Scholarlink

<p>Vibration-based structural health monitoring still lacks compact, interpretable pipelines that robustly decompose nonlinear, nonstationary responses and learn a healthy baseline that can be contrasted against damaged states with sensor-level resolution. In this research, this gap is addressed by integrating improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) and the Hilbert–Huang transform (HHT) to extract instantaneous frequency (IF), instantaneous amplitude, energy, and phase from acceleration responses, and by training a multilayer perceptron (MLP) on healthy intrinsic mode functions to predict these features and compare them against damaged responses. The framework is evaluated on the IASC-ASCE four-story benchmark steel frame using acceleration records under multiple damage scenarios. Quantitatively, the first-mode IF obtained via ICEEMDAN-HHT exhibits proximity-consistent reductions near damaged elements (typically on the order of 4–6% relative to the healthy baseline), whereas locations farther from the damage change more mildly (1–3%). Complementary analyses using VMD-HHT on a neighboring mode show scenario-dependent deviations of similar or larger magnitude (up to order 5–10%, depending on the case), reinforcing sensitivity and the importance of mode selection. Training/validation curves for the MLP display smooth RMSE convergence with a small, stable generalization gap. Overall, the ICEEMDAN-/VMD-HHT-MLP technique yields interpretable, sensor-aware indicators that detect and localize damage with quantitative contrasts, helping to bridge the methodological gap between robust decomposition and practical learning.</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

Adaptive decomposition and baseline learning for structural damage localization in a four-story steel frame: IASC–ASCE benchmark structure

  • 期刊论文

  • 发表日期：

    2026年3月3日

查看全文
  • Moslem Malekmohammadi,
  • Mohamad Mohamadi Dehcheshmeh,
  • Gholamreza Ghodrati Amiri,
  • Pezhman Fazeli Dehkordi

摘要

Vibration-based structural health monitoring still lacks compact, interpretable pipelines that robustly decompose nonlinear, nonstationary responses and learn a healthy baseline that can be contrasted against damaged states with sensor-level resolution. In this research, this gap is addressed by integrating improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) and the Hilbert–Huang transform (HHT) to extract instantaneous frequency (IF), instantaneous amplitude, energy, and phase from acceleration responses, and by training a multilayer perceptron (MLP) on healthy intrinsic mode functions to predict these features and compare them against damaged responses. The framework is evaluated on the IASC-ASCE four-story benchmark steel frame using acceleration records under multiple damage scenarios. Quantitatively, the first-mode IF obtained via ICEEMDAN-HHT exhibits proximity-consistent reductions near damaged elements (typically on the order of 4–6% relative to the healthy baseline), whereas locations farther from the damage change more mildly (1–3%). Complementary analyses using VMD-HHT on a neighboring mode show scenario-dependent deviations of similar or larger magnitude (up to order 5–10%, depending on the case), reinforcing sensitivity and the importance of mode selection. Training/validation curves for the MLP display smooth RMSE convergence with a small, stable generalization gap. Overall, the ICEEMDAN-/VMD-HHT-MLP technique yields interpretable, sensor-aware indicators that detect and localize damage with quantitative contrasts, helping to bridge the methodological gap between robust decomposition and practical learning.