Testing automobile engines electrically

Abstract

Diesel engine parameter prediction became a topic of interest in recent years, along with the development of condition-based maintenance, and is now considered a key instrument for engine diagnosis research. This contribution compares two different approaches for diesel engine performance prediction: thermodynamic modelling and artificial neural networks (ANNs). The thermodynamic modelling was developed using AVL Boost™ software simulating a single-cylinder diesel engine with different engine loads and operating conditions. The ANN modelling was conducted by comparing two efficient training algorithms to achieve the best prediction performance, with the ANN structure parameters determined by network error analysis. Both models’ prediction accuracy was verified by a single-cylinder engine test bench operating under real conditions. The adaptability and robustness of the two approaches was studied for the whole engine load spectrum, comparing predicted values to experimental measurements. Both prediction tools, ANN and thermodynamic modelling, proved to be reliable for engine performance and emissions prediction. In both models brake-specific fuel consumption (BSFC), exhaust gas temperature (Texh), carbon monoxide (CO) and nitrogen oxides (NOx) were predicted using brake mean effective pressure (BMEP) and engine speed as inputs. ANN show higher accuracy for BSFC prediction in all engine loads, and Texh prediction accuracy is better for ANN when dealing with medium to high loads, while the thermodynamic model shows better results when dealing with medium to low loads. CO is better predicted by the thermodynamic model except for the highest engine loads, and NOx predictions present high accuracy in both models, except for the lowest loads. Calculation time is lower for ANN, but the thermodynamic model provides additional performance results (i.e. combustion pressure tracing and associated values).