Research On Behavior Modelling Of RF Power Amplifier Based On Machine Learning Algorithms

With the continuous development of telecommunication industry,the modulation of transmitted signals is becoming more and more complicated,and the bandwidth and peak average power ratio(PAPR)of the signals are also increasing continuously.Radio frequency(RF)power amplifier(PA)as one of the most important devices in the wireless communication system,the high PAPR modulation signal will make the PA work in the saturation area,and the resulting nonlinear characteristics will seriously interfere with the communication quality.In order to reduce the nonlinear distortion,the PA structure is usually improved to increase the back-off region,thus reducing the operating nonlinear strength of the PA;or the PA is linearized by digital predistortion techniques.The most critical step in both schemes is to construct an accurate PA behavioral model to analyze and characterize the nonlinear characteristics of the PA by an accurate model,or to establish an inverse model to achieve the digital predistortion of the PA.Since 1980,machine learning as an independent subject has started to grow rapidly,and powerful machine learning algorithms have achieved widespread success in engineering applications,behavioral models based on machine learning algorithms have received extensive attention.The purpose of this thesis is to investigate the feasibility of behavioral modeling based on machine learning algorithms in various kinds of RF PAs,and to analyze and verify the model performance to establish the foundation for further applications of the subsequent models.Aiming at the strong nonlinear problem of multi-transistor Doherty PA(DPA),this thesis proposes a memory polynomial(MP)-based support vector regression(SVR)model,namely MP-SVR model.This model adds the ability of the model to handle memory effects by including the MP function as an input term.To validate the model,two standard DPAs with different nonlinear characteristics and one multitransistor DPA are used.The experimental results show that the normalized mean square error(NMSE)of the MP-SVR model is improved by at least 14 d B compared to the Volterra series model;by at least 10 d B compared to the SVR model;by at least9 d B compared to the augmented SVR(ASVR)model The above research results have been published in the International Journal of Numerical Modelling:Electronic Networks,Devices and Fields and applied for National Invention Patents.Aiming at the broadband nonlinear problem of ultra-broadband millimeter wave PAs,this thesis proposed an augmented long-short term memory(ALSTM)neural network model.The basic theory of the model and the modeling procedure are introduced in detail.The LSTM parameters are compared and analyzed to select the best parameters of the model and the best optimization algorithm.The established model is also verified by using both a multi-octave single transistor PA and an ultrabroadband millimeter wave PA.The experimental results show that the NMSE of the ALSTM model is improved by at least 3.3 d B compared with the Volterra series model;by at least 3 d B compared with the LSTM model;by at least 2 d B compared with the SVR model;by at least 1.1 d B compared with the ASVR model;by at least 1d B compared with the MP-SVR model.The above research results were published in2021 IEEE International Workshop on Electromagnetics(IWEM),and later were expanded and published in Microwave and optical technology letters.Aiming at the strong nonlinear problem of multi-device active modulation of load modulation balanced amplifier(LMBA),the PSO-DVR model is proposed by combining the decomposed vector rotation(DVR)model with the particle swarm optimization(PSO)algorithm.The PSO algorithm is used to optimize the threshold assignment of the DVR model in the nonlinear region of the PA,thus improving the modeling effect of the model on the strongly nonlinear PA.The performance is verified by modeling the characteristics of load modulated balanced PA and broadband DPA.The experimental results show that the NMSE of the PSO-DVR model is improved by at least 3 d B compared to the Volterra series model;by at least2 d B compared to the DVR model.The above research results have been submitted to IEEE Transactions on Microwave Theory and Techniques.In summary,this thesis presents various methods for modeling the behavior of RF PAs based on machine learning algorithms,analyzes and verifies the nonlinear prediction performance of the models,and lays the foundation for further applications of the subsequent models,as it also provides more options for modeling the behavior of RF PA.