Vector Embeddings to Estimate Angle of Arrival for Wireless Signals

open_access
Open Access
Author:
Misra, Bharavi
Area of Honors:
Computer Engineering
Degree:
Bachelor of Science
Document Type:
Thesis
Thesis Supervisors:
Jinghui Chen, Thesis Supervisor
John Morgan Sampson, Thesis Honors Advisor
Keywords:
Vector Embeddings
Deep Learning
Angle of Arrival Estimation
Signal Processing
Abstract:
This thesis explores the application of deep learning techniques to the problem of angle of arrival (AoA) estimation in wireless communication systems. I propose a novel neural architec- ture that leverages Data2VecAudio, a self-supervised model pre-trained on speech, vision, and text, as a feature extractor for complex antenna array signals. The model is trained end-to-end using synthetic data generated in MATLAB and is evaluated against classical signal processing baselines, including MUSIC and Beamscan. Experimental results demonstrate that the proposed model performs comparably to traditional methods and exhibits strong generalization capabili- ties, despite being trained on a limited and simplified dataset. Notably, the model is able to infer meaningful spatial patterns from raw complex I/Q data, even with variable-length inputs. How- ever, several limitations remain, including reliance on simulated data, simplified noise models, fixed source counts, and a computationally intensive architecture. These findings highlight both the potential and the current challenges of applying general-purpose vector embeddings to wire- less signal processing tasks, and provide a foundation for future research into more efficient and scalable architectures suitable for deployment in real-world environments.