Data-Driven Approach for Diesel Engine Soot Emissions Modelling

Nowadays, control systems development within the automotive industry is not only evolving rapidly for evergreen cost reduction. On one side, emissions regulatory pressure is increasing more now than ever. Also, more aggressive competing market are pushing for shorter Time-to-Market for a new product. Thus, all those points are translated into significantly more complex control systems and additional degrees of freedom in the optimization control parameters. So, with such a complicated system, a costly and time-consuming engine calibration process is needed. To move part of the development process in a virtual test bench, an Automotive Diesel ICE Virtual Calibration has been introduced. However, to obtain such a calibration virtual tool, an accurate engine model is needed. The virtual test bench is defined as the integration of both models of performances and pollutant emissions. The object of this thesis is to provide a predictive model for Soot emissions in a wide range of operating conditions. All the activities have been carried out in collaboration with MODELWAY S.R.L. The design has been performed through a data-driven strategy based on Supervised Machine Learning algorithms. Firstly, a data set of experimental activities on a physical engine has been acquired. Secondly, the observations have been analysed and processed to reduce the high dimensionality of the data. To neglect the irrelevant and redundant engine features concerning Soot emissions, several approaches such as Pearson Correlation Coefficient and two model optimization filters, i.e Neighbourhood Components Features Selection with Regularization and LASSO, have been implemented. This pre-processing has permitted to reduce the number of required model inputs and returns a features ranking. Next, the decreased data set has been utilized to learn the model parameter of a Feed-Forward Neural Network Soot predictor. Training and Testing have been achieved through a K-Fold Cross-Validation, where the net architecture has been optimized. To obtain the optimal hyper-parameters, searching techniques, such as Grid Search, were exploited. Overall, satisfactory results on Soot emissions predictability were achieved. Finally, test outcomes have been examined in terms of statistical reliability to verify the generalization capability of the model.