Implementation of Bartlett matched-field processing using interpretable complex convolutional neural network.

Area of Science:

• Acoustics and Signal Processing
• Artificial Intelligence
• Oceanography

Background:

• Neural networks show promise in underwater source localization, outperforming conventional matched-field processing (MFP).
• However, current neural network approaches often lack physical interpretability, hindering understanding and trust.
• Matched-field processing (MFP) is a traditional method for underwater acoustic source localization.

Purpose of the Study:

• To design an interpretable complex convolutional neural network (CNN) for underwater source localization.
• To provide clear physical meanings for the network's output and structure.
• To establish the interpretability of the proposed network by relating its components to established physical models.

Main Methods:

• Development of a Bartlett processor-based complex convolutional neural network (BC-CNN).
• Analysis of the relationship between BC-CNN convolution weights and the replica pressures used in MFP.
• Conducting simulation experiments with two distinct labeling strategies.

Main Results:

• The designed BC-CNN offers clear physical interpretations for its operations and outputs.
• A strong correlation was found between the convolution weights of the BC-CNN and the replica pressures of Bartlett MFP.
• Simulation results confirmed the equivalence between the Bartlett MFP and the developed BC-CNN.

Conclusions:

• The proposed BC-CNN enhances the interpretability of neural networks for underwater source localization.
• This interpretable approach bridges the gap between data-driven methods and physics-based models.
• The BC-CNN provides a physically meaningful and accurate alternative for underwater acoustic localization.