Reproduce “Adaptive Control Attention Network for Underwater Acoustic Localization and Domain Adaptation” on OpenResearch

About this paper

Quoc Thinh Vo, Joe Woods, Priontu Chowdhury, David K. Han

Researchers from Drexel University developed the Adaptive Control Attention Network (ACA-Net), a multi-branch deep learning model, for robust underwater acoustic range estimation using log-mel spectrogram and GCC-PHAT features. This model achieves an MAE of 0.045 km on the SWellEx-96 S5 VLA dataset, demonstrating superior localization accuracy and strong domain adaptation capabilities with minimal fine-tuning in new acoustic environments.

Key insights

• Combining complementary acoustic features (log-mel spectrogram and GCC-PHAT) within a multi-branch architecture with shared weights significantly improves localization accuracy by leveraging diverse information sources.
• The integration of Conformer blocks, which use a multi-scale attention mechanism, is crucial for capturing both local and global temporal dependencies in complex underwater acoustic signals, leading to enhanced performance.
• An Adaptive Gain Control (AGC) layer effectively enhances model resilience to variations in signal properties, simplifying deployment and improving real-world adaptability by ensuring consistent energy levels in input features.
• Pre-training on a source domain and fine-tuning with as little as 15% of target-domain data enables strong domain adaptation, addressing data scarcity and improving generalization in new, challenging underwater acoustic environments.

Summary by alphaXiv.