基于改进图神经网络图形样本聚合的增量学习模型  

Incremental learning model based on Graph-SAGE

作　　者：刘振柱 侯乔文 兰媛 于磊 牛蔺楷 LIU Zhenzhu;HOU Qiaowen;LAN Yuan;YU Lei;NIU Linkai(School of Mechanical Engineering,Taiyuan University of Technology,Taiyuan 030024,China;Taiyuan University of Technology Key Laboratory of Advanced Sensors and Intelligent Control,Taiyuan 030024,China)

机构地区：[1]太原理工大学机械工程学院,山西太原030024 [2]太原理工大学新型传感器与智能控制教育部重点实验室,山西太原030024

出　　处：《机电工程》2025年第3期549-558,共10页Journal of Mechanical & Electrical Engineering

基　　金：工业和信息化部民用飞机专项科研项目(MJZ3-1N22);山西省基础研究计划(自由探索类)项目(202203021221092)。

摘　　要：针对轴向柱塞泵在故障诊断中对大规模、动态变化数据处理困难,以及故障类型增加导致分类性能下降的问题,提出了一种基于改进图神经网络图形样本聚合(Graph-SAGE)的增量学习模型。首先,将轴向柱塞泵的不同故障振动信号构建为带标签的数据集,并通过数据增强生成了新的数据集;然后,采用K-最邻近法(KNN)分别构建了初始训练阶段和增量训练阶段的图结构数据(其中,初始阶段的图结构用于模型的初始训练,增量训练阶段的图结构用于增量训练);接着,为了确定最适合轴向柱塞泵故障图数据集的聚合方法,在初始训练阶段比较了不同聚合器对故障识别准确率的影响,并在增量训练阶段结合显性知识与隐性知识对模型进行了优化;最后,采用了实验的方式,验证了该模型的可行性,并通过对比实验和鲁棒性测试,对该模型的性能和稳定性进行了评估。研究结果表明:该增量学习模型在应对新增故障类型时表现优异,在轴向柱塞泵的复合故障识别中,平均准确率达到了92.35%,显著优于传统图神经网络在相同条件下的表现;同时,该模型在混合工况下的增量训练准确率达到了95%,展现出较强的适应性和鲁棒性。该方法能够有效应对不同的故障模式和工况条件,准确识别轴向柱塞泵的复合故障。Aiming at the challenges of managing large-scale,dynamically changing data and the decline in classification performance with increasing fault types in axial piston pump diagnostics,a modified graph neural network graph sample and aggregated(Graph-SAGE)-based incremental learning model was proposed.The model was designed to handle evolving data effectively and adapt to new fault types,enhancing accuracy and robustness over time.Firstly,vibration signals of different faults in the axial piston pump were constructed into a labeled dataset,and a new dataset was generated through data augmentation.Secondly,the K-nearest neighbors(KNN)method was used to build graph structure data for both the initial training phase and the incremental training phase,with the graph structure in the initial phase being used for model training and the graph structure in the incremental phase being used for incremental learning.Then,to determine the most suitable aggregation method for the axial piston pump fault graph dataset,the impact of different aggregators on fault recognition accuracy was compared during the initial training phase,and the model was optimized in the incremental training phase by combining explicit and implicit knowledge strategies.Finally,the feasibility of the model was validated through experiments,its performance and stability were assessed via comparative studies and robustness tests.The research results show that the proposed incremental learning model performs well with new fault types,achieving an average accuracy of 92.35%in identifying compound faults in axial piston pumps,significantly surpassing traditional graph neural networks.The model also reaches 95%accuracy in incremental training under mixed conditions,demonstrating strong adaptability and robustness.Furthermore,this method effectively handles various fault modes and working conditions,accurately identifying compound faults in axial piston pumps.

关 键 词：轴向柱塞泵 故障诊断 增量学习 图神经网络图形样本聚合 K-最邻近法 图结构数据 

分 类 号：TH322[机械工程—机械制造及自动化] TP391[自动化与计算机技术—计算机应用技术]