Vehicle Sideslip Angle Estimation Using Deep Reinforcement Learning Combined with Unscented Kalman Filter

Area of Science:

• Vehicle Dynamics and Control
• Machine Learning for Automotive Systems
• State Estimation Techniques

Background:

• Accurate vehicle sideslip angle estimation is crucial for vehicle stability and active safety systems.
• Traditional Unscented Kalman Filter (UKF) methods struggle with adaptive noise parameter adjustment, degrading performance in complex conditions.
• Adaptive optimization of noise covariance matrices (Q and R) is needed for robust estimation.

Purpose of the Study:

• To develop an advanced vehicle sideslip angle estimation method integrating UKF and Deep Reinforcement Learning (DRL).
• To enable dynamic optimization of UKF's process noise covariance matrix (Q) and observation noise covariance matrix (R) using DRL.
• To enhance estimation accuracy and robustness across diverse driving scenarios.

Main Methods:

• Integration of UKF with Deep Reinforcement Learning (DRL) for adaptive parameter tuning.
• Construction of a state space including vehicle motion states and filtering performance metrics.
• Utilization of Proximal Policy Optimization (PPO) algorithm to train the DRL agent for optimizing noise covariance matrices.
• Formulation of a reward function based on estimation errors and uncertainties.

Main Results:

• The proposed UKF-DRL method significantly improves vehicle sideslip angle estimation accuracy under varying speeds, road conditions, and sensor noise.
• Demonstrated a reduction in Root Mean Square Error (RMSE) by over 30% compared to traditional UKF.
• Exhibited strong stability and robustness in complex driving scenarios.

Conclusions:

• The UKF-DRL approach offers a superior solution for accurate vehicle sideslip angle estimation.
• This method provides enhanced vehicle stability control and active safety system development.
• The approach is extendable to autonomous driving and other vehicle safety applications.