Page 151 - 《振动工程学报》2026年第3期
P. 151
第 39 卷第 3 期 振 动 工 程 学 报 Vol. 39 No. 3
2026 年 3 月 Journal of Vibration Engineering Mar. 2026
-
多源域下 MAAN 与 MK MMD 的滚动轴承
跨工况复合故障诊断
王亚萍 , 高圣延 ,王新生 , 王金宝 , 江旭东 1
1
2
1
1
(1. 哈尔滨理工大学先进制造智能化教育部重点实验室,黑龙江 哈尔滨 150080;
2. 国家管网集团西部管道有限责任公司,新疆 乌鲁木齐 830011)
摘要: 针对实际试验中由变工况引起的滚动轴承源域和目标域故障数据分布不平衡,多个相似数据利用不充分,导致模型故障
诊断精度不高的问题,提出一种多源域多头注意力自适应网络的深度迁移学习故障诊断方法。该方法选择多种不同工况下的
数据作为多源域,其他工况下的数据作为目标域;利用由卷积神经网络及多头注意力机制构建的公共特征提取器,对源域及目
标域数据进行原始特征提取及权重分配,再利用特定特征提取器提取特定敏感特征;选择多核最大均值差异作为源域和目标
域的度量函数,度量源域和目标域特征之间的相似性,建立滚动轴承跨工况复合故障诊断模型。经对比验证,在设定的 10 类
任务中,完成了对大量无标签样本的分类识别,证明了多源域多头注意力自适应网络模型的故障诊断效果最好。
关键词: 滚动轴承; 故障诊断; 深度迁移学习; 多源域; 多头注意力机制
+
中图分类号: TH165 .3; TH133.3; TP18 文献标志码: A DOI:10.16385/j.cnki.issn.1004-4523.202403042
Composite fault diagnosis of rolling bearings across operating conditions
with MAAN and MK-MMD in multi-source domains
2
1
1
1
WANG Yaping , GAO Shengyan , WANG Xinsheng ,WANG Jinbao , JIANG Xudong 1
(1.Key Laboratory of Advanced Manufacturing and Intelligent Technology, Ministry of Education,Harbin University of Science
and Technology, Harbin 150080, China; 2.National Pipeline Network Group Western Pipeline Co., Ltd., Urumqi 830011, China)
Abstract: To address the issues of distribution discrepancy between source and target domain fault data due to varying operating
conditions, and the underutilization of multiple similar datasets, which lead to low diagnostic accuracy, a deep transfer learning
fault diagnosis method based on a multi-source domain adaptive attention network (MAAN) is proposed. This method selects data
from various operating conditions as the multi-source domain and data from other operating conditions as the target domain. The
original feature extraction and weight assignment are performed on the source and target domain data using a public feature extrac⁃
tor constructed by a convolutional neural network and multi-head attention. A specific feature extractor is used to extract specific
sensitive features. Multi-kernel maximum mean discrepancy (MK-MMD) is adopted to measure the feature discrepancy between
domains. This facilitates the construction of a compound fault diagnosis model for rolling bearings under cross-operating conditions.
Following comparative validation, the classification and identification of a large number of unlabelled samples was achieved across
10 categories of tasks. This demonstrated that the multi-source domain multi-head attentional adaptive network model was the
most effective in fault diagnosis and fault identification.
Keywords: rolling bearing;fault diagnosis;deep transfer learning;multi-source domain;multi-head attention mechanism
滚动轴承长期处于恶劣的工作条件下,极易引 与训练数据(源域)的分布不同的问题上有着巨大优
起故障,造成不可估量的损失,及时准确地诊断故障 势,具有良好的泛化性以及更高的数据利用率。
至关重要。随着“大数据”时代的到来,深度迁移学 领域自适应是迁移学习的关键,主要目的是将
习故障诊断在处理收集的测试数据(目标域)的分布 源域和目标域的数据映射到同一特征空间,最小化
收稿日期: 2024-03-19; 修订日期: 2024-05-20
基金项目: 国家自然科学基金资助项目(5217550);黑龙江省自然科学基金资助项目(LH2023E082)

