Abstract:To address the issues of low estimation accuracy and weak robustness in array imperfections and low signal-to-noise ratio (SNR) conditions in existing direction of arrival (DOA) estimation algorithms, this paper proposes a DOA estimation algorithm based on efficient channel attention residual network (ECAR-Net). Firstly, a fully connected layer is utilized to map the signal into a high-dimensional space and decouple the nonlinear components within the signal. Secondly, the feature vectors are reshaped to facilitate subsequent convolutional operations. Then, the features are further extracted and refined through efficient channel attention (ECA) residual block. The feature vectors are flattened and fully connected layer outputs the signal with enhanced features. Finally, perform matrix operations on the output signal to generate spatial spectrum, and obtain the DOA estimation value through peak search. The experiments simulated array environments with multiple array imperfections and compared the performance of ECAR-Net, multiple signal classification (MUSIC) method, orthogonal matching pursuit (OMP) method and super-resolution direction of arrival network (SDOA-Net). Simulation results demonstrate that across the tested SNR range, ECAR-Net achieves higher mean DOA estimation accuracy than SDOA-Net, MUSIC and OMP. At -10 dB SNR, the mean DOA estimation accuracy of ECAR-Net has improved by 9.06%, 20.49% and 12.76% compared to SDOA-Net, MUSIC and OMP, respectively. Cumulative distribution function (CDF) evaluation further demonstrates ECAR-Net’s stronger robustness and better estimation error distribution characteristics than SDOA-Net, MUSIC and OMP.