基于快速特征逼近谱图注意力网络的滚动轴承半监督智能故障诊断研究  

Research on Semi-supervised Intelligent Fault Diagnosis of Rolling Bearings Based on Fast Feature Approximation Spectral Graph Attention Network

作　　者：宁少慧[1] 杜越 周利东[1] NING Shaohui;DU Yue;ZHOU Lidong(School of Mechanical Engineering,Taiyuan University of Science and Technology,Taiyuan)

机构地区：[1]太原科技大学机械工程学院,山西太原030024

出　　处：《机床与液压》2025年第6期33-39,共7页Machine Tool & Hydraulics

基　　金：山西省应用基础研究计划(20210302123212)。

摘　　要：基于图注意力网络的诊断模型在故障诊断全监督任务中有较好的表现,但在半监督任务中表现欠佳。针对此问题,构建一种基于快速特征逼近谱图注意力网络的半监督滚动轴承智能故障诊断模型。通过K近邻图方法将振动信号转为可用于诊断的图数据,丰富了数据特征;将图数据输入到构建的诊断模型中,学习故障信息特征,并分析不同的标签比例训练集的诊断结果。同时,分析了Sum、Mean、Max 3种池化方式和超参数对诊断模型的影响;最后,分别在两组实验轴承数据集上进行验证。结果表明:所提模型可以有效地捕获图的全局模式,降低计算复杂度,在全监督诊断任务中其诊断准确率可以保持在99%以上;在标签比例为10%的半监督任务中,其诊断准确率仍能保持在93.5%,所提诊断模型在半监督任务中有良好表现。Diagnostic models based on graph attention networks have good performance in fully supervised fault diagnosis tasks,but their performance in semi-supervised tasks is not satisfactory.To address this issue,a semi-supervised intelligent fault diagnosis model for rolling bearings based on a fast feature approximation spectral graph attention network was proposed.The vibration signals were converted into graph data suitable for diagnosis using the K-nearest neighbor graph method,enriching the data features.The graph data was input into the constructed diagnostic model to learn the characteristics of fault information,and the diagnostic results of training sets with different label ratios were analyzed.Meanwhile,the impacts of three pooling methods(Sum,Mean,Max)and hyperparameters on the diagnostic model were analyzed.Finally,the proposed model was validated on two sets of experimental bearing datasets.The experiments demonstrate that the proposed model can effectively capture the global patterns of the graph and reduce computational complexity.In fully supervised diagnosis tasks,its diagnostic accuracy can be maintained over 99%,and in semi-supervised tasks with a label ratio of 10%,its diagnostic accuracy remains at 93.5%.The proposed diagnostic model performs well in semi-supervised tasks.

关 键 词：轴承 故障诊断 快速特征逼近谱图注意力网络 K近邻图算法 

分 类 号：TH222[机械工程—机械制造及自动化]