کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

High-Resolution Direction of Arrival Estimation using Hybrid Transformer-Graph Neural Networks: A Deep Learning Approach for Uniform Linear

نویسندگان :

Mohammadhossein Sadeghi¹ Morteza Alikhani² Seyed mohammad Karbasi³

1- دانشگاه صنعتی شریف 2- دانشگاه صنعتی شریف 3- دانشگاه صنعتی شریف

کلمات کلیدی :

Direction of Arrival (DOA) Estimation،Graph Neural Network (GNN)،Transformer،Graph Attention Network (GAT)

چکیده :

Direction of Arrival (DOA) estimation serves as a pivotal component in array signal processing, underpinning the functionality of advanced systems in radar, sonar, and wireless communication. While traditional subspace-based methodologies, such as Multiple Signal Classification (MUSIC) and Estimation of Signal Parameters via Rotational Invariance Techniques (ESPRIT), have historically provided high-resolution estimates, their performance is notably constrained by finite snapshots, low signal-to-noise ratios (SNR), and hardware-induced array imperfections. This research report introduces a novel hybrid architecture, the Transformer-Graph Network (TGN), designed to transcend these limitations by integrating spatial-relational modeling and temporal dependency learning. The TGN architecture treats the antenna array as a graph structure, where sensors represent nodes, and utilizes the multi-head attention mechanisms of Transformers to capture complex correlations across multiple signal snapshots. A rigorous data generation framework is established, simulating a 10-element uniform linear array (ULA) subjected to various environmental and hardware stresses. The proposed method is benchmarked against four classical estimation techniques: MUSIC, ESPRIT, Root-MUSIC, and the Capon beamformer. Experimental evaluations demonstrate that the TGN architecture achieves superior precision, yielding a significant reduction in Root Mean Square Error (RMSE) and high Area Under the Curve (AUC) and F1-score values. Specifically, evaluations across a range of SNR levels show that the TGN architecture achieves an RMSE of approximately 1.2° at 10dB, maintaining superior performance even in high-noise environments where traditional benchmarks exhibit significant instability, outperforming traditional benchmarks while maintaining computational robustness. This report provides an exhaustive technical analysis of the network blocks, data labeling processes, and training strategies implemented on high-performance A100 GPU infrastructure.