Machine Learning for Environment-Aware Prediction of Railway Wheel Wear

Abstract

This thesis presents a machine learning approach for predicting railway wheel wear by incorporating environmental data into a per-car analytical pipeline. The pipeline includes data synchronization across three complementary sources, feature bank engineering producing more than 80 candidate predictors from eight weather variables, chronological data splitting, and a systematic comparison of three feature selection and transformation strategies: Variance Threshold (VT), Forward Feature Selection (FFS) using Support Vector Regression, and Principal Component Analysis (PCA). Three regression models are evaluated against a single-feature Huber baseline: Kernel Ridge, Histogram-based Gradient Boosting (HGB), and Random Forest Regression as non-linear baselines. A preliminary univariate analysis reveals that the normalized area wear rate (Area_rate) shows visually discernible correlations with air temperature, dew point temperature, relative humidity, and visibility. Feature selection results show that time-weighted statistics consistently rank among the most influential predictors, indicating that cumulative environmental exposure, rather than instantaneous measurements, affects wear accumulation. At the family level, temperature and precipitation dominate the FFS ranking, whereas PCA distributes importance more evenly across temperature, humidity, and snow depth. Among all evaluated configurations, FFS combined with HGB and FFS combined with Random Forest achieve the most consistent, although the improvements do not reach conventional statistical significance given the limited number of available cars. The proposed framework proved that the wear rate has a correlation with environmental factors and established a foundation for further research in scalable and efficient predictive maintenance solutions using machine learning.