Deep-Learning-Based Interference-Mitigation Method Using Null-Space Beamforming for Hybrid-Array

doi:10.1109/access.2026.3736005

Deep-Learning-Based Interference-Mitigation Method Using Null-Space Beamforming for Hybrid-Array

2026-09-22

0

OA

AI

DOI：10.1109/access.2026.3736005

Original

Original request for help

Save

Abstract

En 中文

Deep-learning-based beamforming methods for spatial-domain interference rejection have been developed for fully digital arrays and cannot be directly applied to the partially-connected hybrid analog–digital architectures commonly used in large-aperture radar systems. We propose a hybrid array deep neural network-based null-space beamformer (HA-DNN-NSB), which extends the DNN-NSB framework to a partially connected hybrid uniformly spaced linear array. We first derive the mathematical formulation for null-space beamforming (NSB) for the hybrid array structure and prove that both the autocorrelation matrix and optimal weights depend solely on relative steering vectors. Building on this property, we introduce a steering angle normalization technique that aligns the analog beamformer with the desired signal direction, yielding a theoretical SINR improvement over the hybrid array NSB without steering angle normalization and resolving the interference-to-signal ratio (ISR) dependency that limits conventional DNN-NSB when the desired and interfering signals have comparable power levels. The proposed HA-DNN-NSB directly infers the digital beamforming weights from the angle-normalized autocorrelation matrix through a fully connected DNN, without requiring prior interference information. Simulation results showed that, under a fixed hybrid configuration, the proposed method maintained near-optimal SINR across the entire ISR range, whereas conventional DNN-NSB exhibited catastrophic failure at ISR $\approx ~0$ dB. Across varying RF chain configurations, HA-DNN-NSB consistently achieved near-optimal SINR performance regardless of the number of RF chains, whereas the recurrent neural network-based counterpart exhibited a performance plateau and failed to exploit the additional spatial degrees of freedom provided by more RF chains. These results confirmed the effectiveness of the proposed HA-DNN-NSB as a practical deep-learning-based beamforming solution for hybrid-array radar systems.

Keywords:

Deep neural network

hybrid array

interference mitigation

null-space beamforming

steering angle normalization

Deep-Learning-Based Interference-Mitigation Method Using Null-Space Beamforming for Hybrid-Array

Abstract

Journal

Organization

Cited Papers