,缓解噪声对故障特征提取的影响;然后,对不同工况样本提取的故障特征引入最大均值差异的正则化约束,实现源域与目标域特征的全局对齐;最后,对无标签的目标工况样本,采用最大最小化分类器差异的对抗学习策略实现不同域特征更细粒度的子领域对齐.采用江南大学轴承数据集对所提方法进行试验验证,结果表明所提方法表现出良好的领域适配能力,具有较高的跨域故障诊断精度.
Abstract
Aimed at the problem that the difference in distribution of vibration data of rolling bearings under different working conditions is large and it is difficult to obtain all fault label samples,resulting in poor generalization ability of fault diagnosis model,a fault diagnosis method under variable working conditions is proposed based on convolutional neural networks-shrink network(CNN-SN)and unsupervised domain adaptation.Firstly,a domain shared one-dimensional convolutional neural network is constructed to extract fault features from vibration signals,a soft threshold learning mechanism is introduced to construct a local feature shrink network,and the impact of noise on fault feature extraction is mitigated.Then,the regularization constraint of maximum mean discrepancy is introduced to fault features extracted from samples under different working conditions,and the global alignment of features in source domain and target domain is achieved.Finally,for unlabeled target condition samples,the adversarial learning strategy of maximizing and minimizing difference of classifiers is used to achieve finer grained sub domain alignment of different domain features.The bearing data set from Jiangnan University is used to verify the proposed method.The results show that the proposed method has good domain adaptability and high cross domain fault diagnosis accuracy.
关键词
滚动轴承/故障诊断/变工况/迁移学习/无监督域适应Key words
rolling bearing/fault diagnosis/variable working condition/transfer learning/unsupervised domain adaptation引用本文复制引用
出版年
2025
NETL
NSTL
万方数据