Loading...
Learning-Based Direction-of-Arrival Estimation With Distributed Sparse Arrays in Underwater Environments
Citations
Altmetric:
Genre
Thesis/Dissertation
Date
2025-08
Advisor
Committee member
Group
Department
Electrical and Computer Engineering
Permanent link to this record
Collections
Research Projects
Organizational Units
Journal Issue
DOI
https://doi.org/10.34944/66zp-te60
Abstract
Source direction-of-arrival (DOA) estimation is an important objective in ocean surveillance. Vertical line arrays with uniform sensor spacing are commonly employed for this purpose. However, for a given number of sensors, these arrays provide limited aperture and lower degrees-of-freedom for DOA estimation as compared to sparse arrays that employ non-uniform sensor spacings. When there are size, weight, power, or cost constraints, distributed configurations of sparse arrays can be the right choice, especially when sonobuoys or unmanned underwater platforms are used for their deployment. Under a centralized processing framework, noncoherent processing relaxes the stringent calibration requirements imposed by coherent algorithms on distributed arrays. As the spatial covariance matrix corresponding to distributed sparse arrays is typically sparse with missing elements, high-resolution subspace-based techniques, such as Multiple Signal Classification (MUSIC), provide degraded performance, especially in the presence of array imperfections due to ocean currents and environmental gradients. In contrast, a pre-trained machine learning (ML) method that can infer the number of sources and estimate their DOAs accurately under array imperfections would be computationally better suited toward real-time operation.
In recent years, research has been conducted using ML methods for DOA estimation with uniform and sparse line array geometries, including distributed configurations. However, the focus of the works on distributed configurations has been primarily on additive white Gaussian noise assumptions. In the context of underwater surveillance, ocean ambient noise often deviates from this assumption as it arises from multiple sources, including ocean turbulance, marine life, geological processes, and commercial or industrial shipping activities. As such, its statistical characteristics vary based on location and time. In addition, multiple sensors in the array may pick up similar noise sources, resulting in spatial correlations. Leveraging existing ML-based schemes designed for white Gaussian noise assumptions, this thesis aims to devise ML-based solutions to estimate DOAs of multiple sources in the presence of underwater ambient noise with distributed sparse arrays. We focus on quiet deep-ocean waters, far from the shipping lanes, where ambient noise can be approximated as spatially-correlated Gaussian noise. Using simulated covariance matrices from a distributed system of three sparse subarrays with sensor imperfections as input, we demonstrate how a white Gaussian noise based ML model can be successfully adapted for spatially-correlated Gaussian noise using warm start or few shot learning. The proposed solutions yield capable pre-trained ML models that offer flexible and lightweight solutions for real-time operation in the field.
Description
Citation
Citation to related work
Has part
ADA compliance
For Americans with Disabilities Act (ADA) accommodation, including help with reading this content, please contact scholarshare@temple.edu