Prediction of emission and performance of internal combustion engine via regression deep learning approach

Strict environmental laws increase the importance of reducing emissions. While; improving three-way catalyst (TWC) technology can help reduce emission levels, attention should also be given to reducing exhaust gases before they enter the TWC. This can be achieved through the development of engine technology and calibration strategies. By doing so, low-cost TWC can be used, and emissions increase less after catalyst aging. In addition to reducing emissions at the engine-out stage before entering the TWC, it is important to reduce emissions after the TWC during the cold start and engine warm-up phase, known as the TWC light-off period. During this stage, the TWC does not reach optimal working efficiency, which can result in higher emissions. Therefore, to comply with environmental regulations, it is necessary to calculate the reduction of emissions while maintaining optimal engine and vehicle performance. Finding the optimal values of control parameters to reduce fuel consumption and emissions simultaneously makes engine calibration a complex multi-objective optimization problem. To meet calibration requirements, it is essential to accurately identify the nonlinear and multivariable behavior of engines. Thus, this study focuses on empirical engine modeling and developing an emissions model for internal combustion engines in both warm and cold engine conditions through an intelligent identification method. To enhance steady state engine modeling in warm conditions, this study proposes a hybrid MLP+CNN method based on the benefits of regression deep neural network. Additionally, the hybrid MLP+CNN+LSTM method adds a long short-term memory (LSTM) neural network, enabling the model to capture the dynamic behavior of emissions during cold start conditions and under the impact of oxygen storage and the temperature in TWC. The results demonstrate that these approaches significantly improve the accuracy of emission modeling when compared to conventional methods. The results demonstrate that using the deep learning approach and dividing the engine emission modeling into two parts, static in warm conditions and dynamic in cold conditions, significantly improve the accuracy of emission modeling compared to conventional methods. Developed models can be used in the model-based calibration due to their high accuracy in emission prediction as well as predicting Torque, BSFC, and other outputs. By coupling the developed model with optimization techniques, calibration of the engine map and cold start can be performed by considering the emissions, torque, etc., simultaneously.