LAUSR

Signal source number detection is an essential issue for the direction of arrival (DOA) estimation in satellite communication systems. The performances of conventional and deep-learning-based signal source number detection methods will deteriorate when the signal-to-noise ratio is low or coherent signals exist. This paper proposes a DOA detection network (DTN) combined with the root weighted subspace fitting (root-WSF) method to tackle this challenge. The DTN uses the constructed deep neural networks (DNN) to denoise the received signals and captures the nonlinear mapping relationship between the received signals and the number of signal sources. The received signals in the complex-valued domain are directly treated as DTN’s input, and the label of DTN is the one-hot encoding of the source number. It solves the issue that the classifier cannot well-handle the coherent signals and extends the values of discrete features to Euclidean space. Accordingly, the trained DTN can detect the signal source numbers with an average detection accuracy of 96.6%, and the root-WSF algorithm is applied as the rear stage of DOA estimation. Compared with the traditional DOA methods, the proposed DTN incorporated with the root-WSF algorithm features superior robustness, high DOA estimation accuracy, and enhanced resolution.