The Effect of Training Conditions on the Accuracy of In-cylinder Pressure Prediction DNN Model Using Engine Block Vibration in a CNG-Diesel Dual Fuel Engine

Abstract

A deep neural network(DNN) model was developed in order to predict in-cylinder pressure. Engine block vibration signal and in-cylinder pressure measured by accelerometer and pressure transducer were used as the input and output of the model, respectively. The test engine was a single cylinder compressed natural gas-diesel(CNG-diesel) dual fuel compression ignition engine. In order to evaluate the effect of training conditions on the accuracy of the model, CNG substitution ratio, diesel injection timing, and intake air pressure were changed while maintaining an engine load of 0.6 MPa. Test conditions were composed of unseen engine conditions in the training conditions to avoid overfitting problems. Basically, the coefficient of determination(R2 ) between the predicted and measured in-cylinder pressure trace was used as an indicator of the model performance. Peak in-cylinder pressure was also included in the analysis. The results showed different prediction accuracies depending on the different intake air pressure and statistically biased test conditions compared to the training conditions.