DOA Estimation of Coherent Sources Using Residual Network-based Subspace Reconstruction

Abstract

Due to the rank deficiency of the covariance matrix (CM) with coherent sources, traditional direction-of-arrival (DOA) estimation methods cannot be used. We propose a DOA estimation method of coherent sources using residual network-based subspace reconstruction. Firstly, This method uses a residual network to learn the CM features of coherent sources in order to reconstruct the noise subspace. The network is optimized using a loss function based on the orthogonality measure between the steering vector and the noise subspace. Subsequently, we construct the spatial spectrum using the reconstructed noise subspace and perform DOA estimation through spectral peak search. The proposed method has model interpretability compared to other deep learning based estimation methods. And it does not require decorrelation preprocessing operations. The experimental results show that the proposed algorithm significantly outperforms the existing methods.