Passive ranging for underwater broadband source using matched field based on physics-informed neural network

Broadband matched field processing enhances the limited resolution of narrowband matched field processing by integrating information from multiple frequency of broadband signals, but its effectiveness is still affected by environmental mismatch. To address this issue, this paper proposes an underwater broadband acoustic source matched field passive ranging method based on physics-informed neural network. The method embeds the Helmholtz equation and the pressure-release surface boundary conditions as physical constraints into the loss function to construct a physics-informed neural network, and introduces a normalized correction loss term. By using sparse spatial coordinates and frequency inputs, combined with a small amount of measured data to correct the known point sound pressure amplitude and interpolate the unknown areas, a broadband full-field corrected replica with reduced amplitude difference from the measured field is generated, thereby alleviating environmental mismatch and improving ranging accuracy. The validation using the SWellEx-96 experiment shows that the proposed algorithm outperforms narrowband/broadband matched field processing, convolutional neural networks combined with measured data, and model-based multitask learning ranging methods in scenarios with scarce training samples, untrained environments, and reduced array elements/aperture.