MULTI-FIDELITY NEURAL NETWORK REGRESSION FOR EFFICIENT TRAINING OF ENERGY-ASSISTED DIESEL ENGINE CONTROL SYSTEM

Low cetane number fuels increase the probability of misfire in diesel engines. Utilizing an ignition assistant, along with a reliable engine-control system can enhance combustion and mitigate misfire. Such a control system requires significant pressure data for training the controller software. For engines operating with low cetane number fuel over a wide range of conditions, traditional training data collection based on experimental data alone is time-consuming and expensive. Our aim is to build a purely data-driven model for predicting average cylinder pressure, while varying cetane number, main injection timing, and ignition assistant power. The parametric space is filled using sparse experimental data, and numerous URANS (Unsteady Reynolds-Averaged Navier-Stokes) simulation results simultaneously. The experimental data has relatively high-fidelity, but is costly to acquire, while the URANS results have lower fidelity, but are much more cost effective to generate. An existing neural-network-based methodology for multi-fidelity regression of bi-fidelity problems is utilized, where two separate artificial neural networks - one for each level of fidelity - are used. The effectiveness of this approach is demonstrated by predicting the in-cylinder pressure profiles at various operating conditions and engine control parameters, which are in good agreement with the experimental data.