改进可视图与图同构网络的变工况轴承故障诊断方法  

Improved Viewable and Graph Isomorphic Network for Fault Diagnosis of Variable Operating Condition Bearings

作　　者：陈驻民[1] 韦继程 CHEN Zhu-min;WEI Ji-cheng(School of Intelligent Manufacturing and Control Engineering,Shanghai Polytechnic University,Shanghai 201209;School of Computer and Information Engineering,Shanghai Polytechnic University,Shanghai 201209)

机构地区：[1]上海第二工业大学智能制造与控制工程学院,上海201209 [2]上海第二工业大学计算机与信息工程学院,上海201209

出　　处：《制造业自动化》2025年第3期156-167,共12页Manufacturing Automation

摘　　要：针对传统深度学习在变工况下轴承故障诊断迁移能力弱的问题,提出一种结合加权有限穿越水平可视图(WLPHVG)与图同构网络(GIN)的轴承故障诊断方法。通过可视图算法将原始轴承振动时序信号转换为图结构数据,利用最大均值差异方法对图节点所连边进行加权以减少噪声对模型精度的影响。同时,对图同构网络聚合层进行调整,以更好地拟合图数据特征。最终,将图数据输入到模型中得到轴承故障诊断分类结果。实验在多个不同的轴承数据集上采用不同模型进行比较,结果显示,模型在不同工况和信噪比条件下的准确率均优于其他对比模型,达到97.27%以上。消融实验验证了加权方法对所提出的模型抗噪性能的有效性,表明模型能充分利用图域数据的内部结构关系,在跨平台与时变转速两种不同工况下展现出良好的可迁移性。To address the limited transferability issue of traditional deep learning in bearing fault diagnosis under variable operating conditions,an innovative approach that combined Weighted Limited Passage Horizontal Visual Graph(WLPHVG)and Graph Isomorphic Network(GIN)was propossed in this paper.The method transformed raw bearing vibration time-series signals into graph-structured data using WLPHVG and employed the Maximum Mean Discrepancy method to weight the edges connecting graph nodes,reducing the impact of noise on model accuracy.At the same time,adjustments were made to the aggregation layer of the Graph Isomorphic Network to better capture the features of graph data.The graph data was processed then by the final model to obtain the classification results for bearing fault diagnosis.Experimental evaluations on multiple bearing datasets using various models demonstrated that the proposed approach achieved superior accuracy(over 97.27%)under different operating conditions and signal-to-noise ratios compared to that of other benchmark models.The deconstructive experiments validated the effectiveness of the weighting method on the noise resistance of the WLPHVG-GIN model,confirming its ability to leverage internal structural relationships in graph domain data and exhibit strong transferability under diverse conditions such as cross-platform and varying speeds.

关 键 词：图同构网络 水平可视化图 故障诊断 

分 类 号：TH133.33[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]