A Unified Framework for Vertical States Estimation: Data-Driven Approach Incorporating Uncertainty

Accurate and real-time knowledge of vertical states plays a critical role in vehicular safety and comfort. It contributes to enhanced cornering performance, detailed modeling of vehicle dynamics, and effective implementation of control strategies for numerous active safety systems, especially semi-active and active suspensions. For this purpose, this study aims to estimate three key dynamical states that are impractical to acquire via direct measurements: tire vertical force, road profile, and unsprung mass velocity. A unified framework tailored to the context of vertical dynamics is proposed, leveraging a state-of-the-art multivariate prediction model known as Neural Basis Expansion Analysis for Time Series (N-BEATS). This model uses existing in-vehicle signals and is trained using data from a simulation platform with a high-fidelity vehicle model. Additionally, the model provides an uncertainty interval of its estimation rather than a single value. Test results reveal that N-BEATS exhibits good tracking performance, outperforming Long Short-Term Memory (LSTM) in terms of estimation errors and inference speed. Finally, we highlight extra functionalities of the developed virtual sensor such as road quality classification and the consistency of the temporal-frequency content between predicted signals and the ground truth labels.