Physics-integrated Segmented Gaussian Process (SegGP) learning for cost-efficient training of diesel engine control system with low cetane numbers

Abstract

Control model training is an essential step towards the development of an engine controls system. A robust controls strategy is required for engines to perform reliably and optimally under challenging conditions, such as using low cetane number fuels (vital to achieving a single fuel concept). Developing such a control strategy through physical experiments, however, can be very costly due to issues such as unexpected engine failure and manufacturing delays. One approach is to rely solely on CFD simulations for control model training, which can be accurate but places significant burden on computing resources to explore the desired control design space. Another approach is via purely data-driven machine learning models, but the training data needed to achieve desirable accuracy can also be prohibitively expensive to generate. To address this, we develop a novel physics-integrated Segmented Gaussian Process (SegGP) model, which integrates fundamental physics on the pressure curve within a flexible probabilistic learning framework. This integration of physics allows for accurate predictive modeling of pressure, heat release and thus control using limited training data, which greatly reduces computational burden. We demonstrate the effectiveness of this approach for quickening the control development training of diesel engines.