Development of a virtual methodology based on physical and data-driven models to optimize engine calibration

Abstract

Virtual engine calibration exploiting fully-physical plant models is the most promising solution for the reduction of time and cost of the traditional calibration process based on experimental testing. However, accuracy issues on the estimation of pollutant emissions are still unresolved. In this context, the paper shows how a virtual test rig can be built by combining a fully-physical engine model, featuring predictive combustion and NOx sub-models, with data-driven soot and particle number models. To this aim, a dedicated experimental campaign was carried out on a 1.6 liter EU6 diesel engine. A limited subset of the measured data was used to calibrate the predictive combustion and NOx sub-models. The measured data were also used to develop data-driven models to estimate soot and particulate emissions in terms of Filter Smoke Number (FSN) and Particle Number (PN), respectively. Inputs from engine calibration parameters (e.g., fuel injection timing and pressure) and combustion-related quantities computed by the physical model (e.g., combustion duration), were then merged. In this way, thanks to the combination of the two different datasets, the accuracy of the abovementioned models was improved by 20% for the FSN and 25% for the PN. The coupled physical and data-driven model was then used to optimize the engine calibration (fuel injection, air management) exploiting the Non-dominated Sorting genetic algorithm. The calibration obtained with the virtual methodology was then adopted on the engine test bench. A BSFC improvement of 10 g/kWh and a combustion reduction of 3.0 dB in comparison with the starting calibration was achieved.

Keywords

• Virtual engine calibration
• Neural network
• Genetic algorithm optimization
• NOx
• Soot emissions