A Hybrid Lateral Dynamics Model Combining Data-driven and Physical Models for Vehicle Control Applications
Zhisong Zhou, Yafei Wang, Qinghui Ji, Daniel Wellmann, Zeng Yifan, Chengliang Yin
Abstract
High-precision lateral dynamics model is essential for vehicle lateral state estimation and stability control. Existing physical models suffer from low accuracy due to simplified modeling, while data-driven models cannot guarantee the robustness. To address these problems for vehicle lateral dynamics modeling, a hybrid lateral dynamics model, which combines data-driven and physical models together, is proposed in this study for vehicle control applications. First, the model parameters of the conventional bicycle model including cornering stiffness are treated as time-varying parameters, and a neural network is adopted to describe the nonlinear relationship between the model parameters and the measurable vehicle states. Then, the neural network and bicycle model are integrated, and a neural network-based bicycle model is established to describe vehicle lateral dynamics. To train the neural network for parameter identification without cornering stiffness labels, a training method, which integrates the bicycle model into the loss function, is proposed. With this method, the neural network can be trained based on the entire hybrid model without providing the true values of the cornering stiffness. Simulations are conducted to verify the effectiveness of the proposed hybrid model.