Research On Linearization Technology Of Short-Wave Power Amplifier Based On Deep Learning

Short-wave communication is an important communication method,it's widely used in aviation,military,broadcasting,and emergency communications.As an important component of wireless transmitters,the working state of short-wave power amplifiers has a great influence on short-wave communication systems.The output power of short-wave power amplifiers is relatively large,so improving the working is very important.However working efficiency and linearity are often incompatible.In the new generation of short-wave communication standards,the use of non-constant envelope modulation puts higher demands on short-wave power amplifiers.The pre-distortion technology can compensate the nonlinear amplification of the power amplifier.After improving the linearity of the power amplifier,the out-of-band spurious radiation can be greatly reduced while efficiency decline avoided.Firstly this paper analyzes the phenomenon of nonlinear amplification of power amplifiers and the causes of out-of-band parasitic radiation.Different modeling ideas and some common model structures are introduced.Common models include memoryless models such as Look-up Table model,Saleh model and Memoryless Polynomial model,as well as memory models such as Volterra series model,Memory Polynomial model and multi-box model.Since deep learning became a research hotspot in 2006,different deep learning networks have been proposed and applied.Compared to traditional methods,deep learning has achieved extraordinary and even rolling performance in many areas.Inspired by deep learning,this paper designs a deep learning model based on LSTM network.The LSTM network-based model consists of three layers: the data processing layer,the LSTM network layer and the fully connected layer.The model can well fit the nonlinearity of power amplifiers because of strong nonlinear fitting ability of LSTM network and fully connected network.The model can deeply reflect the timing of the signal,which is significant for the memory effect description of power amplifiers.In the training process,the MXNet deep learning framework is used to build and train the model.The Adam optimization algorithm is used to calculate the parameters.In the modeling test of the short-wave power amplifier,it is found that NMSE of the LSTM network-based model is significantly smaller than that of the traditional models of Memory Polynomial model(MP)and Augmented Hammerstein model(AH).The test signal is QPSK modulated with 24 k Hz bandwidth at a center frequency of 16 MHz in the short-wave communication band.NMSEs of MP and AH model are about-40 d B.NMSE of the LSTM network-based model is lower than-43 d B.In terms of pre-distortion performance,the LSTM network-based model also significantly works better than the traditional models of MP and AH for the ACPR improvement.The MP and AH model perform similarly at 1.6 MHz,9 MHz,16 MHz,23 MHz and 29.9 MHz.Compared with unpre-distortion,ACPR can be reduce by 12~17 d B,4~14 d B in ±24 k Hz and ±48 k Hz adjacent channels.The ACPR is reduced by 14~20 d B at ±24 k Hz and 10~14 d B at ±48 k Hz adjacent channel when LSTM network-based model used.In contrast,the LSTM network-based model has significantly lower ACPR at each frequency point than that of the MP and AH model.In the modeling and linearization of the short-wave power amplifier,the deep learning method performs better.