ARTIFICIAL INTELLIGENCE BASED ENGINE OUT SOOT PREDICTION FOR A COMPRESSION IGNITION ENGINE

The complexity of the problems to be dealt with, the level of detail and the degree of precision required in all sectors involved in the technological innovation process continue to increase. At the same time, the number of data and therefore the number of information to be managed grow exponentially. The use of Artificial Intelligence systems responds very well to these needs. Given these premises, it is clear that many of the investments made by large multinational companies, which have to manage a huge amount of data on a daily basis, focus on research and development in this sector. This Thesis work was created with the aim of exploiting the potential of Artificial Intelligence systems to predict the concentration of Soot at the Engine Out level emitted by a Compression Ignition Engine. In the first part of the paper, a brief analysis of the operation of a Compression Ignition Engine is therefore reported, with a focus on the processes that lead to the production of the Soot. This is followed by an overview of the world of Artificial Intelligence, Machine Learning and Deep Learning with an in-depth analysis of the algorithms used. In this way we have the means to get to the heart of the Thesis work. It was articulated in a first phase of analysis of the data in possess, and in a second phase of structuring the Artificial Neural Networks useful for the prediction of the Soot. The data analyzed and exploited for the development of the code were provided by AVL Italia. The company has made available the measurements carried out on a dynamometric test bench, in steady and transient conditions. In particular, there are two stationary Datasets, where the number of revolutions varies from 800 rpm to 4500 rpm. These two folders differ from each other because in the first case an EGR Sweep is carried out while in the second case the EGR Rate is maintained at its nominal value. At first the two Datasets were analyzed separately and two distinct Artificial Neural Networks were constructed. Subsequently, various manipulations of the starting Dataset were carried out, combining the two folders, trying each time to improve the results returned by the Network. After having exhausted all the possible evaluation operations on the two stationary Datasets, we went to focus our attention on the transient dataset, representative of a WLTC on the same application. Also in this case an Artificial Neural Network was built using this last folder as a train and test of the system. After optimizing the network and obtaining the best possible results, in order to create a system that can replicate the real driving conditions, several mergers of the stationary and transient Datasets were carried out.