A Hybrid Approach to Side-Slip Angle Estimation With Recurrent Neural Networks and Kinematic Vehicle Models

Torben Gräber, Stefan Lupberger, Michael Unterreiner, Dieter Schramm

Abstract: This paper presents a supervised machine learning state estimation scheme that is able to estimate the current sideslip angle of a vehicle. It consists of a recurrent neural network with gated recurrent units, an additional input projection and a regression head. This structure has been chosen to limit the computational complexity of the model while preserving the expressiveness of the overall system. It will also be shown how equations of a simplified vehicle model can be incorporated to make use of existing domain knowledge. The results show that the neural network is able to reach an excellent estimation quality while generalizing over different tires, surfaces, and driving situations. Comparisons of different model variants on selected data subsets allow us to draw conclusions on the adaptation to varying parameters and show a quality improvement through the physical model. Evaluations with the mean squared error are complemented by more expressive fit and error plots to give a better understanding of the model behavior. All data for this paper have been collected with a Porsche 911 Turbo (Type 991 II) with a series sensor setup and an additional reference sensor.