一维多尺度卷积神经网络及其在滚动轴承故障诊断中的应用  

One-dimensional Multi-scale Convolution Neural Network and its Application in Rolling Bearing Fault Diagnosis

作　　者：张成帆 江泽鹏 曹伟 陈伟[3] 张敏[1,2] ZHANG Chengfan;JIANG Zepeng;CAO Wei;CHEN Wei;ZHANG Min(School of Mechanical Engineering,Southwest Jiaotong University,Chengdu 610031,China;Technology and Equipment of Rail Transit Operation and Maintenance Key Laboratory of Sichuan Province,Chengdu 610031,China;Tribological Design Laboratory of Shield/TBM Equipment,Southwest Jiaotong University,Chengdu 610031,China)

机构地区：[1]西南交通大学机械工程学院,成都610031 [2]轨道交通运维技术与装备四川省重点实验室,成都610031 [3]西南交通大学盾构/TBM装备摩擦学设计实验室,成都610031

出　　处：《机械科学与技术》2022年第1期120-126,共7页Mechanical Science and Technology for Aerospace Engineering

基　　金：中国博士后科学基金项目(2020M673279);国家自然科学基金项目(51675450);四川省科技计划项目(2020JDTD0012);教育部人文社会科学研究青年基金项目(18YJC630255)。

摘　　要：为了有效利用来自实际生产中监测系统的海量数据,并结合一维卷积网络在处理一维数据的优势,提出一种端到端的一维多尺度卷积神经网络滚动轴承故障诊断方法。首先使用两个一维卷积层和池化层将输入振动信号的长度缩减并增加通道数,然后利用多尺度并行一维卷积核对上层输出特征进行不同尺度上的反复提取和重构,最后将提取到的特征输入到一个全连接层进行故障分类。为验证算法的有效性,通过对滚动轴承不同工况、不同训练样本以及与支持向量机、BP神经网络和循环神经网络等算法对比分析。结果表明提出的模型及方法具有较好的识别效果,滚动轴承故障诊断正确率达到99.78%。In order to use the massive data from the monitoring system in actual production effectively,and combine the advantages of one-dimensional convolution network in processing one-dimensional data,a new rolling bearing fault diagnosis method based on end-to-end one-dimension multi-scale convolution neural network is proposed.Firstly,two one-dimensional convolution layers and pooling layers are used to reduce the length of the input vibration signal and increase the number of channels.Then,multi-scale parallel one-dimensional convolution check is used to extract and reconstruct the output features on different scales repeatedly.Finally,the extracted features are input to a full connection layer for fault classification.In order to verify the effectiveness of the method,by comparing and analyzing the different working conditions,different training samples of rolling bearing and other algorithms such as support vector machine,BP neural network and cyclic neural network,the simulation results show that the proposed model and method has better recognition effect,and the accuracy of rolling bearing fault diagnosis reaches 99.78%.

关 键 词：滚动轴承 故障诊断 一维卷积网络 深度学习 

分 类 号：TP277[自动化与计算机技术—检测技术与自动化装置]