Model-based Calibration of Engine Control Units Using Gaussian Process Regression

• Main Author: Tietze, Nils
• Format: Others
• Language: German; en
• Published: 2015
• Online Access: https://tuprints.ulb.tu-darmstadt.de/4572/1/20150506_Dissertation_Tietze.pdf; Tietze, Nils <http://tuprints.ulb.tu-darmstadt.de/view/person/Tietze=3ANils=3A=3A.html> (2015): Model-based Calibration of Engine Control Units Using Gaussian Process Regression.Darmstadt, Technische Universität, [Ph.D. Thesis]

Reducing the number of tests on vehicles is one of the most important requirements for increasing cost efficiency in the calibration process of engine control units (ECU). Here, employing virtual vehicles for a model-based calibration of ECUs is essential. Modelling components for virtual vehicles can be a tedious and time-consuming task. In this context, data-based modelling techniques can be an attractive alternative to physical models to increase efficiency in the modelling process. Data-based models can incorporate unknown nonlinearities encoded in the sampled data, resulting in more accurate models in practice. In combination with automated measurement, data-based modelling can help to significantly accelerate the calibration process. Furthermore, the fast simulation speed of the resulting models allows their implementation into real-time simulation environments, such as Hardware-in-the-Loop (HiL) systems, and thus enables a model-based calibration of the related ECU software function. However, generating appropriate data for learning dynamic models, i.e., the transient Design of Experiments (DoE), is not straightforward, since system boundaries and permissible excitation frequencies are not known beforehand. Thus the training data of the system measurement will be inconsistent and the main challenge of the identification process is to deal with this data to achieve a globally valid model. Furthermore, when dealing with dynamic systems in an automotive context, the Engine Control Unit typically changes operating modes while driving. Thus nonlinearities and changes of physical structures appear, which need to be considered in the model. In this thesis, a modelling system called the Local Gaussian Process Regression (LGPR), is used and adapted in order to receive a flexible modelling approach, which allows an iterative modelling process and obtains robust and globally valid dynamic models. The adapted LGPR approach is employed for the ECU calibration of dynamical automotive systems, which is critical regarding system excitation. Using LGPR, it is possible to measure the system iteratively while exploring the relevant state-space regions and improving the quality of the model step by step. The results show that LGPR is beneficial for iterative modelling of dynamical systems. Compared to the traditional Gaussian Process Regression (GPR) modelling approach, LGPR yields better results regarding the variable system dynamics.