Balanced Semi-Supervised Generative Adversarial Network for Damage Assessment from Low-Data Imbalanced-Class Regime

Oct 23, 2026

Vision-based structural health monitoring increasingly relies on deep learning for damage assessment, yet practical deployment is constrained by insufficient data and imbalanced classes. Existing remedies such as transfer learning and over/under-sampling often yield case-dependent gains. This work presents a balanced semisupervised GAN (BSS-GAN) that combines semi-supervised learning with balanced-batch sampling to address low-data imbalanced-class regimes. Experiments on concrete cracking and spalling classification show improved recall and F-score compared with conventional methods, demonstrating state-of-the-art performance.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages20

Reading Time50 min

AccessFree to download

Background & Motivations

arXiv :2211 . 15961v1 [ cs .LG] 29 Nov 2022 DOI: [https://doi.org/10.1111/mice.12741](https://doi.org/10.1111/mice.12741) ORIGINAL ARTICLE Balanced Semi-Supervised Generative Adversarial Network for Damage Assessment from Low-Data Imbalanced-Class Regime Yuqing Gao 1,2 | Pengyuan Zhai 1,3 | Khalid M. Mosalam* 1,2,4 1Department of Civil and Environmental Engineering, University of California, Berkeley, CA, USA 2Tsinghua-Berkeley Shenzhen Institute (TBSI), Tsinghua University, Shenzhen, China 3Department of Industrial Engineering and Operations Research, University of California, Berkeley, CA, USA 4Paciﬁc Earthquake Engineering Research Center, University of California, Berkeley, CA, USA *Correspondence Khalid M. Mosalam, 723 Davis Hall, University of California, Berkeley, CA 94720-1710, USA. Email: [mosalam@berkeley.edu](mosalam@berkeley.edu) Summary In recent years, applying deep learning (DL) to assess structural damages has gained growing popularity in vision-based structural health monitoring (SHM) . However, both data deﬁciency and class-imbalance hinder the wide adoption of DL in practical applications of SHM. Common mitigation strategies include transfer learning, over-sampling, and under-sampling, yet these ad-hoc methods only provide limited performance boost that varies from one case to another. In this work, we introduce one variant of the Generative Adversarial Network (GAN), named the balanced semisupervised GAN (BSS-GAN) . It adopts the semi-supervised learning concept and applies balanced-batch sampling in training to resolve low-data and imbalancedclass problems. A series of computer experiments on concrete cracking and spalling classiﬁcation were conducted under the low-data imbalanced-class regime with limited computing power. The results show that the BSS-GAN is able to achieve better damage detection in terms of recall and F􀀌 score than other conventional methods, indicating its state-of-the-art performance. KEYWORDS: GAN, Imbalanced-class, Low-data, Semi-supervised learning, Structural damage assessment. 1 BACKGROUND & MOTIVATIONS Nowadays, machine learning (ML) and deep learning (DL) have been at the forefront of a plethora of research activities in many disciplines. Since 2017, there has been an increasing number of DL studies and applications in civil and structural engineering (Amezquita-Sanchez, Valtierra-Rodriguez,& Adeli, 2018; Gao & Mosalam, 2018; Mosalam, Muin,& Gao, 2019; Oh, Kim, Kim, Park, & Adeli, 2017; PerezRamirez et al., 2019; Raﬁei & Adeli, 2017b; Yeum, Dyke, & Ramirez, 2018) . Particularly, for vision-based structural health monitoring (SHM), the convolutional neural network (CNN) is proven to be a promising approach with high practical potential (Cha, Choi, & Büyüköztürk, 2017; Deng, Lu, & Lee, 2020; Gao, Li, Mosalam, & Günay, 2018; Gao & Mosalam, 2018; Jiang & Zhang, 2019; Liang, 2019; Maeda, Sekimoto, & Seto, 2016; A. Zhang et al., 2017; C. Zhang, Chang,& Jamshidi, 2020) . However, the frameworks in past studies are facing three real-world challenges: (1) insuﬃcient data,(2) imbalanced classes, and (3) high computing demands when computing power is limited. Most of the current DL applications in vision-based SHM fall into the supervised learning category, e.g., image classiﬁcation (Gao & Mosalam, 2018), damage localization (Cha, Choi, Suh, Mahmoudkhani, & Büyüköztürk, 2018; Xue & Li, 2018) and segmentation (Yang et al., 2018), which requires substantial amounts of labeled data to reach the desired performance level. Obtaining large-scale labeled structural image datasets is costly and labor-intensive. Compared with popular computer vision (CV) benchmark datasets such as ImageNet (14 million images) (Deng et al., 2009), MNIST (70,000 images) (LeCun, Bottou, Bengio, & Haﬀner, 1998) and CIFAR-10 (70,000 images) (Krizhevsky, 2009), current vision-based SHM datasets, e.g., PEER Hub ImageNet (􀀞 -Net) with 36,413 images (Gao & Mosalam, 2020), are much s

FAQs

What problems does the paper target in vision-based structural health monitoring?

It targets insufficient labeled data, class imbalance between damaged and undamaged states, and high computing demands when GPU resources are limited.

What is the main idea behind BSS-GAN?

BSS-GAN adopts semi-supervised learning and introduces balanced-batch sampling during training to mitigate low-data and imbalanced-class issues.

How was the approach evaluated and what were the results?

The method was tested via computer experiments on concrete cracking and spalling classification under low-data imbalanced-class conditions, achieving better recall and F-score than conventional methods.

Oct 23, 2026

Upload by Mafia Boss