Compressive sensing based downlink channel estimation for mmWave systems using deep learning: Centralized or decentralized

Thanks to the availability of abundant bandwidth, millimeter-wave (mmWave) massive multi-input multi-output (MIMO) systems have numerous applications in wireless communications, in which downlink channel estimation is one of the crucial problems. Through the adaptation of lens antenna array, mmWave channels have a sparsity feature in the beamspace domain, which results in the requirement of a reduced number of RF chains for such a system. Using the sparsity of beamspace channels, the beamspace channel estimation problem can be formulated as a sparse signal recovery problem, which can be solved by the greedy OMP algorithm, where dominant beamspace channel entries are found sequentially one by one, iteratively. To improve the performance of this algorithm, we propose a deep learning (DL)-based OMP algorithm, namely LOMP, in which a previously trained DL model is adopted to choose a set of beamspace dominant entries simultaneously instead of a single dominant entry. As we consider the downlink channel estimation, constructing a DL model well-accepted by all users requires them to send received signals to a centralized entity with high computational power. However, this incurs huge communication burden owing to the signal transmission by all users. To tackle this issue of communication burden as well as preserve the privacy of received signals, we propose a decentralized federated learning (FL)-based OMP algorithm, namely FL-OMP, in which the peer-to-peer (P2P) FL technique is adopted to construct the DL model in the LOMP algorithm, where users share gradients of their local learning models among themselves. Extensive simulations have been conducted to verify the effectiveness of the proposed LOMP and FL-OMP algorithms while comparing with other works in the existing literature.