A GAN-CNN Fusion Framework for Deep Learning-Based DOA Estimation in Low-SNR Environments

Area of Science:

• Signal Processing
• Machine Learning
• Array Signal Processing

Background:

• Direction of Arrival (DOA) estimation performance degrades significantly in low Signal-to-Noise Ratio (SNR) conditions.
• Traditional algorithms and deep learning models struggle with corrupted spatial information and limited training data in low SNR scenarios.

Purpose of the Study:

• To develop a novel two-stage framework for robust DOA estimation in challenging low-SNR and data-scarce environments.
• To enhance signal quality and extract robust spatial features for improved DOA accuracy.

Main Methods:

• A two-stage framework integrating a Generative Adversarial Network (GAN) for signal enhancement and a complex-valued Convolutional Neural Network (CNN) for DOA estimation.
• The GAN utilizes an attention mechanism and a phase-consistent loss function to reduce noise while preserving spatial phase.
• Enhanced signals are converted to covariance matrices and processed by the complex-valued CNN for feature extraction.

Main Results:

• Achieved 72.2% DOA accuracy and 3.9° Root Mean Square Error (RMSE) at -10 dB SNR with 500 snapshots.
• Substantially outperformed conventional and deep learning baseline methods.
• Demonstrated strong robustness with 93.8% accuracy using only 50 snapshots in data-scarce conditions.

Conclusions:

• The proposed framework offers a practical and effective solution for reliable DOA estimation in low-SNR and data-limited scenarios.
• The integration of GAN-based signal enhancement and complex-valued CNNs significantly improves DOA estimation performance.