Prediction of LWST values using ML and ANN techniques from frcition and texture measuring devices based on long term field data

Given the cost, time requirements, and limited accessibility of Locked Wheel Skid Trailer (LWST) testing, this research explores alternative approaches to predicting skid number (SN) using more affordable and readily available devices. These alternatives offer quicker results, which is particularly relevant given the standardized nature of skid number measurements worldwide. The study focuses on predicting the (SN) generated by the LWST test through analyzing data from Dynamic Friction Tester (DFT), British Pendulum Tester (BPT), and Circular Texture Meter (CTM) using statistical and machine learning techniques. The research includes a thorough assessment of various prediction methods and factors influencing skid number accuracy, including strategies for handling missing data. Prediction models employ Multiple Linear Regression (MLR), Support Vector Machine (SVM), Artificial Neural Network (ANN), and Ensemble Bagged Tree (EBT) techniques, utilizing predictors such as Dynamic Friction Number at 20 km/h (DFT20), Dynamic Friction Number at 64 km/h (DFT64), British Pendulum Number (BPN), and Mean Profile Depth (MPD). The development of forty-eight models, including Full, Reduced, and individual predictor models, involved employing data imputation methods and stepwise regression. Significant correlations were observed among friction parameters, with DFT20 showing the highest correlation with skid numbers, while MPD exhibited the lowest. Following data imputation, notable enhancements in BPN’s correlation were noted. The study emphasizes the superiority of Full models in skid number prediction accuracy, with EBT models, especially those enhanced by data imputation, demonstrating outstanding performance. Statistical simulations confirmed the reliability of these models, indicating minimal outliers, overfitting, and high accuracy in skid number estimation. The prioritization of DFT over BPT and CTM is highlighted, with the potential for further exploration of devices to improve prediction reliability.