Machine Learning Evaluation of Traction Force in Locked Wheel Skid Trailer on New Mexico Highways

Abstract: Friction or skid resistance plays a critical role in road safety by controlling the traction force between vehicle tires and pavement surfaces. This study evaluates the performance of various Machine Learning (ML) regression algorithms in predicting traction force and the factors influencing it during Locked Wheel Skid Trailer (LWST) tests on New Mexico Interstate Highways. A total of 18 ML algorithms, including Extra Tree Regressor (ET), Gradient Boosting Regressor, and Random Forest Regressor, were modeled using 10-fold stratified cross-validation for evaluation. Among these, the ET model achieved the lowest RMSE and demonstrated comparable accuracy to traditional physics-based models. Key findings revealed that "Load" was the most significant feature, with a variable importance score of 0.67. The ET model achieved an R² value of 0.968 on the test data, indicating strong predictive performance. Further analysis through hyperparameter tuning and feature importance analysis underscored the ET model's robust accuracy. The model's ability to closely align with real-world values, as shown by predictive plots, highlights its practical application in ensuring safe skid resistance evaluations. The study highlights the effectiveness of ML models in predicting traction force and underscores the importance of load, vehicle speed, and temperature in enhancing road safety through more accurate traction force predictions. Future research could explore additional factors and further refine the model for greater precision in diverse road conditions.