Koopman Operator-based System Identification & Prediction in Vehicular Applications with Nonlinear Dynamics

Abstract

This dissertation presents the development and applications of Koopman operator theory for solving system identification and prediction problems in the domain of vehicular systems and mobility from a data-driven perspective. The research is organized into three thrusts, whose results have been published in three peer-reviewed journal articles published over the past two years. The first article provides a comprehensive survey of Koopman operator theory from over 100 research papers and its applications to vehicular systems (aerospace, automotive, marine, rail, robotic, mining/construction, traffic management, and others), highlighting the potential for reduced-order modeling and control in this domain and gaps in the literature. The second article proposes a data-driven algorithm, i.e., the Hankel Alternative View of Koopman analytic approach, for predicting pre-ignition, a dynamically chaotic phenomenon, and resulting super-knock events in an internal combustion engine without the need for physics-based modeling. This is done within a framework designed for real-time implementation on an engine controller. This application has the potential to improve the operational adaptability of a vehicle, improving safety, performance and cost, by dealing with combustion instability through a ‘perdition and avoidance’ approach rather than the current ‘detection and mitigation’ approach. Simulation results from real engine data show the generation of a learned model in linear form capable of commanding one or more of several provided mitigating actions approximately 2.27 seconds prior to an event. Further validation results use data from low, medium, and high engine speeds within the envelope of low-speed pre-ignition. Finally, the third article discusses the use of Koopman operator theory for model generation of a tethered satellite system subject to unknown disturbances from several exogenous [nonlinear] environmental sources. The resulting state-space model shows excellent matching between the actual and simulated motion in simulation of the mission-critical maneuver of subsatellite deployment from its mothership. Overall, this research demonstrates the potential of the modern data-driven implementations of Koopman operator theory for system identification and control in various vehicular systems, with two specific case studies that illustrate this potential in a novel way.