Answers are grounded in this paper and cite their sources.
Direction of Arrival (DOA) estimation constitutes a fundamental challenge in array signal processing, particularly under low SNR conditions and limited snapshot availability. While traditional subspace decomposition methods suffer from performance degradation in noisy environments, data-driven approaches based on Convolutional Neural Networks (CNNs) face inherent limitations due to training data scarcity. To address this limitation, we propose a novel framework that integrates Generative Adversarial Networks (GANs) for data augmentation to synthesize high-fidelity covariance matrices that preserve the statistical characteristics of physics-based simulated array covariance matrices. Our method employs a Wasserstein GAN with Gradient Penalty (WGAN-GP) to generate synthetic covariance matrices that mimic real data distributions. These matrices are combined with real data to train a CNN optimized for DOA estimation. The proposed GAN-CNN achieves lower Root Mean Square Error and higher prediction accuracy compared to baseline methods (MUSIC, FSS, CS and CNN) across varying SNR levels and snapshot counts. The proposed GAN-CNN framework achieves a 33% reduction in RMSE against the best baseline at -20 dB SNR and achieves an RMSE of 3.45 ∘ in 50 snapshots. This approach mitigates data scarcity issues, improving model generalization and robustness, particularly in challenging environments. The framework offers a practical solution for real-world applications requiring accurate DOA estimation with limited resources.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: