Machine Learning Application to Predict Combustion Phase of a Direct Injection Spark Ignition Engine

Lean-diluted combustion can enhance thermal efficiency and reduce exhaust gas emissions from spark-ignited (SI) gasoline engines. However, excessive lean mixture with external dilution leads to combustion instability due to high cycle-to-cycle variations (CCV). The CCV should be controlled as low as possible to achieve stable combustion, high engine performance, and low emissions. Therefore, a stable combustion control function is required to predict the combustion phase with a low calculation load. A machine learning-based function is developed in this work to predict the 50 % mass fraction burn location (MFB50). Input parameters to the machine learning model consist of 1-, 2-, 3-, and 4-cycle from a three-cylinder production-based gasoline engine operated under stoichiometric to the lean-burn mixture. The results show that the MFB50 prediction model achieves high accuracy when 2-cycle data are used relative to 1-cycle data, which implies that the previous cycle data affects the predicted MFB50 of the next cycle. As a result, the neural network model can predict the cyclic MFB50 error within ± 3 °CA CCV and ± 5 °CA CCV with 70 % and 90 % accuracy, respectively. However, an increasing number of cycle data worsens the prediction accuracy due to model over-learning.