Research On 5G-Oriented Intelligent Channel Estimation And Feedback Technology

In order to achieve the key indicators of the new generation of mobile communication technology,the5 th Generation(5G)communication system adopts technologies such as massive multiple-input multipleoutput(MIMO)to improve system performance.However,the prerequisite for taking full advantage of the massive MIMO technology is to obtain channel state information(CSI)in an accurate manner.This thesis mainly studies 5G-oriented intelligent communication,which focuses on exploring the application of deep learning in multiple tasks such as channel estimation,channel information compression and feedback.In this thesis,classic neural network architectures are fused and optimized to design efficient proposals for channel estimation and channel information compression and feedback.The first chapter of the thesis briefly introduces the 5G mobile communication system at first.Then,the concept and main advantages of the massive MIMO technology,which serves as the key technologies of5 G,are introduced.Subsequently,the importance of CSI to the wireless communication system is explained,and existing channel estimation and CSI feedback technologies for obtaining channel information are investigated.Finally,the development and value of artificial intelligence,machine learning and deep learning are introduced,the general method of parameter adjustment process of neural networks is explained.The three basic network structures of neural networks are introduced in detail,including deep neural network(DNN),convolutional neural network(CNN)and recurrent neural network(RNN).The second chapter of the thesis considers the application of DNN to reduce the demodulation reference signal(DMRS)overhead in equivalent channel estimation.This chapter first introduces channel estimation methods based on reference signals,including the least square(LS)algorithm and the linear minimum mean squared error(LMMSE)algorithm.However,due to the complicated 5G channel environment and a large number of nonlinear factors,the overhead of reference signals in these conventional channel estimation methods is relatively large,and is difficult to meet the high-speed and high-reliability requirements of 5G.Taking advantage of deep learning in dealing with nonlinear problems,the subsequent part of this chapter studies the deep learning scheme for equivalent channel estimation.By performing time-correlated signal preprocessing on the time-correlated equivalent channel coefficient estimates and utilizing DNN for equivalent channel estimation,the DMRS overhead can be effectively reduced and the accuracy of the equivalent channel estimation can be improved.In addition,at the end of this chapter,the pattern learning problem of DNN is explored.For a DNN that has learned a certain fixed pattern,the input structure must conform to the pattern in order to achieve an efficient DNN with good performance.The third chapter of the thesis considers the difficulty to obtain perfect CSI in practical applications.The problem of CSI feedback with non-ideal channel estimation input in massive MIMO systems is studied and a neural network structure for noisy CSI compression feedback named Anci Net is designed.Anci Net adopts a dual-module structure,which consists of two modules: a pre-denoising module and an autoencoderbased feedback module.The pre-denoising module at the front end of the Anci Net preliminarily eliminates the noise before encoding.The feedback module is placed after the pre-denoising module,which is used to suppress the residual noise in the compression and decompression process.The key structure of the Anci Net is the Anci-block,which effectively extracts noise-free feature of the massive MIMO CSI matrix by placing two convolutional layers with smaller convolution kernel size in parallel after the convolutional layer with a larger convolution kernel size.Anci Net is able to restore the original CSI matrix from the noisy input with high accuracy.This chapter also introduces a two-stage training method to optimize the network training of Anci Net and improve CSI reconstruction performance.Simulation results have shown that Anci Net performs well at various conditions.The fourth chapter of the thesis studies the efficient and lightweight design for CSI feedback in massive MIMO systems and devises a neural network named ENet.ENet has the following two characteristics: Firstly,ENet adopts different compression strategies for the CSI matrix in the angular domain and the delay domain to take advantage of the unique correlations of the two transform domains,thereby improving the feedback reconstruction performance of the network.Secondly,by using the correlation similarity between the real and imaginary parts of the complex-valued angular-delay domain channel matrix,only the real part of the CSI matrix is utilized for the network training while both the real and imaginary parts are compressed and fed back using the network with the same trained parameters.Without stacking the real and imaginary parts of the channel coefficients as an entire real-valued input like existing neural networks for CSI feedback,the number of parameters of the ENet can be greatly reduced.While the network size is reduced by nearly an order of magnitude,the ENet exhibits brilliant performance compared to the existing NN-based methods.The fifth chapter reviews the whole thesis and provides suggestions for possible research directions in the future.