Machine Learning-assisted Wireless Positioning And Channel Prediction For Massive MIMO-ofdm Systems

The development of information technology promotes the progress of society.With the increasing demand for wireless communication,future mobile communication systems need to further improve the communication rate and quality.Massive multiple-input multiple-output(MIMO)can fully exploit spatial resources and is still the most potential research hotspot in the context of the development of future mobile communication systems moves to higher frequencies.Integrated sensing and communication(ISAC)aims to integrate the wireless signal sensing and communication in a single system to realize the mutual promotion of the two.Among them,massive MIMO technology plays a key role in realizing ISAC: On the one hand,massive MIMO can obtain higher spatial resolution,thereby improving the communication system’s ability to sense the environment,such as wireless positioning technique;on the other hand,massive MIMO with the spatial resolution capability can sense spatial channel changes in mobile scenarios,such as channel prediction technique,thereby enhancing system communication capabilities.Researches have shown that machine learning has excellent generalization and fitting capabilities,and is particularly suitable for the adaptive design of the wireless physical layer.Therefore,it is of great value to investigate wireless positioning and channel prediction problems by exploiting machine learning in massive MIMO scenarios.In addition,spatial consistency is reflected in the correlation between channels in adjacent spatial positions,which is a prerequisite to verify the effectiveness of wireless positioning and channel prediction algorithms.Therefore,it is necessary to model the channel with spatial consistency.The main work and contributions of this thesis are as follows.Firstly,to address the modeling problem of channel simulator with spatial consistency,we propose a spatially consistent wireless channel generation method based on the transition of reference points.Starting from spatial consistency,we introduce the concept of correlation distance which describes the independence of channel parameters with respect to distance.And based on the circular scattering model,we theoretically prove that the root mean square-delay spread(RMS-DS)of channel changes continuously and smoothly in the space when the number of multipath is sufficiently large.Then,combined with the Quasi Deterministic Radio Channel Generator(Qua DRi Ga),we select the reference points in the target area according to the correlation distance and generate the channels of these reference points independently.Based on the continuity of RMSDS,we construct an optimization problem targeting RMS-DS and channel power changes to determine the change of clusters in the channel when the spatial position changes,and propose a channel transition method between reference points based on continuity of channel varying in the space,and prove the rationality and uniqueness of this channel transition method.Finally,based on the channel transition method,the channel of target point is transitioned from the channels of reference points according to the spatial positions of the target point and the reference points,thereby extending the spatial consistency to the full two-dimensional/threedimensional space.Further to measure spatial consistency,we advocate the use of correlation matrix distance(CMD),power delay profile(PDP)and power angle spectrum(PAS)correlation coefficients.By numerical simulations,the proposed spatially consistent wireless channel simulator is demonstrated to realize spatial consistency in the entire space.Secondly,to address the problem of high-precision wireless positioning in the complex environment,we propose a fingerprint positioning method for massive MIMO systems based on three-dimensional convolution neural network(3D CNN).Starting from the massive MIMO physical channel model equipped with the uniform planar array(UPA),we prove that the angle-delay domain channel power matrix(ADCPM)contains stable and stationary multipath information,e.g.,delay,power,and angles in the vertical and horizontal directions.And we theoretically prove that compared with other positioning fingerprints,ADCPM has a stronger correspondence with spatial position,an has a smaller size than its equivalent fingerprint in the space-frequency domain.Then,by exploiting the sparsity of the angle-delay domain channel,we propose a noise reduction method to realize robust positioning performance of ADCPM in the noisy environment.Finally,taking the ADCPM as the fingerprint,we propose a 3D CNN enabled positioning method to locate the mobile terminal.Specifically,the proposed 3D CNN is composed of a convolutional refine module to learn the low-order feature maps from the ADCPM,three extended Inception modules to learn the high-order feature maps,and a regression module to output the three-dimensional position coordinates.Through intensive simulations,compared with the conventional search-based positioning methods,the proposed 3D CNN positioning method is demonstrated to achieve higher positioning accuracy,lower computational complexity and storage overhead,and more robust performance against noise contamination.Finally,to address the problem of high-precision channel prediction in the mobile environment,we propose a spatio-temporal autoregressive(ST-AR)-based and complex-valued neural network(CVNN)-based channel prediction methods for massive MIMO systems.Starting from the massive MIMO time-varying channel model in the mobile environment,we first obtain the channel matrix in the angle-delay domain from that in the space-frequency domain via inverse discrete Fourier transform(IDFT),and investigate the channel prediction in the angle-delay domain by leveraging the high angle and delay resolution characteristics of wideband massive MIMO systems.We derive the general angle-delay channel characterization and prove that: a)when the number of antennas and bandwidth are sufficiently large,the correlations between the angledelay domain channel response matrix(ADCRM)elements are decoupled significantly;b)when the number of antennas and bandwidth are limited,the decoupling is insufficient and residual correlations between the neighboring ADCRM elements exist.Subsequently,according to the sparse structure characteristics of ADCRM,we propose a method for selecting significant channel elements based on the signal-to-noise ratio(SNR)to alleviate the impact of noise and only implement prediction on the selected significant channel elements.Next,we propose a ST-AR channel prediction method by exploiting the residual temporal correlation between neighboring channel elements.Going beyond the AR model,it comprehensively considers the IDFT leakage caused by the limited number of antennas and bandwidth,thereby improving the accuracy of the prediction.Finally,we propose a CVNN-based channel prediction method which can predict all the significant elements of ADCRM using a single network.Besides,it can adapt to different scenarios when the training is sufficient,achieving accurate channel prediction with excellent generalization capabilities.Numerical simulation results demonstrate that compared with traditional channel prediction methods,the proposed ST-AR and CVNN channel prediction methods can improve channel prediction accuracy.