Graduation Year

2023

Document Type

Dissertation

Degree

Ph.D.

Degree Name

Doctor of Philosophy (Ph.D.)

Degree Granting Department

Industrial and Management Systems Engineering

Major Professor

Trung (Tim) Quoc Le, Ph.D.

Committee Member

Om Prakash Yadav, Ph.D.

Committee Member

Hadi Charkgard, Ph.D.

Committee Member

Susana Lai-Yuen, Ph.D.

Committee Member

Mingyang Li, Ph.D.

Committee Member

Arman Sargolzaei, Ph.D.

Keywords

complex system dynamics, data-driven sensor-based modeling, multi-modal health sensing data analysis, multi-physics integration, multi-resolution systems modeling

Abstract

Nonlinear dynamical systems have been extensively used to model various phenomena in the changing world around us, especially in science and engineering fields. Thanks to breakthrough advancements in sensing technologies, an increasingly high volume of multi-modal sensor data has been collected, which enables us gain better insights into complex systems dynamics and build sophisticated data-driven machine-learning-based dynamic models without having the access to the underlying governing equations. However, integrating domain-specific knowledge in machine learning algorithms remains pivotal for various reasons: it promises enhanced predictive accuracy, better model interpretability, and increased generalizability. This dissertation delves into three core research questions, each highlighting a unique facet of domain-informed machine learning's application in complex, dynamic systems.

The first research question probes the methodology and efficacy of incorporating prior domain knowledge into machine learning models. Traditional data-driven machine learning approaches, while powerful, often overlook the vast body of existing domain expertise. Our research showcases the novel approaches, methodologies, and architectures that allow seamless integration of this domain knowledge. Our results indicate that with appropriate incorporation techniques, machine learning models can achieve significantly improved accuracy, offer richer interpretability, and display superior generalization capabilities across varied datasets.

Addressing the second question, we venture into the intricacies of dynamic models designed to handle multi-scale and multi-physics interactions with high noise robustness. With many real-world systems governed by processes occurring at different temporal and spatial scales, and often influenced by diverse physical phenomena, there's a pressing need for models that capture this complexity. Our study offers advancements in state-of-the-art dynamic models, emphasizing their adaptability and robustness, particularly in noisy environments. Through rigorous testing and validation, we demonstrate how these refined models present a more comprehensive representation of intricate systems, mitigating inaccuracies arising from scale mismatches, multi-physics interplays, and data noise.

The third leg of our investigation ties domain knowledge to the realm of control in complex systems. Control strategies for intricate systems can gain substantially from domain-specific insights, ensuring not just system stability but also optimality in performance. We present frameworks and methodologies to fuse domain knowledge into control design, leading to strategies that are more attuned to the intrinsic characteristics of the systems they govern. Real-world applications, from industrial automation to ecological system management, highlight the profound benefits of these enriched control approaches.

In summation, this dissertation underscores the indispensable role of domain knowledge in shaping the next generation of machine learning models and dynamic systems control. Through focused research on each of our three foundational questions, we hope to chart a path forward for scholars and practitioners aiming to harness the full potential of domain-informed machine learning in understanding and managing complex systems.

Download

Share

COinS