变转速极低标签率下旋转机械故障诊断的图注意力网络  

Entropy-graph attention network for semi-supervised fault diagnosis of rotating machinery under extremely low label rate and variable rotating speed conditions

作　　者：谢俊文 童靳于[1] 郑近德[1] 潘海洋[1] 包家汉[1] XIE Junwen;TONG Jinyu;ZHENG Jinde;PAN Haiyang;BAO Jiahan(College of Mechanical Engineering,Anhui University of Technology,Ma’anshan 243032,China)

机构地区：[1]安徽工业大学机械工程学院,安徽马鞍山243032

出　　处：《振动与冲击》2024年第19期242-248,共7页Journal of Vibration and Shock

基　　金：安徽省高校自然科学研究重点项目(2022AH050315);国家自然科学基金(51975004)。

摘　　要：在极低标签率情况下,现有的图神经网络(graph neural network,GNN)在图构造时存在节点间的关联信息挖掘不充分等问题。工业生产中,旋转机械常工作在变转速工况下,且标记故障样本代价高昂。针对上述两个问题,基于JS(Jenson-Shannon)相对熵和动态图注意力网络(dynamic graph attention network,DGAT),提出了一种熵-图注意力网络,并将其应用于极低标签率下变转速工况的旋转机械半监督故障诊断中。首先,设计了基于JS相对熵的图构造方法,用于充分挖掘GNN中样本间的关联信息。其次,构建基于熵-动态图注意力网络的半监督学习模型,通过动态注意力机制进一步挖掘样本中故障敏感特征。最后,将所提方法在变转速工况下轴承和齿轮箱数据集上进行验证,结果表明所提方法能够在标签率不超过1%的极低情况下准确诊断出旋转机械的不同故障类型,且性能优于其它常用的图神经网络。Under extremely low label rate,the existing graph neural networks(GNN)suffer from insufficient mining of inter-node association information during graph construction.In industrial production,rotating machinery often operates under variable rotating speed conditions,and labeling fault samples is costly.Here,aiming at the above 2 problems,an entropy-graph attention network was proposed based on Jenson-Shannon(JS)relative entropy and dynamic graph attention network(DGAT),and it was applied in semi-supervised fault diagnosis of rotating machinery under extremely low label rate and variable rotating speed conditions.Firstly,a graph construction method based on JS relative entropy was designed to fully explore the correlation information among samples in GNN.Secondly,a semi-supervised learning model based on the entropy-DGAN was constructed to further explore fault sensitive features in samples with dynamic attention mechanism.Finally,the proposed method was verified on bearing and gearbox datasets under variable rotating speed conditions,and the results showed that the proposed method can correctly diagnose different fault types in rotating machinery under extremely low label rate of no more than 1%;its performance is superior to other commonly used GNNS.

关 键 词：旋转机械 故障诊断 相对熵 图神经网络(GNN) 变转速 低标签率 

分 类 号：TH163.5[机械工程—机械制造及自动化]