FREEDOM

Summary

In this project, researchers at Halmstad University collaborate with the software company WirelessCar and Laholm municipality. By analysing data collected from millions of connected vehicles, conclusion and predictions can made in order to develop sustainable vehicles and mobility services. The data analysis is done by using machine learning models and algorithms, developed at Halmstad University.

Background

Today, the transport sector stands for 25 percent of global carbon dioxide emissions, a number that has to go down drastically to reach the Paris agreement. To this end, many initiatives are ongoing: to rethink the need, to adapt behaviour, to change fuel, etc. Connected car data is a surprisingly untapped resource, and machine learning based on it is a crucial tool for making many of these mobility initiatives sustainable. Using vehicle data in the right way has a huge potential. This is something that the FREEDOM project aims to explore in order to decouple pollution and CO2 emissions from the mission of providing the necessary mobility for all.

Innovative transport solutions require accurate insights as input to decision makers. However, car manufacturers lack detailed knowledge of real-world usage for the vehicles they produce; owners and drivers are confused about consequences the decisions they make will have in their particular context, both for sustainability and economy; fleet operators provide inadequate arrangements and inefficient management due to lack of understanding of their distinct needs. All these actors will benefit from the data of millions of connected vehicles once it is analysed in the FREEDOM project. A pipeline from a large-scale car usage data lake maintained at WirelessCar into novel machine learning algorithms developed at Halmstad University will help develop services that lead to sustainable and efficient resource utilisation, while at the same time being realistic in terms of convenience and cost.

Mobility data is characterised by two crucially essential dimensions, namely spatial and temporal aspects. From a technical perspective, obtaining a complete picture requires a framework capable of modelling them simultaneously to take advantage of the insights embedded in the interrelations between the two. Graph Neural Networks (GNNs) is an emerging and promising field of machine learning, in the intersection of deep neural networks and graph theory, that is uniquely suitable to address both aspects. Spatial information is captured by the graph structure, while temporal information is modelled by recurrent neural networks in parallel with the neighbourhood aggregation message passing stage.