Machine learning and system identification for the estimation of data-driven models: an experimental case study illustrated on a tire-suspension system

In this paper, we are investigating the capabilities of both classical system identification and modern machine learning (time regression neural networks) to derive predictive black-box models which can predict the wheel center loads (WCLs) by making use of either the road disturbances acting on the wheel-patch or strain measurements on the suspension components. The WCLs are very important for the durability analyses of a vehicle and their direct measurement requires the use of expensive force transducers. Thus, predictive black-box models can augment data sets by predictions of the WCLs when multiple vehicles need to be tested and where it is infeasible to instrument dedicated transducers in each vehicle. As identification techniques to be used to derive those black-box models, the linear ARX model, the polynomial nonlinear state space (PNLSS) method, and the time regression neural networks (NNs) will be reviewed and benchmarked using experimental data measured on a McPherson suspension test rig.