Machine Learning Based Modeling And Linearization Techniques For 5G Millimeter Wave Transmitters

Millimeter wave(mmWave)band,containing large amount of spectrum resources and playing an important role in the fifth generation(5G)and future communication systems,is going to be widely employed.To overcome the severe path propagation loss of mmWave,mmWave band will be combined with multiple-input multiple-output(MIMO)transmitter architecture and beamforming technique so as to synthesize a beam with highly focused energy and introduce many system functions such as beam steering and multi-beam.Meanwhile,in order to make MIMO transmitters operate in a state of high efficiency and high linearity,it is required to be linearized.As for power amplifiers(PA),which are crucial components in the MIMO transmitter,digital predistortion(DPD)is PAs'popular linearization technique.However,the architecture of the MIMO transmitter will bring huge challenges to the conventional DPD techniques.On one hand,complex nonlinear characteristics will be introduced in MIMO transmitters because of mutual coupling between channels and antennas.On the other hand,MIMO transmitters will operate with dynamic operating parameters in order to make full use of the limited spectrum resources,which means that dynamic nonlinear characteristics will be introduced.To solve the mentioned challenges on modeling and linearization of MIMO transmitters with complicated and dynamic characteristics,several innovative solutions are proposed employing the powerful machine learning technique,with great validation performance achieved on testbenches.Aiming at the issues of modeling MIMO transmitters with complicated and dynamic characteristics,a heterogeneous neural network based behavioral modeling technique for MIMO transmitters is proposed.Employing multilayer perceptron algorithm in supervised learning to achieve regression operation and adding heterogeneous dynamic parameter terms into the input layer of conventional real-valued time delay neural network(RVTDNN),its nonlinear modeling capacity can be highly expanded,which has achieved great experimental validation performance in 27 GHz mmWave MIMO transmitter.Parts of the above results have been published on 2020 IEEE MTT-S International Microwave Symposium(IMS).Aiming at the issues of linearizing MIMO transmitters with complicated and dynamic characteristics,a data-clustering-assisted linearization technique for MIMO transmitter is proposed.Employing clustering algorithm in unsupervised learning and clustering the nonlinear behaviors of an exponential number of MIMO transmitters under multiple dynamic state combinations,the number of DPD models can be significantly reduced with comparable linearization performance,which has achieved great experimental validation performance in 28 GHz mm Wave MIMO transmitter with 16 antenna elements.Parts of the above results have been published on IEEE Transactions on Microwave Theory and Techniques(TMTT).Additionally,with respect to some detailed issues in modeling and linearizing MIMO transmitters with complicated and dynamic characteristics,three researches on detailed problems have also been conducted based on PAs,which are crucial components in MIMO transmitters.Aiming at the neural network parameter extraction problem when modeling complicated nonlinear characteristics,the dynamic iteration process of RVTDNN model(supervised learning)with different activation functions is analyzed and compared,which provides guidance for the selection of activation functions in practical applications,with good experimental comparison achieved in Doherty PA in the frequency band below 6 GHz(sub-6 GHz).Aiming at the fast recognition of dynamic nonlinear behaviors,a pattern sensing-based dynamic linearization technique is proposed,which can effectively recognize the dynamic characteristics and index the corresponding coefficients,which has also achieved great experimental validation performance in sub-6 GHz Doherty PA.Aiming at the hopping nonlinear characteristics,a directional graph navigation based nonlinearity control technique is also proposed.When DPD is hard to be timely updated,the system nonlinearity hopping caused by operating frequency hopping can be effectively offset by controlling operating power,which has achieved great experimental validation performance in the 28 GHz mmWave PA.Parts of the above results have been published in 2019 IEEE International Symposium on Radio-Frequency Integration Technology(RFIT)and 2019 IEEE International Microwave Conference on Hardware and System for 5G and Beyond(IMC-5G),and have been accepted by IEEE Microwave and Wireless Components Letters(MWCL).