Wideband Channel Estimation With a Generative Adversarial Network


Balevi E., Andrews J. G.

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, vol.20, no.5, pp.3049-3060, 2021 (SCI-Expanded) identifier identifier

  • Publication Type: Article / Article
  • Volume: 20 Issue: 5
  • Publication Date: 2021
  • Doi Number: 10.1109/twc.2020.3047100
  • Journal Name: IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, PASCAL, Aerospace Database, Communication Abstracts, Compendex, INSPEC, Metadex, Civil Engineering Abstracts
  • Page Numbers: pp.3049-3060
  • Keywords: Channel estimation, Radio frequency, OFDM, Signal to noise ratio, Frequency estimation, Gallium nitride, Transceivers, Frequency selective channel estimation, GAN, MIMO, terahertz and millimeter wave communication, MASSIVE MIMO, SYSTEMS
  • Middle East Technical University Affiliated: No

Abstract

Communication at high carrier frequencies such as millimeter wave (mmWave) and terahertz (THz) requires channel estimation for very large bandwidths at low SNR. Hence, allocating an orthogonal pilot tone for each coherence bandwidth leads to excessive number of pilots. We leverage generative adversarial networks (GANs) to accurately estimate frequency selective channels with few pilots at low SNR. The proposed estimator first learns to produce channel samples from the true but unknown channel distribution via training the generative network, and then uses this trained network as a prior to estimate the current channel by optimizing the network's input vector in light of the current received signal. Our results show that at an SNR of -5 dB, even if a transceiver with one-bit phase shifters is employed, our design achieves the same channel estimation error as an LS estimator with SNR = 20 dB or the LMMSE estimator at 2.5 dB, both with fully digital architectures. Additionally, the GAN-based estimator reduces the required number of pilots by about 70% without significantly increasing the estimation error and required SNR. We also show that the generative network does not appear to require retraining even if the number of clusters and rays change considerably.