A Hybrid Approach Using Design of Experiment and Artificial Neural Network in a Camless Heavy-Duty Engine

Abstract

Increasingly stricter emission regulations and fleet CO2 targets drive the engine development toward clean combustion and high efficiency. To achieve this goal, planning and conducting experiments in a time- and cost-effective way play a vital role in finding the optimal combinations of all selectable parameters. This study investigated the effects of five engine parameters on two engine-out responses in a camless variable valve actuation (VVA) heavy-duty engine. Five engine parameters were intake valve lift (IVL), inlet valve closing (IVC), injection pressure, start of injection (SOI), and exhaust gas recirculation (EGR). Initially, a design of experiment (DoE) model was generated to predict both engine-out responses: brake-specific fuel consumption (BSFC) and BSNOx emissions. Due to a poor fit of the BSFC regression model from DoE analysis, an artificial neural network (ANN) model was developed to predict BSFC instead. A d-optimal design with five engine parameters at five levels was used to design the experiment. Extra test points together with d-optimal design points were utilized to train the ANN model. The well-trained ANN model for BSFC and DoE model for BSNOx were combined with a genetic algorithm (GA) to generate the Pareto-optimal front. The results proved the concept of using a hybrid statistical approach (DoE + ANN) with GA as an effective tool to generate a range of compromise design solutions. By extracting designs along the Pareto-optimal front, the impact of engine parameters on the system can be explained.