An Underwater Wideband Sound Source Localization Method Based on Light Neural Network Structure

For specific wideband sound source localization tasks performed in underwater environments, high precision, low computational complexity, and high robustness are required to meet the demand for accurate location in real-time processes. Traditional sound source localization methods that rely on models and signal processing techniques perform poorly in complex but common scenarios where noise, reverberation, and array deviations exist. Sound source localization systems based on deep neural networks (DNNs) are proposed recently and show superiority over conventional localization methods. However, lighter neural network structures for common wideband signals localization have been underexplored, which are crucial for improving the operational efficiency and practicability in underwater localization systems. In this article, we propose a light neural network structure to locate the typical underwater wideband sound sources in communication and localization systems. The network mainly consists of convolutional and residual blocks. We train the network with simulated single tones and finetune it with multiple sets of single tones received in the experiment. As for the training feature, we choose the phase component of the covariance matrix and reshape the upper triangle into one column to reduce the feature dimension. The test signals include chirps, single-carrier modulation quadrature phase shift keying (SC-QPSK), multicarrier modulations such as multitone and orthogonal frequency division multiplexing (OFDM). Both the simulation and experiment results have verified the high accuracy and robustness of our proposed method. Compared with the state-of-the-arts, our method obtains superior performance, especially in scenarios with low signal-to-noise ratio (SNR).