Diagnosis of Calibration State for Massive Antenna Array via Deep Learning

Abstract

Time division duplex mode provides an attracting advantage called channel reciprocity to avoid excessive downlink (DL) pilot overhead, where the DL channel state information can be obtained from the uplink pilot. Unfortunately, such an ideal reciprocity would be broken easily because of hardware glitches from the transmit or receive radio frequency branches. In this letter, we develop a novel learning-based calibration state diagnosis network, where the linear pilot matrix and the non-linear reconstruction mapping decoder are learned jointly through training measure samples. With the assist of proposed network, a learning-based diagnosis procedure is designed to pick out the antennas that need to be calibrated. Experimental results demonstrate that both the normalized mean square error and accuracy are improved compared with the existing compressive sensing-based method.