Millimeter Wave Communication Parameter Estimation And Resource Allocation Based On Machine Learning

Machine learning is considered to be one of the key technologies to solve the problems of high energy consumption,high hardware cost and high computational complexity of existing millimeter wave(mm Wave)multiple-input and multiple-output(MIMO)communication systems.Parameter estimation and resource allocation for mm Wave massive MIMO systems is studied by adopting machine learning in this thesis,and the detailed work is concluded as follows.The channel estimation is studied for wideband mm Wave massive MIMO system.By transmitting frequency-domain pilot symbols and beam scanning,that time-domain channels is observed to exhibit delaydomain sparsity and space-domain block sparsity,hence the time-domain channel estimation problem is modeled as a block sparse reconstruction problem.Then a deep learning based time-domain channel estimation scheme is proposed,which includes non-zero block location estimation and channel reconstruction.The scheme first adopts machine learning methods to predict the positions of all non-zero blocks,and then uses the least squares estimation method to reconstruct the channels.The multiuser beam allocation and power allocation problem for wideband mm Wave massive MIMO systems is also studied.By introducing the interference-free achievable rate,the analog beamforming and the digital beamforming is decoupled for the beam allocation problem.Then the beam allocation is treated as a multi-label classification problem and a deep learning-based beam allocation(DLBA)scheme is proposed.The scheme first uses machine learning methods to predict the beam allocation results of all users online,and then in order to avoid the beam conflict and maximize the sum-rate,a rule to avoid the beam conflict is also proposed.Simulation results demonstrate that the DLBA scheme can substantially reduce the computational complexity with a marginal sacrifice of sum-rate performance compared to the existing schemes.On this basis,the multiuser power allocation problem is further studied.The problem is first modeled as an optimization problem with inequality constraints,and then the KKT conditional method and Lagrangian function are used to find the optimal solution of this optimization problem.Since the specific analytical expression cannot be obtained,the water filling(WF)algorithm is adopted to find the approximate solution,hence the result of power allocation is obtained.Simulation results show that the WF-based scheme has better sum-rate performance compared with other existing schemes.In order to reduce the computational complexity of the WF algorithm,a deep learning based power allocation(DLPA)scheme is proposed.The scheme firstly uses machine learning method to predict the power allocation results of all users online,and then in order to satisfy the constraint of total transmit power,a constraint redistribution rule is also proposed.The simulation results show that the DLPA scheme can greatly reduce the computational complexity on the premise that the sum-rate performance is basically unchanged compared with the WF-based scheme.