An Interpretable Method for Asphalt Pavement Skid Resistance Performance Evaluation Under Sand-Accumulated Conditions Based on Multi-Scale Fractals

The skid resistance of asphalt pavement is vital for traffic safety and reducing accidents. Existing research using only wavelet transforms or fractal theory to study the pavement surface texture-skid resistance relationship has limitations. This paper integrates a wavelet transform and fractal theory to extract the multi-scale fractal features of pavement texture. It proposes an interpretable machine learning model for skid resistance assessments of sand-accumulated pavements. The three-dimensional (3D) texture of asphalt pavements is decomposed at multiple scales, and fractal and multifractal features are extracted to build a dataset. The performance of mainstream machine learning models is compared, and the eXtreme Gradient Boosting (XGBoost) model is optimized using the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) algorithm. The SHapley Additive exPlanations (SHAP) method is used to analyze the optimal model's interpretability. The results show that asphalt concrete with a maximum nominal particle size of 13 mm (AC-13) has the most concentrated fractal dimension, followed by open-graded friction course with a maximum nominal particle size of 9.5 mm (OGFC-10), with stone matrix asphalt with a maximum nominal particle size of 16 mm (SMA-16) being the most dispersed. The singular intensity difference of the multifractal (Δ<i>α</i>) changes oppositely to the fractal dimension (<i>D</i>), and the fractal dimension difference of the multifractal (Δ<i>f</i>) decreases with the number of wavelet decomposition layers. The CMA-ES-XGBoost model improves R² by 27.1%, 9%, and 3.4% over Linear Regression, Light Gradient Boosting Machine (LightGBM), and XGBoost, respectively. The 0.4-0.8 mm range fractal dimension most significantly impacts the model output, with complex interactions between features at different scales.