基于改进卷积神经网络CNND的电主轴轴承故障诊断方法  

Fault Diagnosis Method of Motorized Spindle Bearing Based on CNND

作　　者：石怀涛 乔思康[1] 丁健华 王子男 白晓天 SHI Huaitao;QIAO Sikang;DING Jianhua;WANG Zinan;BAI Xiaotian(School of Mechanical Engineering,Shenyang Jianzhu University,Shenyang,China,110168;National Engineering Laboratory of High-grade Stone Material Numerical Machining Equipment and Technology,Shenyang Jianzhu University,Shenyang,China,110168)

机构地区：[1]沈阳建筑大学机械工程学院,辽宁沈阳110168 [2]沈阳建筑大学高档石材数控加工装备与技术国家地方联合工程实验室,辽宁沈阳110168

出　　处：《沈阳建筑大学学报（自然科学版）》2020年第2期361-369,共9页Journal of Shenyang Jianzhu University：Natural Science

基　　金：国家自然科学基金项目(51705341,51905357,51675353);国家重点研发计划项目(2017YFC0703903);辽宁省自然科学基金项目(2016010623);沈阳市科技计划项目(F17-231-1-28、F16-096-1-00)。

摘　　要：目的解决深度学习方法在建立电主轴轴承故障诊断模型时出现的过拟合现象,提高电主轴轴承故障诊断准确率.方法提出一种基于改进卷积神经网络的诊断方法,该方法在卷积神经网络的训练过程中融入Dropout优化方法,使整个故障诊断模型按照一定的比例随机“关闭”隐藏层中的神经元,减少模型在每一次训练过程中所需要调整的参数数量.结果将Dropout优化方法与卷积神经网络相结合所建立的电主轴轴承故障诊断模型是可行的,其平均诊断准确率能够达到99.012%,远高于基于CNN、CNN+L2和BPNN这3种神经网络诊断方法的诊断准确率.卷积神经网络方法相比于传统的“基于信号处理提取到的特征和机器学习模型”方法,更适用于电主轴轴承故障诊断.结论提出的CNND方法实现了卷积神经网络与Dropout优化方法的有机结合,对原始数据进行降维处理使模型学习到的特征更利于电主轴故障的分类,同时根据故障数据的特点确定相关参数的初始值,克服一般深度学习方法在进行电主轴故障诊断时出现的过拟合现象,提高诊断准确率.A diagnosis method based on an improved convolutional neural network for fault diagnosis of motorized spindle bearing was proposed to solves over-fitting phenomenon in the deep-learning method occured when the fault diagnosis model of the motorized spindle bearing is established which improves its fault diagnosis accuracy.In this method the dropout optimization method was integrated into the convolutional neural network training process to make the entire fault diagnosis model randomly“Turn off”neurons in the hidden layer according to a certain proportion.So the number of parameters adjusted were reduced in the model during each training session.The results show that the fault diagnosis model based on the combination of dropout optimization method and convolutional neural network is feasible.The average diagnostic accuracy can reach 99.012%,which is much higher than other fault diagnosis methods which are CNN,CNN+L2 and BPNN.The convolutional neural network method is more suitable for the fault diagnosis of the motorized spindle bearing than the traditional method whose feature and machine learning model were obtained based on signal processing.The method proposed in this paper is the organic combination of convolutional neural network and Dropout optimization method.The feature of CNND obtained base on the original data processed by dimensionality reduction is more conducive to classify the fault of the motorized spindle bearing.Simultaneously,the initial value of the relevant parametersdeterminedbase on the fault data characteristics can improve the diagnostic accuracy by avoiding over-fitting phenomenon which happens in the fault diagnosis of spindle by general deep learning methods.

关 键 词：故障诊断 电主轴 深度学习 卷积神经网络 Dropout优化方法 

分 类 号：TH502.15[机械工程]