Dual-Stream Reconstruction Network-Aided ESPRIT Algorithm for DoA Estimation in Coherent Scenarios

Conventionally, direction of arrival (DoA) estimation of coherent signals relies on pre-processing approaches to eliminate the rank loss of the signal covariance matrix (SCM). However, almost all existing methods utilize the empirical covariance matrix (ECM) as a practical approximation of the SCM, and can not make full use of the information in received signals. In this paper, we propose a deep learning-aided estimation of signal parameters via the rotational invariance technique (ESPRIT) algorithm for DoA estimation in coherent scenarios. Specifically, a dual-stream reconstruction network (DsRNet) is proposed for reconstructing the full-rank SCM directly from received signals, which overcomes the fundamental dependence on the ECM and exploits all available information in received signals. Next, the logarithmic eigenvalue-aided deep residual network (LogDRNet) is proposed for the source number detection, which improves the efficiency of network training and the detection performance, especially with low signal-to-noise ratio (SNR) and limited number of snapshots. Subsequently, we determine the signal subspace using the source number estimated by the LogDRNet and estimate the DoA referring to the ESPRIT algorithm. The proposed deep learning-based framework replaces some crucial aspects of traditional algorithms with customized neural networks, which mitigates the dependence on ideal model assumptions, does not require taking the source number as a prior and improves the estimation performance. Simulation results show that the proposed method not only performs better than existing methods in source number detection and DoA estimation, but also shows superior robustness to various kinds of array imperfections that are widespread in practical systems.