USID - Unsupervised Identification of the Driver for Vehicle Comfort Functions

Abstract

Modern vehicles must meet the rapidly increasing customer requirements for in-vehicle comfort. Moreover, the consumer’s demand for personalized comfort functions keeps growing accordingly. Vehicle comfort functions are designed to ensure comfort with a more pleasant and convenient driving experience, including thermal comfort, seating comfort, noise reduction, and entertainment systems. They are highly customized for individual drivers to enhance their effectiveness. To achieve precise adjustment for each driver of a vehicle, it is imperative to identify the driver accurately.The distinctive signature of each driver’s driving style is embedded within the time series data gathered from various vehicle sensors. Therefore, there is no need for integrating novel driver identification sensors, as the identification task can be accomplished by utilizing existing and standardized vehicle data from the On-Board Diagnostic II system (OBD II).This research paper investigates the viability of precise driver identification based on unsupervised machine learning methods. For this purpose, the USID (Unsupervised Identification of the Driver) concept is introduced. Existing driver identification systems are functional, yet they exhibit limitations in precision. For example, a vehicle could detect the presence of a driver’s smartphone but fail to distinguish whether the individual is the actual driver or a passenger. In contrast to the state-of-the-art, the USID approach proposes a seamless and precise solution for driver identification. Due to its unsupervised characteristic, USID offers significant scalability, as its models don’t rely on predefined classes to distinguish and identify drivers. Our USID approach is inspired by the Knowledge Discovery in Databases (KDD) process and comprises several stages. In the initial phase of our process pipeline, we perform data quality checks, identifying and addressing outliers to avoid any potential distortions in model inferences. Given that time series data represents a flow of data points, we apply moving windows to split the time series sequence into variable-length segments. The choice of window length is intended to capture relevant events for predictive purposes. In order to ensure the relevance of our time sequence comparisons across a diverse range of driving contexts, drivers, and vehicles, we utilize global normalization techniques incorporating physical signal boundaries. Our approach integrates feature selection methods, such as information gain, signal-to-signal correlation analysis, and principal component analysis (PCA), facilitating the extraction of the most informative features while effectively reducing input dimensionality. As a final step in transformation, we propose an optional feature space transformation to align with the selected model’s input space. The recorded real OBD II time series of 24 drivers and their driving cycles are used for the purpose of training the models in USID. The dataset comprises 10 driving cycles from a single vehicle driven by 10 different individuals and an additional 14 driving cycles from 14 vehicles, each driven by distinct drivers. Consequently, this paper also explores the feasibility of accurately distinguishing between different drivers operating various vehicles.Unsupervised machine learning methods used in USID models are K-Means, Autoencoders, Self-Organizing maps, and Density-based spatial clustering (DBSCAN). In the end, the models are tested, and the results are outlined with the comparison of the quality of models for driver identification.