基于图特征学习的泵类设备多组件故障诊断方法  

Multi-Component Fault Diagnosis Method for Axial Flow Pumps Based on Graph Feature Learning

作　　者：张兆波 张金鹏 杜灿阳 曾庚运 朱传古 李晓春 刘颉[4] ZHANG Zhao-bo;ZHANG Jin-peng;DU Can-yang;ZENG Geng-yun;ZHU Chuan-gu;LI Xiao-chun;LIU Jie(Guangdong Yuehai Pearl River Delta Water Supply Co.,Ltd.,Guangzhou 511453,China;Nanjing NARI Water Resources and Hydropower Technology Co.,Ltd.,Nanjing 211106,China;Guangdong Hydropower Planning and Design Institute Co.,Ltd.,Guangzhou 510635,China;School of Civil and Hydraulic Engineering,Huazhong University of Science and Technology,Wuhan 430074,China)

机构地区：[1]广东粤海珠三角供水有限公司,广东广州511453 [2]南京南瑞水利水电科技有限公司,江苏南京211106 [3]广东省水利电力勘测设计研究院有限公司,广东广州510635 [4]华中科技大学土木与水利工程学院,湖北武汉430074

出　　处：《水电能源科学》2024年第11期145-149,共5页Water Resources and Power

基　　金：国家自然科学基金项目(52205104)。

摘　　要：针对轴流泵等泵类设备故障位置分散、传感器安装不便等难题,采用水听器装置采集水下声信号,并提出一种基于图特征学习的水听器信号故障识别方法。具体地,提出了针对一维声信号的无向图构建方法,包括节点构建、边构建、权重构建。基于线性加权策略构建加权图,使得图上同类型样本的近似程度能够得到更精准描述,提高后续诊断模型的识别精度。该方法采用图卷积网络构建故障诊断模型,并以加权图为模型输入。通过三层图卷积挖掘水听器信号中潜藏的各类故障信息,并构建SoftMax分类器得到样本的预测标签。通过自主搭建水泵试验台开展相应试验,试验结果表明,所提方法能够实现对轴流泵各部件多类故障的精确诊断,并与其他机器学习方法对比验证了其优越性。For the problems of scattered fault locations and inconvenient sensing installation of axial flow pump equip-ment,a hydrophone device was used to collect underwater acoustic signals,and a graph feature learning based fault iden-tification method for hydrophone signals was proposed.Specifically,the undirected graph construction method was pro-posed for one-dimensional acoustic signals,including node construction,edge construction,and weight construction.And then the weighted graph was constructed based on a linear weighting strategy,which makes the approximation degree of the same type of samples in the graph can be more accurately described and improves the recognition accuracy of the sub-sequent diagnosis model.The proposed method constructed the fault diagnosis model based on graph convolutional net-work,and took the weighted graph as the model input.The various types of fault information hidden in the hydrophone signals can be mined through three-layer graph convolution.The predicted labels of the samples can be obtained with con-structing a SoftMax classifier.The experimental results show that the proposed method can achieve accurate diagnosis for multiple types of faults of axial flow pump components,and its superiority is verified by comparing with other machine learning methods.

关 键 词：加权图 图卷积网络 水听器 故障诊断 

分 类 号：TH312[机械工程—机械制造及自动化]