DATA-DRIVEN PREDICTION, DESIGN, AND CONTROL OF SYSTEM BEHAVIOR USING STATISTICAL LEARNING

Loading...
Thumbnail Image

Files

Publication or External Link

Date

2021

Citation

Abstract

The goal in this dissertation is to develop new data-driven techniques for prediction, design, and control of the behavior of a variety of engineering systems. The data used can be obtained from a variety of sources, including from offline, high-fidelity system’s simulation, physical experiments, and online, sparse measurements from sensors. Three inter-related research directions are followed in this dissertation. Following the first direction, the author presents a multi-step-ahead prediction technique for evaluating a single-response (or single-output of the) system’s behavior through an integration of the data obtained offline from the system’s high-fidelity simulation, and online from single sensor measurements. With regard to the first research direction, the key contribution includes a reasonably fast and accurate prediction strategy that can be used, among others, for online, multi-step ahead forecasting of the system’s operational behavior. Building on the work from the first direction, the author follows a second research direction to present a multi-step ahead prediction technique, this time for a multi-response system’s behavior, that can be used for evaluating various system’s designs and corresponding operations. Data in this case is obtained from the offline, high-fidelity system’s simulations, and online sparse measurements from multiple sensors (or limited number of physical experiments). The main contribution for this second direction is in construction of a new data-driven, multi-response prediction framework that has a robust predictive capability. Along the third research direction, a data-driven technique is used for prediction and co-optimization of a system’s design and control. The data in this case is obtained from sensor measurements or a simulator. The main contribution achieved through the third direction is a new data-driven reinforcement learning-based prediction and co-optimization approach. The methods from this dissertation have numerous applications, including those demonstrated here: (i) assessment of safe aircraft flight conditions (Chapters 2 and 3), (ii) evaluation of design and operation of a robotic appendage (Chapter 3), and (iii) design and control of a traffic system (Chapter 4).

Notes

Rights