Hybrid of Neural Network and Physics-Based Estimator for Vehicle Longitudinal Dynamics Modeling Using Limited Driving Data

An accurate longitudinal dynamics model is essential for state estimation and control of autonomous vehicles. However, existing physical models suffer from limited working conditions and large errors, while pure data-driven models require massive amounts of driving data to cover working conditions fully. To address these issues, a hybrid architecture composed of a neural network-based traction model, a recursive least square-based parameter estimator, and a physics-based dynamics model is proposed for longitudinal dynamics modeling, in which the parameter estimator is used for mass and modified rolling friction coefficient estimation. Under this architecture, the longitudinal dynamics model can be established using limited driving data collected on a test field with a given load, and achieve precise vehicle dynamics characterization under various roads and loads. To design the neural network for traction description, the dynamics of vehicle powertrain and braking systems are analyzed, and a physics-guided neural network, which fully considers the traction transmission characteristics, is formulated. For model training, a two-stage hybrid model training method is proposed, which can train the hybrid model with the co-existence of unknown network and physical parameters. Results demonstrate that the proposed hybrid model can realize accurate parameter estimation and vehicle longitudinal dynamics modeling using limited driving data collected at a test field under a given load, especially with excellent generalization performance under different loads and roads.