Over the years, engines have been continuously improved to operate across a wide range of operating conditions and meet various requirements. This has led to an increase in actuators and their corresponding control inputs. Traditionally, to control the engine, maps are developed by human calibration. However, as the number of operating conditions and control inputs increases, this can be challenging. To overcome this, data-driven model-based frameworks have been proposed, assuming the existence of a dataset for model development. The accuracy of the developed controller depends on the accuracy of the model, and that in turn depends on the dataset used. However, developing extensive datasets sweeping across the entire range of operating conditions and control inputs can be challenging. In this paper, a sampling framework called the Iso-surface Sampler (ISS) is proposed for developing datasets with outputs within a desirable range, making them suitable for controller development. The proposed sampling framework uses knowledge of the underlying system and Bayesian optimization to build datasets. An experimental study with a computational fluid dynamics (CFD) based engine model was used to demonstrate the effectiveness of the proposed ISS framework. A dataset spanning varying fuel properties and engine speed was generated. Using the generated dataset, feedforward (FF) control maps for combustion phasing were developed and validated using the CFD engine model. The sampled dataset had 73.4% of points within the desired output range. During the validation of FF control maps, an RMSE of 1.78 crank angle degrees (CAD) was obtained, with 63.91% of the locations in FF maps having an error of less than 1 CAD.
Raju et al. (Mon,) studied this question.