State of Health estimation in EV batteries through deep neural networks using Transfer Learning approach

Modern environmental problems have highlighted the need to curb pollution, especially greenhouse gas emissions, to try to slow global warming, which has now reached alarming levels. The point of no return identified by the scientific community is a global average temperature increase of about 2°C over pre-industrial levels, a threshold beyond which it will be virtually impossible to avoid very serious and permanent damage to the environment. One of the areas where research is doing the most to reduce CO2 emissions is transportation, in particular, road transport being the most impactful, as well as the one where the prospect of achieving zero emissions is most feasible. The European Union has arranged a legislative package to incentivize the transition to electric or hybrid vehicles to reduce road transport emissions by 55% by 2030 and 100% by 2035. The increase in demand for electric vehicles has as an obvious consequence the increase in demand for electric batteries, especially lithium-ion batteries, being the most widely used technology. Such batteries are mainly described by two parameters, the SOC and the SOH: the former is an indicator of how much current is left in the battery and can be used before the battery is completely discharged; the latter is an indicator of battery aging and is used to understand when it needs to be disposed of. SOH estimation is a complex problem and can be approached in several ways. In this thesis work, first, the literature is explored and the various methods used are analyzed. Next, the work focuses on the development and analysis of deep neural networks, performing a comparative analysis highlighting which is the most performable. The datasets used for training and testing the networks are mainly two: the first contains real data collected by a private company on a Volkswagen eGolf, while the second contains simulated data created through a Simulink model. The simulated dataset is intended to complement the real dataset, as the real dataset does not cover the entire battery life cycle, and is used to perform an initial training step. Tested networks range belong to various types, from feedforward networks to recursive and convolutional networks, each time going for hyperparameter optimization.