Eric Carlsson och Oskar Hellqvist, Elektroteknik

​Titel: Explainable machine learning for state-of-health prediction using electric vehicle histogram data
Examinator: Torsten Wik

The popularity of electric vehicles is rapidly growing, owning much to advancements made in battery technology that allow them to compete with traditional Internal Combustion Engine vehicles in terms of price and performance. However, a key problem that remains for automotive manufacturers and prospective electric vehicle buyers is that batteries age with and time and use. A number of different methods have been used previously to predict battery state-of-health (SOH), ranging from detailed electrochemical models to semi-empirical parametric models. With the increased availability of large-scale vehicle data, attempts have also been made with purely statistical models. A major issue with this approach is how to collect and store the data with the limited hardware resources that are available inside vehicles. Historically, this has been solved by accumulating data over time into a histogram format that is easier to store.

This thesis aims to investigate how state-of-health prediction models are affected by the proprietary format of battery data that is commonly found within the automotive industry. This is done by utilizing explainability methods such as feature importance and SHAP, to see how the behaviour of prediction models reacts to changes in the dataset distribution and format. This is done using real-world electric vehicle data, collected during service visits for a large fleet of Volvo customer Plug-in Hybrid Electric Vehicles (PHEVs), that is processed into a usable format with a novel data pre-processing pipeline. In order to make the dataset and resulting models comparable to other, publicly available datasets, a novel method for transforming time series battery data into a format that is similar to the proprietary vehicle data format is presented.

The results are inconclusive, but indicate that the proprietary data format has some degree of impact on the state-of-health prediction models. Furthermore, the results indicate that correlation between features and feature-time coupling may have an even more considerable impact.

Eric, Oskar and Torsten
Kategori Studentarbete
Plats: Landahlsrummet, conference room, Hörsalsvägen 9, EDIT trappa F, G och H
Tid: 2022-06-07 15:30
Sluttid: 2022-06-07 16:00

Sidansvarig Publicerad: ti 24 maj 2022.