Distributed Sensor-Assisted DNN-Based Massive-MIMO Channel Reconstruction

Modern massive multiple-input multiple-output (MIMO) systems allow for a great increase in cell throughput by means of spatial multiplexing. However, urban environmental conditions, such as low signal-to-noise ratio (SNR) and non-line-of-sight (NLoS) cell areas, can limit their full potential. To overcome these issues, techniques such as beam sweeping and beam tracking may be used to establish the correct beamforming for signal reception. Nevertheless, beam sweeping can induce overhead in terms of the time required to scan the site area, and beam tracking can only be performed after the initial connection, what may restrict the practical application of these methods. In this paper, we propose to utilize the set of passive sensors equipped with small antenna arrays to receive the uplink pilot signals and apply the state-of-the-art deep neural network transformer architecture for the base station (BS) channel state information (CSI) reconstruction task. We show that it is possible to recover the BS high-dimensional (128 ports) CSI in the negative SNR region with the help of passive synchronized sensors equipped with 4 cross-polarized antenna arrays. The developed method allows for the reconstruction of the 1 transmission layer precoding matrix with an accuracy of approximately 84% for more than 70% of UEs deployed in the NLoS area, while being trained on only 10% of samples.