Design and Implementation of a Neural Network Based Predistorter for Enhanced Mobile Broadband

Abstract

Digital predistortion is the process of correcting for nonlinearities in the analog RF front-end of a wireless transmitter. These nonlinearities contribute to adjacent channel leakage, degrade the error vector magnitude of transmitted signals, and often force the transmitter to reduce its transmission power into a more linear but less power-efficient region of the device. Most predistortion techniques are based on polynomial models with an indirect learning architecture which have been shown to be overly sensitive to noise. In this work, we use neural network based predistortion with a novel neural network training method that avoids the indirect learning architecture and that shows significant improvements in both the adjacent channel leakage ratio and error vector magnitude. Moreover, we show that, by using a neural network based predistorter, we are able to achieve a 42% reduction in latency and 9.6% increase in throughput on an FPGA accelerator with 15% fewer multiplications per sample when compared to a similarly performing memory-polynomial implementation.

Residual Neural Networks for Digital Predistortion

“…In contrast to model-based DPD approaches, deep learning techniques such as neural networks (NNs) have recently been proposed for DPD [7]- [14]. Among them, the multilayer perceptron (MLP) is the most commonly chosen type of NNs for DPD [9]- [14] because of the simple implementation and training algorithm.…”

“…In contrast to model-based DPD approaches, deep learning techniques such as neural networks (NNs) have recently been proposed for DPD [7]- [14]. Among them, the multilayer perceptron (MLP) is the most commonly chosen type of NNs for DPD [9]- [14] because of the simple implementation and training algorithm. Based on the MLP, [9] proposed a realvalued time-delay neural network (RVTDNN) that separates the complex-valued signal into real in-phase and quadrature components to use a simple real-valued training algorithm.…”

“…However, while these additional components have been shown to improve performance, they also significantly increase the network complexity, which pushes more pressure on the power consumption of DPD. [14] considered a different approach to connect the input and output layer by a linear bypass, which makes the NN focus on the nonlinear relation. However, this approach is infeasible for a memory input, which limits its performance on PAs with memory.…”

“…However, this approach is infeasible for a memory input, which limits its performance on PAs with memory. Moreover, the performance comparison between NN with and without shortcuts for DPD is not discussed in [14].…”

“…The identity shortcuts introduce no new parameters as well as negligible computational complexity (one element-wise addition). Unlike [14], which excludes memory inputs, the R2TDNN considers memory inputs by applying identity shortcuts between two neurons of the current instantaneous input-output, which also solve the dimension difference between the input and output layer. We apply the proposed R2TDNN to DPD.…”