Tensor Blind Receivers For Multi-Users 3-D MIMO Communication Systems

Multi-input multi-output(MIMO)that uses the multiplex antennas at both transmitter(source)and receiver(destination),is well known as an enabling technology for fifth generation(5G)communication systems.MIMO enables to send and receive more than one signal over the same radio channel,by using multi path propagation.This makes full use of space resources,and offers many advantages,such as high transmission performance,large capability,significant reduction of latency,a simplified multiple access layer,and robustness to interference mobile reception.However,to get the benefits of MIMO systems,accuracy of the channel state information(CSI)obtained at the transmitter is required,such as the use of pre-coding and beam-forming techniques at the transmitter generally require the instantaneous CSI knowledge to carry out transmit optimization.Therefore,besides the users data,CSI also has to be estimated in the real-life communication systems.In this thesis,we propose a family of tensor-based receivers that joint blind estimate the users data and CSI in multi-users MIMO communication systems.We show that the multi-user MIMO received signals can be expressed as a three-way(3-D)tensor model,where the three factor matrices of tensor model are corresponding the user symbols and CSI(including direction-of-arrival(DOA)and delay).Such hybrid tensorial modelling enables a jointly and separately estimation of users symbols and CSI without requiring of the training data.Further,to compare with the traditional matrix receivers,such tensor receivers enable to explicitly take into account the structure information by effectively capturing the multi linear interactions among multiple latent factors.Moreover,to overcome the disadvantages of the traditional alternating least squares(ALS),a basic learning algorithm in the existing tensor receivers,we propose some new learning algorithms,namely:?Delta bilinear ALS(DBALS)algorithms that exploit the predictions between two iterations to obtain predictions,refine these predictions by using the enhanced line search and use these refined values to initialize for two factor matrices.This avoids the random initializations found in the existing ALS algorithm.Further,we take into account the potential orthogonal and Vandermonde structures in the factor matrices for DBALS.This improves the accuracy of latent factors recovery,and even provides a better uniqueness results for the use of tensor models.? Tensor dictionary learning algorithms that combine the tensor factorizations and compressed sensing theory.This enables us to take into account the potential sparse of signals,which also improves the accuracy of latent factors recovery.Specially,this offers a two-way(2-D)DO A estimation for CSI.? Variational inference algorithms that automatic recover the tensor rank that is difficult to acquire and known to be NP-hard,and robustness to the outliers in measurements such as ubiquitous impulsive noise in sensor arrays networks and images.This avoids the requirement of the knowledge of tensor rank and the lack robustness of outliers in the existing ALS algorithm.Moreover,we also take into account the potential orthogonal and sparse structures in the factor matrices to improve the accuracy of latent factors recovery.? Tensor-decomposed neural networks algorithm which is able to evaluate the systems status(such as CSI,interference and noise,etc),and behavior of users,to give the suggestions and adjust parameters for improving and obtaining a higher transmission quality.This includes self-learning(interference and noise),self-adapting(networks status)and self-optimizing(capacity,resources)in 5G networks.? Finally,we extend to the 3-D massive MIMO systems that promise of meeting future huge capacity challenge in 5G communication systems,and the 3-D cooperative MIMO relay systems that offer a better communication reliability,by combining three source-relay-destination nodes to increase the channel gains and space diversity.