Neural Roll Angle Estimation in a Model Predictive Control System

This article deals with the integration of a neural state estimator into a control environment. A pre-trained recurrent neural roll angle estimator is analyzed in a closed loop environment of a model predictive control. The roll estimator consists of long short-term memory cells and predicts the angle based on the lateral acceleration, the steering wheel angle and the outputs of the control system, which influence active stabilizers and semi-active dampers. The analysis is done in a simulation environment. The artificial neural network that was trained on standard driving maneuvers is tested with independent data, i.e. a drive on a racetrack. To estimate the quality of the neural estimator, it is compared to a classical non-linear roll model that was also used as estimator in the control system when constructing the training data for the network. The results show that the neural estimator can interact with the controller. Especially by comparing the neural and the classical estimator, it can be seen that using the recurrent network to predict the roll angle leads to better control results than using the non-linear roll estimator.