A Machine Learning Approach for Forecasting Rail Degradation in Railway Systems

Abstract

In railway systems, the contact between rolling stock wheels and rails causes continuous wear, requiring periodic rail replacement to ensure operational safety and reliability. Traditional methods to estimate rail lifetime often fail to account for the specific conditions of each railway, leading to inaccurate forecasts. This paper proposes a method based on machine-learning algorithms to estimate the remaining lifetime of rails using historical data from the railway itself. The proposed approach was applied to the Vitória-Minas Railway (EFVM), in Brazil, using a dataset collected over 6 years and 8 months, totaling 1,275,034 records. Two algorithms were evaluated: artificial neural networks (ANN), with 54 configurations tested, and K-nearest neighbors (k-NN), with 12 configurations. The models were trained and validated with 1,100,000 records and tested with 175,034 records. Four categories were defined for rail lifetime estimation: month, quarter, semester, and year. Results showed that k-NN outperformed ANN in all tested configurations, reaching Accuracy and F1-score values up to 0.97427 and 0.97329, respectively. These findings confirm the potential of data-driven models for supporting maintenance planning in railways.