Combining analytic and data-driven machine-learning based methods for efficient adaptive dynamic modelling
Abstract
In many industrial application areas, such as the aerospace industry, obtaining a personalized dynamic mathematical model of each vehicle which can reliably describe the flight characteristics in an automated and low cots fashion has paramount importance. This importance is based on the possibility to achieve the operational limits of the vehicle in terms of maneuvering performance, fuel efficiency, etc., based on flight controllers, tuned with this accurate model knowledge. Inspired by these practical challenges, in this research, we focus on establishing a new theory and novel algorithmic tools to cope with the data-driven modeling problem of general systems, including vehicles, where imperfect prior knowledge of the dynamics exists and an automated modelling process is required which is capable to adapt the model even online based on measurement data from the system. Specifically, the research will focus on the case, when the model mismatch is significant w.r.t. a given physically well understood baseline model of the system dynamics, either because of manufacturing tolerances, imprecise knowledge of complicated aspects of the dynamics, abrupt changes of the dynamics during operation. To cope with this challenge, we aim to achieve efficient fusion of the given prior knowledge on the dynamics with measured data from the system in order to provide rapid adaptation capabilities of the base-line model. For this purpose, the team proposes novel model-augmentation methods that build on the the state-of-the-art approaches of machine-learning based system identification in terms of state-space artificial neural network (SS-ANN) models via deep-learning and Gaussian process (GP) based states-pace models using Bayesian estimation algorithms. Building on the augmented model structures, a novel control design methodology is also aimed to be realized for constructing efficient, optimal, feedback controllers that can be implemented for fast real-time operations. We propose to extend the theory of predictive and linear parameter-varying control synthesis and establish a theoretical background for deriving mathematical guarantees for safety and high operational performance in a robust fashion.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 22, 2024
- Source ID
- FA86552317061
Entities
People
- Roland Toth
Organizations
- Air Force Office of Scientific Research
- United States Air Force