Date of Publication

12-2024

Document Type

Master's Thesis

Degree Name

Bachelor of Science in Electronics Engineering (Honors) - Ladderized

Subject Categories

Electrical and Computer Engineering

College

Gokongwei College of Engineering

Department/Unit

Electronics And Communications Engg

Thesis Advisor

Gerald P. Arada

Defense Panel Chair

Edwin Sybingco

Defense Panel Member

John Anthony C. Jose
Lawrence Y. Materum

Abstract (English)

The integration of deep learning into array signal processing has shown promise, particularly in Direction-of-Arrival (DOA) estimation. Traditionally, the DOA estimation problem is addressed through multi-target regression or multiclass classification. However, regression-based approaches impose a fixed target number, limiting adaptability in dynamic environments where source counts vary. Meanwhile, multiclass classification, despite mapping the angular spectrum, restricts predictions to discrete angles, introducing quantization errors. This study proposes an end-to-end convolutional neural network (CNN) model capable of 32-bit precise real-valued DOA estimation with flexible cardinality. A novel peak-finding algorithm was developed, leveraging non-discrete information from generated data to adjust detected peaks in the angular spectrum. This solution integrates a DOA estimation network for spectrum generation and a source-number classification network to predict the number of signals present. The model's performance was rigorously benchmarked against two established models across diverse scenarios, including near-endfire, closely spaced sources, position, gain, and phase perturbations, and scenarios with well-separated DOAs at varying source counts.

Key results demonstrated a high angular resolution at each angular grid with a mean RMSE of 0.16 degrees, consistently below 0.2 degrees for any source number, surpassing discrete classifiers that tend to saturate under favorable signal-to-noise ratio (SNR) and number of snapshots conditions. Despite not being explicitly trained on perturbed data, the model yielded error rates comparable to MUSIC algorithm. Additionally, it also maintained its performance in challenging spectrum edge cases with a 0.22-degree RMSE. The model could also resolve sources separated by as little as 5 degrees and was able to predict DOAs with cardinalities exceeding its training range.

These findings imply that the proposed model could offer more versatile and accurate DOA estimation in real-world applications, especially in dynamic and challenging environments where source numbers fluctuate, or physical conditions introduce perturbations. By addressing the limitations of traditional methods, this approach has the potential to advance robust signal processing capabilities in areas such as wireless communications, radar systems, and sonar, where high-resolution, continuous-valued DOA estimations are essential for enhanced situational awareness and performance.

Abstract Format

html

Abstract (Filipino)

None

Abstract Format

html

Language

English

Format

Electronic

Keywords

Signal processing

Upload Full Text

wf_yes

Embargo Period

12-2026

Download

Available for download on Tuesday, December 01, 2026

Share

COinS