Research on Neural Network Digital Pre-distorter Based on Transformer

The design of 5G communication systems requires wider transmission bandwidth and higher modulation levels, which necessitates the linearization of power amplifiers (PAs) to mitigate their nonlinearity. Neural network (NN) models, such as back propagation (BP) and long short term memory (LSTM), have been applied in the field of digital predistortion (DPD) and have become a new solution for radio frequency power amplifier (RF PA) linearization. This paper proposes a Transformer-based NN digital predistorter that utilizes the multi-head self-attention mechanism of the model to simultaneously capture the global temporal relationship and the dependency between signal features, constructing better signal nonlinear characteristics and memory effects for the current signal in the lookahead term. The model improves the prediction performance and parallelism of the model, accelerating the training speed. The experimental results show that the model improves the adjacent channel power ratio (ACPR) index by about 12dB compared to the output signal without DPD processing, and reduces the parameter volume by 57% compared to the LSTM network with similar accuracy. Meanwhile, the model's normalized mean square error (NMSE) index reaches -39dB, which is about 5dB higher than that of the traditional BP NN model.

I. Introduction

With the continuous development of communication technology, the realization of high-speed data transmission has become an important goal of 5G communication system design. In order to achieve gigabit per second (Gbps) transmission rates, researchers are working to improve transmission bandwidth and modulation. However, the wider transmission bandwidth and higher modulation system put forward more demanding requirements for the linearity of radio frequency (RF) power amplifier (PA). As a key component in the 5G communication system, the RF power amplifier is responsible for amplifying the signal and driving the antenna for transmission. In RF communication, RF power amplifiers usually need to operate in a nonlinear region to produce a gain in a compressed state. However, this causes the RF power amplifier output signal to produce alternating and harmonic components on the spectrum. If these crossmodulated signals fall into the band, the bit error rate of the signal will be increased when the signal is demodulated at the receiving end, which will cause distortion in the signal band. At the same time, because the out-of-band crossmodulated signal is close to the in-band signal, it is difficult to suppress by band-pass filtering, so these crossmodulated components (i.e., spectrum proliferation) will seriously affect the signal transmission quality of adjacent channels. In order to solve this problem, researchers began to focus on the linearization technology of power amplifiers [1].