Development of a Machine Learning code for predicting Soot Tail Pipe in a Compression Ignition Engine

One of the most important problems for our planet is the pollution produced in the transport sector. Over the years, increasingly stringent regulations have been imposed on the pollutants production because of the damage that these bring to the human health and to the environment. In this thesis, I analyzed the production of Particulate Matter (PM) within Compression Ignition Engines (IC) and the after-treatment systems (ATS). In particular, my work is based on the construction of a virtual sensor based on Machine Learning algorithms for the On-Board Driving (OBD) prediction of the Soot Tail pipe produced in a Diesel Engine. Such system was completed, building a Predictive Artificial Neural Network (ANN) in python, using calculation models belonging to the Deep Learning branch. The artificial intelligence systems are adequate for the resolution of this type of problems thanks to their high levels of precision and because they can deal with a large amount of data. This analysis was possible thanks to the data provided by AVL Italia S.r.l containing some measurements carried out with a diesel engine on a roller bench in stationary and transient conditions. The different datasets contain measurements made under different operating or environmental conditions and within them there is the trend of twenty-one features including the Soot Tail Pipe. Thanks to this, the predictive algorithm was implemented in supervised learning so that the model can collect input and output data from these sheets and then, through a training phase, it finds a rule which is useful for the generation of a desired output even for input values that it has never seen before. Specifically, I focused my attention on three Stationary Datasets with an engine speed variation between 800 rpm and 4500rpm, which differ from each other in ambient temperature and EGR conditions. At first, three Fully Connected Feed Forward Neural Networks (FFNN) for the prediction of the Soot Tail Pipe in stationary conditions were constructed. To make the neural networks as efficient as possible, numerous analyzes and tests were carried out, initially to study and understand the dataset’s behavior and then to detect the features of greatest relevance for the prediction of the soot through the tuning of the hyperparameters (feature importance). Subsequently, I improved the network parameters with appropriate algorithms for the upgrade of the performances and the minimization of the error between real and predicted values by a numerical evaluation of the Mean Square Error (MSE) and Determination Coefficient 〖(R〗^2). After that, I passed to focus my attention on the Transient Dataset, containing some measurements made on the WLTC normative guide cycle. Also, in this case a new Feed Forward Neural Network was build carrying out the same optimization processes as in the previous networks for the prediction of Soot Tail Pipe.