Proximity Multi-Target Resolution Based on Deep Neural Network

Traditional Direction of Arrival (DOA) estimation methods often suffer from performance degradation under low signal-to-noise ratio (SNR) and limited snapshot conditions. Furthermore, existing deep learning models exhibit limited adaptability to complex electromagnetic environments. To address these challenges, this paper proposes a deep neural network (DNN) based DOA estimation method leveraging multidimensional feature fusion of the covariance matrix. By extracting real parts, imaginary parts, magnitudes, phases, and eigenvalues from the complex-valued covariance matrix, a high-dimensional real-valued feature vector is constructed, effectively preserving the spatial coherence characteristics of the array signal. A funnel-shaped fully connected feedforward network architecture is designed, employing an adaptive learning rate scheduling strategy to learn the nonlinear mapping from spatial features to angular parameters. Simulation results demonstrate that, at an SNR of 10 dB with 100 snapshots, the proposed algorithm achieves an overall root mean square error (RMSE) of 0.35°, representing an approximate 62.5% improvement over the conventional MUSIC algorithm. When the SNR increases to 30 dB, the estimation accuracy reaches 0.35°, with the absolute deviation for each source being less than 0.15°. The proposed method overcomes the dependency of traditional algorithms on high SNR and numerous snapshots, providing an effective solution for DOA estimation in proximate multi-target environments.