Comparison of Rail Deterioration Prediction Models

Worldwide, railroads are essential for transporting passengers and cargo. Any failure during this infrastructure can have a significant social, environmental, and economic impacts. Therefore, a better understanding of this failure and prediction of railroad deterioration can lead to more efficient maintenance and rehabilitation, as well as safer operations. Recent advances in railway data collection techniques provide an opportunity to develop more accurate models for predicting track deterioration using machine-learning techniques. This research uses machine-learning models to forecast track deterioration to handle large and automatically collected railroad condition datasets. The study was developed in four steps: The open dataset of track characteristics and defects made available by the “INFORMS 2015 Railway Applications Section Problem Solving Competition” was used to validate the models. Defect length, amplitude, type, and tag were set as targets in separate models. Random forest and XGBoost were compared for each. The machine-learning models developed were able to determine current tag with an accuracy of 99% and predict future tags with 78% accuracy. As defects are predicted and maintained before they exceed certain thresholds, these models can improve track performance, reduce costly downtime and ensure continued safe operations.