Label propagation-based unsupervised domain adaptation for intelligent fault diagnosis | |
Wang, Huanjie1,2; Li, Yuan1,2; Bai, Xiwei2; Li, Jingwei1,2; Tan, Jie1,2; Liu, Chengbao2 | |
刊名 | JOURNAL OF INTELLIGENT MANUFACTURING |
2023-08-23 | |
页码 | 18 |
关键词 | Intelligent diagnosis Rotating machinery Label propagation Subpopulation shift Unsupervised domain adaptation |
ISSN号 | 0956-5515 |
DOI | 10.1007/s10845-023-02186-1 |
通讯作者 | Tan, Jie(jie.tan@ia.ac.cn) |
英文摘要 | Current unsupervised domain adaptation methods for intelligent fault diagnosis focus on learning domain-invariant representations under covariate shift assumption. However, the covariate shift assumption is usually unsatisfied when each fault class in different domains consists of multiple modes with skewed proportions, which is common in industrial scenarios. Imbalanced data from multiple modes cause the presence of within-domain class imbalance and between-domain label distributional shift. This paper introduces a novel subpopulation shift that further considers the domain shift from a subpopulation perspective, i.e., that covariate shift and label distributional shift across domains are caused by shifts in the multiple modes. To address this issue, a label propagation-based unsupervised domain adaptation is proposed based on a realistic expansion assumption. We apply the theoretical analysis of the proposed method with a bi-level optimization strategy adapted from meta-learning. Using joint optimization of a teacher model and a student model, the label propagation-based model-agnostic meta-learning (LPMAML) not only propagates supervision information from the source to the target but also adjusts the teacher's strategy throughout the student's learning process. To alleviate the noise caused by label distributional shift, we integrate a sampling-based alignment method that aligns the empirical label distributions across the two domains into LPMAML. Experimental results on three bearing datasets show that the proposed method has impressive generalization ability under covariate, label distributional, and subpopulation shifts. The proposed method offers consistent improvements to unsupervised domain adaptation (UDA) methods. Compared with the vanilla UDA methods, the average diagnosis accuracies of the proposed method on the label distributional shift benchmark and subpopulation shift benchmark are improved by 8.21% and 7.63%, respectively. |
资助项目 | This work was supported in part by the National Key Research and Development Program of China under Grant 2022YFB3304602 and the National Nature Science Foundation of China under Grant 62003344.[62003344] ; National Key Research and Development Program of China ; National Nature Science Foundation of China ; [2022YFB3304602] |
WOS研究方向 | Computer Science ; Engineering |
语种 | 英语 |
出版者 | SPRINGER |
WOS记录号 | WOS:001053541000001 |
资助机构 | This work was supported in part by the National Key Research and Development Program of China under Grant 2022YFB3304602 and the National Nature Science Foundation of China under Grant 62003344. ; National Key Research and Development Program of China ; National Nature Science Foundation of China |
内容类型 | 期刊论文 |
源URL | [http://ir.ia.ac.cn/handle/173211/54116] |
专题 | 中科院工业视觉智能装备工程实验室 |
通讯作者 | Tan, Jie |
作者单位 | 1.Univ Chinese Acad Sci, Sch Artificial Intelligence, Beijing 100049, Peoples R China 2.Chinese Acad Sci, Inst Automat, Beijing 100190, Peoples R China |
推荐引用方式 GB/T 7714 | Wang, Huanjie,Li, Yuan,Bai, Xiwei,et al. Label propagation-based unsupervised domain adaptation for intelligent fault diagnosis[J]. JOURNAL OF INTELLIGENT MANUFACTURING,2023:18. |
APA | Wang, Huanjie,Li, Yuan,Bai, Xiwei,Li, Jingwei,Tan, Jie,&Liu, Chengbao.(2023).Label propagation-based unsupervised domain adaptation for intelligent fault diagnosis.JOURNAL OF INTELLIGENT MANUFACTURING,18. |
MLA | Wang, Huanjie,et al."Label propagation-based unsupervised domain adaptation for intelligent fault diagnosis".JOURNAL OF INTELLIGENT MANUFACTURING (2023):18. |
个性服务 |
查看访问统计 |
相关权益政策 |
暂无数据 |
收藏/分享 |
除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。
修改评论