No abstract

The knock phenomenon is one of the major hindrances for enhancing the thermal efficiency in spark-ignited engines. Due to the stochastic behavior of knocking combustion, analytical cycle studies are required. However, there are many problems to be addressed with regard to the individual cycle analysis of in-cylinder pressure data. This study thus proposes novel, comprehensive and efficient methodologies for evaluating the knocking combustion in the internal combustion engine. The proposed methodologies include a filtering method for the in-cylinder pressure, the determination of the knock onset, and the calculation of the residual gas fraction. Consequently, a smart knock onset model with high accuracy could be developed using a supervised deep learning that was not available in the past. Moreover, an improved zero-dimensional (0D) estimation model for the residual gas fraction was developed to obtain better accuracy for closed system analysis. Finally, based on a cyclic analysis, a knock prediction model is suggested; the model uses 0D ignition delay correlation under various experimental conditions including aggressive cam phase shifting by a dual variable valve timing (VVT) system. Using the proposed analysis method, insight into stochastic knocking combustion can be obtained, and a faster combustion speed can lead to a higher knock intensity in a steady-state operation.

No abstract

Recently, deep learning has played an important role in the rise of artificial intelligence, and its accuracy has gained recognition in various research fields. Although engine phenomena are very complicated, they can be predicted with high accuracy using deep learning because they are based on the fundamentals of physics and chemistry. In this research, models were built with deep neural networks for gasoline engine prediction. The model consists of two sub-models. The first predicts the knock occurrence, and the second predicts performance, combustion, and emissions. This includes maximum cylinder pressure, crank angle at maximum cylinder pressure, maximum pressure rise rate, and brake mean effective pressure, brake-specific fuel consumption, brake-specific nitrogen oxides, and brake-specific carbon oxide, which are representative results of the engine (for normal combustion cases without knock). Model input parameters were selected considering engine operating conditions, and physically measurable sensor values. For test cases, the accuracy of the first model for knock classification is 99.0%, and the coefficient of determination (R2) values for the second model are all above 0.99. Test times of both models were approximately 2 ms. The robustness of all the models was verified using K-fold cross-validation. A sensitivity study of accuracy, according to the amount of training utilized, was also conducted to determine how many data points are required to effectively train the deep learning model. Accordingly, a deep learning approach was applied to predict the steady-state conditions of a gasoline engine. Achieved model accuracies and robustness proved deep learning to be an effective modeling approach, and test time was recognized to be able to apply for the real-time prediction. The sensitivity analysis can be applied for the preliminary study to define the number of experimental points for the deep learning model.

No abstract