Transformer-Enabled Extended Kalman Filter for Robust Sideslip Angle Estimation in Autonomous Vehicles

Accurate sideslip angle estimation is a prerequisite for effectively implementing vehicle active safety control algorithms, particularly in autonomous vehicles (AVs). To increase estimation accuracy while ensuring robustness under varying conditions, we propose a hybrid method that integrates an adaptive extended Kalman filter (EKF) with a transformer network. The core idea is to first utilize the transformer to obtain pseudo-sideslip angle measurements and the associated prediction errors, and then incorporate them into the EKF framework to facilitate adaptive tuning of the means and covariances of process and measurement noises. Additionally, to further increase the prediction accuracy of the transformer, a mixed training scheme is introduced by combining simulation-based pre-training and real-vehicle finetuning. In this scheme, multidimensional active excitation curves are designed to efficiently generate a transformer training dataset with exhaustive coverage of diverse operating conditions. Finally, driver-in-the-loop simulation and real-vehicle test demonstrate that the proposed hybrid method outperforms the baseline approaches on sideslip angle estimation. These findings provide a preliminary reference for researchers exploring the application of transformer networks to motion control in AVs.