摘要
针对相控阵超声检测(Phased Array Ultrasonic Testing,PAUT)图像缺陷的人工识别结果准确性和一致性难以保证的问题,本文提出一种基于多层神经网络的PAUT图像缺陷识别方法。在碳钢平板试件中设置长度为2mm~4mm,倾斜角度为0°~60°的裂纹,以及直径为2mm~5mm的横通孔,并进行PAUT检测与S扫描图像采集。构建用于训练和测试神经网络的数据集,其中训练集包含13500个样本,测试集包含1874个样本。基于多层卷积神经网络,对PAUT图像实施缺陷搜索,并区分体积型和面积型。根据分类结果加载不同参数进行语义分割,实现缺陷特征还原。研究结果表明,该方法的缺陷识别准确率达到94.53%,缺陷特征还原像素准确率(Class Pixel Accuracy,CPA)平均值与交并比(Intersection of Union,IoU)平均值分别为99.92%与0.979,且训练和预测过程十分高效。
To overcome the difficulty in inaccurate and inconsistent manual identification of the flaws from phased array ultrasonic testing(PAUT)images,a defect identification method is proposed based on the multilayer neural network.Artificial cracks with 2 mm~4 mm in size and 0°~60°in inclined angle and side-drilled holes with 2 mm~5 mm in diameter were set in carbon steel plate specimen.PAUT inspection was performed to collect S-scan images.The dataset was established for neural network training and testing,which included 13500 samples for training and 1874 samples for testing.Based on the multilayer neural network,defects were searched from PAUT images and distinguished into volumetric or planar flaws.According to classification results,different parameters were loaded to implement semantic segmentation for restoring flaw characteristics.The results show that the accuracy of defect identification is 94.53%,the average class pixel accuracy(CPA)of flaw feature restoration is 99.92%,and the average intersection of union(IoU)is 0.979.Moreover,training and prediction with the multilayer neural network are quite efficient.
作者
程燕岷
林莉
廖静瑜
张晓峰
杨会敏
张东辉
金士杰
CHENG Yanmin;LIN Li;LIAO Jingyu;ZHANG Xiaofeng;YANG Huimin;ZHANG Donghui;JIN Shijie(NDT&E Laboratory,Dalian University of Technology,Dalian 116024,China;China Nuclear Industry 23 Construction CO.,LTD.,Beijing 101300,China)
出处
《无损探伤》
2022年第6期6-10,共5页
Nondestructive Testing Technology
关键词
相控阵超声检测
缺陷识别
语义分割
多层神经网络
Phased Array Ultrasonic Testing(PAUT)
Defect Identification
Semantic Segmentation
Multilayer Neural Network