焊缝缺陷超声图谱的卷积神经网络分类研究

Ultrasonic Image Classification of Weld Defects Based on Convolution Neural Network

针对传统的工件焊缝缺陷分类方法分类准确率低、分类速度慢的问题,提出了基于迁移学习的卷积神经网络实现对工件焊缝缺陷图谱的智能分类。首先使用超声相控阵探伤仪对焊缝试块进行扫描实现数据采集,然后将采集好的各类缺陷图谱数据按9:1的比例分为训练集与测试集,最后使用基于迁移学习的ResNet-34、MobileNet-v2、AlexNet三种卷积神经网络模型训练。其中Resnet-34网络模型最高准确率可达到98.6%,MobileNet-v2网络模型的最高准确率达到84.5%,AlexNet网络模型的最高准确率达到96.5%。试验结果表明,使用基于迁移学习的卷积神经网络不仅可以有效地提高件焊缝缺陷图谱的分类准确率,而且分类的速度也远比人工分类的速度要快,有效的加快了缺陷图谱的分类速度。

Aiming at the problems of low accuracy and slow classification speed of traditional classification methods about work⁃piece weld defects.In this paper,a convolutional neural network based on transfer learning was proposed to realize the intelligent classification of weld defect atlas.Firstly,the ultrasonic phased array detector was used to scan the weld block to realize data ac⁃quisition.Then,the collected data are divided into training set and test set according to the ratio of 9:1.Finally,three convolu⁃tional neural network models,ResNet-34,MobileNet-v2 and AlexNet,are used for training.The highest accuracy of Resnet-34 network model can reach 98.6%,MobileNet-v2 network model can reach 84.5%,and the highest accuracy of AlexNet network model was 96.5%.The results show that the convolution neural network based on transfer learning can not only improve the classi⁃fication accuracy of weld defect atlas effectively,but also is much faster than manual classification,which the classification speed about defect atlas was speed.