Multi-layer residual beamforming for underwater acoustic source bearing estimation.

Deconvolution beamforming maintains robustness against array mismatch and limited snapshots in underwater source localization, yet its resolution and accuracy require improvement. Inspired by deep learning architectures, this work proposes a multi-layer residual beamforming framework that processes array snapshots through 2 K layers of alternating phase compensation and source cancellation to estimate directions of arrival for K sources. Simulations validate the proposed method's robustness to array mismatch and its capacity for accurate source localization under challenging conditions, achieving estimation accuracy comparable to LASSO (i.e., the least absolute shrinkage and selection operator) while outperforming deconvolution beamforming. Experimental validation using the SWellEx-96 dataset confirms the method's practical feasibility.