Three-Dimensional Defect Characterization of Ultrasonic Detection Based on GCNet Improved Contrast Learning Optimization

Author:

Wang Xinghao1ORCID,Wang Qiang1,Zhang Lei1,Yu Jiayang1,Liu Qiuhan1

Affiliation:

1. Equipment Management and Unmanned Aerial Vehicle Engineering School, Air Force Engineering University, Xi’an 710051, China

Abstract

In order to automate defect detection with few samples using unsupervised learning, this paper, considering materials commonly used in aircraft, proposes a phased array ultrasonic detection defect identification method using non-defect samples for training, and three-dimensional characterization is completed on this basis. A phased array ultrasonic device was used to detect two typical structures: a carbon fiber composite cylinder structure and a metal L-shaped structure. No damage label image was required, and the non-damaged sample was used as the the network training input. Based on contrast learning and the cross-registration loss of common features, a feature-matching network was constructed to extract the common features of undamaged detection data, and the performance was optimized by combining STN and GCNet modules. When the detection data of the sample were input to the aforementioned network, the defect distribution representing the location and rough shape of the defect was obtained through Mahalanobis distance calculation. The length was estimated using the S-scan image sequence sampling method. Additionally, the depth of the hole was estimated by combining the B-scan data with line recognition. According to the original model of the sample, the 3D characterization of defects was completed by pyautocad. In the experimental stage, three ablation experiments were carried out to verify the necessity of each module, and performance comparisons were mainly evaluated by F1 score and visualization using four existing well-known anomaly detection methods.

Publisher

MDPI AG

Subject

Electrical and Electronic Engineering,Computer Networks and Communications,Hardware and Architecture,Signal Processing,Control and Systems Engineering

Cited by 1 articles. 订阅此论文施引文献 订阅此论文施引文献,注册后可以免费订阅5篇论文的施引文献,订阅后可以查看论文全部施引文献

同舟云学术

1.学者识别学者识别

2.学术分析学术分析

3.人才评估人才评估

"同舟云学术"是以全球学者为主线,采集、加工和组织学术论文而形成的新型学术文献查询和分析系统,可以对全球学者进行文献检索和人才价值评估。用户可以通过关注某些学科领域的顶尖人物而持续追踪该领域的学科进展和研究前沿。经过近期的数据扩容,当前同舟云学术共收录了国内外主流学术期刊6万余种,收集的期刊论文及会议论文总量共计约1.5亿篇,并以每天添加12000余篇中外论文的速度递增。我们也可以为用户提供个性化、定制化的学者数据。欢迎来电咨询!咨询电话:010-8811{复制后删除}0370

www.globalauthorid.com

TOP

Copyright © 2019-2024 北京同舟云网络信息技术有限公司
京公网安备11010802033243号  京ICP备18003416号-3