Hybrid Methods in Vehicle Dynamics State Estimation and Control – Exploiting Potentials and Ensuring Reliability of Artificial Intelligence

Within the scope of this thesis, two hybrid methods are developed and presented, which allow to exploit the potentials and to ensure the reliability of artificial intelligence for vehicle dynamics state estimation and control systems. The two hybrid methods are applied for an implementation of a central predictive vehicle dynamics control system. The objectives are to increase the safety of vehicles and to improve the ride comfort. The implementation of vehicle dynamics control systems is a well-researched field, but still offers unused potential. To achieve the aforementioned objectives, hybrid methods are utilized to exploit the great potential of artificial intelligence.

The first hybrid method focuses on the task of state estimation. By default, physical models resulting from theoretical modeling are used for state estimation. Due to assumptions and simplifications such models always feature a limited accuracy. In contrast, there are models resulting from experimental modeling. These also include models based on artificial intelligence. Such models are not based on explicit assumptions or simplifications, which generally yields a higher potential of accuracy. The increased accuracy, however, is accompanied by a loss of reliability and safety. The hybrid method of state estimation addresses and solves this issue. Within the thesis, artificial neural networks are combined with simple physical models and thus secured by them. The hybrid method determines a confidence level, which represents the confidence in the artificial neural network. The combination is done by a Kalman filter, the confidence level thereby affects the covariances of transition and measurement. Due to this novel method, the artificial neural network is only trusted stronger in well trained areas. If unknown input data are present, the reliable physical model is completely trusted. This ultimately results in an increase of estimation accuracy, while providing a reliable and valid state estimation throughout. In principle, the hybrid method thus provides the foundation for a legally approvable implementation of artificial intelligence based models for state estimation in vehicles.

The second hybrid method relates to the implementation of the control itself. With the objective of increasing vehicle safety and ride comfort, a central predictive approach for the control of the vehicle dynamics is pursued in order to exploit unused synergies within the vehicle. The class of model-based predictive control algorithms addresses the required aspects of a central predictive control system. In principle, this class of algorithms exhibits an excellent control performance, but it often results in a strongly increased computational effort. This is also confirmed in the context of the thesis. The increased effort arises from the use of non-linear models for the prediction as well as the subsequent optimization. Since the optimization within this vehicle dynamics control is performed numerically, a non-real-time capable system is present. One possibility to address these issues is the limitation of iteration steps or the use of linear models, which however result in a reduction of the excellent control quality. The hybrid method for control is an option to drastically reduce the computational effort while preserving the control quality. The original non-linear model-based predictive control is reproduced by a neuro-fuzzy system. Through the use of artificial intelligence, a fuzzy inference system becomes trainable and can thus reproduce the desired behavior. After training, a pure fuzzy inference system remains, which is much easier to comprehend than, for example, an artificial neural network. By using this hybrid method, the computational effort of the central predictive vehicle dynamics control is drastically reduced. At the same time, the original control quality is preserved. Furthermore, the fuzzy inference system is based on a direct and not iterative working principle. Thus, this hybrid method opens up the possibility to implement control systems which originally have a high computational effort, are non-real-time capable or feature both aspects.

Whereas the hybrid methods are presented in this thesis for one application in vehicle dynamics control systems, they both offer the prerequisite for the use in various fields of application.


Citation style:
Could not load citation form.


Use and reproduction:
All rights reserved