Combining Model-Based and Data-Driven Observer Designs for Sideslip Angle Estimation

Abstract

The vehicle side slip angle represents a key indicator of dynamic stability. Elevated values of the side slip angle may indicate a loss of stability or undesired vehicle behaviors such as understeering or oversteering. With the increased use of advanced driver assistance systems (ADAS), the need for accurate estimation of the side slip angle has become increasingly critical. This quantity in general needs to be indirectly measured or estimated, with the latter often representing a more cost-effective and more reliable approach. This is usually done by simple observer design, e.g., Kalman filter, which requires a well-parameterized system dynamics model. In this work we explore Machine Learning techniques in combination with a budget hardware inertial measurement unit to estimate the sideslip angle. This is done independently of the actual vehicle configuration, i.e., vehicle load and tires used. We model the system dynamics with a traditional Luenberger Observer, Long-short-term memory, Gated recurrent unit neural networks, and their combination, and investigate possible performance benefits when incorporating well-known physical relations. The results demonstrate that a well-designed combination of model-based and data-driven approaches can achieve high estimation accuracy even without the need for large datasets, which are typically required when employing purely data-driven methods. The performance of the proposed sideslip angle estimator under different driving conditions and tire configurations is validated with real-world measurement data.