Single-snapshot DOA estimation using DU-SBLnet
Xiaowei Tu, Xiaolong Su, Panhe Hu, Lida He, Zehao Wang, Cheng Chen · IET conference proceedings. · 2026
Single-snapshot direction-of-arrival (DOA) estimation poses considerable challenges in real-time applications due to the constraint of having only one snapshot. Conventional subspace-based methods often fail to achieve reliable performance under such limitations, while sparse recovery (SR) techniques offer improved resolution by leveraging signal sparsity. However, SR methods are often hampered by high computational cost, sensitivity to parameters, and slow convergence. This paper introduces a Deep Unfolding Sparse Bayesian Learning Network (DU-SBLnet), which integrates model-based inference with data-driven learning. The iterative process of sparse Bayesian learning (SBL) is unrolled into a trainable neural network structure, enabling simultaneous optimisation of the steering dictionary and noise variance. A tailored loss function is constructed to promote sparsity and preserve reconstruction fidelity, allowing effective gradient-based training. Extensive simulation results show that DU-SBLnet achieves superior estimation accuracy and interference robustness, especially in low signal-to-noise ratio (SNR) conditions. Moreover, the network exhibits faster convergence with fewer layers and strong generalisation to varying numbers of sources. These characteristics make DU-SBLnet a promising solution for high-resolution, real-time DOA estimation applications.