Online in-cylinder pressure and temperature prediction using a modeling approach

Sammanfattning

The automotive industry is evolving to meet stricter emissions standards, such as Euro 6 and the upcoming Euro 7, while enhancing engine performance. Improving the Engine Management System (EMS), which relies on sensor data, models, and an Electronic Control Unit (ECU), is key to optimizing the performance of Spark-Ignited Internal Combustion Engines (SI-ICEs) while reducing emissions. However, engine sensors can be costly, slow, prone to defects, and difficult to place which is creating the need for alternative parameter estimation methods. This work is a comparative proof-of-concept study to investigate whether estimating in-cylinder pressures and exhaust parameters in real-time using EMS sensor data is a possible solution to the sensor problem. The project employed a combination of data-driven black-box modeling and physical modeling. Two regression models, a multivariate linear regression (MvLR) and an artificial neural network (ANN), were used to predict in-cylinder pressures during an engine cycle in a spark-ignited ICE. These predictions were based on data collected from steady-state experiments in an engine test cell. The predicted pressures were subsequently used as inputs for two of the three temperature models presented. These two models are physics-based and utilize the predicted pressures to estimate exhaust parameters, including residual gases and exhaust temperatures. The third model is a standalone linear regression model constructed to predict the exhaust manifold temperature. The predicted pressures were validated against cell-specific sensors measuring the in-cylinder pressure. The estimated temperatures were validated against a mean value temperature sensor located before the turbocharger in the exhaust manifold. The ANN model was the most accurate pressure model and performed a mean squared error (MSE) of 0.2126 Bar2 and a R2-score of 0.9973 across all estimated points. Of the three temperature models, the regressive model predicted the temperature in the exhaust manifold most accurately with an MSE of 34.06◦C2 and a R2-score of 0.9863. The best performing physical model performed an MSE of 138.4◦C2 and a R2-score of 0.9442 which is a significant performance reduction. Despite promising results, none of the models were implemented due to time constraints, although all were designed with the feasibility of real-time applications in mind.