An easy-to-use machine learning approach for predicting train-induced vibrations: application to a world heritage site

Abstract

Train-induced vibrations can negatively affect nearby structures and human comfort, yet traditional methods of assessing vibrations are either complex or limited. This research aims to provide a user-friendly, coding-free machine learning (ML) tool to accurately predict these vibrations, empowering civil engineers to conveniently implement advanced ML approaches in practical applications. Using XLSTAT, a coding-free ML platform integrated into MS Excel, multiple regression algorithms were tested. The Extreme Gradient Boosting (XGBoost) algorithm was selected based on superior predictive accuracy. Two datasets were analyzed: Dataset-1, obtained from published literature, and Dataset-2, collected experimentally at Shalimar Gardens, a UNESCO World Heritage site. Predictions were compared with results from the empirical Federal Transit Administration (FTA) method. XGBoost significantly outperformed the empirical FTA method. For Dataset-1, XGBoost achieved an R² of 0.9092 compared to 0.8618 from FTA. For Dataset-2, XGBoost notably excelled with an R² of 0.9622, whereas the FTA method reached only 0.1907. Additionally, the XGBoost algorithm demonstrated higher accuracy than the previously used GRNN and BPNN models from the literature. A user-friendly, coding-free ML approach using XLSTAT effectively predicts train-induced vibrations with high accuracy, substantially surpassing traditional empirical methods. This accessible tool facilitates practical vibration prediction and supports civil engineers in making informed, data-driven decisions to mitigate adverse impacts of structural vibrations.