Transformer Neural Networks for Maximum Friction Coefficient Estimation of Tire-Road Contact using Onboard Vehicle Sensors

For the optimization of advanced driver assistance systems (ADAS) and the implementation of autonomous driving, the perception of the vehicles environment and in particular the maximum friction coefficient μ max is crucial. Since μ max cannot be measured directly via existing serial sensors, estimating this coefficient based on available sensors is an area of research. In this paper, μ

_max estimation is presented using transformer neural networks (TNN) based on the input data measured by onboard vehicle sensors. The TNN is applied to both a simulative dataset created with IPG CarMaker and an experimental dataset recorded on a test track, each using a sports utility vehicle (SUV) as the test vehicle. Both datasets contain typical longitudinal and lateral driving maneuvers on different road surfaces. On an independent test dataset, the data-based TNN approach shows improved results in estimating μ max compared to the model-based approach of an unscented Kalman filter (UKF) and to two other data-based approaches using recurrent artificial neural networks (RANN s) from previous works. In particular, the TNN responds faster and more accurate to jumps of μ

_max, especially during lateral driving maneuvers. Moreover, the TNN has both less parameters, and training epochs compared to the RANN.