不平衡训练数据下的基于生成对抗网络的轴承故障诊断  

Bearing Fault Diagnosis Based on Generative Adversarial Network on Imbalanced Data

作　　者：李忠智 尹航 左剑凯 刘鹤丹 LI Zhong-zhi;YIN Hang;ZUO Jian-kai;LIU He-dan(School of Computer Science,Shenyang Aerospace University,Shenyang 110136,China;School of Information Technology,Zhongkai University of Agriculture and Engineering,Guangzhou 510230,China;Department of Computer Science and Technology,Tongji University,Shanghai 201804,China)

机构地区：[1]沈阳航空航天大学计算机学院,沈阳110136 [2]仲恺农业工程学院信息科学与技术学院,广州510230 [3]同济大学计算机科学与技术系,上海201804

出　　处：《小型微型计算机系统》2021年第1期46-51,共6页Journal of Chinese Computer Systems

基　　金：国家航空基金项目(2015ZB54007)资助;辽宁省教育厅科技基金项目(L201704)资助。

摘　　要：近年来,随着深度学习模型及其衍生模型在故障诊断领域中的成功应用,基于深度学习的故障诊断方法开始成为研究主流.但是当训练数据不均衡时,通过深度学习从不平衡的数据中提取的故障特征是不准确的,训练得到的神经网络模型的分类结果往往倾向多数类,极大影响了分类效果.针对这种情况,本文结合卷积神经网络设计了一种新的生成对抗网络模型(Convolutional Wasserstein Generative Adversarial Network,CWGAN).首先卷积神经网络从故障样本中提取故障特征,并将其作为对抗网络的输入,然后由解码器网络解码来自生成器的故障特征向量来生成故障样本,同时将提取的故障特征和训练过程中的故障诊断误差添加至生成器训练的损失函数中.实验表明本文提出的方法相比于基线模型(GAN-CNN)的平均F1值提高4%,较好地解决数据不平衡的分类问题.In recent years,with the successful application of deep learning models and their derivative models in the field of fault diagnosis,deep learning-based fault diagnosis methods have become the mainstream of research.How ever,when the training data is imbalanced,the fault features extracted from the imbalanced data through deep learning are inaccurate.The classification results of the trained neural netw ork models tend to favor most classes,which greatly affects the classification effect.In view of this situation,a new generative adversarial netw ork model(Convolutional Wasserstein Generative Adversarial Netw ork,CWGAN)is designed in this paper in conjunction with convolutional neural netw orks.First,the convolutional neural netw ork extracts fault features from the fault samples and uses them as input to the adversarial netw ork.Then the decoder netw ork decodes the fault feature vectors from the generator to generate fault samples.The fault diagnosis error is added to the loss function trained by the generator.Experiments show that the method proposed in this paper improves the average F1 value by 4%compared w ith the baseline model(GAN-CNN),which can better solve the problem of imbalanced data classification.

关 键 词：卷积神经网络 生成对抗网络 不平衡数据集 故障诊断 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]