Scholarlink

<p>Ensuring the operational safety of high-speed trains during earthquakes is a core challenge for China’s extensive high-speed rail network. While machine learning (ML) -based seismic response assessment has become a mainstream approach, conventional ML methods suffer from limitations such as heavy training data demands, poor interpretability, and over-reliance on deterministic predictions. This study proposes an interpretable dynamic ensemble learning model integrated with sample augmentation to predict extreme seismic responses of vehicle-track-bridge (VTB) systems. The framework combines Generative Adversarial Networks (GAN) for data generation, the Kepler Optimization Algorithm (KOA)—chosen for its superior convergence speed and optimization performance over classical algorithms—for hyperparameter tuning, and a dynamically weighted ensemble of Long Short-Term Memory (LSTM)-Attention and Support Vector Machine (SVM). A 3D nonlinear VTB model under bidirectional seismic excitation serves as the physical basis, with GAN-based augmentation mitigating data imbalance. Comprehensive validation against traditional ML models confirms significant accuracy gains, marked by reduced Mean Absolute Error (MAE) and coefficient of determination (<i>R</i><sup>2</sup>) values consistently exceeding 0.97. SHapley Additive exPlanation (SHAP) analysis identifies key input features affecting wheel-rail interaction parameters, and Gaussian probabilistic interval prediction quantifies predictive uncertainty with adaptive confidence bounds. The findings offer references for seismic prediction and safety risk assessment of high-speed railways.</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 new interpretable dynamic ensemble learning model with sample augmentation for predicting seismic responses of vehicle-track-bridge systems

  • 期刊论文

  • 发表日期：

    2026年6月29日

查看全文
  • Zi-yi Wang,
  • Zhi-peng Lai,
  • Li-zhong Jiang

摘要

Ensuring the operational safety of high-speed trains during earthquakes is a core challenge for China’s extensive high-speed rail network. While machine learning (ML) -based seismic response assessment has become a mainstream approach, conventional ML methods suffer from limitations such as heavy training data demands, poor interpretability, and over-reliance on deterministic predictions. This study proposes an interpretable dynamic ensemble learning model integrated with sample augmentation to predict extreme seismic responses of vehicle-track-bridge (VTB) systems. The framework combines Generative Adversarial Networks (GAN) for data generation, the Kepler Optimization Algorithm (KOA)—chosen for its superior convergence speed and optimization performance over classical algorithms—for hyperparameter tuning, and a dynamically weighted ensemble of Long Short-Term Memory (LSTM)-Attention and Support Vector Machine (SVM). A 3D nonlinear VTB model under bidirectional seismic excitation serves as the physical basis, with GAN-based augmentation mitigating data imbalance. Comprehensive validation against traditional ML models confirms significant accuracy gains, marked by reduced Mean Absolute Error (MAE) and coefficient of determination (R2) values consistently exceeding 0.97. SHapley Additive exPlanation (SHAP) analysis identifies key input features affecting wheel-rail interaction parameters, and Gaussian probabilistic interval prediction quantifies predictive uncertainty with adaptive confidence bounds. The findings offer references for seismic prediction and safety risk assessment of high-speed railways.