Field Data–Driven Machine-Learning Approach for Estimating Pavement Skid Numbers from LWST Measurements for Improved Asphalt Pavement Safety

This study investigated the feasibility of predicting skid numbers (SNs) measured by a locked-wheel skid trailer (LWST) using more-accessible surface friction and texture data from a dynamic friction tester (DFT), British pendulum tester (BPT), and circular texture meter (CTM). Five years of field data from asphalt cement (AC) pavement sections were used to develop and compare predictive models. Initial approaches—including simple regression, multiple linear regression (MLR), and nonlinear regression (NLR)—exhibited limited performance (R2<0.70), highlighting challenges in modeling nonlinear relationships. To address this, an adaptive neurofuzzy inference system (ANFIS) was developed using a hybrid learning algorithm combining least-squares optimization and gradient descent. The model was trained on 80% of the data set and tested on the remaining 20%, and achieved high prediction accuracy (R2>0.91) with a low RMS error (RMSE). Gaussian membership functions were optimized over 100 epochs to reduce prediction error. The ANFIS model was benchmarked against other advanced techniques, including artificial neural networks (ANNs), support vector machines (SVMs), Gaussian process regression (GPR), bagged trees, and random forests (RFs). Among all models, ANFIS demonstrated the best overall accuracy and robustness across evaluation metrics. Variable importance analysis identified DFT20 as the most significant predictor, followed by BPT and CTM features. Additionally, three-dimensional (3D) surface plots were used to visualize interaction effects among input variables. Results indicate that ANFIS is a scalable and effective tool for estimating LWST-based skid resistance from simpler field measurements, thereby assisting in pavement safety assessments and network-level maintenance planning.