Auxiliary-enhanced neural networks for electric vehicle dynamics: Advancing tire force and state estimation during wheel slip

Accurately perceiving vehicle states and tire forces is crucial for the advanced safety and control technologies in vehicles. However, due to the highly complex and nonlinear characteristics of both vehicles and tires, extracting these states remains a significant challenge. This paper introduces a novel approach that combines Long Short-Term Memory (LSTM) neural networks with vehicle dynamics to address these challenges. Using real vehicle data, the study develops an attention-enhanced LSTM network to estimate tire forces. The architecture incorporates an attention mechanism (AM) to strengthen relevant feature extraction, with Bayesian optimization employed for hyperparameter tuning to effectively map sensor signals to tire forces. . A decoupled dual Unscented Kalman Filter (UKF) is also employed to estimate lateral and longitudinal velocities. It incorporates normalized tire forces, estimated by a neural network, as observations to enhance the convergence and accuracy of the UKF, providing a comprehensive vehicle velocity estimate. Finally, the performance of the algorithm was validated through DLC and slalom experiments. Under these conditions, the proposed method reduced the RMSE of the estimated longitudinal and lateral tire forces by up to 60.63% and 44.26%, respectively, compared to other methods. Additionally, RMSE of the estimated longitudinal and lateral velocity decreased by 41.31%, 50.46%, 0.97%, and 59.22%, respectively.