基于U-Net图像分割算法的相控阵超声缺陷图像定量检测方法  

Quantitative detection method for phased array ultrasonic defect images based on U-Net image segmentation algorithm

作　　者：靖珍珠 赵玉琦 涂思敏 陈振华[2] JING Zhenzhu;ZHAO Yuqi;TU Simin;CHEN Zhenhua(AECC Commercial Aircraft Engine Co.,Ltd.,Shanghai 201306,China;Key Laboratory of Nondestructive Testing of Ministry of Education,Nanchang Hangkong University,Nanchang 330063,China;Nuclear Industry Engineering Research and Design Co.,Ltd.,Beijing 101300,China)

机构地区：[1]中国航发商用航空发动机有限责任公司,上海201306 [2]南昌航空大学无损检测技术教育部重点实验室,南昌330063 [3]核工业工程研究设计有限公司,北京101300

出　　处：《无损检测》2025年第2期13-19,共7页Nondestructive Testing

基　　金：江西省重点研发计划(20212BBE51006)。

摘　　要：相控阵超声检测结果需经检测人员评定,存在主观性强、效率低、可靠性差等问题。据此,提出一种基于图像分割算法的相控阵超声缺陷智能定量方法。首先,采集平底孔缺陷图像并对其进行扩增处理以形成训练数据库;其次,构建并训练U-Net智能缺陷分割模型,使其能够在检测图像中自动分割缺陷和背景;接着,提出一种基于分割后的二值图像定量方法用于测量缺陷尺寸;最后,验证U-Net模型对相控阵缺陷图像的定量检测能力。结果表明,基于U-Net算法的缺陷分割模型的平均定量检测误差小于6%,能够达到与-6 dB法相似的定量检测能力,且具有效率高、智能化、易操作的优势。The results of phased array ultrasonic testing require the evaluation by inspectors,which poses some issues such as strong subjectivity,low efficiency,and poor reliability.To address these issues,an intelligent quantitative method for defect detection in phased array ultrasonic testing based on image segmentation algorithms was proposed.Firstly,defect images of flat-bottom holes were collected and enlarged to form a training database.Secondly,a U-Net intelligent defect segmentation model was constructed and trained to automatically segment defects from the background in the inspection images.Thirdly,a quantitative method based on the post-segmentation binary image was proposed for measuring defect dimensions.Finally,the quantitative detection capability of the U-Net model for phased array defects was verified.The results showed that the average quantitative detection error based on U-Net defect segmentation model was less than 6%,which could achieve the same quantitative detection ability as the-6 dB method,and had the advantages of high efficiency,intelligence and easy operation.

关 键 词：相控阵超声检测 智能定量 图像分割 缺陷 

分 类 号：TG115.28[金属学及工艺—物理冶金]