A multi-task learning convolutional neural network for source localization in deep ocean.

Area of Science:

• Ocean Acoustics
• Signal Processing
• Machine Learning

Background:

• Accurate underwater acoustic source localization is crucial for naval and oceanic research.
• Traditional methods like matched field processing (MFP) face challenges with environmental uncertainties and array imperfections.

Purpose of the Study:

• To develop and validate a robust multi-task learning (MTL) method using a convolutional neural network (CNN) for estimating acoustic source range and depth in deep ocean environments.
• To assess the impact of environmental parameter uncertainty, particularly array tilt, on localization performance.

Main Methods:

• A CNN model was designed to process normalized sample covariance matrices from broadband acoustic data received by a vertical line array.
• Training and validation data were generated using an acoustic propagation model accounting for environmental parameter variations.
• Sensitivity analysis was performed to identify key environmental factors affecting localization accuracy.

Main Results:

• The proposed CNN with MTL demonstrated superior performance and robustness to array tilt compared to conventional MFP.
• Environmental uncertainty, especially array tilt, significantly impacts localization accuracy.
• The method successfully estimated ranges and depths in scenarios where MFP failed, validated by simulations and real-world data from the South China Sea.

Conclusions:

• The CNN-based MTL approach offers a more reliable and robust solution for underwater acoustic source localization in challenging deep-ocean conditions.
• The method's ability to handle environmental uncertainties, particularly array tilt, makes it a valuable advancement over existing techniques.