Lightweight Deep Neural Network-Assisted Real-Time Pattern Synthesis for Large-Scale Spherical Array Antennas Based on a High-Efficiency Data Acquisition Method

Authors:
Publisher website
Google Scholar
Abstract
Pattern synthesis for large-scale spherical array antennas (SAA) has traditionally been computationally intensive. To enable flexible beam pattern control in SAAs, a real-time pattern synthesis method integrating the fast spherical Fourier transform (FSFT) with a lightweight deep neural network (LDNN) is proposed in this paper. The framework comprises three phases: sample database generation, neural network training, and online prediction. To efficiently collect training data, the direct and inverse Fourier transform relationship between the excitation distribution and the array pattern of a SAA is established using spherical harmonics, thereby extending the traditional fast Fourier transform (FFT) pattern synthesis algorithm, which is typically applied only to planar arrays, to SAAs. Numerical simulation results demonstrate that the proposed method based on the FSFT achieves significantly lower computational complexity compared with existing methods. In the LDNN training phase, an end-to-end model is introduced to approximate the input-output mapping of the FSFT-based method. An iterative training-pruning mechanism is employed to minimize the redundancy of the neural network. Simulation results show that the obtained LDNN achieves a computational complexity reduction of over 50% in the online prediction process. For a SAA with 1196 elements, the focused beam patterns are synthesized in just 35 μs.
Funding Information
  • National Natural Science Foundation of China (62431006)
  • Plan for Strengthening Foundation
  • China National Postdoctoral Program for Innovative Talents (BX20250348)
  • National Natural Science Foundation of China (2025M780496)

This publication has 22 references indexed in Scilit: