Super-Resolution Estimation of Signal Direction Based on Unsupervised Learning

Abstract

Target direction estimation is one of the main research directions in the field of array signal processing. In this paper, unsupervised learning method is adopted to study the multi-target direction estimation ability of Deep Neural Network (DNN), under low SNR without using a large amount of training data. The method in this paper is designed to estimate target direction by nonlinear least square spectrum estimation. It is found that when the SNR is -10dB, the precision rate of the DNN can still reach about 92%. Compared with the Conventional Beam Forming (CBF) method, the DNN has a significantly narrow main lobe, and the parameters obtained have the characteristics of sparse. In addition, when we explore whether adjacent targets have an impact on the results, we find that the method in this paper also has the ability of super-resolution. The above findings provide new ideas and experience for the further study of the relationship between array signals and deep learning. As well as for the design and improvement of relevant algorithms on this basis.