PREDICTIVE MODELING OF TRACK QUALITY INDEX WITH NEURAL NETWORKS

Abstract

The Track Quality Index (TQI) was used as a primary metric to assess the structural and geometric condition of railway track infrastructure, which is directly related to traffic safety, ride comfort, and maintenance prioritization. This paper presents a predictive modeling approach aimed at forecasting future TQI values by incorporating historical measurements, geometric characteristics, and operational parameters. The analyzed dataset consisted of 821 observations of track sections collected over several measurement cycles. It included explanatory variables such as previous TQI levels, train operating speed, traffic volume, and infrastructure elements like tunnels, bridges, and records of ballast tamping. A General Regression Neural Network (GRNN) was trained using 657 data samples and validated on 164 samples, achieving a coefficient of determination (R²) of 0.699, a root mean square error (RMSE) of 4.70, and a mean absolute error (MAE) of 3.55. The results confirmed that neural networks are suitable for short-term predictions of track conditions, facilitating more effective planning and optimization of maintenance interventions. The proposed framework supports engineering decision-making related to condition-based maintenance of railways.