• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2026, Vol. 62 ›› Issue (17): 361-372.doi: 10.3901/JME.260358

• 机器人及机构学 • 上一篇    

扫码分享

基于类别感知增强和原型约束的齿轮箱不均衡跨域故障诊断

刘杰1, 李世新1, 李朝峰2, 杨娜1   

  1. 1. 沈阳工业大学机械工程学院 沈阳 110870;
    2. 东北大学机械工程与自动化学院 沈阳 110819
  • 收稿日期:2025-10-22 修回日期:2026-02-06 发布日期:2026-11-20
  • 作者简介:刘杰(通信作者),女,1980年出生,博士,副教授,博士研究生导师。主要研究方向为故障诊断与寿命预测。E-mail:starliujie@126.com
  • 基金资助:
    辽宁省教育厅面上项目基金(LJ212410142026)和辽宁省科技厅面上项目(2025-MS-119)资助项目。

Imbalanced Cross-domain Fault Diagnosis of Gearboxes Based on Class-aware Learning and Prototype Constraint Enhancement

LIU Jie1, LI Shixin1, LI Chaofeng2, YANG Na1   

  1. 1. School of Mechanical Engineering, Shenyang University of Technology, Shenyang 110870;
    2. School of Mechanical Engineering and Automation, Northeastern University, Shenyang 110819
  • Received:2025-10-22 Revised:2026-02-06 Published:2026-11-20

摘要: 在实际工业场景中难以获取充足且同分布的故障样本,不同工况下采集的样本往往呈现严重的类别不均衡特性,健康样本数量远超故障样本,这给现有的智能故障诊断研究带来新的挑战。为此,提出一种基于类别感知增强和原型约束的不均衡跨域故障诊断方法,面向类别不均衡与故障样本稀少的挑战,实现多源域下不同工况间的跨域故障诊断。首先,设计一种类别感知跨域增强方法,可以充分挖掘多源域样本信息。该方法通过计算不同域间同类别样本特征中心的余弦相似度,量化域间分布差异,并根据相似度自适应确定插值权重范围,避免盲目跨域插值导致的特征退化。其次,定义一种基于原型网络的质量约束机制,可以灵活有效地评估增强样本与目标类别原型的相似性,确保生成样本的类别一致性,进一步提高模型在不同工况下的泛化性。通过设置不同的跨域诊断任务和不均衡比例,与其他先进的跨域故障诊断方法进行对比验证。结果表明,相比现有方法,所提方法可以更有效地在不均衡挑战下实现不同工况间的齿轮箱跨域故障诊断。

关键词: 齿轮箱故障诊断, 类别感知学习, 原型网络, 数据增强, 不均衡数据

Abstract: In practical industrial scenarios, samples collected under different working conditions often exhibit severe class imbalance, with healthy samples significantly outnumbering faulty samples. This imbalanced distribution of samples poses new challenges for intelligent fault diagnosis. To address this issue, an imbalanced cross-domain fault diagnosis method based on class-aware enhancement and prototype constraints is proposed. This method is designed to tackle class imbalance and fault sample scarcity challenges and enables cross-domain fault diagnosis across different working conditions in multi-source domains. First, a class-aware cross-domain enhancement method is designed to fully exploit information from multi-source domain samples. This method quantifies inter-domain distribution discrepancies by calculating the cosine similarity of feature centers for identical classes across different domains and adaptively determines interpolation weights to prevent feature degradation from blind cross-domain interpolatio. Second, a quality constraint mechanism based on prototype networks is defined to flexibly and effectively evaluate the similarity between enhanced samples and target class prototypes. This ensures category consistency of generated samples and further improving model generalization across different working conditions. Different cross-domain diagnosis tasks and imbalance ratios are configured for comparisons against other advanced cross-domain fault diagnosis methods. The results demonstrate that the proposed method can more effectively achieve cross-domain fault diagnosis of gearboxes across different working conditions under imbalanced challenges compared to existing methods.

Key words: gearbox fault diagnosis, class-aware learning, prototype network, data augmentation, imbalanced data

中图分类号: