卷积神经网络的缺陷类型识别分析  

Recognition and analysis of defect types by convolutional neural network

作　　者：高子洋 师芳芳 张碧星[1,2] 苏业旺 GAO Ziyang;SHI Fangfang;ZHANG Bixing;SU Yewang(University of Chinese Academy of Sciences,Beijing 100049,China;State Key Laboratory of Acoustics,Institute of Acoustics,Chinese Academy of Sciences,Beijing 100190,China;State Key Laboratory of Nonlinear Mechanics Institute of Mechanics,Chinese Academy of Sciences,Beijing 100190,China;School of Engineering Science,University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院大学,北京100049 [2]中国科学院声学研究所声场声信息国家重点实验室,北京100190 [3]中国科学院力学研究所非线性力学国家重点实验室,北京100190 [4]中国科学院大学工程科学学院,北京100049

出　　处：《应用声学》2022年第2期301-309,共9页Journal of Applied Acoustics

基　　金：中国科学院声学研究所青年英才计划项目(QNYC201736);中国科学院基础前沿科学研究计划(ZDBS-LY-7023);非线性力学国家重点实验室开放基金资助项目,声场声信息国家重点实验室开放课题研究基金项目(SKLA202109)。

摘　　要：该文提出一种基于卷积神经网络直接对阵列超声检测原始信号进行缺陷类型识别的方法,该方法无需对超声回波原始信号进行特征提取。文章研究对比了不同卷积神经网络及其优化的识别性能。首先采用超声相控阵系统对不同试块上的平底孔、球底孔、通孔三种缺陷进行超声检测,然后利用LeNet5、VGG16和ResNet三种卷积神经网络对一维和二维数据分别进行缺陷识别,并使用Leaky ReLU、Dropout、Batch Normalization等来对网络进行优化,对比分析识别准确率和效率等。研究结果表明,虽然一维卷积在训练速度上快于二维,但是二维卷积在识别准确度方面更高,同时网络模型结构如果过于复杂会导致准确率的下降,数据增强和优化方式有助于加快收敛速度、提高准确率。In this paper,a method of the defect type recognition based on the convolutional neural network is proposed by the original signal from the array ultrasonic testing.The method does not require feature extraction of the original ultrasonic echo signals.The recognition performance and optimization characteristic of different convolutional neural networks are studied and analyzed.Firstly,the experiments are conducted by the ultrasonic phased array system to collect detection signals which comes from the flat-bottom hole,spherical-bottom hole and via hole.Then,Lenet5,VGG16 and Resnet convolutional neural networks are used to identify the defects in one and two dimensions,and the network is optimized by Leaky ReLU,Dropout and Batch Normalization respectively.Subsequently,the identification accuracy and efficiency are analyzed.It is shown that 2-d convolution has higher recognition accuracy although its training speed is slower than that of 1-d convolution.At the same time,the identification accuracy will decline if the network model structure is too complex,and the data enhancement and optimization methods can help to accelerate the convergence speed and improve the accuracy.

关 键 词：卷积神经网络 超声检测 缺陷类型识别 

分 类 号：O429[理学—声学]