Research On Convolutional Neural Network Based Channel Estimation

Orthogonal Frequency Division Multiplexing(OFDM)and millimeter-wave multiinput and multi-output(MIMO)is the key signal transmission technology in current wireless communication system.OFDM is the most dominant format for transmitting signals in current wireless communication systems and is the basis for waveform design for fifth generation(5G)communications.Millimeter-wave massive MIMO provides significant performance gains for communications by leveraging the large bandwidth of millimeter waves and large arrays of massive MIMO,hence its common use in 5G communication technologies and future wireless communications.In a rapidly changing wireless environment,obtaining accurate Channel State Information(CSI)is critical to ensure high data throughput,and in OFDM systems,the receiver recovers signals that have been interfered with during transmission based on CSI.For millimeter-wave massive MIMO systems,the maximum benefit is obtained when different transmitreceive antenna pairs experience independently fading channel coefficients,and the achievement of this optimal performance relies on the availability of accurate CSI.CSI can be obtained directly through channel estimation,and the accuracy of the channel estimation affects the accuracy of CSI and thus the performance of the communication system.Currently,there are certain challenges in channel estimation in communication systems:(1)Traditional channel estimation algorithms have the drawbacks of high computational complexity,high guide frequency overhead,and poor robustness.(2)The high cost of hardware and power consumption becomes unaffordable in millimeter-wave massive MIMO,where each antenna uses a dedicated RF chain.The beamspace channel model and the lensed antenna array-based architecture reduce the number of RF chains,but the channel estimation for beamspace millimeter-wave massive MIMO systems is extremely challenging.To address the above two challenges,this thesis proposes a novel convolutional neural network-based channel estimation algorithm using the results achieved by convolutional neural networks in image processing.First we verify the effectiveness of the convolutional neural network in a simple OFDM system,and then apply it in a beamspace millimeter-wave massive MIMO system.(1)To address challenge one,this thesis proposes a convolutional neural networkbased channel estimation algorithm to optimize the channel estimation performance and improve the channel estimation accuracy by using attention mechanism and residual network.The residual network is used to incorporate the hopping connection,and this constant mapping makes the convolutional neural network back-propagation more convenient.The attention mechanism is used to model the interdependence of all channels in the feature graph.The importance of each feature channel is automatically obtained through learning.According to the importance,the weight of features useful to the task is increased,while the weight of features useless to the task is reduced.Simulation results show that the proposed algorithm can improve the accuracy of channel estimation algorithm and has strong robustness in different channel environments.(2)To address challenge two,this thesis proposes a channel estimation algorithm based on denoising convolutional neural networks.The channel is viewed as a lowresolution image by taking advantage of the sparsity of the channel in wave velocity space.Accurate CSI is obtained by using approximate message passing algorithm and denoising neural network.Simulation results show that our proposed algorithm still has high channel estimation accuracy even if the receiver is equipped with a small number of RF chains.