Representation learning for unsupervised heterogeneous multivariate time series segmentation and its application

Abstract

While driving a vehicle, data are collected from a huge number of sensors that generate both categorical and continuous variables with varying scales. In order to understand the status of the vehicles and the drivers’ behaviors, it is crucial to segment and identify different phases within this time series data. However, data often lacks labels to denote different phases, rendering supervised learning based segmentation methods as futile. Consequently, distance based time series segmentation method is a realistic solution for detecting different phases in the sensor data. However, there is no universal distance measure that utilizes both categorical and continuous variables simultaneously to segment the multivariate data.In this paper, we propose a novel unsupervised time series segmentation framework for heterogeneous multivariate data. By applying the distributed representation of the word embedding methods, we transform multivariate heterogeneous data into continuous vectors, allowing them to be segmented by conventional distance metrics such as Euclidean or Cosine distance. Subsequently, similar segments are clustered to generate general patterns. Without any labels or feature engineering, our framework successfully segments and discovers insightful driving patterns from heterogeneous sensor data collected from actual vehicles.