Fusing line spectrum enhancement, physics-informed features, and data-driven models: Underwater weak acoustic source...

Area of Science:

• Ocean Acoustics
• Signal Processing
• Machine Learning

Background:

• Underwater acoustic source ranging is challenging due to signal attenuation, multipath interference, and noise.
• Accurate localization of weak acoustic sources is critical for marine applications.

Purpose of the Study:

• To develop an integrated framework for accurate underwater weak acoustic source ranging.
• To evaluate the performance of a Transformer-based model against conventional and other machine learning methods.

Main Methods:

• Proposed an integrated framework combining line spectrum enhancement, conventional beamforming, and data-driven learning.
• Utilized a 24-element array for data collection during a 2025 sea trial with GPS-synchronized ground truth.
• Generated 30,000 training samples using Bellhop simulation with seven key parameters.

Main Results:

• The Transformer-based framework achieved the best performance with Mean Absolute Error (MAE) of 477.09 m, Root Mean Square Error (RMSE) of 809.93 m, and Mean Absolute Percentage Error (MAPE) of 8.90%.
• The Transformer model reduced RMSE, MAE, and MAPE by 9.60%, 10.77%, and 8.06% respectively, compared to a Convolutional Neural Network (CNN).
• Experimental results and arrival angle perturbation tests confirmed the framework's effectiveness.

Conclusions:

• The integrated framework demonstrates superior performance in weak acoustic source ranging in challenging marine environments.
• The data-driven learning approach, particularly using Transformers, offers significant improvements over conventional methods and CNNs.