A road adhesion coefficient estimation method based on feature screening and ensemble learning

Abstract

The road adhesion coefficient is an important parameter that affects the acceleration, steering, and braking performance of vehicles. Accurately estimating the road adhesion coefficient can ensure the performance of vehicle active safety systems and reduce the risk of accidents such as rear-end collisions and skidding during vehicle operation. Existing estimation methods do not take into account the overfitting problem that is easily caused by high dimensionality of features, as well as the degradation of estimation accuracy due to different intervals of superiority of the performance of a single model structure. To address these issues, this paper proposes a method for estimating the road adhesion coefficient based on feature screening and ensemble learning. Firstly, dynamic parameters related to the road adhesion coefficient are determined through dynamic analysis. The F-test regression is used to select dynamic parameters that have a strong correlation with the road adhesion coefficient, reducing the feature dimensionality to improve model training speed and generalization. Secondly, to combine the estimation results of multiple models, a stacked generalization (SG) model is established using three basic models: gated recurrent neural network (GRU), long short-term memory networks (LSTM), and convolutional neural network (CNN). A multilayer perceptron is used as the meta-learner to estimate the road adhesion coefficient. Finally, the Bayesian optimization algorithm (BO) is introduced to optimize the hyperparameters of the meta-model network, such as the number of layers, optimizer, and dropout rate, in order to construct an optimized stacked generalization model and improve the estimation accuracy of the road adhesion coefficient. Simulation results demonstrate that, compared to a single model, the optimized stacked generalization model can effectively improve the estimation accuracy of the road adhesion coefficient while ensuring real-time estimation. It also exhibits strong robustness and generalization ability in different operating conditions.