基于生成对抗网络的轴承不平衡数据研究  

Generate Adversarial Net-Based Unbalanced Data Analysis for Bearings

作　　者：韩辉[1] 徐赫 杨森林 程德权 Han Hui;Xu He;Yang Senlin;Cheng Dequan

机构地区：[1]沈阳理工大学,辽宁沈阳110159 [2]华晨鑫源重庆汽车有限公司,重庆408102

出　　处：《一重技术》2020年第5期40-46,共7页CFHI Technology

摘　　要：数据类别不平衡对轴承故障的诊断有很大影响,而生成对抗网络(Generate Adversarial Net)作为最热门的神经网络之一,在数据生成方面有巨大优势,对解决数据不平衡问题有很大帮助。通过GAN的生成器G和判别器D使生成数据无限接近于真实数据,而且数据分布与真实数据类似。再利用卷积神经网络(Convolutional Neural Networks)中的诊断分类对GAN生成的数据进行验证。利用西安交通大学的轴承数据集作为不平衡数据样本所做实验表明:由GAN模型生成的数据与真实数据的差别很小,能够有效解决实验样本数据量不足,种类不全,数据偏差较大等问题。Data category unbalance has great negative influence on the diagnosis of bearings failure.Generate Adversarial Net(GAN)as one of the most-popular neural network has significant advantage in data generation that can do more for solution of data unbalance problems.The GAN generator(G)and Discriminator(D)make the generated data infinitely near to true data and distributing similar to true data.After that,GAN generated data can be verified with the diagnosis and classification function of Convolutional Neural Networks.The test with the bearings data base of Xi'an Jiaotong University as unbalanced data shows that GAG generated data almost have no difference with true data and can solve problems about insufficient experimental sample data,incomplete data categories and large data difference.

关 键 词：轴承 生成对抗网络 不平衡分类 数据扩充 卷积神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TH133.3[自动化与计算机技术—控制科学与工程]