InceptCodeNet Based CSI Feedback in Massive MIMO Systems

Abstract

Utilization of spatial multiplexing and diversity gain in massive multiple-input multiple-output (MIMO), requires availability of the downlink channel state information (CSI) at the base station. Frequency division duplex (FDD) systems use different uplink and downlink channels and it limits the use of reciprocity. Downlink precoding computations requires channel responses of the downlink to be estimated and the base station (BS) is fed back with those estimated channel responses. The matrix carrying channel state information is usually large due to massive number of antennas, resulting in considerable feedback overhead. Most of the conventional algorithms use compressed sensing which depends on the channel sparsity level. Recent approaches use deep learning (DL), which compresses the CSI into a codeword with low dimensionality to recover the original channel matrix at the base station. This paper proposes a novel deep learning convolutional network called InceptCodeNet, which is the combination of the concept of inception network and autoencoder, hence the name InceptCodeNet. The network compresses the channel response matrix at the user equipment (UE) side. This is reliably recovered at the base station. InceptCodeNet (ICN) shows superior performance compared to existing techniques in terms of cosine similarity and normalized mean square error (NMSE) metrics. The proposed method provides an improvement of 4.47 dB in NMSE for recovery of channel matrix for indoor scenario and an improvement of 1.89 dB is observed for outdoor scenario compared to CsiNet.