Estimation of Maximum Friction Coefficient Using Recurrent Artificial Neural Networks

Abstract

Knowledge of vehicle dynamics and in particular the maximum friction coefficient is required for the optimization of advanced driver assistance systems and the implementation of autonomous driving. Since the maximum friction coefficient cannot be measured directly, estimating this coefficient based on available sensors is a field of interest. In particular, model-based approaches based on Kalman filter derivates are used. However, their accuracy is limited by the accuracy of the physical model. Due to the real-time capability required, more detailed modeling is not possible. In addition, system identification and a robust filter design are needed. As a result, data-based approaches are gaining popularity in vehicle dynamics, which are also suitable for estimating the maximum friction coefficient. In this paper, maximum friction coefficient estimation is presented using recurrent artificial neural networks (RANN) based on vehicle sensors. In order to avoid incorrect estimations during low vehicle excitation, an excitation monitoring approach is introduced. Typical longitudinal and lateral driving maneuvers on different road surfaces simulated in IPG CarMaker with a Dacia Duster are used for training, validation and testing of the RANN. Finally, the data-based approach shows improved results compared to the model-based approach of a sensitivity-based unscented Kalman filter (sUKF) from previous works.