Deep-learning source localization using multi-frequency magnitude-only data.

Area of Science:

• Ocean acoustics
• Signal processing
• Machine learning

Background:

• Accurate localization of underwater acoustic sources is crucial for various applications.
• Ocean bottom parameters significantly impact acoustic propagation and source localization.
• Existing methods often struggle with environmental uncertainties.

Purpose of the Study:

• To develop a robust deep learning approach for localizing broadband acoustic sources using a single hydrophone.
• To address the challenge of uncertain ocean bottom parameters in source localization.
• To validate the proposed method with both simulated and experimental data.

Main Methods:

• Utilized deep learning, specifically 50-layer residual neural networks, trained on extensive sound field replicas.
• Implemented a two-step training strategy involving coarse and fine grid discretization for source range and depth.
• Employed magnitude-only, multi-frequency data for training and testing.

Main Results:

• The deep learning models successfully handled uncertainties in ocean bottom parameters for source localization.
• The two-step training strategy improved the performance of the deep learning models.
• The approach demonstrated effectiveness in simulated uncertain environments.

Conclusions:

• Deep learning offers a powerful solution for acoustic source localization in complex ocean environments.
• The proposed method provides accurate source localization despite unknown bottom parameters.
• Experimental validation in the China Yellow Sea confirms the practical applicability of the approach.