Automated detection of weld defects in TOFD images for steel bridges using generative adversarial networks

Time-of-flight diffraction (TOFD) has been widely adopted for weld defect detection in bridge steel quality assurance due to its harmlessness to the human body, real-time performance, and satisfactory detection accuracy. Most deep learning-based methods for automated weld defect recognition rely hea...

Full description

Saved in:
Bibliographic Details
Main Authors: Yanfeng Gong, Zihao Chen, Hong Zhang, Meng Xu, Wen Deng
Format: Article
Language:English
Published: Elsevier 2025-07-01
Series:Case Studies in Construction Materials
Subjects:
Online Access:http://www.sciencedirect.com/science/article/pii/S2214509525006394
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1850182871645421568
author Yanfeng Gong
Zihao Chen
Hong Zhang
Meng Xu
Wen Deng
author_facet Yanfeng Gong
Zihao Chen
Hong Zhang
Meng Xu
Wen Deng
author_sort Yanfeng Gong
collection DOAJ
description Time-of-flight diffraction (TOFD) has been widely adopted for weld defect detection in bridge steel quality assurance due to its harmlessness to the human body, real-time performance, and satisfactory detection accuracy. Most deep learning-based methods for automated weld defect recognition rely heavily on defect samples. In practice, such samples are scarce in steel bridge applications because modern welding technologies significantly reduce defect occurrence. To address this limitation, we propose a two-stage defect detection method for TOFD weld images based on an enhanced generative adversarial network (GAN) that operates without requiring defect-containing samples. In the first stage, the Region of Interest (ROI) which contains potential defects is localized using YOLOv8. In the second stage, the extracted ROI is sliced into patches and analyzed by a GAN architecture enhanced with a self-attention mechanism, which improves the encoding and aggregation of local defect features. The integration of the self-attention GAN with the slicing strategy further enhances defect recognition performance. The proposed method is evaluated on a self-constructed TOFD dataset of steel bridge welds. Experimental results demonstrate that our approach achieves an AUC of 86 %, outperforming existing state-of-the-art methods by 5 %.
format Article
id doaj-art-aa067ba628a4420cb3951a70d05fedba
institution OA Journals
issn 2214-5095
language English
publishDate 2025-07-01
publisher Elsevier
record_format Article
series Case Studies in Construction Materials
spelling doaj-art-aa067ba628a4420cb3951a70d05fedba2025-08-20T02:17:29ZengElsevierCase Studies in Construction Materials2214-50952025-07-0122e0484110.1016/j.cscm.2025.e04841Automated detection of weld defects in TOFD images for steel bridges using generative adversarial networksYanfeng Gong0Zihao Chen1Hong Zhang2Meng Xu3Wen Deng4School of Shipping and Naval Architecture, Chongqing Jiaotong University, Chongqing 400074, ChinaSchool of Shipping and Naval Architecture, Chongqing Jiaotong University, Chongqing 400074, ChinaState Key Laboratory of Mountain Bridge and Tunnel Engineering, Chongqing Jiaotong University, Chongqing 400074, China; Correspondence to: Chongqing Jiaotong University, Chongqing 400074, China.Jianshe Industry Group Yunnan Co.,Ltd., Chongging 400054, ChinaSchool of Shipping and Naval Architecture, Chongqing Jiaotong University, Chongqing 400074, ChinaTime-of-flight diffraction (TOFD) has been widely adopted for weld defect detection in bridge steel quality assurance due to its harmlessness to the human body, real-time performance, and satisfactory detection accuracy. Most deep learning-based methods for automated weld defect recognition rely heavily on defect samples. In practice, such samples are scarce in steel bridge applications because modern welding technologies significantly reduce defect occurrence. To address this limitation, we propose a two-stage defect detection method for TOFD weld images based on an enhanced generative adversarial network (GAN) that operates without requiring defect-containing samples. In the first stage, the Region of Interest (ROI) which contains potential defects is localized using YOLOv8. In the second stage, the extracted ROI is sliced into patches and analyzed by a GAN architecture enhanced with a self-attention mechanism, which improves the encoding and aggregation of local defect features. The integration of the self-attention GAN with the slicing strategy further enhances defect recognition performance. The proposed method is evaluated on a self-constructed TOFD dataset of steel bridge welds. Experimental results demonstrate that our approach achieves an AUC of 86 %, outperforming existing state-of-the-art methods by 5 %.http://www.sciencedirect.com/science/article/pii/S2214509525006394Weld defect detectionTOFDGANSlicing technique
spellingShingle Yanfeng Gong
Zihao Chen
Hong Zhang
Meng Xu
Wen Deng
Automated detection of weld defects in TOFD images for steel bridges using generative adversarial networks
Case Studies in Construction Materials
Weld defect detection
TOFD
GAN
Slicing technique
title Automated detection of weld defects in TOFD images for steel bridges using generative adversarial networks
title_full Automated detection of weld defects in TOFD images for steel bridges using generative adversarial networks
title_fullStr Automated detection of weld defects in TOFD images for steel bridges using generative adversarial networks
title_full_unstemmed Automated detection of weld defects in TOFD images for steel bridges using generative adversarial networks
title_short Automated detection of weld defects in TOFD images for steel bridges using generative adversarial networks
title_sort automated detection of weld defects in tofd images for steel bridges using generative adversarial networks
topic Weld defect detection
TOFD
GAN
Slicing technique
url http://www.sciencedirect.com/science/article/pii/S2214509525006394
work_keys_str_mv AT yanfenggong automateddetectionofwelddefectsintofdimagesforsteelbridgesusinggenerativeadversarialnetworks
AT zihaochen automateddetectionofwelddefectsintofdimagesforsteelbridgesusinggenerativeadversarialnetworks
AT hongzhang automateddetectionofwelddefectsintofdimagesforsteelbridgesusinggenerativeadversarialnetworks
AT mengxu automateddetectionofwelddefectsintofdimagesforsteelbridgesusinggenerativeadversarialnetworks
AT wendeng automateddetectionofwelddefectsintofdimagesforsteelbridgesusinggenerativeadversarialnetworks