Algorithmic vehicle data analysis as business enabler of electric and distributed mobility

authored by

Leonie von Wahl

supervised by

Wolfgang Nejdl

Abstract

In recent years, the massive collection of mobility data has become possible due to the installation of sensors in vehicles and the usage of connected devices like smartphones. This data reveals opportunities for helpful applications, such as reducing carbon emissions, improving vehicle quality and run time, making mobility more accessible to everyone, and introducing autonomous driving. With the increase in mobility data collection, we have also experienced growth in data analysis methods, such as machine learning. While neural networks were able to detect relatively simple patterns a decade ago, modern deep learning will overturn each aspect of daily life in the future and already does so in the present. Leveraging mobility data results in unique challenges. First, we must pay attention to how we collect the data. We have to make sure to obtain unbiased coverage of the observed space. Since deep learning models tend to amplify the bias found in the training data, it is crucial to collect meaningful data. Moreover, we need preprocessing methods that make the real-world data usable for training. Second, data privacy concerns have become increasingly important since multiple abuse possibilities were unearthed in the past. Mobility data is especially critical as it can betray the user’s location, behaviour, or potential trespassing of traffic rules. Therefore, we need to find privacy-preserving mechanisms to allow individuals to govern their data. Third, data transfer causes huge communication costs, hindering effective classic training by companies gathering user data. This thesis presents novel algorithms to address these challenges. In particular, we examine how vehicle data analysis can be used as a business enabler for electric mobility use cases. Thus, we suggest a novel routing algorithm that covers an entire urban road network to collect environmental sensor data in Chapter 3. This routing algorithm is designed as a side business for ride-hailing fleets. Then, we examine an asynchronous federated learning mechanism in Chapter 4 to enable predictive maintenance for vehicle or aeroplane transportation fleets. Federated learning protects user data and reduces communication costs. Our asynchronous aggregation scheme overcomes the challenges of temporal unavailability and data disparity that often occur in the transportation sector. Moreover, we develop the first early stopping rule for asynchronous federated learning. Last, we turn to a mobility data application, the placement of charging infrastructure in an urban road network in Chapter 5. For the first time, we adopt reinforcement learning to find an optimal placement and charger type configuration based on charging demand data. In this setting, we take the previously neglected private charging infrastructure into account.

Organisation(s)

L3S Research Centre

Type

Doctoral thesis

No. of pages

108

Publication date

22.01.2025

Publication status

Published

Sustainable Development Goals

SDG 11 - Sustainable Cities and Communities

Electronic version(s)

https://doi.org/10.15488/18347 (Access: Open)