ACGAN-Based Multi-Target Elevation Estimation with Vector Sensor Arrays in Low-SNR Environments.

Area of Science:

• Signal Processing
• Acoustics
• Machine Learning

Background:

• Direction-of-arrival (DOA) estimation accuracy degrades significantly in low signal-to-noise ratio (SNR) conditions and with multiple interfering sources.
• Traditional methods struggle to maintain performance under these challenging acoustic environments.

Purpose of the Study:

• To propose a novel deep learning architecture, the Auxiliary Classifier Generative Adversarial Network (ACGAN), for robust DOA estimation.
• To enhance the ACGAN with a Squeeze-and-Excitation (SE) attention mechanism and a Multi-scale Dilated Feature Aggregation (MDFA) module.

Main Methods:

• Utilized a vector hydrophone array to capture particle velocity (vx,vy,vz) and acoustic pressure (p) signals.
• Integrated an SE attention mechanism for improved feature sensitivity.
• Employed the MDFA module to extract multi-scale features and capture cross-scale patterns for weak target enhancement.
• Incorporated an auxiliary classification branch in the discriminator for joint optimization of generation and classification tasks.

Main Results:

• The proposed ACGAN architecture demonstrated improved DOA estimation accuracy in low-SNR scenarios.
• The MDFA module effectively enhanced the representation of weak targets in beamforming maps, mitigating interference bias.
• The auxiliary classification branch facilitated better identification and separation of multiple labeled sources.
• Experimental results confirmed the network's effectiveness across diverse acoustic scenarios.

Conclusions:

• The proposed ACGAN with SE attention and MDFA module offers a robust solution for DOA estimation in challenging low-SNR and multi-source environments.
• The integrated approach effectively addresses limitations of existing methods, providing enhanced accuracy and source separation capabilities.