Machine Learning Enhanced CSI Acquisition and Training Strategy for FDD Massive MIMO

Massive Multiple Input Multiple Output (MIMO) is able to boost the system throughput. A key challenge is a large overhead of Channel State Information (CSI) feedback with the increased number of antenna ports in Frequency Division Duplexing (FDD) massive MIMO systems. Conventional methods apply either compressed sensing or beamformed reference signal to reduce the CSI overhead. However, there are still other probems such as additional overhead, user's implementation complexity, or performance limitation. We propose a machine learning enhanced CSI acquisition and training solution for FDD massive MIMO. It can efficiently recover the CSI with more ports than those of the CSI feedback. Furthermore, a practical training strategy is developed, which shows the feasibility of using uplink dataset to train the neural network for the downlink use in FDD.