| With the continuous growth of mobile Internet and Internet of Things application requirements,driving the rapid development of wireless communication,now the fifth-generation mobile communication system(5G)has entered the commercial stage,of which Massive Multiple Input Multiple Output(MIMO)has become an important enabling technology in 5G mobile communication system.Massive MIMO systems have great improvements in spectral efficiency and energy efficiency;however,more accurate channel estimation and signal detection are required when utilizing the gain brought by multiple antennas.Among them,the channel estimation performance for the uplink not only affects the performance of uplink signal detection,but also affects the performance of downlink multiuser MIMO precoding,and the signal detection also faces the challenge of high computational complexity.In recent years,deep learning technology has achieved a lot of results in computer vision,natural language processing and other fields,and researchers in the field of wireless communication are also looking forward to applying it to communication systems to meet people’s ever-changing demands for data transfer rates.In this thesis,deep learning is applied to two modules of channel estimation and signal detection in massive MIMO systems.In the research of channel estimation,the performance of channel estimation is improved by extracting the spatial correlation and frequency correlation of the channel matrix through a deep learning model for denoising.In the research of signal detection,the conjugate gradient method is expanded into a deep network model to reduce the computational complexity of detection and improve a certain detection performance.Among them,the main contributions and innovations of the thesis include:1.A channel estimation algorithm based on K-S test and deep learning is proposed.The innovative proposal uses traditional transform domain denoising before deep learning model denoising,and uses the removed time domain noise to estimate the noise variance in the frequency domain.Then,the initial denoised channel matrix and the estimated frequency domain noise variance are input into the convolution denoising network,and the spatial and frequency correlations are used for further denoising.The simulation results show that the performance of the algorithm is better than the traditional channel estimation.algorithm and some deep learningbased channel estimation algorithms.In addition,the traditional transform domain denoising algorithm does not deal with the noise in the cyclic prefix,and an algorithm based on KS hypothesis test is proposed to use the characteristics of the noise distribution brought by multiple antennas to determine whether the removed noise satisfies the noise distribution.The noise in the prefix is processed,and the simulation results show that its performance is better than the traditional threshold-based algorithm under both DFT and DCT transforms.2.It is proposed to expand the iterative conjugate gradient algorithm into a deep network model.By expanding the iterative conjugate gradient method into a multi-layer deep network,setting the iterative step size as a learnable parameter,and then using a large amount of data for learning,the model no longer needs to calculate the iterative step size when it is deployed online,reducing the need for It has a certain computational complexity,and can achieve basically the same performance as the MMSE algorithm.Secondly,the iterative step size is set from a scalar to a vector,and a nonlinear Hadamard product is added to improve the performance of signal detection without increasing the complexity of the algorithm and only increasing the parameter storage space.Simulations show that the algorithm can bring better detection performance and lower complexity than MMSE and other algorithms under BPSK and QPSK modulation. |
