Development of a machine learning model for in-motion train weight estimation on railway bridges

Tóm tắt

This paper presents a fast, accurate approach for estimating train weight during bridge crossings by coupling a validated train–track–bridge dynamic interaction model with an XGBoost regressor optimized via Bayesian hyperparameter tuning. A high-speed railway bridge in China is used as a case study. We generate a dataset of 1000 labeled samples from finite-element dynamic simulations driven by Latin Hypercube Sampling, and validate the interaction model against available experimental data. The learned model maps measured bridge responses to train loads and is suitable for real-time deployment. Results show high predictive accuracy, with coefficients of determination R² ≈ 0.996 for motor cars and 0.993 for trailer cars. Gain-based feature importance indicates that the bridge’s maximum vertical displacement during passage is the most influential input, substantially outweighing other variables. The proposed framework demonstrates a practical path toward in-service train load estimation and monitoring without intrusive onboard sensors or extensive instrumentation.