An Experimental Selection of Deep Neural Network Hyperparameters for Engine Emission Prognosis

Abstract

This research presents results and discussion on prognostic deep learning models developed for the prediction of emission parameters of internal combustion engines. The resulting models were trained to predict various engine emission parameters from engine vibrations and the type of fuel used. To train and test the models, a dataset was created, where the input data was obtained from the measured engine vibrations at certain moments of time and the type of fuel used. Using obtained input data, the output data were calculated using a prognostic model. 162 models of multilayer perceptron network architecture were created and trained using different combinations of training parameters. The accuracy of each model was measured using the mean absolute percentage error and mean squared error metrics. The paper analyzes the influence of each selected parameter on the accuracy of the model. The best accuracy was achieved by a multilayer perceptron neural network model with 1 hidden layer and 50 neurons, which was trained for 20 epochs with a batch size of 16. The accuracy of the model on the testing dataset was 0.02161 and 0.00014, respectively, based on the mean absolute percentage error and mean squared error metrics.