A data-driven approach to improve Battery Management and predict State Of Health: analysis of heavy commercial vehicles leveraging telematic data

In the next few years electric transport will become increasingly widespread. In fact it will be of paramount importance to reduce the polluting emissions produced by traditional engines. So, while waiting for electric trucks to take hold, it is important to start analysing the battery. Indeed, it is a key component even in non-electric vehicles: it provides the jolt of electricity necessary to power all the electrical components which are becoming more and more numerous. This thesis focuses on the application of a data-driven methodology to improve Battery Management and predict State Of Health (SOH) in heavy commercial vehicle’s. The analysis will exploit telematic data of trucks produced by IVECO: a vehicle tracking device is installed on board to send, receive and store telemetry data. In particular, the vehicles life cycle is split into two distinct periods: stock and travelling. Regarding the first phase, the aim is to extract simple advice to follow in order to preserve the battery of trucks in stock. This will avoid a preventive battery replacement by reducing IVECO's costs. Instead, once the vehicle has been sold, we define a forecast model to be able to predict the battery State Of Health after a fixed time horizon. During this blind window, the user, in the event of a deteriorated battery prediction, will be able to carry out timely repairs to the battery. In fact, a sudden breakdown of this component, and therefore of the vehicle, causes a considerable economic loss for the shipping industry. The data-driven methodology followed is based on a data preparation and cleansing phase, a step of Feature Engineering and Feature Selection. Finally, the model is trained using only the top predictors selected. This approach requires the model training in distinct scenarios, characterized by an input and blind window and different regression algorithms. In conclusion, the optimal performance resulting from the data-driven methodology is compared with an approach based on Neural Networks.