Robust DOA Estimation for Coprime Array via a Complex-valued Convolutional Attention Network

Traditional direction-of-arrival (DOA) estimation algorithms heavily rely on underlying model assumptions and suffer from performance degradation due to array manifold mismatches when applied to coprime array signals. To address these challenges, this paper proposes a robust DOA estimation approach for coprime array based on a Complex-valued Convolutional Attention Network (CV-CAN). The proposed network employs multiple layers of complex-valued convolutional operations to deeply extract spatial features, while an attention module is introduced to dynamically enhance key informative features. Finally, complex-valued fully connected layers map the extracted features to a discretized angular domain. Simulation results demonstrate that, compared to MUSIC, Root-MUSIC, and DNN methods, the proposed CV-CAN exhibits superior estimation performance, particularly under low signal-to-noise ratio (SNR) and limited snapshot conditions.