Convolutional Neural Network for Array Size Selection of a Dual-band Reconfigurable Array

Abstract

A convolutional neural network (CNN) is designed and trained to partially control a dual-band, large uniform rectangular array of reconfigurable radiating elements. The CNN selects the number of active elements and switch states needed to achieve a desired beam shape. Both pattern multiplication and finite element method (FEM) are used to simulate the radiation patterns of a PIN-diode square-spiral antenna array. After training on radiation pattern images of arrays calibrated for both phase and gain imbalance and mutual coupling, the CNN achieves 97 percent validation accuracy. Then, using the resulting size and switch states, the patterns are simulated with and without mutual coupling using the pattern multiplication model and FEM, respectively. The mean beam steering and 3-dB beamwidth errors without mutual coupling are less than 5.5 degrees and up to 12.3 degrees with mutual coupling.