Multi-Source 2D DOA Estimation of Uniform Circular Array Based on Deep Learning

To deepen the integration of two-dimensional Direction of Arrival (DOA) estimation problems with deep learning techniques, this paper considers the limitations of traditional 2D-DOA estimation schemes, such as low precision and high computational time overhead in scenarios with low Signal-to-Noise Ratio (SNR) and small angular intervals. Therefore, we propose the use of Multilayer Perceptron(MLP) and one-dimensional Convolutional Neural Network(CNN) network structures to accurately estimate the number of sources and to acquire spatial parameters from the data received by a uniform circular array. Additionally, we introduce a novel low-complexity angle matching method for automatic matching output. Simulation results indicate that the proposed source number estimation network, (Square Eigenvalue)SE-MLP, outperforms source number estimation methods based on statistical models under conditions of low SNR (below 0 dB) and small number of snapshots. The proposed UCA-1D-CNN prediction network demonstrates superior performance to subspace-based DOA estimation algorithms across a wide range of SNR values. Lastly, the proposed angle matching method based on reconstructed signal covariance matrix matching maintains a high accuracy rate of 94.6% even at an SNR as low as −10 dB and has been proven to surpass the currently mainstream neural network-based angle matching schemes in terms of performance.