• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2021, Vol. 57 ›› Issue (1): 157-167.doi: 10.3901/JME.2021.01.157

• 机械动力学 • 上一篇    下一篇



李栋1, 刘树林2, 孙欣2   

  1. 1. 常州大学石油工程学院 常州 213164;
    2. 上海大学机电工程与自动化学院 上海 200072
  • 收稿日期:2019-11-28 修回日期:2020-07-31 出版日期:2021-01-05 发布日期:2021-02-06
  • 通讯作者: 李栋(通信作者),男,1981年出生,博士,副教授,硕士研究生导师。主要研究方向为设备智能故障诊断方法。E-mail:lidong@cczu.edu.cn
  • 作者简介:刘树林,男,1963年出生,博士,教授,博士研究生导师。主要研究方向为复杂设备故障诊断。E-mail:lsl346@shu.edu.cn;孙欣,女,1994年出生,博士研究生。主要研究方向为设备智能故障诊断方法。E-mail:sunxin52@shu.edu.cn
  • 基金资助:

A Continual Learning Fault Diagnosis Method for Discontinuous Time-varying Sample Space

LI Dong1, LIU Shulin2, SUN Xin2   

  1. 1. School of Petroleum Engineering, Changzhou University, Changzhou 213164;
    2. School of Mechatronic Engineering and Automation, Shanghai University, Shanghai 200072
  • Received:2019-11-28 Revised:2020-07-31 Online:2021-01-05 Published:2021-02-06

摘要: 智能故障诊断方法是保障机械设备安全可靠运行的重要手段,然而,现有智能故障诊断方法多属于批量学习模式,缺乏连续学习能力,无法高效地处理不连续时变样本空间的故障诊断问题。在生物免疫系统利用记忆细胞快速识别二次入侵抗原,以及记忆细胞随入侵抗原进化而进化等智能机理的启发下,提出一种针对不连续时变样本空间的具有连续学习能力的故障诊断方法,当样本空间与时间无关时,其退化为一般的监督学习故障诊断方法。在诊断过程中通过对样本的不断学习,持续更新记忆细胞,利用亲和度阈值识别未参与训练以及发生间断后再次出现的故障样本。通过20个标准数据集的仿真分析了当其退化为一般的监督学习故障诊断方法时的性能;利用西安交通大学的XJTU-SY滚动轴承加速寿命试验数据集分析了其处理不连续时变样本空间故障诊断问题时的性能。试验结果表明与经典故障诊断方法相比,此故障诊断方法对时不变样本空间的故障诊断问题具有良好的故障诊断性能,对于不连续时变样本空间的故障诊断问题具有更好的故障诊断性能。

关键词: 时变样本, 连续学习, 故障诊断, 生物免疫机理, 人工免疫算法, 滚动轴承

Abstract: Intelligent fault diagnosis methods play an important role in ensuring the equipment safely and reliably operating. However, they cannot effectively classify the discontinuous time-varying sample space, for most of them belong to batch learning and lack continual learning ability. A continual learning fault diagnosis method for discontinuous time-varying sample space based on the artificial immune system, CLFDMDTVS, is proposed. It is inspired by the intelligent mechanism that memory cells of the biological immune system can recognize previous invaders when they attack again very fast and more efficiently, and these memory cells can evolve with the evolution of previous invaders. CLFDMDTVS will degenerate into a common supervised learning fault diagnosis method when all data independent of time. Memory cells were continuously updated by learning testing data during the fault diagnosis stage, the samples which are not used to train and the reappeared samples which are discontinuous can be diagnosed by affinity threshold. To assess the performance and possible advantages of CLFDMDTVS, the experiments on 20 well-known datasets from the UCI repository and XJTU-SY rolling element bearing accelerated life test datasets were performed. Results show that CLFDMDTVS has better fault diagnosis performance for time-invariant data, and outperforms the other methods for the problems with discontinuous time-varying sample space.

Key words: time-varying sample, continual learning, fault diagnosis, biological immune mechanism, artificial immune algorithm, rolling element bearing