Large-scale MIMO channel state information compression and reconstruction method based on deep learning attention mechanism

Abstract

The invention discloses a large-scale MIMO channel state information compression and reconstruction method based on a deep learning attention mechanism. The large-scale MIMO channel state informationcompression and reconstruction method comprises the following steps of 1, performing DFT on an input channel matrix; 2, constructing a DS-NLCsiNet model; 3, training the model to obtain parameters ofeach layer of the model; and 4, performing inverse DFT on the output of the model. According to the large-scale MIMO channel state information compression and reconstruction method based on the deep learning attention mechanism, the relevance of long-distance channel information can be efficiently extracted, meanwhile, deeper features are extracted with fewer parameters, and the channel matrix feature extraction efficiency is greatly improved.

Description

technical field [0001] The present invention relates to the technical field of massive MIMO systems, in particular to a massive MIMO channel state information compression and reconstruction method based on a deep learning attention mechanism. Background technique [0002] The massive MIMO (multiple-input multiple-output) system has been recognized as one of the key technologies of the 5G wireless system. This technology forms multiple independent channels in the space domain by configuring a large number of antennas at the base station, thereby greatly increasing the wireless communication. system throughput. However, in a frequency division duplex (FDD) MIMO system, the channel is not reciprocal, and the downlink CSI obtained at the user end needs to be sent to the base station through the feedback link, and the feedback of complete CSI will bring huge resource overhead. In addition, as the number of antennas increases greatly, traditional feedback reduction schemes such ...

AI Technical Summary

Problems solved by technology

[0005] However, traditional models generally only use 3x3 or 5x5 convolution kernels, feature extraction is limited to a certain local neighborhood, and their receptive field is increased through continuous iteration, but this iterative process is very inefficient and relies on long-range The extraction effect is poor, requiring huge computing resources, and the output of each layer is only passed to the next layer, requiring more parameters, and the efficiency of feature extraction is low

When there are many layers, there will be gradient dispersion, which greatly reduces the transmission efficiency and accuracy

Images