Model and Data-Driven Off-Grid Sparse Bayesian Learning Network for DOA Estimation
IEEE Transactions on Aerospace and Electronic Systems · 2026 · 62 · 13526-13541
DOI: 10.1109/taes.2026.3709261 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Aerospace and Electronic Systems, 2026. Verfasst von Cong Liu und 2 weiteren Autor:innen. Die Arbeit nennt 38 Literaturangaben. Gefördert von: National Natural Science Foundation of China, System of Systems and Artificial Intelligence Key Laboratory of SICHUAN Province pioneer fund, Natural Science Foundation of Guangdong Province.
Abstract (Originalsprache)
Traditional on-grid methods are confronted with the grid mismatch (GM) problem when estimating the directions-of-arrival (DOA) of off-grid sources, and their performance deteriorates significantly under challenging conditions among which are low signal-to-noise ratios (SNRs). Compressive sensing-based algorithms, taking advantage of the sparsity exhibited by sources in the spatial domain, they can achieve better estimation in such scenarios. Among them, the traditional off-grid sparse Bayesian learning (OGSBL) method is a notable solution. However, traditional methods typically require multiple iterations to converge, leading to high computational complexity, and their estimation accuracy still needs improvement. Herein, we develop a hybrid model-and-data-driven off-grid sparse Bayesian learning network (OGSBL-Net) for superior and efficient off-grid DOA estimation. Specifically, the iterative OGSBL framework is unfolded into a series of cascaded OGSBL layers. Each layer retains the traditional posterior computation and incorporates three hyperparameter update network modules: a $\alpha$-Net for precisely estimating the sparsity parameters, a $\alpha _{0}$-Net for estimating the noise precision, and a $\beta$ -Net for estimating the off-grid gaps. By synergizing the strengths of deep networks and traditional models, the proposed OGSBL-Net accelerates convergence, captures the underlying angle sparsity more effectively, and generates superior hyperparameter update rules. A well-designed loss function and a training dataset with mixed SNRs are employed to train a high-performing model. Simulation findings indicate that the proposed method exhibits superior performance. Specifically, at SNR=10 dB, the RMSE is reduced by approximately 35% compared to the traditional OGSBL. Furthermore, OGSBL-Net achieves successful angular resolution at $3.4^\circ$ (SNR=0 dB), while maintaining an online inference time of only 0.331 s—representing a 4.3-fold improvement in efficiency over traditional iterative algorithms.
Förderung
- National Natural Science Foundation of China
- System of Systems and Artificial Intelligence Key Laboratory of SICHUAN Province pioneer fund
- Natural Science Foundation of Guangdong Province
- Science and Technology Planning Project of Key Laboratory of Advanced IntelliSense Technology
- Guangdong Science and Technology Department
Zitieren
APA
Liu, C., Wu, Y., & Zhang, L. (2026). Model and Data-Driven Off-Grid Sparse Bayesian Learning Network for DOA Estimation. IEEE Transactions on Aerospace and Electronic Systems, 62, 13526-13541. https://doi.org/10.1109/taes.2026.3709261
Vancouver
Liu C, Wu Y, Zhang L. Model and Data-Driven Off-Grid Sparse Bayesian Learning Network for DOA Estimation. IEEE Transactions on Aerospace and Electronic Systems. 2026;62:13526-13541. doi:10.1109/taes.2026.3709261
BibTeX
@article{Liu2026,
author = {Liu, Cong and Wu, Yifeng and Zhang, Lei},
title = {Model and Data-Driven Off-Grid Sparse Bayesian Learning Network for DOA Estimation},
journal = {IEEE Transactions on Aerospace and Electronic Systems},
year = {2026},
volume = {62},
pages = {13526--13541},
doi = {10.1109/taes.2026.3709261}
}
EUVIMED