Research On Wideband Digital Predistortion Technology Based On Machine Learning

The future 5G communication will become the main direction of the wireless communication industry.New requirements in terms of speed,capacity and delay have been put forward in the fifth generation of mobile communication.Therefore,5G communication system needs to increase spectrum efficiency and spread spectrum bandwidth.There are many factors that affect the communication quality of wireless communication systems,in which power amplifier performance is one of the most critical factors.Transmitting that kind of signals through nonlinear power amplifier results in in-band and out-of-band distortions,which will increase the bit error rate and cause adjacent channel interference(ACI),so broadband wireless communication systems require high linearity power amplifiers.Power amplifier pre-distortion technology is the mainstream technology to solve this problem.However,the traditional predistortion method has limited effect.This thesis mainly explores how to improve the power amplifier baseband digital pre-distortion linearization technology based on machine learning support vector regression from several aspects such as improving generalization ability,improving modeling effect and modeling speed.Firstly,the thesis introduces the nonlinear distortion characteristics of the power amplifier and related evaluation criteria in detail,and then studies the common nonlinear behavior models of power amplifiers and the commonly used digital pre-distortion structures.The baseband digital pre-distorter in this thesis is the power amplifier inverse model.The support vector machine(SVM)algorithm,which belongs to machine learning,is introduced into the model building of power amplifiers and the study of pre-distorters.At the same time,based on the traditional delay structure,a method for establishing the amplitude and phase enhanced time delay power amplifier model is proposed,and carries out related experiments to verify that the structure has better modeling effect than traditional the delay structure.In this thesis,the SVR,TSVR,LSSVR algorithms and the proposed modeling structure are applied to the power amplifier models respectively.The acquired inverse model is the pre-distorter.These algorithms are used for power amplifier behavior model building and simulation of pre-distorter.Compared with traditional algorithms,the support vector regression algorithm is more accurate in modeling.Ordinary support vector machines have the disadvantages of solving complexities and many supporting vectors.In order to solve those problems,we use a simplified kernel matrix decomposition least square support vector regression(CLSTSVR)method to solve it directly in the original space.After the preliminary training of large-scale training samples,we find the basic set and obtain the approximate matrix through matrix cholesky decomposition.This method can simplify the computational complexity of LSTSVR,reduce the number of support vectors,and solve the problem of the traditional least square method with many support vectors.At the same time,large-scale training samples can be used to greatly improve the modeling effect of the power amplifier model.In order to verify the effectiveness of the proposed method,a digital pre-distortion verification platform is built using a vector signal generator and a vector signal analyzer.The performance of the proposed algorithm is verified by single-device gallium nitride(GaN)Fclass PA and dual-tube GaN Doherty PA respectively.The experimental signals are 10 M bandwidth WCDMA signal and 10 M bandwidth LTE signal.The experimental results show that the proposed modeling method has very accurate ability and very effective nonlinear correction ability than the GMP algorithm.