Simulation of Performance and Emissions Related Parameters in a Thermal Engine Using a Deep Learning Approach

Abstract

The work carried out in this paper focused on "Simulation of Performance and Emissions Related Parameters in A Thermal Engine using A Deep Learning Approach". The goal of this work is to develop a neural network model of a thermal engine and to make a prediction of parameters related to engine management and directly impacting pollutant emissions and fuel consumption. The novelty of this work is the use of a particular type of neural network to learn long sequence of data, obtained from a running engine, and to predict internal parameters related to engine performance and emissions with a good precision. For it, data were taken from an experimentation engine, the L15B7 1.5 L, a gasoline engine with direct lateral injection from the manufacturer Honda, fitted to Honda Civic vehicles. These data made it possible to make maps of its operation. These maps enabled the calibration of a Simulink model of a thermal engine. Through a system identification approach, the temporal response of the motor was estimated and made it possible to develop a database that was used for training the LSTM artificial neural network. The work carried out showed that the learning phase of the neural network proceeded consistently (overall decrease in cost functions) and converged towards a value of RMSE = 1.09 better than those observed in the literature. The resulting neural engine model made it possible to predict several variables (fuel mass flow rate and pollutant mass flow rates) with acceptable residual errors. These results reveal that the neural model obtained correctly predicts the said variables and can, therefore, be used in closed-loop simulations of the operation of a vehicle or for a context of simulation of the operation of the engine.