• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2026, Vol. 62 ›› Issue (9): 191-200.doi: 10.3901/JME.260415

• 机械动力学 • 上一篇    

扫码分享

基于多任务孪生核矩阵机的机械故障诊断方法

潘海洋, 陈春安, 郑近德, 童靳于, 程健   

  1. 安徽工业大学机械工程学院 马鞍山 243032
  • 收稿日期:2025-05-10 修回日期:2025-10-29 发布日期:2026-07-08
  • 作者简介:潘海洋,男,1989年出生,博士,副教授,硕士研究生导师。主要研究方向为设备状态监测与故障诊断、信号处理和模式识别。E-mail:pansea@ahut.edu.cn;陈春安,男,2001年出生,硕士研究生。主要研究方向为设备状态检测与故障诊断。E-mail:19818718153@163.com;郑近德(通信作者),男,1986年出生,博士,教授,博士研究生导师。主要研究方向为机械健康检测和故障诊断、信号处理和复杂性理论。E-mail:lqdlzheng@126.com;童靳于,女,1987年出生,博士,高级实验师,硕士研究生导师。主要研究方向为统计信号处理、振动信号分析和测量。E-mail:jytong@ahut.edu.cn;程健,男,1995年出生,博士,讲师。主要研究方向为机械健康检测与故障诊断、信号处理。E-mail:chengjian@ahut.edu.cn
  • 基金资助:
    安徽省高校杰出青年科研(2022AH020032)和安徽省高校自然科学研究重点(2022AH050292)资助项目。

Mechanical Fault Diagnosis Method Based on Multi-task Twin Kernel Matrix Machine

PAN Haiyang, CHEN Chunan, ZHENG Jinde, TONG Jinyu, CHENG Jian   

  1. School of Mechanical Engineering, Anhui University of Technology, Ma'anshan 243032
  • Received:2025-05-10 Revised:2025-10-29 Published:2026-07-08

摘要: 在进行机械设备多目标诊断时,往往需要建立多个具有独立性的单任务模型,其忽略了多目标任务间的关联性信息,致使诊断模型不准确。基于此,提出了一种基于多任务孪生核矩阵机(Multi-task twin kernel matrix machine, MTTKMM)的机械故障诊断方法。在MTTKMM中,首先设计核增强项,实现对多任务间数据的同步处理,有助于挖掘不同任务间的关联性特征,从而提高模型利用共性信息的能力;然后,利用内置偏移量描述数据间的非线性关系,使模型在处理多目标数据时更具灵活性,提高模型分类的稳定性;最后,考虑不同任务间数据具有较大差异问题,设计了泛化损失项,优化模型在多任务数据间的拟合能力,减少过拟合现象。为了验证MTTKMM在多任务机械故障诊断中的有效性,采用滚动轴承、齿轮等数据集进行实验验证,实验结果表明:MTTKMM在多目标诊断中具有优越的分类性能。

关键词: 故障诊断, 多任务孪生核矩阵机, 多任务学习, 滚动轴承, 齿轮

Abstract: Multi-object joint diagnosis often requires the establishment of multiple independent single task models, which leads to the omission of correlation information between multiple objects in the diagnostic model, resulting in inaccurate diagnostic models. Based on this, a mechanical fault diagnosis method based on multi-task twin kernel matrix machine (MTTKMM) is proposed. In MTTKMM, the synchronization processing of multi-task data is first achieved by designing kernel enhancement terms, which help to explore the correlation features between different tasks and improve the model's ability to utilize common information. Then, by utilizing the built-in nonlinear offset to capture the nonlinear relationships between data, the model becomes more flexible when dealing with complex data of multiple objects, further improving the stability of model classification. Finally, considering the significant differences in data between different tasks, a generalization loss is designed to optimize the model's fitting ability on multi-task data and reduce overfitting. To verify the effectiveness of MTTKMM in multi-task mechanical fault diagnosis, experimental verification is conducted using datasets such as rolling bearings and gears. The experimental results prove that MTTKMM has superior classification performance in multi-object diagnosis.

Key words: fault diagnosis, multi-task twin kernel matrix machine, multi-task learning, rolling bearing, gear

中图分类号: