基于机器学习和超声成像的缺陷识别与分析  

Defect recognition and analysis based on machine learning and ultrasonic imaging

作　　者：李灏天 刘晓宙 何爱军[3] Li Haotian;Liu Xiaozhou;He Aijun(School of Physics,Nanjing University,Nanjing,210023,China;Key Laboratory Modern Acoustics,MOE,Institut Acoustics,Nanjing University,Nanjing,210093,China;School of Electronic and Science Engineering,Nanjing University,Nanjing,210023,China)

机构地区：[1]南京大学物理学院,南京210093 [2]近代声学教育部重点实验室,南京大学声学研究所,南京210093 [3]南京大学电子科学与工程学院,南京210023

出　　处：《南京大学学报（自然科学版）》2022年第4期670-679,共10页Journal of Nanjing University（Natural Science）

基　　金：国家重大科学研究计划(2020YFA0211400,2016YFF0203000);国家自然科学基金(12174192,11774167);2021年大学生创新创业训练计划(202110284062Y)。

摘　　要：为了量化分析样本中的缺陷,利用卷积神经网络(Convolutional Neural Network,CNN)结合阈值分割和深度优先搜索算法实现了对超声检测图像中样品内部缺陷的识别,将实际图像输入神经网络模型中,成功完成了对缺陷的标记,验证了模型的可靠性.利用Field Ⅱ对全矩阵捕获(Full Matrix Capture,FMC)过程及对数据以全聚焦成像方法(Total Focus Method,TFM)进行成像的过程进行了仿真模拟,获得了可用于机器学习的数据集.基于方向梯度直方图(Histogram of Oriented Gradient,HOG)提取了全聚焦成像结果图的图像特征,利用改进的支持向量机(Support Vector Machine,SVM)获得由图像预测缺陷半径的模型并对该模型进行了评价.结果表明,利用上述方法提取的缺陷半径信息精准度在0.1 mm,能够应用于缺陷半径的量化分析,预测误差主要来源于数据集两端,可以通过预处理算法进一步提升检测精度.In order to achieve quantitative analysis of defects in the samples,the convolution neural network combined with threshold segmentation and depth-first search algorithm was used to realize the identification of internal defects of samples in ultrasonic detection images. The actual images were input into the neural network model and the defects were marked successfully,which verified the reliability of the model. Field Ⅱ was used to simulate the process of full matrix capture and the process of data imaging by total focus method,and the data set that can be used for machine learning was obtained. Based on histogram of oriented gradient to extract the imaging result focused image characteristics,all of them,and the method of using improved support vector machine(SVM) of machine learning to predict defect radius from the model and the model was evaluated, the results show that using the above methods to extract the defect radius information accuracy in the order of magnitude 0.1 mm,can be applied to the quantitative analysis of defects radius,The prediction errors mainly come from the two ends of the data set,and the detection accuracy can be further improved by the preprocessing algorithm.

关 键 词：超声相控阵 卷积神经网络 方向梯度直方图 支持向量机 机器学习 

分 类 号：O426.9[理学—声学]