Real-Time 2D+t Low-Order Flow Field Reconstruction from Sparse Measurements for Improved AUV Navigation
Files
TR Number
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This thesis presents the development of a real-time, low-order data assimilation framework for reconstructing local ocean flow fields from sparse observations. The central challenge is that an autonomous underwater vehicle (AUV) typically observes only one-dimensional vertical velocity profiles, while the surrounding flow of interest is a higher-dimensional, spatially continuous field. Recovering that field from such limited data is therefore fundamentally underdetermined. The objective of this work is to develop a reconstruction methodology that is both computationally efficient enough for onboard, real-time execution and robust enough to preserve physically meaningful flow structure under sparse sampling conditions.
The proposed framework addresses this challenge by combining sparse measurement data with low-order flow representations and additional physical constraints, including temporal coherence and divergence-free structure. In this thesis, the notation ``2D+t'' denotes a sequence of two-dimensional transverse-plane reconstructions updated using recent measurement history. It does not denote direct time integration of a governing flow equation over the full reconstruction domain. Temporal information is introduced through local corrective forcing term derived from the local temporal history rather than through a full transient PDE solver. These constraints are necessary because the measurements alone do not uniquely determine the surrounding flow field. Instead, prior structural information and embedded physical consistency must be introduced to regularize the problem and produce stable, interpretable reconstructions. In particular, the divergence-free constraint improves the physical realism of the recovered field by enforcing incompressibility during interpolation.
To improve robustness and overcome rank-deficiency issues associated with sparse one-dimensional sampling, this thesis introduces a hybrid methodology that couples a planar flow identification model with a divergence-free radial basis function (DF-RBF) reconstruction framework. The planar flow identification model acts as a low-order, non-uniform surrogate model for the local flow field by combining a uniform background component with a dominant singular/coherent component. The planar model captures the dominant low-order rotational flow structure, while the DF-RBF stage uses that structural information to reconstruct a continuous field that remains physically consistent. Together, these components enable stable two-dimensional flow field reconstruction from limited vertical profile measurements.
The reconstruction problem is formulated through coupled linear and constrained optimization systems, including least-squares estimation. These systems are solved within a Fortran-based computational framework using LAPACK routines to achieve performance appropriate for real-time application.
The broader significance of this work is that reliable onboard flow reconstruction may provide a useful data-assimilation layer for future three-dimensional, time-resolved (3D+t) forefield prediction. A predictive forefield model can only be as reliable as the assimilated flow information used to initialize or constrain it. For that reason, the present reconstruction framework is developed not as an isolated mapping tool, but as the foundational data assimilation layer required for future real-time 3D+t ocean flow prediction around an AUV. Such capability would support improved navigation, environmental awareness, and mission planning in dynamic ocean environments.