Estimating the in-cylinder pressure of a diesel engine fuelled with a blend of IBE (isopropanol-nbutanol-ethanol) and petroleum diesel under varying injection strategies and unmeasured operating conditions by hybrid deep learning methods

Abstract

In recent years, concerns about environmental and air pollution have raised issues such as restricting or even banning the use of diesel engines, which produce high levels of polluting emissions. Therefore, research is being conducted on many fuel and engine technologies to ensure their continued use while meeting emission standards; this leads to a large number of experiments and consequently high time and energy consumption. Alternatively, making predictions for other situations using CNN-based deep learning techniques with a small amount of experimental data has become quite popular due to increasing experimental costs. In this study, the in-cylinder pressure values of a mixture of isopropanol-n-butanol-ethanol (IBE) alcohol fuel with petroleum-based diesel fuel (IBE30) under different injection strategies were predicted using five (5) CNN-based hybrid machine learning architectures (CNN + LSTM, ANN + LSTM, CNN + RF, CNN + SVR, and CNN + XGBoost). For training and testing these models, in-cylinder pressure values from a diesel engine operated under constant load-speed conditions and different fuel injection timings were used. The innovative aspect of this study is its application, for the first time, to predict the in-cylinder pressure of a diesel engine operating with a mixture of isopropanol-nbutanol-ethanol (IBE) and petroleum-based diesel fuel (IBE30) under different injection strategies, using CNNbased hybrid machine learning architectures, one of the most powerful architectures for designing datasets containing both spatial and temporal features. Furthermore, another innovative aspect of the study is the prediction of pressure traces under unmeasured operating conditions and the SHAP-based interpretability assessment that links model outputs to physically meaningful input features. According to the predicted results, CNN + RF showed the best performance based on R2 (0.9968), RMSE (0.93 bar) and MAPE (2.24%) while the lowest coefficient of determination (R2 = 0.8532) was obtained by ANN + LSTM along with very high error percentages compared with other architectures. Furthermore, SHAP-based interpretability analysis confirmed that the model predictions are physically coherent and driven primarily by local pressure history. After estimating the measured data with high accuracy, the trained models were used to predict pressure traces at unmeasured injection timings, yielding physically consistent results in both cases. It is concluded that hybrid CNN + ML models, when trained on a carefully structured combined dataset, can serve as reliable and interpretable predictive tools for combustion analysis as a virtual sensor in next step.