Observation of stable components of the sound field in Lake Kinneret using the autoproduct transform

TL;DR

This paper demonstrates how the autoproduct transform and quantum-inspired methods can reconstruct and analyze underwater sound fields in Lake Kinneret, enabling improved sound source localization with neural networks trained on synthetic data.

Contribution

It introduces a novel application of the autoproduct transform and coherent state method for reconstructing underwater sound fields and uses this for neural network-based source localization.

Findings

Reconstructed sound fields match theoretical models in a multipath-free environment.

The intensity distribution is stable against environmental variations.

Neural networks trained on synthetic data effectively localize sound sources.

Abstract

An analysis was conducted of broadband sound pulses received by a vertical array in Lake Kinneret (Israel). For most frequencies within the pulse frequency bands, the array is sparse. The application of the autoproduct transform made it possible to approximately reconstruct the signals that would be received after the emission of pulses at low frequencies for which the array is dense. Using the coherent state method developed in quantum theory, a transition has been made from representing the reconstructed field as a function of depth and time to its distribution in the 'depth-angle-time' phase space. Due to the absence of multipath, the intensity distribution in this space should be weakly sensitive to variations in environmental parameters. In accordance with this expectation, the distribution found is close to the result of its calculation using an idealized (range-independent) waveguide model. It has been shown that this intensity distribution can be used as input data for a neural network when solving the problem of sound source localization in an underwater waveguide. In the examples considered, the neural network is trained on synthetic data, i.e., data obtained from theoretical calculations.