Scholarlink

<p>Structural health monitoring (SHM) of reinforced concrete (RC) structures increasingly relies on deep learning models for accurate and automated damage identification. Among these, deep learning classification algorithms have shown significant potential in identifying and classifying damage. However, their functionality is primarily limited to damage type recognition rather than making finer distinctions within the same category, leading to difficulty in broader applicability in complex scenarios. Additionally, while the richness and volume of datasets are critical for improving deep learning models, practical constraints in constructing large-scale datasets, particularly in terms of quantity and diversity, limit their generalization capabilities. To address these challenges, this study presents an integrated SHM framework combining Unmanned Aerial Vehicles (UAVs) with a data augmentation–enhanced deep learning architecture, termed DamageNet, for multiclass damage detection in RC structures. The proposed framework features a hierarchical classification strategy: the first level detects damage presence, while the second level utilizes multiple parallel sub-networks to differentiate damage types and severity levels. This modular structure facilitates scalable category expansion without retraining the entire model. To overcome dataset limitations, both classical augmentation techniques (e.g., rotation, flipping, cropping) and Generative Adversarial Networks (GANs) are employed to synthetically enrich training data. Results demonstrate that GAN-based synthetic images can effectively supplement real-world data, enhancing model robustness while reducing the burden of manual data collection. The proposed approach offers a flexible and extensible solution for automated SHM and contributes to advancing deep learning applications in real-world infrastructure monitoring.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A scalable UAV-based structural health monitoring framework using augmented deep learning for multilevel damage classification

  • 期刊论文

  • 发表日期：

    2026年1月28日

查看全文
  • Jiehui Wang,
  • Yong Yang,
  • Zelin Wang,
  • Leyang Lv,
  • Tamon Ueda

摘要

Structural health monitoring (SHM) of reinforced concrete (RC) structures increasingly relies on deep learning models for accurate and automated damage identification. Among these, deep learning classification algorithms have shown significant potential in identifying and classifying damage. However, their functionality is primarily limited to damage type recognition rather than making finer distinctions within the same category, leading to difficulty in broader applicability in complex scenarios. Additionally, while the richness and volume of datasets are critical for improving deep learning models, practical constraints in constructing large-scale datasets, particularly in terms of quantity and diversity, limit their generalization capabilities. To address these challenges, this study presents an integrated SHM framework combining Unmanned Aerial Vehicles (UAVs) with a data augmentation–enhanced deep learning architecture, termed DamageNet, for multiclass damage detection in RC structures. The proposed framework features a hierarchical classification strategy: the first level detects damage presence, while the second level utilizes multiple parallel sub-networks to differentiate damage types and severity levels. This modular structure facilitates scalable category expansion without retraining the entire model. To overcome dataset limitations, both classical augmentation techniques (e.g., rotation, flipping, cropping) and Generative Adversarial Networks (GANs) are employed to synthetically enrich training data. Results demonstrate that GAN-based synthetic images can effectively supplement real-world data, enhancing model robustness while reducing the burden of manual data collection. The proposed approach offers a flexible and extensible solution for automated SHM and contributes to advancing deep learning applications in real-world infrastructure monitoring.