Cross-Domain Underwater Sound Source Localization Based on Cross-Attention and Adversarial Knowledge Distillation

Deep learning algorithms are widely employed in the field of underwater sound source localization. However, in scenarios with limited data, most models exhibit low accuracy in cross-domain localization. In this paper, we propose a cross-domain underwater sound source localization algorithm based on cross-attention and adversarial knowledge distillation. Firstly, we apply cross-attention to fuse features in source domain and target domain, focusing on the correlation between features of different domains to achieve better feature alignment. Secondly, we apply adversarial knowledge distillation for mapping constraints, forcing the model to learn and distinguish domain-specific and domain-invariant representations, which enables the neural network to extract features that are robust to domain changes, and improve the accuracy of cross-domain underwater sound source localization. The experimental results show that compared with the comparison algorithms, the proposed algorithm has better localization accuracy in different waters.