基于改进ResNet模型的轴承沟道表面加工缺陷分类  

Classification of Bearing Raceway Surface Processing Defects Based on Improved ResNet Model

作　　者：刘浩翰[1] 王钰涛 贺怀清[1] 孙铖 LIU Haohan;WANG Yutao;HE Huaiqing;SUN Cheng(Department of Computer Science & Technology,Civil Aviation University of China,Tianjin 300300,China)

机构地区：[1]中国民航大学计算机科学与技术学院,天津300300

出　　处：《轴承》2021年第7期52-58,共7页Bearing

摘　　要：轴承沟道表面缺陷具有细节丰富但特征不突出的特点,传统的特征提取方法进行缺陷分类时存在建模困难、分类准确率低的弊端,因此使用改进残差网络(ResNet)实现轴承沟道表面缺陷的高精度分类。以卷积神经网络为基础模型架构,使用残差块作为主要特征计算方法,在深层网络中融入Inception模块进行特征降维和拼接以获取更多的图像细节特征;同时,在特征计算中引入批量标准化(BN)进行数据正则化处理以加速模型收敛。对轴承沟道表面缺陷数据集的测试结果表明,改进ResNet模型比经典的LeNet5模型和ResNet模型精度更高且收敛性更好,准确率可达到98.84%,能够满足实际需要。The bearing raceway surface defects are characterized by abundant details but not prominent features.The traditional feature extraction method for defect classification has the disadvantages of difficult modeling and low classification accuracy,the improved Residual Network(ResNet)is used to realize the high-precision classification of bearing raceway surface defects.The convolutional neural network is adopted as basic model architecture,and the residual block is used as main feature calculation method.Inception module is integrated into deep network for feature reduction and splicing to obtain more detailed image features.At the same time,Batch Normalization(BN)is introduced into characteristic calculations for data regularization to accelerate model convergence.The test results of bearing raceway surface defect data set show that the improved ResNet model has higher accuracy and better convergence than classical Lenet5 model and ResNet model,and the accuracy reaches 98.84%,which can meet the actual needs.

关 键 词：滚动轴承 沟道 表面缺陷 分类 卷积神经网络 残差 

分 类 号：TH133.33[机械工程—机械制造及自动化] TP391[自动化与计算机技术—计算机应用技术]