The invention discloses a large-scale
MIMO hybrid beam forming method capable of learning a multi-scale network, and belongs to the technical field of
hybrid beam forming. The invention provides a learnable multi-scale
hybrid beam forming network based on a
frequency division duplex system, and an enhanced multi-channel
convolution attention module and a learnable channel enhancement network are jointly designed. The EMCAM introduces multi-scale
feature extraction and a dual-channel attention mechanism into a channel
estimation and feedback link, so that the robustness of the
system under the conditions of low
signal-to-
noise ratio and limited
pilot frequency is remarkably improved, and CSI representation has higher robustness and
compressibility under
noise disturbance. The LCEN projects an antenna domain channel to an angular domain representation with higher sparsity by using a predefined angular
domain mapping matrix, and applies complex rotation to an angular domain coefficient through a phase combination generated by the network; and the obtained equivalent channel has stronger directional energy aggregation and higher feature distinguishability, so that the robustness and prediction performance of the downstream HB under a limited feedback condition are improved.