Data Driven Methods to Predict Vertical Profiles of the Rails: A Case Study

Predicting the value of railroad track geometry is needed to reduce the cost and time of rail condition measurements that have been carried out periodically every three months by railroad operators in Indonesia (PT Kereta Api Indonesia/PT KAI). Four track geometry parameters - vertical profile, horizontal alignment, twist, and gauge - are the reference measurements that will be mapped with alternative measurement results using accelerometers installed on regular trains (in-service trains). In this paper, eight prediction algorithms from the machine learning model are applied to map the vertical profile with 73 kilometers of measurement data in the Jakarta to Padalarang railroad segment, precisely on the Jatinegara-Cikampek measuring trip conducted from April 2019 to June 2022. The evaluation performance of the prediction results has been measured with the GadientBoosting Regressor algorithm is the best algorithm with MSE = 0.791, R^2 = 0.507 and feature importance on the Train speed and Az attributes. The research can be continued to detect the correlation of three other track geometry parameters referring to the dataset of rail geometry condition measurement results through accelerometers on regular trains (in-service trains).