Machine learning approach for modelling of a virtual sensor for the prediction of the rail pressure on a diesel injection system

Abstract

Diesel fuel injection systems have become increasingly complex over time to meet stringent emissions regulations, with the common rail system being used in modern diesel engines. This system depends on several factors to function properly, of which fuel pressure in the common rail is one of the most important. The main objective of this work is to develop a machine learning approach to predict the fuel pressure in the rail. A method for predicting the rail pressure is important as it provides comparable values for checking the condition of the physical sensor, helping to ensure that the system as a whole functions properly. Real data from a validation vehicle, consisting of multivariate time series, is used to build the models, and three different algorithms for predicting rail pressure are tested: multilayer perceptron (MLP), convolutional neural network (CNN), and long-short term memory (LSTM). The algorithms were fed with the preprocessed data and outputted 4 points of the target variable: the present, 10 ms in the future, 20 ms in the future and 30 ms in the future. The forward selection technique was used to find a suitable subset of features to feed the models and perform the experiments. The results show that LSTM and CNN are capable of producing better results, but MLP still delivers accurate predictions. Since the MLP is a simpler model, which makes it a better candidate for being deployed in the ECU, further experiments were performed with it. The effect of look back in time in the quality of the results from the algorithm was analyzed. To evaluate its generalization capability, the K-Fold validation technique was used. Since over 99% of the predicted values had a prediction error smaller than ±100.000 hPa (±100 bar), the initial accuracy goal is considered fulfilled. Finally, it can be concluded that a machine learning approach is efficient on modelling and predicting the fuel rail pressure.