Location

Collaborative Design Environment (CoDE) Weber Space Science and Technology Building (SST II)

Paul Wang
(Advisor: Prof. Dimitri Mavris]

will propose a doctoral thesis entitled,

Domain Adaptable Digital Twins for Dynamic Environments

On

Monday, April 21 at 12:00 p.m. EDT 
Collaborative Design Environment (CoDE)

Weber Space Science and Technology Building (SST II)
and

Online: Click here to join the meeting

Abstract
Engineered systems are deployed in ever-changing operating conditions, contexts, and environments. These can include changes in terrain, atmosphere, the forces a system experiences during operation, or a specific system configuration, such as its sensor suite and payload. The growing complexity of these systems makes it costly and difficult to physically test for all possible conditions and configurations before deployment. In response, engineering organizations are rushing to digital twins to reduce costs and virtually evaluate system performance.

A digital twin is a virtual construct that imitates a physical system, consisting of a computational model and a bidirectional data feedback loop. However, digital twins are validated for a finite number of environments, conditions, and configurations, since it is prohibitively expensive to obtain test data on a system for all possible domains. The predictive capability of digital twins suffers when twins are operating outside their validated domain because the twin encounters out-of-distribution data.

Existing work has explored the application of various unsupervised domain adaptation techniques on digital twins, but these implementations do not account for evolving and unknown domains. This thesis proposes an online domain adaptation method for digital twins by aligning the subspaces of the validation and operational domains, using a subspace tracking method to incrementally update the subspace of an incoming operational data stream to a digital twin. After aligning the validation and operational domain subspaces, a domain adapted model can be learned in an aligned subspace representation. 

The proposed method preserves the predictive accuracy of a digital twin experiencing domain shifts, enabling an online approach to domain adaptation in digital twins that is currently not explored in the literature.

Committee

  • Prof. Dimitri N. Mavris – School of Aerospace Engineering (advisor)
  • Prof. Graeme J. Kennedy – School of Aerospace Engineering
  • Prof. Kyriakos G. Vamvoudakis – School of Aerospace Engineering
  • Dr. Olivia J. Pinon Fischer – School of Aerospace Engineering