Source-linked AI summary
BeamFocusNet: Beamforming-based Explicit Spatial Signal Focusing for Robust DoA Estimation under Low-SNR and Single-Snapshot Conditions
Xuyao Deng, Qisheng Xu, Shuo Liu, Yong Dou, Wen Zhang, Kele Xu
TL;DR
Neural filter methods can couple signal focusing and noise suppression, limiting robustness under low SNR and few snapshots. BeamFocusNet uses explicit spatial focusing, and the paper reports strong average performance and robustness across challenging conditions.
Problem
Neural filter methods implicitly combine signal focusing and noise suppression, while low SNR and limited snapshots challenge robust DoA estimation.
Source: §1 · p. 1, §2.2 · p. 2Method
BeamFocusNet learns filters using the sparse spatial structure of BA as an explicit focusing target rather than implicitly coupling focusing with noise suppression.
Source: §1 · p. 1, §2.2 · pp. 2–3Results
BeamFocusNet achieves the best average performance across various conditions and is particularly robust under low SNR and limited snapshots.
Source: §5 · p. 4Takeaways & Limitations
The paper recommends the BA paradigm for angle estimation and the BX paradigm for source number estimation.
Source: §4.3 · p. 4, §5 · p. 4BeamFocusNet is relatively weak at source number prediction, for which applying B to X is more suitable than applying it to A.
Source: §4.3 · p. 4
Abstract
from arXiv · showhide
DoA estimation plays a crucial role in signal processing. Inspired by beamforming, recent works employ neural networks to estimate filters for filtering received signals to achieve DoA estimation. These methods typically generate filters by implicitly focusing signals and suppressing noise. However, this couples the two objectives, making training difficult to balance and resulting in poor robustness, especially under low SNR and limited snapshots. To address this issue, we propose the BeamFocusNet method, which generates filters through explicit signal focusing, introducing a new paradigm for neural network-based filter generation. Extensive experiments under various challenging conditions demonstrate the superiority and robustness of the proposed method in DoA estimation. Code is available at https://github.com/colaudiolab/BeamFocusNet.
1. INTRODUCTION
BeamformNet’s implicit coupling of signal focusing and noise suppression, together with snapshot-dependent inputs, limits robustness under low SNR and few snapshots. BeamFocusNet instead uses explicit spatial focusing, with experiments reporting robustness particularly in those challenging settings.
1. INTRODUCTION
- BeamformNet couples signal focusing and noise suppression, while snapshot-dependent inputs make spatial inference difficult and can degrade resolution when snapshots are limited.The coupled objective forces a trade-off in low-SNR environments.Source: §1 · p. 1
- BeamFocusNet anchors filter learning to explicit spatial focusing, using the sparse structure of BA rather than implicitly targeting BX = S.Its contributions also include covariance-matrix input to address dimensionality collapse under single-snapshot conditions.Source: §1 · p. 1
- Experiments report that BeamFocusNet is robust and superior in DoA estimation, particularly under single-snapshot and low-SNR conditions.The conclusion also reports best average performance across various conditions.Source: §1 · p. 1
2. SPARSE SIGNAL MODEL AND MOTIVATIONS
The section sets out a sparse array-signal model and explains how BeamformNet’s snapshot-dependent inputs and coupled objective motivate BeamFocusNet’s explicit focusing target and covariance input.
2.1. Sparse Signal model
- For K narrowband far-field signals on an M-element array, the section models observations as a sparse signal over direction grids and writes the discrete model as X = AS + N.The setup assumes a half-wavelength ULA and K < M.Source: §2.1 · p. 2
2.2. Motivations
- BeamformNet’s snapshot-dependent input can collapse temporal features at one snapshot, while its filters operate on signals where noise and spatial information are intermingled.Its implicitly coupled focusing and noise-suppression objective also creates a low-SNR trade-off.Source: §2.2 · p. 2
- BeamFocusNet shifts filter learning from BX = S to the sparse spatial structure of BA, whose active-source columns peak at source directions and approach zero elsewhere.This target is independent of noise and is used to focus signals without explicit noise modeling.Source: §2.2 · pp. 2–3
- Replacing X with the sample covariance matrix fixes the network input dimension at M × M, independent of snapshot count, and supplies second-order array-manifold information.The paper presents this as a way to avoid input and layer dimension collapse in the single-snapshot case.Source: §2.2 · p. 3
3. BEAMFOCUSNET
BeamFocusNet uses the steering dictionary and signal covariance to generate an input-adaptive filter, then explicitly evaluates the filter on the dictionary to form a spatial energy spectrum. Thresholding and peak search convert that spectrum into DoA estimates.
3. BEAMFOCUSNET
- The network separates real and imaginary inputs, processes them with bidirectional RNNs, and uses attention to retrieve spatial information for constructing the filter B.Three linear networks form value, key, and query vectors; an RNN and linear network then produce B.Source: §3 · p. 3
- Applying B to steering dictionary A and averaging column-wise energy produces a spectrum that emphasizes source directions.The target is high energy at source directions and low energy elsewhere.Source: §3 · p. 3
- Training normalizes the spectrum to [0,1] with tanh and uses binary labels with Asymmetric Loss to mitigate sparse-signal class imbalance.Source: §3 · p. 3
- The method thresholds tanh(P) at 0.5, then selects the top K peaks when source count is known and fills any missing directions by descending energy.Source: §3 · p. 3
4. NUMERICAL EXPERIMENTS
The experiments compare BeamFocusNet with established baselines across low-SNR, varying-source, limited-snapshot, and coherent-signal conditions. Results favor BeamFocusNet overall, while ablations indicate that explicit BA focusing contributes more than covariance input and that its advantage persists at low embedding capacity.
4.1. Experiments Setup
- Experiments use RMSPE to compare BeamformNet, DA-MUSIC, DeepSSE, DeepSSE, SubspaceNet, lowSNRNet, IQRes-Net, and MUSIC.The model and experimental settings are listed in Table 1.Source: §4.1 · p. 3, Table 1 · p. 3
4.2. Results
- BeamFocusNet achieves the best average performance across conditions, including the best DoA performance at T = 1 and satisfactory performance at SNR ≤ −10 dB.Fig. 2 reports experimental comparisons; the reported scenarios include varying source counts, SNRs, snapshot counts, and coherent versus non-coherent environments.Source: §4.2 · pp. 3–4, Figure 2 · p. 4
4.3. Ablation Study
- Under low SNR and limited snapshots, average performance ranks BeamFocusNet, BA-X-Variant, BX-C-Variant, then BeamformNet, with explicit BA focusing contributing more than covariance input.The ablations compare BA versus BX paths, covariance versus raw-signal input, and a joint variant equivalent to BeamformNet.Source: §4.3 · p. 4
- At matched embedding dimensions, BeamFocusNet outperforms BeamformNet, especially with low embedding dimensions and parameter counts.The authors identify capacity-constrained embedded devices as a more suitable setting for BeamFocusNet than BeamformNet.Source: §4.3 · p. 4
- For source-number estimation at low SNR, BeamFocusNet is relatively weak; BX is more suitable for source counting, whereas BA is recommended for DoA estimation.The authors attribute the distinction to BX's more robust active/inactive separation and BA's sharper angular peaks.Source: §4.3 · p. 4
5. CONCLUSION
BeamFocusNet achieves the best average performance across varied conditions and is more robust under low SNR and limited snapshots. The authors report that BX suits source-number estimation while BA suits angle estimation, and identify their fusion as future work.