中国机械工程 ›› 2022, Vol. 33 ›› Issue (24): 2990-2996,3006.DOI: 10.3969/j.issn.1004-132X.2022.24.010

• 智能制造 • 上一篇    下一篇

基于shapelets时间序列的多源迁移学习滚动轴承故障诊断方法

李可1,2;燕晗1,2;顾杰斐1,2;宿磊1,2;苏文胜3;薛志钢3   

  1. 1.江南大学江苏省食品先进制造装备技术重点实验室,无锡,214122
    2.江南大学机械工程学院,无锡,214122
    3.江苏省特种设备安全检验监督研究院无锡分院,无锡,214071
  • 出版日期:2022-12-25 发布日期:2023-01-12
  • 通讯作者: 顾杰斐(通信作者),男,1992年生,副教授。研究方向为故障诊断与智能检测。E-mail:jfgu@jiangnan.edu.cn。
  • 作者简介:李可,男,1978年生,教授、博士研究生导师。研究方向为故障诊断与振动分析。E-mail:like_jiangnan@163.com。
  • 基金资助:
    国家自然科学基金(51775243,52175096);中国博士后科学基金(2021T140279);江苏省市场监督管理局科技计划(KJ196043)

Research on Multi-source Transfer Learning Bearing Fault Diagnosis Based on Shapelets Time Series

LI Ke1,2 ;YAN Han1,2;GU Jiefei1,2 ;SU Lei1,2; SU Wensheng3;XUE Zhigang3   

  1. 1.Jiangsu Key Laboratory of Advanced Food Manufacturing Equipment and Technology,Jiangnan University,Wuxi,Jiangsu,214122
    2.School of Mechanical Engineering,Jiangnan University,Wuxi,Jiangsu,214122
    3.Jiangsu Province Special Equipment Safety Supervision Inspection Institute Branch of Wuxi,Wuxi,Jiangsu,214071
  • Online:2022-12-25 Published:2023-01-12

摘要: 针对滚动轴承故障诊断在工程实际中故障数据稀缺的问题,提出一种基于shapelets时间序列的多源迁移学习滚动轴承故障诊断方法。首先利用典型故障信息丰富、标记样本充足的滚动轴承数据构建多源域数据集,使用不同源域的数据对源域特征提取器与分类器进行预训练;然后利用基于动态时间规整的shapelets学习算法提取源域与目标域的shapelets作为判别结构,通过度量判别结构优化源域数据,对源域网络进行微调以得到诊断模型;最后根据每个源域与目标域的shapelets之间的差异,利用自适应域权重对各分类器的结果进行聚合得出诊断结果。实验结果表明,该方法在小样本与强噪声的情况下具有较高的故障诊断准确率。

关键词: 滚动轴承, 故障诊断, shapelets时间序列, 多源迁移学习

Abstract: Aiming at the problems that the available fault data of rolling bearing fault diagnosis were scarce in industrial productions, a multi-source transfer learning bearing fault diagnosis method was proposed based on shapelets time series. Firstly, source domain sets were constructed by the laboratory data which included abundant typical fault information and sufficient label information, and the source domain feature extractor and classifier were trained using the training data of each source domain. Then, the shapelets learning algorithm based on dynamic time warping (DTW) was used to extract the shapelets of the source domain and the target domain as the discriminant structure, the source domain data was optimized through the measurement discriminant structure, and the source domain network classifier was fine tuned to obtain the diagnostic model. Finally, according to the difference between the shapelets of each source domain and the target domain, the results of each classifier were aggregated by using the adaptive domain weight to obtain the diagnosis results. Experimental results show that the proposed method has good fault diagnosis performance in the case of few shot and high noise.

Key words: rolling bearing, fault diagnosis, shapelets time series, multi-source transfer learning

中图分类号: