Model-Based Deep Learning for Underdetermined DOA Estimation Exploiting High-Order Difference Co-Arrays

A model-based (MB) deep learning (DL) method is proposed to address the challenge of underdetermined direction of arrival (DOA) estimation exploiting high-order difference co-arrays. Specifically, a deep neural network (DNN) is employed to perform decorrelation and reconstruct the covariance matrix, followed by subspace-based methods incorporated into backpropagation for DOA estimation. Unlike existing DL-based methods, the proposed method utilizes high-order cumulants as the network input, leading to improved performance in multitarget scenarios. Moreover, the proposed method demonstrates strong generalization to unseen number of sources, which remains challenging for existing DL-based methods. Simulation results demonstrate that the proposed method, termed High-Order SubspaceNet, consistently outperforms existing methods under various conditions, particularly in scenarios with limited number of snapshots, low signal-to-noise ratios (SNRs), and multiple targets.