• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2019, Vol. 55 ›› Issue (7): 35-43.doi: 10.3901/JME.2019.07.035

• 基于深度学习的机械装备故障预测与健康管理 • 上一篇    下一篇

基于终止准则改进K-SVD字典学习的稀疏表示特征增强方法

王华庆, 任帮月, 宋浏阳, 董方, 王梦阳   

  1. 北京化工大学机电工程学院 北京 100029
  • 收稿日期:2018-05-28 修回日期:2018-11-16 出版日期:2019-04-05 发布日期:2019-04-05
  • 通讯作者: 王华庆(通信作者),男,1973年出生,博士,教授,博士研究生导师。主要研究方向为机械装备健康监测及故障智能诊断、信号处理与特征提取。E-mail:hqwang@mail.buct.edu.cn
  • 作者简介:任帮月,女,1994年出生,硕士研究生。主要研究方向为机械故障诊断、稀疏表示。E-mail:rby_buct@163.com;宋浏阳,女,1988年出生,博士后。主要研究方向为机械故障诊断。E-mail:xq_0703@163.com;董方,男,1993年出生,硕士研究生。主要研究方向为机械故障诊断、动力学仿真。E-mail:doveel1028@163.com;王梦阳,男,1992年出生,硕士研究生。主要研究方向为机械故障诊断、盲源分离。E-mail:mengyang1992@163.com
  • 基金资助:
    国家自然科学基金资助项目(51675035,51805022)。

Sparse Representation Method Based on Termination Criteria Improved K-SVD Dictionary Learning for Feature Enhancement

WANG Huaqing, REN Bangyue, SONG Liuyang, DONG Fang, WANG Mengyang   

  1. School of Mechanical and Electrical Engineering, Beijing University of Chemical Technology, Beijing 100029
  • Received:2018-05-28 Revised:2018-11-16 Online:2019-04-05 Published:2019-04-05

摘要: 针对传统K奇异值分解(K-Singular value decomposition,K-SVD)算法在稀疏表示过程中,由于目标信号稀疏度难以确定以及字典原子受噪声干扰大导致稀疏表示效果较差的问题,结合变分模态分解(Variational mode decomposition,VMD)算法,提出了基于VMD与终止准则改进K-SVD字典学习的稀疏表示方法。借助VMD算法剔除信号中的干扰分量,依据相关分析与峭度准则选择最优模态分量;采用终止准则改进的K-SVD字典学习算法对最优分量的特征信息进行学习,优化目标函数与约束条件,在无需设置稀疏度的前提下,构造出准确匹配故障冲击成分的字典;此外,构建一种残差阈值改进的正交匹配追踪算法(OMPerr)实现稀疏重构及微弱故障特征增强。通过仿真及试验信号进行验证,结果表明:基于VMD与改进K-SVD字典学习的稀疏表示方法在字典原子构建、稀疏重构精度以及故障特征增强等方面均优于传统K-SVD稀疏表示方法,可以有效实现微弱故障的诊断。

关键词: K奇异值分解, 故障特征增强, 特征提取, 稀疏表示

Abstract: A sparse representation method based on VMD and termination criteria improved K-SVD dictionary learning algorithm is proposed to solve the issues about the choice of signal sparsity and the interference of noise. With the aid of VMD algorithm, the interference components can be removed. According to correlation analysis and kurtosis criterion, the optimal modal component can be selected successfully. Then the characteristic information of the optimal component is learned by termination criteria improved K-SVD algorithm, optimizing the objective function and constraints, and the sparse representation dictionary matched the fault impact components can be constructed without setting the sparsity. In addition, an improved orthogonal matching pursuit algorithm with residual error threshold is constructed to achieve sparse reconstruction and weak fault feature enhancement. Verification by simulated and experimental signals show that the sparse representation method based on VMD and modified K-SVD could effectively diagnose the weak fault, which outperformed the traditional K-SVD algorithm in terms of the construction of dictionary atom, sparse reconstruction accuracy and fault feature enchantment.

Key words: fault feature enhancement, feature extraction, K-singular value decomposition, sparse representation

中图分类号: