A Multitask Learning Framework for Pilot Decontamination in 5G Massive MIMO

Abstract

Reference signals enable the acquisition of channel state information (CSI) for purposes such as channel estimation, beam selection, precoding, and symbol detection in 5G massive multiple-input multipleoutput (MAMIMO) systems. Eventually, as more and more users and cells are added, orthogonal reference signals become few which leads to pilot contamination. Pilot contamination limits the performance and occurs when non-orthogonal reference signals occupy time-frequency resources that are alike. Learning-based techniques have been proposed to alleviate it. However, each can only learn to perform a single task namely pilot assignment, power allocation, pilot design, or de-noising for pilot decontamination. In addition, each learner can only be successful if postulated conditions are met. This study proposes a multitask learning framework that can be trained to dynamically select from the multitude of deep learning models which have been suggested for pilot decontamination. Under all signal-to-noise (SNR) ratios, experiments conducted on the deep residual learning aided channel estimator using the multitask learning framework showed minimum channel estimation errors compared to single-task learning.