Dipole Source Reconstruction By Convolutional Neural Networks

Abstract

Equivalent dipole moments are widely used for noise source reconstruction in radio frequency interference (RFI) study. The equivalent dipole sources are usually extracted from measured near-field pattern. This paper introduces a machine learning based method to extract the dipole moments. A convolutional neural network is trained to perform a multi-label classification to determine the type of dipole moments. The locations of the dipole moments are obtained from the global averaging pooling layer. Then the magnitude and phase of the dipoles can be calculated from least square (LSQ) optimization. The proposed method is tested on simulated near-field patterns. The comparison between reconstructed field pattern and original field pattern is given.