张玉彦, 张金龙, 文笑雨, 等.基于DBI-小波包分解和改进BP神经网络的轴承故障诊断方法研究[J].河南理工大学学报（自然科学版）,2023,42(1):116-123.

ZHANG Y Y, ZHANG J L, WEN X Y, et al.Study on bearing fault diagnosis method based on DBI-wavelet packet decomposition and improved BP neural network[J].Journal of Henan Polytechnic University(Natural Science) ,2023,42(1):116-123.

基于DBI-小波包分解和改进BP神经网络的轴承故障诊断方法研究

张玉彦, 张金龙, 文笑雨, 李浩, 孙春亚, 王昊琪, 乔东平

郑州轻工业大学 河南省机械装备智能制造重点实验室，河南 郑州 450002

摘要:针对轴承故障诊断问题，以振动信号为分析对象，提出DBI-小波包分解和改进BP神经网络的诊断方法。采用4层小波包分解获取振动信号的不同频带特征，引入戴维森堡丁指数（Davies-Bouldin index，DBI）定量评价小波包分解结果，确定小波基函数为FK22时达到最佳分解结果。采用改进的BP神经网络对不同频带特征识别，引入弹性梯度下降法解决传统BP神经网络收敛速度慢和梯度消失等问题，提高网络训练速度。同时，针对BP神经网络隐藏层层数及各隐藏层节点个数难以确定的问题，设计正交实验对不同参数组合效果进行验证，选出最佳参数，避免盲目低效调参。对电机滚动轴承进行验证，结果表明平均故障识别准确率达到98.833%。

关键词:轴承故障诊断;小波包分解;BP神经网络;戴维森堡丁指数

doi:10.16186/j.cnki.1673-9787.2021060096

基金项目:国家自然科学基金资助项目（52105536，51905494）；河南省重点研发与推广专项（科技攻关）项目（212102210072）；郑州轻工业大学博士科研基金资助项目（JDG20200109）；教育部人文社会科学研究青年基金资助项目（19YJCZH185）；河南省科技攻关项目（202102210088）；河南省高等学校重点科研计划项目（20A460029）

收稿日期:2021/06/28

修回日期:2021/08/20

出版日期:2023/01/25

Study on bearing fault diagnosis method based on DBI-wavelet packet decomposition and improved BP neural network

ZHANG Yuyan, ZHANG Jinlong, WEN Xiaoyu, LI Hao, SUN Chunya, WANG Haoqi, QIAO Dongping

Henan Key Laboratory of Intelligent Manufacturing Mechanical Equipment，Zhengzhou University of Light Industry，Zhengzhou 450002，Henan，China

Abstract:Aiming at bearing fault diagnosis，taking vibration signals as study objects，a novel method based on wavelet packet decomposition （WPT）and BP neural network was proposed.Vibration signals were fed into four-layer WPT for obtaining sub-frequency bands and Davies-Bouldin index（DBI） was employed to quantitatively evaluate the results of WPT.Optimal decomposition results were produced by using FK22 wavelet basis function.Improved BP neural network was used to recognize these sub-frequency band features.An elastic gradient descent method was introduced into BP neural network for alleviating the problems of slow convergence and gradient vanishing.Meanwhile，in order to determine the number of hidden layers and nodes number，orthogonal experiment was designed to verify different parameter combinations.Experiments were conducted on motor bearing and the results showed that average fault diagnosis accuracy reached as high as 98.833%.

Key words:bearing fault diagnosis;wavelet packet decomposition;BP neural network;Davies-Bouldin index

 014_2021060096_张玉彦_H.pdf