航空学报 > 2024, Vol. 45 Issue (17): 530611-530611   doi: 10.7527/S1000-6893.2024.30611

基于联邦图网络的转子系统故障诊断方法

李晖1,2, 陈银超1, 孙绍山1, 梁兆鑫1, 毛刚2, 乔彬2, 李永波2()   

  1. 1.中国航空工业集团公司 成都飞机设计研究所,成都 610091
    2.西北工业大学 航空学院,西安 710072
  • 收稿日期:2024-04-24 修回日期:2024-05-20 接受日期:2024-06-03 出版日期:2024-09-15 发布日期:2024-06-07
  • 通讯作者: 李永波 E-mail:yongbo@nwpu.edu.cn
  • 基金资助:
    国家自然科学基金(12172290);深圳市自然科学基金(JCYJ20220530161801003)

Fault diagnosis method of rotor system based on federated graph network

Hui LI1,2, Yinchao CHEN1, Shaoshan SUN1, Zhaoxin LIANG1, Gang MAO2, Bin QIAO2, Yongbo LI2()   

  1. 1.Chengdu Aircraft Design & Research Institute,AVIC,Chengdu 610091,China
    2.College of Aerospace,Northwestern Polytechnical University,Xi’an 710072,China
  • Received:2024-04-24 Revised:2024-05-20 Accepted:2024-06-03 Online:2024-09-15 Published:2024-06-07
  • Contact: Yongbo LI E-mail:yongbo@nwpu.edu.cn
  • Supported by:
    National Natural Science Foundation of China(12172290);Shenzhen Science and Technology Program(JCYJ20220530161801003)

摘要:

转子系统运行环境恶劣、多源监测数据融合困难且易产生数据孤岛现象,对转子系统健康监测带来巨大挑战。本文提出了基于遗传进化构图的联邦图卷积神经网络(FGCN)转子系统故障诊断方法。首先,利用联邦学习和图神经网络,构建联邦图网络框架,通过在每个本地客户端上进行局部模型训练,并通过联邦加权平均算法聚合这些局部模型得到全局模型,这种数据不动、模型动的方式,不仅实现了数据的本地化处理,还确保了模型参数的安全与隐私;此外,为了解决多源传感器数据融合自适应差的问题,提出了遗传进化构图方法,该方法通过模拟生物进化过程中的自然选择和遗传变异机制,在训练过程中自动调整图结构节点之间的连接关系和权重;极大地提高了多源传感器构图的自适应性和灵活性,进而提升故障诊断的准确性。最后,通过在转子故障实验台数据集上进行验证,实验结果表明,所提出的方法能够更充分利用多传感器监测数据,在客户端包含故障种类数量不同的诊断场景中达到了95%以上的故障诊断准确率。

关键词: 图神经网络, 联邦学习, 故障诊断, 遗传进化构图, 隐私保护

Abstract:

Considerable harsh operating environments of rotor systems, combined with the difficulty in fusing monitoring data from multiple sources and a tendency to form data islands, present substantial challenges to the health monitoring of rotor systems. This paper proposes a rotor system fault diagnosis method utilizing Federated Graph Convolutional neural Networks (FGCN) based on a genetic evolutionary composition. First, a federated migration learning framework, employing federated learning and graph neural networks, is established. The global model is derived by training local models on individual clients and aggregating them via a federated weighted average algorithm. This method facilitates data localization while securing the privacy and integrity of model parameters. Furthermore, to address the challenge of inadequate adaptive integration of multi-source sensor data, a genetic evolution composition method is introduced. This method dynamically adjusts the connectivity and weights among graph nodes during training, emulating the mechanisms of natural selection and genetic variation found in biological evolution. This approach significantly enhances the adaptability and flexibility of the multi-source sensor composition, thereby improving the accuracy of fault diagnosis. In conclusion, experimental validation on the rotor failure testbed dataset demonstrates that the proposed method effectively utilizes limited target domain data, achieving over 95% accuracy in fault diagnosis scenarios where the clients contain different number of faults classes.

Key words: graph neural networks, federated learning, fault diagnosis, genetic evolutionary graph construction, privacy preservation

中图分类号: