Adaptive early-exit lightweight unfolding for sparse bayesian learning-based DOA estimation

Abstract Direction of Arrival (DOA) estimation is a core technology in array signal processing, which is widely applied in radar, communications and sonar systems. Traditional subspace methods suffer from poor performance under low signal-to-noise ratio (SNR) and limited snapshots, while Sparse Bayesian Learning (SBL) methods have high computational complexity. Existing deep unfolding networks typically adopt fixed iteration depths, leading to considerable computational redundancy, while many existing early-exit strategies rely on black-box decision mechanisms with limited interpretability. In this paper, we propose an adaptive early-exit lightweight unfolding network with inter-layer parameter tying for DOA estimation. An interpretable residual-guided early-exit mechanism is designed based on the monotonic stabilization behavior of SBL, and a weighted multi-exit training strategy combined with iteration bounds is adopted to improve estimation performance and computational efficiency. Simulation results under the considered ideal ULA conditions show that the proposed method achieves improved estimation accuracy and reduced inference time compared with the evaluated baselines. These findings indicate computational-efficiency potential in simulated array-processing scenarios.