Frequency-Domain Second-Order Decorrelation with Compact Time-Domain Regularization for Convolutive Underwater...

Area of Science:

• Underwater Acoustics
• Signal Processing
• Machine Learning

Background:

• Classical algorithms struggle with long-delay multipath and low signal-to-noise ratios in underwater acoustic mixtures.
• Spectral overlap from mechanical and biological sources further complicates signal separation.
• Existing methods like FastICA and JADE are limited by degraded higher-order statistical cues.

Purpose of the Study:

• To adapt frequency-domain second-order decorrelation (FSD) for convolutive underwater acoustic mixtures.
• To introduce a simulated benchmark (ShipsEarBSS) for evaluating source separation algorithms.
• To optimize FSD configurations for improved underwater acoustic signal recovery.

Main Methods:

• Adapted FSD using multi-block joint diagonalization of cross-power spectral density matrices in the short-time Fourier transform domain.
• Implemented compact time-domain regularization for demixing filters.
• Utilized the ShipsEarBSS benchmark with simulated multichannel mixtures and known reference sources.

Main Results:

• An optimized compact FSD configuration demonstrated improved performance compared to baseline methods (frozen FSD, PCA-SVD, AuxIVA).
• Compact time-domain regularization positively impacted the FSD operating point under the ShipsEarBSS protocol.
• Filter length, CPSD block count, and output ordering were found to be empirical configuration choices.

Conclusions:

• Compact time-domain regularization is beneficial for FSD in underwater acoustic environments.
• The effectiveness of specific FSD configuration parameters (filter support, block count, output ordering) is context-dependent.
• The ShipsEarBSS benchmark provides a valuable tool for evaluating underwater acoustic source separation techniques.