Neural Network Algorithm for Water Dew Point Detection in Diesel Passenger Car's Exhaust Pipeline

In this work a research problem is going to be discussed. The hosting company is Bosch, one of the main software and ECU supplier of the world, being in particular the reference for all the main Italian OEMs. Bosch engineering team supports the customers, providing calibration services for different software functions that ranges from thermal models to rail and engine components government. This project has the goal of developing a proof of concept model, based on Machine Leaning (ML) architecture, showing the capabilities of the technology in improving the performances of the actual software function chosen for this study. The name of the function is "Dew Point Detection", which has the aim of predicting the presence of condensed liquid water inside the exhaust pipeline, fundamental to ensure the proper operation of the pollutant sensors (NOx, PM) mounted along the ATS (After Treatment System) exhaust pipeline. The basic concept is that every combustion process produces discharge gasses, among which vapor water (steam) is a relevant fraction. Hot gasses travel along the pipe and if the vehicle is operating in cold environmental temperature conditions, water condenses on the pipe wall. The issue is that modern combustion engines are equipped with many exhaust sensors which are in charge to monitor the efficiency of the catalysts and produce the feedback signal employed in the ECU to implement closed loop control on the combustion process itself. These sensors require external active electrical heating (ECU controlled), which sets the actual operating temperature of the sensors itself around 750 °C. If liquid water drops hit the sensor in these specific working conditions, thermal shock happens on the sensing surface (made on ceramic materials) causing the destruction of the device. To avoid this, the so called "sensors release time" is feasible after all the liquid water is evaporated (pipe dry), thanks to the heat exchange process that naturally happens with the hot gasses when the engine has reached its rated operating temperature. From this point on, no more water is able to condense on the pipe surfaces and it's safe to operate the sensors normally. The motivations that have pushed the development of this solutions are many, but the fundamentals are: 1. Need of performance gain in terms of accuracy and reliability: the more precise is the dry-condition estimation, the better is the control of the pollutant emissions (in particular fo homologation requirements). 2. Economical cost saving compared to the standard calibration procedure, enabled by a self supervised algorithm that is able to collect and manage both the learning and the testing phase. This thesis project is gonna study only to the 1. motivation, because the preliminary goal is to evaluate the possible outcomes coming from the core estimation algorithm. The structure of this work tries to stick as much as possible the scientific and engineering best practices which involves data collection, methodologies and critical analysis of the results. In the first chapter the needed background to give context to the problem is provided, focusing on the details of the basic principle that characterize the phenomenon. Then the procedure followed to deal with the data, later employed to train and test the model, including also the details of the technical implementation of the solution. Finally the results are presented highlighting of performance reached, fine tuning the procedure and discussing future improvements.