A hybrid physics-data driven approach for vehicle dynamics state estimation

Autonomous electric vehicles (AEVs) are equipped with numerous advanced control systems that rely on measurements of longitudinal velocity, yaw rate, lateral speed, and sideslip angle. However, one of the main challenges is that mass-produced vehicles cannot accommodate overly expensive sensors. This paper proposes a novel hybrid physics-data driven observer (HPDD-Observer) for vehicle dynamics state estimation (VDSE). HPDD-Observer aims to provide comprehensive and cost-effective information about the vehicle dynamics state using low-cost onboard sensors with high sampling frequency. This approach leverages the power of hybrid modeling. Firstly, it creates a linear relationship between the estimated states and the sensor vector using ridge regression. Secondly, it designs a Long Short-Term Memory (LSTM) network with dual stage attention mechanism as the data-driven component. Then, it integrates the output of ridge regression with the data-driven component, enhancing the accuracy and reliability of the deep learning model with the scientific knowledge of the physics model. Lastly, the proposed HPDD-Observer was validated through simulation tests using MATLAB/Simulink and CarSim software, followed by real-world vehicle testing. Experimental results validate that the proposed HPDD-Observer effectively combines the strengths of deep learning and physics models without any adverse effects.