Performance Analysis of Data-Driven Plant Models on Embedded Systems

Data-driven plant models are well established in engine base calibration to cope with the ever increasing complexity of today’s electronic control units (ECUs). The engine, drive train, or entire vehicle is replaced with a behavioral model learned from a provided training data set. The model is used for offline simulations and virtual calibration of ECU control parameters, but its application is often limited beyond these use cases. Depending on the underlying regression algorithm, limiting factors include computationally expensive calculations and a high memory demand. However, development and testing of new control strategies would benefit from the ability to execute such high fidelity plant models directly in real-time environments. For instance, map-based ECU functions could be replaced or enhanced by more accurate behavioral models, with the implementation of virtual sensors or online monitoring functions. This paper focuses on Gaussian process regression models, a Bayesian modeling framework with practical advantages regarding achievable accuracy and usability. An approach for a more compact model expression is shown and evaluated to meet the real-time requirements of embedded systems without a significant loss in model quality. The particular systems under investigation are: a rapid prototyping target (RP), a development engine ECU, and a Hardware-in-the-Loop (HiL) system. Performance is then validated with an identification example from an internal combustion engine. This paper outlines all necessary steps to port the developed plant models onto the particular real-time target.