Deep transfer learning for underwater direction of arrival using one vector sensor.

Area of Science:

• Acoustics
• Signal Processing
• Machine Learning

Background:

• Direction of Arrival (DOA) estimation is crucial for source localization in various applications.
• Traditional methods often struggle with complex acoustic environments and limited training data.
• Convolutional Neural Networks (CNNs) show promise but require extensive domain-specific data.

Purpose of the Study:

• To propose a Deep Transfer Learning (DTL) method for enhanced DOA estimation using a single-vector sensor.
• To adapt a CNN trained on synthetic data to perform accurately with real-world, at-sea measurements.
• To improve the reliability of DOA estimates, particularly in the presence of interfering acoustic sources.

Main Methods:

• A CNN was trained using synthetic data representing acoustical pressure and particle velocity cross-spectra.
• Transfer learning was employed by copying initial convolutional layers from the pre-trained CNN to a new target CNN.
• The target CNN's remaining layers were randomly initialized and trained on available at-sea data for domain adaptation.

Main Results:

• The DTL method demonstrated more reliable DOA estimates compared to a conventional CNN.
• Performance improvements were particularly notable when dealing with moving surface ships and interfering sources.
• The adaptation of a source domain (synthetic data) to a target domain (at-sea data) proved effective.

Conclusions:

• DTL offers a robust approach for DOA estimation with single-vector sensors, overcoming limitations of traditional methods.
• The proposed DTL strategy effectively leverages synthetic data for training and real-world data for adaptation.
• This technique significantly enhances DOA estimation accuracy and reliability in challenging acoustic scenarios.