Data-driven Modeling, State Classification, and Performance Monitoring of Complex Systems

Abstract

This work is focused on using data-driven methods to investigate state behavior of complex systems. A complex system consists of several interacting elements and system behavior cannot be trivially inferred from the collective behavior of the individual elements. Important research questions in the study of complex systems include how to characterize individual agent behavior, and how system behavior may emerge from the collective behavior of the multi agent system. The work in this dissertation focuses on understanding and addressing modeling and analysis challenges in complex systems at both an agent- and system-level for manufacturing and biological systems. My work in the manufacturing space is focused on using event-based modelling and machine learning to understand the interaction between working stations in a cylinder head production line to enhance performance monitoring. Companies strive to increase efficiency, improve quality and reduce costs by reducing downtime and improving productivity. In a manufacturing system, performance monitoring must consider the behavior of both individual machines and the interactions between these machines at the system level. Additionally, environmental factors, including: machine health, maintenance schedule, supply of raw materials and customer demands must be considered. Therefore, to develop an intelligent performance monitoring system, it is essential to understand both machine- and system-level interactions. However, current methods are either focused on monitoring single machines or simplified systems. To address this gap, a state-based model was developed to describe the performance of an individual machine, and a data-driven modeling approach was proposed to monitor system-level interactions, enabling the analysis of local disruptions on the overall performance of the system. Elements of the framework developed in the manufacturing space were then applied to a compelling biological system, humpback whales. It is of interest for biologists to understand humpback whales' behavior in the wild, especially underwater behavior when direct observation is not available. Human activities including commercial fishing and shipping are threats to humpback whales in the wild. For example, fishing gear entanglement is the leading cause of injury and deaths for humpback whales. To help understand whale behavior and mitigate these threats, it is important to gain a better understanding of the animals in their natural habitat. In this part of my work, data-driven methods were used to combine information from multiple data sources to investigate the behavior of humpback whales in the Gulf of Maine. These results have improved or understanding of animal movement, foraging ecology, and the temporal and spatial distribution of behavioral states at day scale. Importantly this new knowledge will directly inform mitigation strategies that seek to reduce animal entanglement and vessel collision. The ability to provide enhanced system-level classification and understanding of complex systems has the potential to impact several fields. This research provides new frameworks for state classification and characterization that can be used to understand agent behavior and system-level interactions within a manufacturing system. This framework was also applied for biological characterization to identify behavior patterns across day-scale time intervals. While these fields are drastically different, there are interesting aspects including state transition maps from each system that can lead to new insights in how systems interact across time and spatial domains.