Model-Driven Federated Learning for Channel Estimation in Millimeter-Wave Massive MIMO Systems

This paper investigates the model-driven federated learning (FL) for channel estimation in multi-user millimeter-wave (mmWave) massive multiple-input multiple-output (MIMO) systems. Firstly, we formulate it as a sparse signal recovery problem by exploiting the beamspace domain sparsity of the mmWave channels. Then, we propose an FL-based learned approximate message passing (LAMP) channel estimation scheme, namely FL-LAMP, where the LAMP network is trained by an FL framework. Specifically, the base station (BS) and users jointly train the LAMP network, where the users update the local LAMP network parameters by local datasets consisting of measurement signals and beamspace channels, and the BS calculates the global LAMP network parameters by aggregating the local network parameters from all the users. The beamspace channel can thus be obtained in real time from the measurement signal based on the parameters of the trained LAMP network. Simulation results demonstrate that the proposed FL-LAMP scheme can achieve better channel estimation accuracy than the existing orthogonal matching pursuit (OMP) and approximate message passing (AMP) schemes, and provides satisfactory prediction capability for multipath channels.