航空学报 > 2026, Vol. 47 Issue (7): 232642-232642   doi: 10.7527/S1000-6893.2025.32642

固体力学与飞行器总体设计

基于图域泛化的直升机尾传动系统故障诊断方法

陈国旺1,2, 唐倩1(), 何刘1   

  1. 1. 重庆大学 机械与运载工程学院,重庆 400044
    2. 长江师范学院 机器人工程学院,重庆 408100
  • 收稿日期:2025-07-31 修回日期:2025-10-09 接受日期:2025-12-25 出版日期:2025-12-31 发布日期:2025-12-29
  • 通讯作者: 唐倩
  • 基金资助:
    国家自然科学基金(52335006); 国家自然科学基金(52275518); 重庆市自然科学基金(CSTB2023NSCQ-LZX0137)

Fault diagnosis method of helicopter tail transmission systems based on graph domain generalization

Guowang CHEN1,2, Qian TANG1(), Liu HE1   

  1. 1. College of Mechanical and Vehicle Engineering,Chongqing University,Chongqing 400044,China
    2. School of Robotics Engineering,Yangtze Normal University,Chongqing 408100,China
  • Received:2025-07-31 Revised:2025-10-09 Accepted:2025-12-25 Online:2025-12-31 Published:2025-12-29
  • Contact: Qian TANG
  • Supported by:
    National Natural Science Foundation of China(52335006); Chongqing Natural Science Foundation(CSTB2023NSCQ-LZX0137)

摘要:

针对直升机尾传动系统高精度故障数据难获取和传统迁移学习对未知故障的知识泛化不足问题,提出了用于直升机尾传动系统的自适应权重多级融合图域泛化(AMFGDG)故障诊断框架。首先,结合集中质量法与有限元法,生成直升机尾传动系统不同健康状态下的多节点振动数据,并用实验验证了该方法的有效性。基于所生成数据方法构建多源域的带有限元节点的图样本,同时以不同传感器的测试数据作为目标域的图样本。然后,采用自注意力机制完成图网络边权重自适应更新,实现节点间影响强度精准表征。引入混合专家门控策略完成跨图层级的信息融合,缓解多层图网络堆叠导致的过平滑现象。最后,结合分类损失、深度聚类损失及局部分布对齐损失实现诊断模型更新,完成故障诊断分类。基于直升机尾传动实验台私有数据集和西储大学轴承公开数据集的诊断结果对比和消融实验,验证了提出方法具有低数据依赖性和高泛化能力,诊断精度高达95.4%。

关键词: 故障诊断, 直升机尾传动系统, 图卷积网络, 域泛化, 动力学建模

Abstract:

To address the challenges of limited access to high-quality fault data and the poor generalization performance of traditional transfer learning methods on unseen fault types, this paper proposes an Adaptive-weight Multi-level Fusion Graph Domain Generalization (AMFGDG) framework for fault diagnosis of helicopter tail transmission systems. Specifically, multi-node vibration signals under various health conditions are generated by integrating the concentrated mass method with the finite element method, and their validity is experimentally verified. Graph-structured samples with finite element nodes are constructed from the generated data for multiple source domains, while test data from different sensors are formulated as target domain graphs. A self-attention mechanism is employed to adaptively update edge weights, enabling precise modeling of inter-node influence. Additionally, a mixture-of-experts gating strategy is introduced to fuse information across graph layers and mitigate the over-smoothing effect in deep GNNs. The model is optimized via a joint loss function combining classification loss, deep clustering loss, and local distribution alignment loss. Diagnostic results and ablation studies on a private helicopter tail transmission test bench dataset and the public Case Western Reserve University (CWRU) bearing dataset demonstrate the proposed method’s low data dependency and strong generalization capability, achieving an accuracy of up to 95.4%.

Key words: fault diagnosis, helicopter tail transmission system, graph convolutional networks, domain generalization, dynamic modeling

中图分类号: