Scholarlink

<p>Near-fault pulse-like ground motions (PLGMs) are characterized by intense velocity pulses that can impose severe demands on structures, making their reliable identification crucial for seismic hazard assessment and structural design. This research establishes a comprehensive framework that significantly advances an energy-based classification framework for discriminating single-pulse, double-pulse, and non-pulse ground motion signals. A dataset of 57 near-fault earthquakes is utilized, with records exhibiting pronounced geographical clustering in pulse characteristics, suggesting underlying source and path effects. We extract a set of physically interpretable features, including energy ratio metrics and conventional intensity measures (peak ground velocity, PGV, and pulse period, <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(T_p\)</EquationSource> </InlineEquation>), using Butterworth zero-phase (non-causal) low-pass filtering and a zero-crossing pulse extraction method. Both supervised and semi-supervised machine learning classifiers are evaluated: in particular, Self-Organizing Maps (SOMs) as a semi-supervised approach, and various supervised learners such as extreme gradient boosting (XGBoost), LightGBM, random forests, <i>k</i>-nearest neighbors, decision trees, logistic regression, support vector machines, and naive Bayes. Comprehensive performance metrics (Accuracy, Area Under Curve (AUC), Recall, Precision, F1-score, Cohen’s Kappa, Matthews Correlation Coefficient (MCC)) are computed for each model alongside training time (TT). Empirical validation demonstrates that tree-based ensemble methods achieve the highest classification accuracy (up to 93.4%), while the SOM yields the highest AUC (0.987) despite slightly lower accuracy. The SOM’s ability to map high-dimensional feature patterns without full supervision highlights its utility in exploratory seismic data analysis. All models exhibit rapid training (on the order of 1&#xa0;s), making them feasible for real-time or large-scale application. By emphasizing the most effective PLGM classification strategies, these findings support more reliable ground-motion selection and stronger hazard assessments.</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

Comparative efficacy of supervised and semi-supervised machine learning classifiers for energy-based discrimination of near-fault pulse-type ground motion

  • 期刊论文

  • 发表日期：

    2026年5月7日

查看全文
  • Sayed S. R. Moustafa,
  • M. Sami Soliman,
  • Shimaa H. Elkhouly,
  • Mohamed Yassien

摘要

Near-fault pulse-like ground motions (PLGMs) are characterized by intense velocity pulses that can impose severe demands on structures, making their reliable identification crucial for seismic hazard assessment and structural design. This research establishes a comprehensive framework that significantly advances an energy-based classification framework for discriminating single-pulse, double-pulse, and non-pulse ground motion signals. A dataset of 57 near-fault earthquakes is utilized, with records exhibiting pronounced geographical clustering in pulse characteristics, suggesting underlying source and path effects. We extract a set of physically interpretable features, including energy ratio metrics and conventional intensity measures (peak ground velocity, PGV, and pulse period, \(T_p\) ), using Butterworth zero-phase (non-causal) low-pass filtering and a zero-crossing pulse extraction method. Both supervised and semi-supervised machine learning classifiers are evaluated: in particular, Self-Organizing Maps (SOMs) as a semi-supervised approach, and various supervised learners such as extreme gradient boosting (XGBoost), LightGBM, random forests, k-nearest neighbors, decision trees, logistic regression, support vector machines, and naive Bayes. Comprehensive performance metrics (Accuracy, Area Under Curve (AUC), Recall, Precision, F1-score, Cohen’s Kappa, Matthews Correlation Coefficient (MCC)) are computed for each model alongside training time (TT). Empirical validation demonstrates that tree-based ensemble methods achieve the highest classification accuracy (up to 93.4%), while the SOM yields the highest AUC (0.987) despite slightly lower accuracy. The SOM’s ability to map high-dimensional feature patterns without full supervision highlights its utility in exploratory seismic data analysis. All models exhibit rapid training (on the order of 1 s), making them feasible for real-time or large-scale application. By emphasizing the most effective PLGM classification strategies, these findings support more reliable ground-motion selection and stronger hazard assessments.