Unsupervised Multivariate Time Series Segmentation: Driving Scene Identification Use Case

In today's data-driven world, extracting meaningful insights from complex datasets is crucial for progress and innovation. This challenge takes on new dimensions with multivariate time series, where multiple signals intertwine to describe a phenomenon. The thesis tackles this challenge in the context of Toyota Motor Europe, focusing on segmenting Controller Area Network (CAN) data, a cornerstone of modern vehicles. Motivated by the need to understand driver behavior better, this thesis presents a novel pipeline for unsupervised segmentation of multivariate CAN time series. Leveraging the potential of data-driven approaches, the system automatically identifies significant events within the data, eliminating the need for manual labeling or prior knowledge. As a proof of concept, the pipeline addresses the Driving Scene Identification problem. Given a CAN data stream representing driving trips, the system segments it into variable-length segments corresponding to distinct driving maneuvers. These segments are then linked to specific driving scenes through statistical analysis. This thesis delves into the underlying technologies and algorithms, providing a comprehensive explanation of the pipeline's architecture and functionalities. The effectiveness of the system is demonstrated through the Driving Scene Identification application, showcasing the extracted segments and their association with meaningful driving behaviors.