Deep Learning Method for Prediction of Planar Radiating Sources from Near-Field Intensity Data

Abstract

A deep learning method, cascaded convolutional neural networks, is investigated as a means for the prediction of frequency-dependent intensity distribution of planar radiating sources from frequency-dependent, near-field intensity data. More specifically, two convolutional neural networks are utilized as follows. The first one uses as input the available near-field amplitude data to predict the amplitude and phase of radiated fields on a plane in closer proximity to the radiating sources. Using the obtained distribution as input, the second one estimates the intensity of the planar radiating sources. The proposed method exhibits very good accuracy in the prediction of the radiating source distribution over the frequency range used for the training of the convolutional neural networks.