Research On Emitter Power Amplifier Modeling And Digital Predistortion Based On Deep Learning

When the signal is amplified by the power amplifier in the transmitter,there is nonlinear distortion and spectrum leakage,which affects adjacent channels and violates the out-of-band radiation standard.The corresponding in-band distortion will affect the communication quality.With the update and iteration of the communication system,the signal modulation mode becomes more and more complex,the peak to average ratio becomes higher,and the linearization of power amplifier becomes more important and arduous.In addition,deep learning shows excellent performance in various fields.It can not only process and extract the multi-dimensional features implied in various business or research scenarios,but also has the flexibility of structural optimization and the mobility of model application.Therefore,the power amplifier modeling and predistortion problems also gradually adopt the deep learning model to get better performance.Considering the Volterra series model is difficult to solve the model parameters when modeling the power amplifier with strong memory and nonlinearity,a method based on gated recurrent neural network(BGDNN)is proposed.It is easier to get the parameter of model,which is successfully to solve the problem that the traditional model is difficult to obtain the parameter in complex communication system.Simulation results show that compared with traditional models and RNN models,the proposed BGDNN model not only improves the modeling performance,but also reduces the amount of parameters.According to the experimental basis that the deep learning model can be well applied to the modeling of power amplifier,a predistortion structure of power amplifier based on deep learning is proposed.The power amplifier model and predistortion model can share the same neural network structure,which can effectively solve the problem that the traditional model is difficult to reverse in the indirect learning structure.This structure can not only effectively use mature algorithms to obtain the parameters of power amplifier model and predistortion model in parallel,but also avoid the problem of long time of using closed-loop iteration to get the parameters.Next,considering that the traditional model is difficult to accurately model and linearize the power amplifier with strong dynamic nonlinearity,a convolution cyclic spatiotemporal neural network(CRSNN)is proposed.In this model,the expanded convolution is used to learn the static information,such as short-term memory and spatial distribution information;Then BGRU is utilized to learn the implicit long-term memory information,which can better characterize the strong dynamic characteristics of the power amplifier.Finally,through the multi-dimensional comparison between the proposed CRSNN and the traditional power amplifier predistortion methods and the predistortion methods based on other neural network models such as CNN and RNN,it is verified that the proposed CRSNN has a good linearization correction effect,significantly alleviates the nonlinear distortion and memory effect,and has a good ability to suppress the spectrum leakage out of band.Finally,the content and progress of the research are sorted out,and the directions that can be optimized and explored in the next step are analyzed.