摘要
以轴承为例,对卷积神经网络在故障诊断领域中的可解释性进行了探讨,采用Grad-CAM方法,基于可视化的角度建立了神经网络的重点激活区域与目标类别之间的联系,并且利用凯斯西储大学的轴承数据库,分别从时域和频域的角度对LeNet,AlexNet和ResNet-18这3种应用较广的卷积神经网络结构进行了验证,结果表明,卷积神经网络在轴承故障诊断领域中对于样本的分类识别与人为的认知规律存在基本的相似性,可以为卷积神经网络在故障诊断领域的工程应用提供参考。
Taking the bearings as examples,the interpretability of convolutional neural network(CNN)in the field of fault diagnosis is discussed.The Grad-CAM method is applied to establish the relationship between key activation area of CNN and target category based on perspective of visualization.By utilizing bearing dataset from Case Western Reserve University,the validation is carried out from perspective of time domain and frequency domain via three widely used CNN structures:LeNet,AlexNet and ResNet-18.The results indicate that there is a basic similarity between classification and identifucation of samples by CNN and human recognition rules in the field of bearing fault diagnosis,which provide a reference for application of CNN in the field of fault diagnosis.
作者
张俊鹏
杨志勃
陈雪峰
翟智
刘一龙
ZHANG Junpeng;YANG Zhibo;CHEN Xuefeng;ZHAI Zhi;LIU Yilong(State Key Laboratory for Manufacturing Systems Engineering,Xi′an Jiaotong University,Xi′an 710049,China;School of Mechanical Engineering,Xi′an Jiaotong University,Xi′an 710049,China)
出处
《轴承》
北大核心
2020年第7期54-60,共7页
Bearing
基金
国家自然科学基金项目(51875433)。
关键词
滚动轴承
故障诊断
卷积神经网络
可视化
rolling bearing
fault diagnosis
convolution neural network
visualization