Vehicle Dynamics Estimator Utilizing LSTM-Ensembled Adaptive Kalman Filter

Abstract

Active safety systems, such as the electronic stability control (ESC), have been widely utilized in modern vehicles. The feedback control of these systems requires accurate vehicle states information such as vehicle longitudinal velocity and sideslip angle. In this article, a novel vehicle dynamics estimator based on adaptive Kalman filter utilizing long short-term memory neural network (LSTM-AKF) is proposed to observe longitudinal velocity and sideslip angle. A planar vehicle kinematics model is adopted for constructing the state-space equation of the LSTM-AKF. Two virtual sensors are designed to obtain the longitudinal and lateral measurements for the LSTM-AKF, respectively. The first virtual sensor is designed for longitudinal measurements, utilizing wheel speeds and acceleration signals to estimate the longitudinal velocity. The second virtual sensor, based on an LSTM neural network, estimates lateral velocity to derive the sideslip angle. To enhance the robustness of state observation, the LSTM-AKF estimator considers the uncertainties of measurements by incorporating measurement noise covariance adjustments through two LSTM neural networks. The training labels for these two networks are designed based on a feedback method. A dataset based on real vehicle experiments is constructed to train the networks. Finally, the performance of the LSTM-AKF estimator is examined through longitudinal and lateral field test scenarios under high-adhesion and low-adhesion roads. The results demonstrate that the LSTM-AKF estimator exhibits higher estimation accuracy compared with baseline methods.