基于改进多图卷积网络的液压泵小样本故障诊断  

Hydraulic pump fault diagnosis of small samples based on an improved multi-graph convolutional network

作　　者：郑直 赵文博 李克 朱占辉 刘彤谣 孙杨 林帅恒 ZHENG Zhi;ZHAO Wenbo;LI Ke;ZHU Zhanhui;LIU Tongyao;SUN Yang;LIN Shuaiheng(College of Mechanical Engineering,North China University of Science and Technology,Tangshan 063210,China)

机构地区：[1]华北理工大学机械工程学院,河北唐山063210

出　　处：《振动与冲击》2024年第24期59-67,83,共10页Journal of Vibration and Shock

基　　金：河北省自然科学基金资助项目(E2022209086);唐山市科技创新团队培养计划项目(21130208D);河北省科技重大专项项目(22282203Z)。

摘　　要：多图卷积网络模型(multi-graph convolutional network, M-GCN)可将图像转为特征向量,并可利用图卷积操作增强同类节点聚集。由于受到现场空间和经济条件限制,无法有效地采集液压泵充足故障样本,导致小样本问题;当引入M-GCN模型对液压泵的故障进行诊断时,该模型特征表达存在区分度不足和信息单一等问题。因此,该文提出了一种改进多图卷积网络模型,即MMH-GCN模型。首先,为解决模型特征提取区分度不足问题,引入掩码自编码器(masked autoencoder, MAE)降低编码维度并提取关键图像特征,提升模型的小样本诊断精度;然后,为解决模型特征信息单一问题,引入异构图注意力网络(heterogeneous graph attention network, HAN)提取更丰富和全面的图结构数据特征,以提升模型的小样本诊断精度和效率。通过液压泵实测故障试验验证分析可知,该文所提MMH-GCN模型较原模型具有明显的高效性和优越性,在诊断精度和效率方面分别提升了12.14%和14.63%。Images can be converted into feature vectors by multi-graph convolutional network(M-GCN),and the aggregation of similar nodes can be enhanced by graph convolution operation.Due to the limitation of field space and economic conditions,it is impossible to collect sufficient fault samples of hydraulic pump,resulting the problem of small samples.When M-CCN is introduced to diagnose fault of hydraulic pumps,there are problems such as insufficient discrimination and single information in feature expression.Therefore,an improved multi-graph convolutional network,namely MMH-GCN,was proposed in this paper.In order to solve the problem of insufficient discrimination of model feature extraction,masked autoencoder(MAE)was introduced to reduce the encoding dimension and extract key image features,so as to improve diagnostic accuracy based on small sample size.For solving the problem of single feature information,the heterogeneous graph attention network(HAN)was introduced to extract much abundant and comprehensive features of graph structure data,so as to improve diagnostic accuracy and efficiency based on small sample size.Through the measured fault experimental verification and analysis of hydraulic pumps,it can be seen that the MMHGCN proposed in this paper has obvious efficiency and superiority compared with the original M-GCN,and the diagnostic accuracy and efficiency are increased by 12.14%and 14.63%,respectively.

关 键 词：多图卷积网络 掩码自编码器 异构图注意力网络 小样本 

分 类 号：TH137[机械工程—机械制造及自动化] TP277[自动化与计算机技术—检测技术与自动化装置]