Deep Learning-Based Near-Field Source Localization Without a Priori Knowledge of the Number of Sources

Abstract

In this paper, we propose a high resolution grid-based deep learning source localization that precisely estimates the locations of near-field sources without a priori knowledge of the number of sources. The proposed method consists of a principal component analysis network (PCAnet) and a spatial spectrum network (Sp2net). The proposed PCAnet calculates the noise spaces of the received signals by convolutional layers without a priori knowledge or the estimation of the number of sources and has the lower complexity than eigenvalue decomposition (EVD). The proposed Sp2net calculates the spatial spectrum with a linear layer from the output of the PCAnet and classifies dense location grids with a convolutional neural network (CNN). From the spatial spectrum, this paper also proposes an activation function to enlarge the values at the grid points where the near-field sources exist, which are differentiable for all input values. Then, the direction of arrivals (DOAs) and the ranges of the near-field sources are estimated with high resolution. Computer simulations demonstrated that the proposed method had better DOA and range estimation performances than those of the conventional methods.