• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2026, Vol. 62 ›› Issue (9): 182-190.doi: 10.3901/JME.260414

• 机械动力学 • 上一篇    

扫码分享

小波赋能联邦学习的机械故障诊断

李志农1, 余紫莹1, 王奉涛2, 李喆3   

  1. 1. 南昌航空大学无损检测技术教育部重点实验室 南昌 330063;
    2. 汕头大学智能制造技术教育部重点实验室 汕头 515063;
    3. 中国航发沈阳黎明航空发动机有限责任公司 沈阳 110043
  • 收稿日期:2025-10-18 修回日期:2026-01-29 发布日期:2026-07-08
  • 作者简介:李志农,男,1966年出生,博士,教授,博士研究生导师。主要研究方向为智能检测与信号处理,机械故障诊断。E-mail:lizhinong@tsinghua.org.cn;余紫莹,女,1999年出生,硕士研究生。主要研究方向为人工智能与故障诊断。E-mail:yzy18042305@163.com;王奉涛(通信作者),男,1974年出生,教授,博士研究生导师。主要研究方向为机械故障诊断。E-mail:ftwang@stu.edu.cn;李喆,男,1989年出生,高级工程师。主要研究方向为智能检测与评价。E-mail:Lizhesylm@163.com
  • 基金资助:
    国家自然科学基金(52075236)、智能制造技术教育部重点实验室(汕头大学)开放课题基金(STME2024002)、福建省特种智能装备安全与测控重点实验室开放基金(FJIES2024KF06)和国家自然科学基金(11202128)资助项目。

Research on Mechanical Fault Diagnosis Using Wavelet Empowered Federated Learning

LI Zhinong1, YU Ziying1, WANG Fengtao2, LI Zhe3   

  1. 1. Key Laboratory of Nondestructive Testing of the Ministry of Education, Nanchang Hangkong University, Nanchang 330063;
    2. Key Laboratory of Intelligent Manufacturing Technology (Shantou University), Ministry of Education, Shantou 515063;
    3. AECC Shenyang Liming Aero-Engine Co., Ltd., Shenyang 110043
  • Received:2025-10-18 Revised:2026-01-29 Published:2026-07-08

摘要: 基于联邦学习的机械故障诊断方法由于缺乏可解释性,常被视作黑盒模型,这在一定程度上降低了客户端的信任度。针对此不足,提出了一种面向机械故障诊断的小波赋能联邦学习模型。在提出的模型中,利用小波系数替代传统联邦学习中的局部模型参数,在中央服务器上聚合小波系数并重构全局模型。由于小波具有平移尺度特性,所构造的模型极大地增强了可解释性。同时利用小波系数代替局部模型的参数,显著减少了参数的计算量,使收敛速度明显加快。实验结果验证了所提出方法的有效性。与传统的联邦学习模型相比,小波赋能联邦学习的故障识别性能更好。而对比小波深度学习网络,两个模型都拥有可解释性基础上,小波赋能联邦学习的鲁棒性更强。

关键词: 故障诊断, 联邦学习, 可解释性, 小波分解

Abstract: Since the mechanical fault diagnosis method based on federated learning lacks interpretability and is often regarded as a black box model, which reduces the trust of users to a certain extent. Based on this deficiency, a wavelet empowered federated learning (WEFL) model for mechanical fault diagnosis is prposed. In the proposed model, the wavelet coefficients are innovatively used to replace the local model parameters in traditional federated learning, and the wavelet coefficients are aggregated on the central server to reconstruct the global model. Due to the translation scale characteristics of wavelets, the constructed model greatly improves interpretability in federated learning. Simultaneously, local model parameters are replaced with wavelet coefficients, greatly reduce the computational cost of parameters and significantly accelerates the convergence. The experimenta test verify the effectiveness of the proposed model. Compared with the traditional federated learning model, the WEFL model has more transparency and interpretability in the decision-making process of mechanical fault diagnosis. Compared with the wavelet deep learning network (e.g. WaveletKernelNet), the WEFL model can effectively protect data privacy and solve the problem of data island.

Key words: fault diagnosis, federated learning, interpretability, wavelet decomposition

中图分类号: