Dealing with the curse of dimensionality in Twin-in-the-Loop observer design

Many vehicle dynamics controllers require the knowledge of unmeasured signals, for instance, the sideslip angle in electronic stability control. For this reason, vehicles are usually equipped with several observers running in parallel in different electronic control units. The Twin-in-the-loop approach represents an effective alternative paradigm, in which a single complex Digital Twin is run on-board and a data-driven correction matrix is employed to adjust the estimate of the whole vehicle state in real-time. However, such a complex observer might require the tuning of (too) many parameters if no prior knowledge is available. In this work, we propose an unsupervised learning approach to reduce the dimensionality of the problem, so as to deal also with numerically intractable problems. The strategy is experimentally tested on speed/yaw rate estimation for a real-world vehicle setup.