Deep Learning Methods in Direction-of-Arrival Estimation: Recreating a Novel Architecture in Establishing Robust, Super-resolution Estimates of Angle of Arrival in Single Snapshot Environments

Thumbnail Image

Files

URL

Journal Title

Journal ISSN

Volume Title

Sähkötekniikan korkeakoulu | Bachelor's thesis
Electronic archive copy is available locally at the Harald Herlin Learning Centre. The staff of Aalto University has access to the electronic bachelor's theses by logging into Aaltodoc with their personal Aalto user ID. Read more about the availability of the bachelor's theses.
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.

Authors

Date

2025-05-09

Department

Major/Subject

Digital Systems and Design

Mcode

ELEC3056

Degree programme

Aalto Bachelor's Programme in Science and Technology

Language

en

Pages

30+5

Series

Abstract

In recent years, an adoption of various machine learning (ML) methods has been prevalent in the field of signal processing, mainly motivated by the rapid increase of computational resources, ease of access to cloud computing, and ever-increasing amount of data. Therefore, the aim of this thesis is to evaluate the potential of the fully connected neural network (FCNN) for direction of arrival (DoA) estimation by replicating a study by [BGT19] in order to compare it against existing neural network (NN) architectures. NNs are generally trained from multichannel data of an antenna array output and are able to predict angular directions using a single snapshot. This thesis determines ways for learning-based methods to address and handle common issues faced by classical model-based methods as well as the drawbacks that accompany end-to-end methods. The review of state-of-art methods, anchored in ML, highlights the importance of modeling assumptions, training, validation and testing process, origin of training data. Experimental results demonstrate a significant contribution and advancement potential,however, due to the infancy of the DL method approaches within DoA estimation context, many results must be vehemently validated, as they are anchored in concrete assumptions. Such assumptions are not always realistic, applicable, or barely met to realize the empirically offered gains by the efforts of researchers. That is, the performance of DL may look great on paper, such as in this case with up to 4 sources, but higher number of sources with smaller angular separation results in degraded performance. Manyofthesubtleties of utilizing DL methods in DoA estimation are occasionally omitted or concealed. Such omissions result in an unattainable recreation of what hypotheses, propositions and empirical results are demonstrated. In this validation study, the importance of recreatibility is highlighted. Upon finalizing the experiment, the network yielded satisfactory yet did not attain the performance demonstrated in [BGT19]. Reasons for unattainable performance are discussed in the end of the thesis.

Description

Supervisor

Ylirisku, Salu

Thesis advisor

Ollila, Esa

Keywords

direction-of-arrival, maximum likelihood estimation, model order determination, neural network, parameter estimation, single snapshot

Other note

Citation

Permanent link to this item

https://urn.fi/URN:NBN:fi:aalto-202506245209

Collections

[kand] Sähkötekniikan korkeakoulu / ELEC