Research On Linearization Of Beamforming System And Neural-Network-Based Broadband Digital Predistortion

The nonlinearities are the internal characteristics of radio-frequency power ampli-fiers(PAs).The distortions caused by PA nonlinearities are the most influential radio-frequency impairments in modern wireless communication systems.PAs' nonlineari-ties will increase the bit error rate(BER)and cause interferences on adjacent channels.Digital predistortion(DPD)is the mainstream method of compensating for PAs' nonlin-earities.Due to its excellent linearization performance,flexible reconfiguration,simple implementation and low cost,DPD becomes an important part of modern wireless com-munication systems.With the increasing demands for high data transmission speed,large capacity and low latency of wireless communications in daily life and industrial production,modern wireless communication system is gradually developing towards the trend of high frequency band and large bandwidth.In order to improve the data transmission rate and the system capacity,the MIMO technology and the beamform-ing technology are also applied in wireless communication systems.Due to the high PAPR caused by wideband signals,the nonlinear distortions of PAs are aggravated.In addition,with the application of high frequency band,such as millimeter wave,the mod-ern base stations present the trend of miniaturization or microminiaturization,and the volume and power consumption of each base station are greatly reduced.Therefore,aiming at the above problems,this paper studies DPD in MIMO beamforming archi-tecture,neural network modeling of broadband PAs and low-complexity DPD adaptive algorithm.This paper focuses on the DPD technology in MIMO beamforming architecture.Firstly,the system model of single-user beamforming DPD is analyzed,and the user-based or beam-based DPD architecture is introduced.The anti-beamforming module is set on the feedback loop of the architecture to approximate the far-field beam sig-nal.Then,based on the single-user beamforming system model,the nonlinear model of multi-user beamforming system is derived.Aiming to reduce the high hardware complexity of the feedback architecture with all PAs,a multi-user beamforming DPD method based on a single PA feedback is proposed.The single PA feedback is used for forward modeling,and the extracted model is used to approximate all the PAs in the array,so as to estimate the output of all PAs and the far-field beam signals which is used to extract DPD parameters.This architecture can linearize the multi-user beamforming systems with a low complexity.Furthermore,aiming at the problem of inconsistent PA characteristics,a low-complexity PA difference compensation method is proposed based on pretraining,which can make the multi-user beamforming DPD method based on a single PA feedback available in the non-ideal actual system with diferrent PA char-acteristics.On the basis of the DPD method with PA difference compensation,accord-ing to the nonlinear characteristics of PAs,a simple power compensation coefficient is proposed to modify the differential coefficient of PAs to deal with the dynamic power changes.On the premise of avoiding cumbersome pretraining,effective DPD lineariza-tion can be realized with this kind of power compensation coefficient.As for the neural network-based modeling of broadband PAs,this paper mainly studies a novel vector decomposed recurrent neural network.The nonlinearity of the PAs is often determined by the envelope of the signal,and its nonlinearity is mainly de-termined by the amplitude of the baseband signal.The output of the amplitude nonlinear function is weighted by the phase to get the output of the PA.This mechanism is called vector decomposition mechanism.However,most of the nonlinear operations of the tra-ditional neural network model are directly conducted on the I/Q inputs,which does not conform to the nonlinear mechanism of PAs.Moreover,the traditional DPD neural net-work models based on multi-layer perceptron has insufficient abilities to model memory effects,so they can not be applied to the modeling of broadband PAs.Therefore,this paper utilizes the excellent memory effect modeling ability of recurrent neural network and introduces the vector decomposition mechanism to design a novel long short-term memory(LSTM)model based on vector decomposition mechanism,named VDLSTM model.Compared with the traditional model,this model has better modeling ability of broadband PAs.Furthermore,to solve the problem of the numerous parameters of the transfer matrix corresponding to the hidden state of LSTM,a new hidden state with a lower dimension is designed based on the physical mechanism of PAs,together with a new network unit based on the new hidden state,and consequently a simplified VDL-STM model named SVDLSTM model is proposed.This model can effectively reduce the number of parameters and keep a comparable performance compared to the VDL-STM model.The low-complexity DPD adapatation architecture is also a hot research topic of DPD.In the process of DPD adaptation,considerable multiplication and addition op-erations are often needed.Compared with the addition operation,the multiplication operation often requires a longer calculation cycle and a higher power consumption.In this paper,a low-complexity DPD adaptive algorithm is designed to solve this prob-lem.Referring to the signed regressor least mean square algorithm,this algorithm quan-tizes the regression matrix of the second-order Gauss Newton algorithm based on direct learning structure with a single bit,so as to reduce most of the multiplications in the calculation process.In order to solve the problem that the regression matrix generated by the commonly used model of DPD will have the same basis function with the 1-bit quantization,the orthogonal transformation based on principal component analysis is introduced to conducted on the regression matrix and the orthogonal regression ma-trix is used to be quantized instead.This method can avoid solving unstable problems.Furthermore,the correlation matrix between the signed orthogonal regression matrix and the orthogonal regression matrix is approximated to be a diagonal matrix,so that the adaptive algorithm based on the signed orthogonal regressor can extract each DPD parameter independently.Then,this paper combines the basis function reduction and signed regressor algorithm to reduce the computational complexity of DPD adaptation from two aspects,which are the number of basis functions and the operation type of the adaptive process.In particular,the DOMP algorithm is introduced to be combined with the signed orthogonal regressor algorithm.