Railway track stiffness estimation from instrumented vehicle data using frequency features and machine learning

Abstract

Railway track defects modify structural stiffness, influencing dynamic response, safety, and ride quality, thereby necessitating accurate stiffness estimation for effective defect detection and maintenance. In this study, track parameters derived from the discrete support model (DSM) and validated with field measurements were incorporated into a multibody vehicle-track dynamic model, whose modal responses were further verified against the DSM to ensure reliability. Vehicle response signals were simulated in Simpack^® using sensor configurations consistent with instrumented railway vehicles (IRVs). Track stiffness prediction was performed using six regression-based machine learning models, including ensemble and boosting algorithms. Feature engineering in the frequency domain using statistical metrics, along with optimized sensor configuration and hyperparameter tuning, enhanced model performance. Among the models, the CatBoostRegressor achieved an R² of 0.9814 using axle box acceleration data of the front bogie and 0.9685 with a single leading axle box accelerometer, demonstrating robustness even with minimal sensor input. Furthermore, a CatBoostClassifier trained on frequency-domain features accurately classified track stiffness variations, achieving over 90% accuracy across multiple sensor configurations and proving reliable under varying speed profiles. The proposed approach demonstrated strong predictive accuracy, generalizability, and effective spatial mapping of track quality, highlighting its potential for real-world railway infrastructure monitoring and maintenance planning.