Assessment of railway bridge pier settlement based on train acceleration response using machine learning algorithms. (June 2023)

• Record Type: Journal Article
• Authors: Abdu, Danladi Mamman; Wei, Guo; Yang, Wang
• Is Part Of: Structures. Volume 52(2023)
• Page Start: 598
• Page End: 608
• Publication Date: 2023-06
• Subjects: Bridge pier settlement -- Train vertical acceleration -- Machine learning -- Random forest regressor -- Extra tree regressor -- Gradient boost regressor; Structural engineering -- Periodicals; 624.1
• DOI: 10.1016/j.istruc.2023.03.167
• Languages: English
• ISSNs: 2352-0124

Abstract

Abstract: To meet up with the safety and design requirements of High-Speed Railways, bridge structure has become an essential part of the lines occupying up to 90% of the mileage. Bridge pier settlement has proven to be an inevitable phenomenon, despite its critical effect on the train-track-bridge system, affecting passengers' comfort and endangering operation safety. For this reason, different theories and techniques have been employed to assess bridge pier settlement. However, few attempts have been made to utilise train vertical acceleration, the most sensitive index, to estimate pier settlement, which has the potential to save time and overcome all the other methods' geographical and economic challenges. In this study, a CRH380A high-speed train acceleration response dataset is obtained using train-track-bridge interaction simulation in Simpack. The dataset is used to train three machine learning models selected based on PyCaret. The model test results proved that pier settlement can be accurately predicted using train vertical acceleration response. Gradient boost regressor outperformed the other models on the test result with an R-squared of 99%, a mean absolute error of 0.39, and a mean squared error of 0.24. Extra tree regressor is the second-best model on the test result, with an R-squared of 98%, a mean absolute error of 0.56, and a mean squared error of 0.47. The random forest regressor test result has an R-squared of 97%, a mean absolute error of 0.73, and a mean squared error of 0.75. Additionally, over 70% of the gradient boost test prediction error is below 0.5 mm, and the maximum error is 1.5 mm, demonstrating the accuracy of the machine technique to predict pier settlement using the vertical acceleration response of the train.