Scholarlink

<p>Structural damage detection is critical for ensuring the safety of civil infrastructure, yet traditional methods face limitations in model generalization and cross-domain adaptability. This paper proposes a novel structural damage detection framework integrating bridge population modeling and continuous cross-domain transfer learning, with the following innovations: (1) A multi-stage transfer learning strategy enables progressive knowledge transfer from numerically simulated bridge populations to experimental and real bridges, overcoming dynamic heterogeneity across structural types; (2) A hybrid CNN architecture synergistically combines 1D CNN, 2D CNN, and energy entropy branches to extract complementary time-domain, frequency-domain, and global energy features, enhancing damage-sensitive feature representation; (3) A meta-learning-enhanced adaptive mechanism optimizes initial parameters using model-agnostic meta-learning (MAML) and decision-level fusion, allowing rapid adaptation to new bridge types with minimal healthy-state data; (4) Physics-guided feature engineering leverages Hilbert energy spectra and energy entropy to extract interpretable damage fingerprints, improving noise robustness. Experimental results demonstrate a detection accuracy of 95.3% across heterogeneous bridges (steel, concrete), outperforming direct transfer methods by 24%, validating the framework’s practicality for large-scale infrastructure health 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

Population-Driven Meta-transfer Learning: A Physics-Guided Framework for Multi-stage Damage Detection in Heterogeneous Bridge

  • 期刊论文

  • 发表日期：

    2025年10月29日

查看全文
  • Zhihua Wu,
  • Shuai Teng,
  • Xianling Wang,
  • Yinghou He,
  • Shaodi Wang

摘要

Structural damage detection is critical for ensuring the safety of civil infrastructure, yet traditional methods face limitations in model generalization and cross-domain adaptability. This paper proposes a novel structural damage detection framework integrating bridge population modeling and continuous cross-domain transfer learning, with the following innovations: (1) A multi-stage transfer learning strategy enables progressive knowledge transfer from numerically simulated bridge populations to experimental and real bridges, overcoming dynamic heterogeneity across structural types; (2) A hybrid CNN architecture synergistically combines 1D CNN, 2D CNN, and energy entropy branches to extract complementary time-domain, frequency-domain, and global energy features, enhancing damage-sensitive feature representation; (3) A meta-learning-enhanced adaptive mechanism optimizes initial parameters using model-agnostic meta-learning (MAML) and decision-level fusion, allowing rapid adaptation to new bridge types with minimal healthy-state data; (4) Physics-guided feature engineering leverages Hilbert energy spectra and energy entropy to extract interpretable damage fingerprints, improving noise robustness. Experimental results demonstrate a detection accuracy of 95.3% across heterogeneous bridges (steel, concrete), outperforming direct transfer methods by 24%, validating the framework’s practicality for large-scale infrastructure health monitoring.