基于格拉姆角场和迁移深度残差神经网络的滚动轴承故障诊断  

Rolling bearing fault diagnosis based on Gram angle field and transfer deep residual neural network

作　　者：古莹奎[1] 吴宽 李成[1] GU Yingkui;WU Kuan;LI Cheng(School of Mechanical and Electrical Engineering,Jiangxi University of Science and Technology,Ganzhou 341000,China)

机构地区：[1]江西理工大学机电工程学院,江西赣州341000

出　　处：《振动与冲击》2022年第21期228-237,共10页Journal of Vibration and Shock

基　　金：国家自然科学基金(61963018);江西省自然科学基金(20181BAB202020);江西省研究生创新资金项目(YC2020-S466)。

摘　　要：针对应用传统卷积神经网络进行故障诊断时存在的过拟合现象及传统灰度图编码存在的时间信息损失等问题,提出一种基于格拉姆角场图像编码和迁移深度残差神经网络相结合的滚动轴承故障诊断方法。依据格拉姆角场图像编码方法对时间序列编码映射的唯一性,将原始振动信号转化为格拉姆角差场图和格拉姆角和场图,并将在ImageNet上预训练好的ResNet18模型参数,迁移到以格拉姆角场图作为输入的ResNet18中,进行不同故障模式下格拉姆角场图的特征提取和分类,从而达到故障诊断的目的。分析结果表明,所提方法相比于传统灰度图编码,更能突出不同故障模式的内在特征,与传统卷积神经网络模型相比,提出的方法具有更高的识别精度,达到99.30%,且具有更快的收敛速度和更强的鲁棒性。Here,aiming at problems of over-fitting occurring when using traditional convolutional neural network for fault diagnosis and time domain information losing in traditional gray image coding,a rolling bearing fault diagnosis method based on combination of Gram angle field image coding and transfer deep residual neural network was proposed.According to the uniqueness of Gram angle field image coding method to time series coding mapping,the original vibration signal was converted into Gram angle difference field diagram and Gram angle sum field diagram.ResNet18 model parameters pre-trained on ImageNet were transferred to ResNet18 with Gram angle field diagram as input to do feature extraction and classification of Gram angle field diagram under different fault modes,andrealize fault diagnosis.The analysis results showed that compared with traditional gray image coding,the proposed method can better highlight intrinsic characteristics of different fault modes;compared with traditional convolutional neural network model,the proposed method can havea higher recognition accuracy of 99.30%,and a faster convergence speed and stronger robustness.

关 键 词：格拉姆角场 图像编码 迁移深度残差神经网络 滚动轴承 故障诊断 

分 类 号：TH17[机械工程—机械制造及自动化]