CSI Compression And Restoration In IRS-assisted MIMO Systems Based On Deep Learning

Intelligent Reflecting Surface(IRS)is regarded as one of the key technologies for next generation communications due to its low power consumption and improved communication quality.In massive multiple input multiple output(MIMO)communication systems,the use of deep learning methods to compress and reconstruct channel state information(CSI)is currently a hot academic research topic.However,most existing research works have been conducted in the point-to-point MIMO systems,and CSI compression and reconstruction for IRS-assisted MIMO communication systems has not been sufficiently studied.In this thesis,the CSI of all channels in IRS-assisted Frequency Division Duplex(FDD)MIMO orthogonal frequency division multiplexing(OFDM)communication systems is compressed and reconstructed using deep learning techniques.Besides,in order to investigate the impact of quantization error cancellation on deep learning networks,this thesis further investigates the quantization error cancellation in a point-to-point FDD MIMO-OFDM system.The main contents and innovations of this thesis are as follows.(1)To address the problem that current deep learning networks suffer from low reconstruction accuracy when compressing and reconstructing the CSI matrix of all channels under IRS-assisted FDD MIMO-OFDM systems,this thesis proposes a new deep learning network called Inception-Attention-Residual-Net(IARNet),which is a multi-resolution convolutional residual network based on the attention mechanism.The IARNet adopts the modules of multi-convolutional feature fusion,hybrid attention model and residual network on top of the convolutional neural network.Simulation results show that as compared to two existing CSI compression networks,the IARNet can significantly reduce the normalized mean square error between the original CSI and the reconstructed CSI under the model training scheme based on the warm-up method in terms of model parameters.In addition,the warm-up method-based model training scheme used in this system is also better than the other three schemes.(2)To address the CSI quantization error problem generated by the system during quantization and inverse quantization in a point-to-point FDD MIMO-OFDM system,this thesis proposes a quantization error refine module based on a multilayer perceptron to help the system eliminate quantization error and further improve the CSI reconstruction accuracy.In addition,to improve the reconstruction quality,an encoding-decoding network based on the Selective Kernel Networks(SKNet)is also proposed.Simulation results show that the proposed quantization error recovery module can help reduce the quantization error and improve the reconstruction quality of the system under the conditions of quantization bits of2 and 4.Simulation results also show that the SKNet-based coding-decoding network has more advantages in reconstruction accuracy,as compared to other existing networks.