Applying deep learning for underwater broadband-source detection using a spherical array.

Area of Science:

• Underwater Acoustics
• Signal Processing
• Machine Learning

Background:

• Passive detection of underwater broadband sources is crucial for sonar and surveillance.
• Accurate source-detection and direction-of-arrival (DOA) estimation remain challenging, especially for broadband signals.

Purpose of the Study:

• To develop a novel deep neural network (DNN)-based method for passive detection and DOA estimation of underwater broadband sources using a spherical array.
• To enhance the robustness and accuracy of underwater acoustic detection systems.

Main Methods:

• Utilized a spherical array to capture underwater acoustic signals.
• Applied Spherical Fourier Transform to convert element pressure signals into spherical Fourier coefficients for DNN input.
• Designed DNN labels using a Gaussian distribution with a spatial-spectrum-like form.
• Developed a physical model simulating underwater acoustic propagation and spherical array signals for DNN training.
• Incorporated white noise into training data to improve detection and reduce false estimations.

Main Results:

• The DNN method demonstrated enhanced detection capability and effective suppression of false estimations.
• Performance was evaluated using detection rate at a constant false alarm rate.
• The model achieved broadband detection capability across varying signal-to-noise ratios.
• The DNN showed potential for multisource detection even when trained on a single source.
• Simulation and experimental results validated the proposed method's effectiveness.

Conclusions:

• The developed DNN-based method offers a robust solution for passive detection and DOA estimation of underwater broadband sources.
• The approach does not require prior knowledge of source spectral features, increasing its applicability.
• The method shows promise for real-world applications in underwater acoustics and surveillance.