基于两阶段域泛化学习框架的轴承故障诊断方法  

A Two-stage Domain Generalization Learning Framework for Fault Diagnosis of Bearings

作　　者：谢刚 韩秦 聂晓音 石慧[1,2] 张晓红 田娟 XIE Gang;HAN Qin;NIE Xiao-Yin;SHI Hui;ZHANG Xiao-Hong;TIAN Juan(School of Electronic and Information Engineering,Taiyuan University of Science and Technology,Taiyuan 030024;Advanced Control and Industrial Intelligence Shanxi Provincial Key Laboratory,Taiyuan 030024)

机构地区：[1]太原科技大学电子信息工程学院,太原030024 [2]先进控制与工业智能山西省重点实验室,太原030024

出　　处：《自动化学报》2024年第11期2271-2285,共15页Acta Automatica Sinica

基　　金：山西省科技重大专项计划“揭榜挂帅”项目(202201090301013);山西省重点研发计划项目(202102020101002,202202100401002,202202150401005);山西省基础研究计划(自由探索类)面上项目(20210302123206,202203021211194,202203021221142)资助。

摘　　要：设备在实际运行过程中工况复杂多变,导致振动信号分布存在较大差异.现有的多数方法通过添加度量指标来约束特征提取过程,提取源域和目标域的相似特征以解决从单一源域到目标域的诊断问题.然而,实际运行过程往往包含多个源域数据,且目标域信息在不同源域中存在较大差异,难以有效学习不同域之间的域不变特征.针对上述问题,提出了一种基于两阶段域泛化学习框架的轴承故障诊断方法.在第一阶段,利用大尺寸卷积特征提取模型对多视图振动信号进行预训练,提取多个源域数据之间的初级故障特征.在第二阶段,将初级故障特征输入动静双态融合的时空图卷积模型中,捕捉随时间变化的动态特征和全局时空特征.通过两阶段的学习,将多个源域的数据映射到一个共有特征空间,提取判别性和泛化性特征.实验结果表明,该方法在多源域轴承故障诊断任务中具有较高的诊断精度和较强的泛化能力.During the actual operation of the equipment,the working conditions are complex and changeable,resulting in large differences in vibration signal distribution.Many existing methods constrain the feature extraction process by incorporating measurement metrics,aiming to extract similar features from both the source and target domains to address diagnostic problems from a single source domain to a target domain.However,the actual operational process often involves data from multiple source domains,and the target domain information exhibits significant differences across these various source domains,making it difficult to extract the domain invariant feature.In response to the above problems,this paper proposes a two-stage domain generalization learning framework for fault diagnosis of bearings.In the first stage,the large-scale convolutional feature extraction model is used to pre-train multi-view vibration signals to extract primary fault features between multiple source domain data.In the second stage,the primary fault features are input into the spatial-temporal graph convolutional model for dynamic and static two-state fusion combining dynamic and static states to capture the dynamic features and global spatiotemporal features that change over time.Through two-stage learning,data from multiple source domains are mapped to a common feature space,and discriminative and generalization features are extracted.Experimental results show that this method has high diagnostic accuracy and strong generalization ability in multi-source domain bearing fault diagnosis tasks.

关 键 词：故障诊断 域泛化 多源域 多视图 时空关系 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TH133.3[自动化与计算机技术—控制科学与工程]