Advanced powertrains for modern vehicles require the optimization of conventional combustion engines in combination with tailored electrification and vehicle connectivity strategies. The resulting systems and their control devices feature many degrees of freedom with a large number of available adjustment parameters. This obviously presents major challenges to the development of the corresponding powertrain control logics. Hence, the identification of an optimal system calibration is a non-trivial task.
To address this situation, physics-based control approaches are evolving and successively replacing conventional map-based control strategies in order to handle more complex powertrain topologies. Physics-based control approaches enable a significant reduction in calibration effort, and also improve the control robustness. However, due to the requirement of real-time capability, physical models have to be formulated via simplified mean value approaches, which in turn limits the control accuracy.
To eliminate the constraints of a purely physics-based control approach, the underlying physical process model can be augmented by an additional data-driven model. For this purpose, an artificial neural network or a Gaussian Process model can be considered amongst others. Data driven models can provide high model accuracy, but they usually show a poor predictive robustness in regions which have not extensively been trained with data beforehand. Consequently, an alternative modeling strategy is utilized, where the general process tendencies are estimated by a physical model, while the data-driven model corrects the estimation in well-known operation regions to maximize the overall process model accuracy. This modeling approach is commonly referred to as “Hybrid semi-parametric modeling” or “Hybrid Modeling Technique” (HMT). It follows the general idea of combining advantageous attributes of a physical model (predictive robustness and low calibration effort) with the benefits of a data-driven model (high model accuracy). Besides the achievement of an improved process model accuracy, HMT further enables model updates by re-training of the data-driven model, when the process behavior changes as a consequence of e.g. hardware drifts, aging or different process boundary conditions.
To evaluate the performance of the given HMT, the approach is exemplarily applied to derive an adaptive, real-time capable Diesel ignition delay and a NOx raw emission model. Available physics-based models are used as a baseline. In both application examples, HMT achieves an increased model accuracy in standard conditions by means of an artificial neural network and a Gaussian process model respectively. In parallel, thanks to the contribution of the physics-based model, a high predictive robustness is maintained in operating conditions that have not been considered during model training.