Design of channel state information compression technique with deep learning algorithm

Abstract

The massive Multiple Input Multiple Output (MIMO) technique is widely used in the fifth-generation communication system, which improves spectrum efficiency and energy efficiency significantly. Relatively, it also challenges the existing channel state information (CSI) feedback system due to the use of a large number of antennas. Considering the balance between the huge overhead and the requirements for highly-accurate communication, we plan to use a deep learning-based method to compress and reconstruct the CSI. In this dissertation, we consider a network that combines deep neural network (DNN) and a convolutional neural network with two parallel convolutional kernels, which is called Anciblock. In the encoding process, the channel matrix will be processed by several convolutional layers, an Anciblock and a fully-connected layer in order. After reshaping into a M-dimensional vector, the vector will be transmitted into the decoder as the codeword. The decoder module consists of several FC layers. The codeword will be reconstructed to the channel matrix. We evaluate the performance of the CSI compression system by comparing the output matrix of the decoder with the initial matrix. After training with the COST 2100 MIMO channel dataset and 2020 NIAC dataset, the DNN with Anciblock performs better than simple DNN and CSInet.