Digital Twins with Uncertain and Imperfect Data

Abstract

Digital Twin is one of the promising digital technologies being developed to support decision making and has gained considerable attention in both industry and academia in recent years. This has resulted in an increasing variety of definitions that threatens to dilute the concept and lead to ineffective implementations. There is a need for a consolidated and generalized definition, with clearly established characteristics to distinguish what constitutes a Digital Twin and what does not. Furthermore, a comprehensive framework for the implementation of a Digital Twin considering the generalized characterization of its key components is essential to facilitate its effective use. Additionally, Digital Twins are being used to represent engineering systems that are increasingly complex, often with uncertain and imperfect input data being the only basis for the update of these models. As such, Digital Twin implementations must also consider the integration of uncertainty quantification during the model updating process. The research effort in this dissertation is divided into four objectives. The first objective focuses on establishing a clear and concise definition and characterization of a Digital Twin, generalized so that it can be applied in all use cases. In the second objective, using this generalized definition and characterization, a framework for a Digital Twin implementation is developed and demonstrated. The third objective investigates the incorporation of abstracted data which occurs when raw data has been reduced to a simplified representation of portions or the entirety of the raw dataset. The fourth objective considers the incorporation of uncertain data arising from machine learning models, specifically categorical data from classification models which occurs when observations of the system result in uncertain and/or ambiguous interpretations. Several computational methods are proposed to overcome different challenges with these uncertain and imperfect data sources and a Bayesian approach is pursued for the incorporation of this uncertainty. The consideration of uncertain and imperfect data in the implementation of Digital Twins are presented for several applications which examined use cases for monitoring and managing decisions concerning degradation mechanisms such as corrosion and fatigue in marine vessels.