EGR Prediction of Diesel Engines in Steady-State Conditions Using Deep Learning Method

Most of the parameters needed to predict Nitrogen Oxides (NOx) emissions, for example, combustion temperature, oxygen concentration and in-cylinder composition ratio, can be predicted by phenomenological 0-D prediction model if accurate EGR rates are provided. However, it is difficult to predict the EGR rate itself accurately by the model, so the EGR rate is predicted by the temperature measurement method. Although this method predicts EGR rates very accurately and quickly, there are some problems such as thermocouple failures and the difficulty in applying to mass production engines, so it is necessary to predict EGR rates by another method. The deep learning method follows an inductive methodology that extracts common characteristics of data based on a lot of data themselves. Therefore, although it requires a lot of experimental data, it has an advantage of high accuracy that can be obtained without any feature engineering. In this study, the EGR rate, which was difficult to predict in the past, was predicted by making various models using the deep learning method. Finally, EGR rate was predicted with a high accuracy of R-square 0.9994 and root mean squared error 0.0692 using a deep learning method at 1500 rpm and bmep 4, 6 and 8 bar. This study can be used as a basic study to predict EGR rates in transient and RDE conditions.