基于振动图像和卷积神经网络的滚动轴承故障诊断方法研究  

Research on Rolling Bearing Fault Diagnosis Method Based on Vibration Images and Convolutional Neural Networks

作　　者：李智皓 鲁殿君 吕荣水 刘伟 Lu Dianjun;Lyu Rongshui;Li Zhihao

机构地区：[1]中车青岛四方机车车辆股份有限公司

出　　处：《工程机械》2023年第5期60-66,I0014,共8页Construction Machinery and Equipment

基　　金：国家工信部智能制造综合标准化与新模式应用:高速动车组智能工厂运行管理标准(2016213)。

摘　　要：机械振动会反应出机械内部的运行状态,因此,可以通过振动分析识别设备健康状态和使用寿命。采用振动分析的方法对滚动轴承的运行状态进行评估,提出基于振动特征图像结合卷积神经网络(Convolutional Neural Networks,CNN)的滚动轴承故障诊断方法。通过移动重排方法,将一维振动信号表示为二维振动特征图像,通过逐层特征提取将数据样本从原空间的特征变换到一个新的特征空间来表示初始数据,并结合CNN方法构建滚动轴承转动精度评定的深层学习模型,实现样本深层次特征的提取和分类。研究结果表明,采用振动信号结合网格筛选(Grid Search,GS)优化后的浅层学习支持向量机(Support Vector Machines,SVM)模型测试集的准确率和F值分别为89.5%和89.33%;振动特征图像结合深度学习C NN模型测试集的准确率和F值可达1 0 0%。一维振动信号转化为二维图像后结合深度学习方法可以较好地对滚动轴承的振动信号进行信息提取和判别模型构建。Mechanical vibration reflects the operating conditions within the machinery,so vibration analysis can be used to identify the health status and service life of the equipment.A vibration analysis method is used to evaluate the operating conditions of rolling bearings,and a rolling bearing fault diagnosis method based on vibration feature images and Convolutional Neural Networks(CNN) is proposed.By movement and rearrangement,one-dimensional vibration signals are represented as two-dimensional vibration feature images.Through layer-by-layer feature extraction,the data samples are transformed from the original space features to a new feature space to represent the initial data,and in conjunction with CNN method,a deep learning model for evaluating the rotation accuracy of rolling bearings is constructed to achieve the extraction and classification of deep features of the samples.The research results show that the accuracy and F-value of the test set of the shallow learning Support Vector Machine(SVM) model optimized by combining vibration signals with GridSearch(GS) are 89.5% and 89.33% respectively,while the accuracy and F-value of the test set of the combination of vibration feature images and the deep learning CNN model are up to 100%.The method of transforming onedimensional vibration signals into twodimensional images and then combining deep learning can well extract the information of and construct the discriminative model of the vibration signals of rolling bearings.

关 键 词：振动信号 振动特征图像 卷积神经网络 故障诊断 

分 类 号：TH133.33[机械工程—机械制造及自动化] TP277[自动化与计算机技术—检测技术与自动化装置]