Machine learning-driven investigation of environmental effects on dynamic behavior of railway noise barriers based on long-term field test

The passage of trains by railway noise barriers induces vibrations that may affect their fatigue performance and reduce their service life. However, long-term field monitoring of noise barriers under complex environmental and operation conditions remains rare. This study develops an interpretable machine learning (ML) framework to investigate the aerodynamic pressure and dynamic behaviors of noise barriers based on a nine-month long-term field monitoring campaign, yielding 12810 train runs over 105 valid days. Input variables include train type, speed, temperature, wind speed and direction, relative humidity, and air pressure, while the target responses cover train-induced aerodynamic pressure, stress near the base of the steel post, and displacement at the post top. Eight ML models, including four traditional and four ensemble algorithms, were used and systematically compared to evaluate their predictive capabilities and robustness. Ensemble models, particularly Gradient Boosting Decision Tree (GBDT), Light Gradient Boosting Machine (LGBM), and Extreme Gradient Boosting (XGBoost), achieved the best predictive performance, with R 2 values exceeding 0.935 for stress and displacement, and 0.895 for pressure. XGBoost, offering a strong balance of predictive accuracy and computational efficiency, was selected for SHapley Additive exPlanations (SHAP)-based interpretability analysis to uncover the physical relationships behind the data-driven predictions. Results reveal that aerodynamic pressure was the most challenging response to predict, given its higher sensitivity to turbulent airflow and environmental fluctuations, whereas stress and displacement exhibited more stable and predictable patterns. SHAP analysis identified train speed and type as the most influential factors across all responses. While environmental factors had comparatively lower influence, temperature and instantaneous wind direction consistently showed higher importance among them. Relative humidity has a moderate effect on aerodynamic pressure but a minor impact on dynamic behavior. Air pressure and wind speed exhibit limited influence on all outputs. These findings highlight the novelty and effectiveness of integrating long-term monitoring data, ML methods, and SHAP-based interpretability, offering new insights into the dynamic behavior of railway noise barriers. • Long-term field tests of 12810 train runs capture the dynamic behaviour of noise barrier. • Multiple machine learning models are used to predict the dynamic responses of barriers. • SHAP analysis is used to reveal the environmental effects on monitoring responses. • Train parameters affect responses more strongly than environmental factors. • Temperature and instant wind direction are the most important environmental factors.