Research On Modelling Of Dynamic And Trajectory Tracking Control Method Of Intelligent Vehicle Based On Data-driven Method

Intelligent vehicles are an indispensable component of intelligent transportation systems.Establishing accurate vehicle dynamics model is an important basis for realizing path tracking control and lateral stability of vehicle.The connotation of vehicle dynamics has been developed from physical dynamics to data-driven dynamics under the intelligent system.Building upon this foundation,the analysis and modeling problem of vehicle dynamics can be reframed from the perspective of "data processing and analysis-mining statistical patterns-modeling and reproducing data patterns." This approach aims to transform the analysis and modeling of vehicle dynamics into the exploration and modeling of statistical patterns observed in the data.By leveraging the self-learning capabilities of neural networks,it seeks to elucidate the evolutionary patterns of complex dynamics in intelligent vehicles,characterize their dynamic behaviors and underlying mechanisms.This has the potential to offer novel theories and methodologies for analyzing similar complex nonlinear systems.This study is conducted with the support of the National Natural Science Foundation of China under projects U20A20333 on the autonomous perception and behavioral cognitive mechanisms of intelligent vehicle based on complex network theory,and51875255 on the mixed characteristics and model predictive control of lateral dynamics in autonomous driving vehicles.The focus of this research includes the establishment of a vehicle dynamics dataset,data-driven modeling of intelligent vehicle dynamics,and trajectory tracking control.The key findings and conclusions of this study are summarized as follows:(1)Firstly,a vehicle dynamics dataset is established.A vehicle dynamics model is built using Matlab/Simulink,and the longitudinal load transfer at high speeds and the tire hysteresis response at low speeds are analyzed.With the constructed physics model,a virtual dataset based on the physics model is generated.Car Sim/Simulink is utilized to simulate real-world driving scenarios and a 27-degree-of-freedom virtual dataset of vehicle dynamics is collected.Lastly,real-world data is acquired by an intelligent driving vehicle through sensors.These data provide support for subsequent experimental research.(2)Second,a data-driven vehicle dynamics model and trajectory tracking control algorithm are designed.To address the limitations of simplification and linearization techniques in capturing the essence of vehicle dynamics,a vehicle dynamics model based on time-delay neural networks is proposed.The model is trained using the collected dataset,and the test results on both virtual and real vehicle data demonstrate its superiority and generalization.And it achieves accurate predictions under different conditions.A longitudinal controller based on feedforward-feedback control method is designed,and a trajectory tracking lateral controller is developed using model predictive control algorithm.Through the Car Sim/Simulink software,the proposed control strategy is validated under double lane change conditions.Compared to nonlinear model predictive control,the trajectory tracking control method achieved an 8% reduction in peak lateral error,and compared to linear time-varying model predictive control,the peak lateral error is reduced by 48%,demonstrating smaller trajectory tracking errors and higher stability.(3)Finally,to address the issue of model-based controllers being susceptible to unpredictable errors caused by model mismatch,an innovative approach combining traditional machine learning methods and deep learning methods for vehicle dynamics modeling is proposed.This approach achieved complementary advantages between different models by introducing a branch for predicting model input uncertainties in the model structure.To expedite model training,a Gaussian negative log-likelihood loss function is incorporated into the cost function.Additionally,constraints are imposed on the model output uncertainties in the cost function.Through simulation on the "8"-shaped road,it is demonstrated that compared to nonlinear model predictive control methods,the maximum lateral position error is reduced by 40%.Moreover,through experimental validation on a double lane change scenario with variable speeds,the robustness of the proposed control method is confirmed.