DACL-Net: A Dual-Branch Attention-Based CNN-LSTM Network for DOA Estimation

Area of Science:

• Signal Processing
• Artificial Intelligence
• Deep Learning

Background:

• Deep learning methods are common for direction of arrival (DOA) estimation.
• Existing methods often fail to optimize input features, limiting accuracy improvements from attention mechanisms.

Purpose of the Study:

• To propose a novel spatio-temporal fusion model, DACL-Net, for enhanced DOA estimation.
• To improve the accuracy of DOA estimation by optimizing input features and employing attention mechanisms.

Main Methods:

• A spatial branch utilizes a 2D Fourier transform (2D-FT) on the covariance matrix, converting it into a magnitude spectrum where angles appear as peaks.
• A convolutional neural network (CNN) with an attention module focuses on these bright-spot components.
• A spectrum attention mechanism (SAM) enhances temporal feature extraction in the time branch.
• The model integrates spatial and temporal branches for simultaneous learning.

Main Results:

• DACL-Net demonstrates superior accuracy compared to existing DOA estimation algorithms.
• Achieved a Root Mean Square Error (RMSE) of 0.04° at a Signal-to-Noise Ratio (SNR) of 0 dB.
• The proposed feature transformation and attention mechanisms effectively enhance DOA estimation performance.

Conclusions:

• DACL-Net offers a significant advancement in DOA estimation accuracy.
• The spatio-temporal fusion approach combined with optimized feature representation is effective.
• The model provides a robust solution for DOA estimation in various signal conditions.