A dual-encoder U-net architecture with prior knowledge embedding for acoustic source mapping

Area of Science:

• Acoustics
• Signal Processing
• Machine Learning

Background:

• Deconvolution is standard for acoustic source mapping but computationally intensive.
• Current deep learning methods lack feature diversity and PSF variability handling, reducing localization accuracy.

Purpose of the Study:

• To develop a supervised learning framework for high-resolution acoustic source mapping.
• To improve localization accuracy by addressing limitations of existing methods.

Main Methods:

• A dual-encoder U-net architecture is proposed, processing delay-and-sum and functional beamforming maps.
• A contrastive loss function ensures consistent latent feature learning.
• Frequency and position encoders incorporate prior knowledge of source characteristics and spatial locations.

Main Results:

• The proposed model outperforms existing methods across four metrics on simulation and MIRACLE datasets.
• Demonstrated generalization across varying numbers of sound sources and frequencies.
• Achieved higher resolution mapping of true source strength distribution.

Conclusions:

• The dual-encoder U-net framework offers a significant advancement in acoustic source mapping.
• The method effectively handles PSF variations and improves computational efficiency.
• This approach provides a robust solution for accurate acoustic source localization.