Research On Digital Predistortion Technology Of 5G Broadband Power Amplifier Based On Deep Learning

With the rapid development of mobile communication systems,the bandwidth of baseband signals is continuously widening,the modulation mode of signals is becoming more and more complex,the dynamic nonlinear phenomenon reflected by RF power amplifiers is becoming more and more serious,and the memory effect is also particularly obvious.The linearity requirements of power amplifiers are becoming higher and higher.Digital predistortion,as an important technical means to study the linearization of RF power amplifiers,has also put forward more stringent requirements.Facing the broadband signal of 5G system over 100 MHz,the strong memory effect in the broadband RF power amplifier seriously reduces the linearization performance of the digital pre-distorter based on traditional nonlinearity,and it is necessary to seek a power amplifier model with strong nonlinear characteristics;With the rapid advancement of artificial intelligence and deep learning and the amazing results achieved in recent years,it has won the favor of all walks of life.Compared with traditional methods,the self-learning ability of neural networks is particularly prominent.In terms of RF power amplifier modeling capabilities,deep learning will also take advantage of its powerful self-learning.This topic is mainly based on the Long Short-Term Memory neural network to perform strong dynamic nonlinear behavior modeling and linearization research on RF power amplifiers operating under large-bandwidth signals.The timing of the power amplifier signal and the memory effect of the power amplifier form a good match.This article first analyzes the differences in the nonlinearity of the power amplifier when different bandwidth signals are applied to the RF power amplifier.From the AM / AM nonlinear characteristic diagram,it can be intuitively found that the wider the signal bandwidth,the greater the dynamic range of the nonlinear characteristic.Based on the existing research,the article makes a PA model based on recurrent neural network by combining the memory depth of the RF power amplifier and the number of leading terms,and builds an experimental platform based on R & S instruments,and conducts a series of experimental comparison analysis with traditional power amplifier models.Finally,the 100 MHz test signal of 5GNR was used to verify the Doherty amplifier.The experimental results show that: under the same memory depth,the Generalized Long Short-Term Memory neural network model has the ability to model the strong dynamic nonlinear behavior of the RF power amplifier.In the pre-distortion performance experiment,its suppression of parasitic radiated power of adjacent channels is also good,and its out-of-band parasitic radiated power suppression ratio is improved by up to about 10.8dB compared with that without DPD.