Machine Learning Augmentation of Wideband Spectrum Sensors for Polarimetry

A cascaded neural network architecture is discussed to classify the polarization of a signal that is incident on a dual-polarized log periodic antenna. The antenna system is designed for spectrum sensing but retrofitted to add polarimetry through the use of machine learning. The cross-polarized contamination in the system causes raw estimation to barely exceed 13% classification accuracy, as opposed to 25% accuracy with blind guessing. However, the use of machine learning enables over 90% accuracy when classifying. Three neural networks are deployed. The first separates linear from circular, followed by specialized networks to discriminate between horizontal and vertical as well as left- and right-hand circular polarizations.