引用本文:金江涛,许子非,李春,缪维跑,肖俊青,孙康.基于深度学习与混沌特征融合的滚动轴承故障诊断[J].控制理论与应用,2022,39(1):109~116.[点击复制]
JIN Jiang-tao,XU Zi-fei,LI Chun,MIAO Wei-pao,XIAO Jun-qing,SUN Kang.Rolling bearing fault diagnosis based on deep learning and chaotic feature fusion[J].Control Theory and Technology,2022,39(1):109~116.[点击复制]
基于深度学习与混沌特征融合的滚动轴承故障诊断
Rolling bearing fault diagnosis based on deep learning and chaotic feature fusion
摘要点击 2440  全文点击 621  投稿时间：2021-02-28  修订日期：2021-10-28
查看全文  查看/发表评论  下载PDF阅读器
DOI编号  10.7641/CTA.2021.10177
  2022,39(1):109-116
中文关键词  卷积神经网络  长短期记忆网络  支持向量机  混沌  特征融合  轴承  故障诊断
英文关键词  convolutional neural network  long short term memory networks  support vector machine  chaos  feature fusion  bearing  fault diagnosis
基金项目  国家自然科学基金项目(51976131, 51676131, 52006148), 上海市“科技创新心动计划”地方院校能力建设项目(19060502200)资助.
作者单位E-mail
金江涛 上海理工大学 lese0808@163.com 
许子非 上海理工大学  
李春* 上海理工大学 leseusst@163.com 
缪维跑 上海理工大学  
肖俊青 上海理工大学  
孙康 上海理工大学  
中文摘要
      本文针对现有滚动轴承智能故障诊断方法在面向大噪声背景下鲁棒性能差的问题. 基于混沌理论, 提出采 用相空间重构方法还原并丰富轴承振动的动力学特性, 通过卷积神经网络(CNN)提取混沌序列中的高级抽象特征, 又考虑故障信号具有长程相关性, 将低维抽象故障特征引入长短期记忆网络(LSTM), 以灰狼算法优化的支持向量 机(OSVM)作为分类器, 提出CCNN (Chaotic CNN)–LSTM–OSVM智能故障诊断方法. 试验结果表明, 在处理信噪比 为??6 dB信号时, 该方法仍具有89.96%的准确率, 相比以Softmax作为分类器的CNN–LSTM和CCNN–LSTM方法分 别高出15.36%和5.21%, 且在收敛速度方面亦有较大优势.
英文摘要
      This paper addresses the problem of poor robustness of existing rolling bearing intelligent fault diagnosis methods under the background of large noise. Based on chaos theory, it is proposed to use the phase space reconstruction method to restore and enrich the dynamic characteristics of bearing vibration, to extract the high-level abstract features in the chaotic sequence through the convolution neural network (CNN), and to consider the long-range correlation of the fault signal. The low-dimensional abstract fault features are introduced into the long and short-term memory network (LSTM), and the support vector machine (OSVM) optimized by the gray wolf algorithm is used as the classifier, and the CCNN (Chaotic CNN)-LSTM-OSVM intelligent fault diagnosis method is proposed. The experimental results show that when the SNR is ??6 dB, the accuracy of this method is still 89.96%, which is 15.36% and 5.21% higher than the CNN–LSTM and CCNN–LSTM methods with Softmax as the classifier respectively. In addition, the convergence speed also has a great advantage.