Deep-learning source localization using autocorrelation functions from a single hydrophone in deep ocean.

Area of Science:

• Ocean acoustics
• Signal processing
• Machine learning

Background:

• Acoustic source localization in the deep ocean often relies on multi-path time delays.
• Conventional methods face challenges in manually assigning extracted delays to specific propagation paths.
• Ocean bottom absorption can cause some multi-path arrivals to be undetectable.

Purpose of the Study:

• To develop an automated method for acoustic source localization that bypasses the need for artificial path determination.
• To enhance the robustness of localization techniques against missing multi-path arrival data.

Main Methods:

• A convolutional neural network (CNN) was employed for source localization.
• Autocorrelation functions of acoustic signals were used directly as input features for the CNN.
• A data augmentation technique, based on a ray propagation model, was developed to address missing arrivals.

Main Results:

• The proposed CNN method successfully localized acoustic sources without manual path assignment.
• The data augmentation strategy improved localization performance when multi-path arrivals were obscured.
• Validation was performed using both simulated and real-world deep ocean acoustic data.

Conclusions:

• The CNN-based approach offers an effective and automated solution for deep ocean acoustic source localization.
• The integration of data augmentation enhances the method's reliability in challenging acoustic environments.
• This technique represents a significant advancement in underwater acoustic signal processing and localization.