Scholarlink

<p>This paper pioneers a deep learning framework for adaptive beamforming in multi-user active electronically scanned array (AESA) systems, addressing the computational intensity of conventional optimization methods like linearly constrained minimum variance (LCMV) and second-order cone programming (SOCP). We propose a novel approach where convolutional neural networks (CNNs) directly estimate optimal beamforming weights from synthesized beam pattern images. A key innovation is the construction of two distinct types of beam pattern datasets: one generated purely via traditional SOCP and the other using the enhanced SOCP with LCMV constraints, specifically proposed to improve model performance in complex multi-user environments. Experimental results compellingly demonstrate that our CNN model can rapidly generate high-quality beam patterns that achieve deep nulls and narrow main lobes comparable to traditional solvers, while significantly reducing computation time, especially in the case of the proposed hybrid SOCP+LCMV-based CNN. This work establishes the viability of integrating deep learning with model-based beamforming to enable low-latency, interference-aware signal transmission for dynamic multi-user AESA applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep learning-based beamforming for multi-user active electronically scanned array systems

  • 期刊论文

  • 发表日期：

    2025年9月26日

查看全文
  • Nguyen Thi To Uyen,
  • Do Dinh Hung,
  • Le Thi Binh,
  • Pham Thanh Hiep,
  • Nguyen Thu Phuong

摘要

This paper pioneers a deep learning framework for adaptive beamforming in multi-user active electronically scanned array (AESA) systems, addressing the computational intensity of conventional optimization methods like linearly constrained minimum variance (LCMV) and second-order cone programming (SOCP). We propose a novel approach where convolutional neural networks (CNNs) directly estimate optimal beamforming weights from synthesized beam pattern images. A key innovation is the construction of two distinct types of beam pattern datasets: one generated purely via traditional SOCP and the other using the enhanced SOCP with LCMV constraints, specifically proposed to improve model performance in complex multi-user environments. Experimental results compellingly demonstrate that our CNN model can rapidly generate high-quality beam patterns that achieve deep nulls and narrow main lobes comparable to traditional solvers, while significantly reducing computation time, especially in the case of the proposed hybrid SOCP+LCMV-based CNN. This work establishes the viability of integrating deep learning with model-based beamforming to enable low-latency, interference-aware signal transmission for dynamic multi-user AESA applications.