Predicting road adhesion coefficient with a fusion strategy of SHAP dynamic parameters

Accurately estimating the road adhesion coefficient is essential for ensuring vehicle driving safety. This study proposes a method that integrates a seven-degree-of-freedom vehicle model and the magic formula tire model with the eXtreme Gradient Boosting algorithm, leveraging Shapley Additive exPlanations values to determine feature importance. Based on this, a road adhesion coefficient estimation model is constructed, incorporating a Bayesian neural network and a Support vector regression to enhance interpretability. To address potential issues such as the model predicting road adhesion coefficients beyond the range present in the training data, a data fusion strategy is designed to improve generalization. Simulation results demonstrate that the fusion estimator outperforms individual models in terms of robustness, real-time performance, and prediction accuracy, especially under varying vehicle speeds and road conditions. Compared to a single estimator, the data fusion model reduced the mean square error by an average of 16.01% and 18.89%, respectively, while the root mean square error was reduced by an average of 11.33% and 16.04%, respectively. This approach provides valuable theoretical support to advance the development of vehicle active safety systems.