A Data-Driven Framework for Car-in-the-Loop Test Bench Control and Optimization

The mega-trends in automotive industry, from advanced driver assistance systems to electrification and autonomous driving, are driving the further development of the "Road-to-Rig" concept. This results in the emergence of the multi-dynamic complete vehicle test benches, which simulate dynamic driving scenarios in different degrees of freedom simultaneously under laboratory conditions. The Car-in-the-Loop (CiL) concept, which has its advantages in terms of compactness and modularity, belongs to this class of test benches. For the purpose of proof-of-concept, a quarter-vehicle CiL prototype was built in the laboratory of the Institute for Mechatronic Systems at the Technical University of Darmstadt. Previous work on the CiL control demonstrates the potential of the model-based controllers for the control of the test bench longitudinal dynamics. However, the control strategy that is implemented results in undesired oscillations on the side shaft of the vehicle under test (VUT) during dynamic driving scenarios that capture the tire slip and load change phenomena. Moreover, the drive torque from the VUT, which is not necessarily available in practice, is assumed known. To address these issues, a data-driven framework called AI4CiL (A Data-drIven framework for Car-in-the-Loop control and optimization) is proposed in this work. It combines the test bench control with machine learning techniques. In AI4CiL, different model-based controllers that have the potential to improve the reference tracking quality and disturbance rejection behaviour are integrated. A generalized performance evaluation function is suggested, serving as the objective function for the core algorithm. To be specific, in AI4CiL, a Gaussain Process provides a model that maps the optimization variables to the performance evaluation function. Bayesian Optimization utilizes this model to acquire the next iteration location which is supposed to maximize the performance improvement. In this work, a variety of AI4CiL applications are implemented for the CiL control and optimization. Among others, AI4CiL is used to learn performance-driven models, which are optimized directly under closed-loop conditions with the aim to improve the control performance. Furthermore, safeOpt-type algorithm is integrated into AI4CiL to safely optimize the controllers directly on the CiL prototype. The optimized controllers are experimentally validated with dynamic driving scenarios on the CiL prototype. It is demonstrated that different controllers are best suitable for different scenarios. For example, the LQG controller shows the best disturbance rejection behaviour, without inducing oscillations on the side shaft of the VUT. Finally, all the controllers demonstrate good reproducibility and robustness by performing tests on the CiL prototype at different steering positions. In summary, with the help of AI4CiL, the control performance of the CiL concept for dynamic driving scenarios is significantly improved. Further development of controllers or optimization algorithms can be easily integrated into the existing structure. In addition, due to its practicability and flexibility, the proposed framework AI4CiL can be transferred for the control and optimization of other multi-dynamic complete vehicle test benches.