Experimental Demonstration of a Machine Learning-based Piece-wise Digital Predistortion Method in 5G NR systems

Aug 4, 2026

Upload by Noah

The paper presents a piece-wise digital predistortion (PW-DPD) approach for power amplifier linearization in 5G NR systems, where a digital predistorter is built by classifying operational states using machine learning. The method derives modeling advantages by extracting key features from 5G NR signal statistics and the PA operating point, enabling a favorable performance–complexity tradeoff versus a conventional single pruned Volterra model. Laboratory validation confirms up to 3.5 dB EVM improvement for a class A PA at 28 GHz.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages4

Reading Time10 min

AccessFree to download

INTRODUCTION

THE PROPOSED DPD METHOD

ML Classification Model for Region Partition

Volterra model

Experimental Demonstration of a Machine Learning-based Piece-wise Digital Predistortion Method in 5G NR systems S. S. Krishna Chaitanya Bulusu, Bilal Khan, Nuutti Tervo, Marko E. Leinonen, Mikko J. Sillanpää, Olli Silvén, Markku Juntti, and Aarno Pärssinen University of Oulu, Oulu, FI-90014, Finland. Abstract—This paper demonstrates a piece-wise digital predistortion (PW-DPD) for a power amplifier (PA) in 5G new radio (NR) systems. It involves modeling the digital predistorter based on the machine learning (ML) classification of the operational states. The experimental results demonstrate that by extracting some key features from 5G NR signal statistics and the PA operating point can offer better PA linearization performance/complexity tradeoff than the conventional approach based on a single pruned Volterra model. The proposed approach is validated by laboratory experiments and shows up to 3.5dB error vector magnitude (EVM) improvement over the conventional approach for a class A PA at 28GHz. Keywords—5G New Radio (NR), behavioral modeling, digital predistortion, linearization, machine learning, power amplifiers. I. INTRODUCTION Digital predistortion (DPD) of wireless transmitters requires accurate modeling of power amplifiers (PAs) . However, PA modeling with a single polynomial that is valid over the entire range of the communication signal power dynamics is a daunting task. Moreover, as the input data, modulation order, power, and bandwidth may be varying overtime, a constant adaptation of a large set of model coefficients is often required to achieve sufficient modeling accuracy. To overcome the challenge, piece-wise (PW) polynomial-based DPD approaches have recently gained attention in the literature as they are very effective in modeling the PAs with strong non-linearities and memory effects [1], [2] . In general, PW approaches require dividing the model into multiple regions, for example, based on the instantaneous power of the signal samples. However, different waveforms, such as 5G NR signals, have different probabilities for different instantaneous power levels around the average power of the signal. In other words, while the highest peaks of the signal suffer the most from the non-linearity, their probability is also the lowest. In this paper, we utilize the input signal statistics for the region partition of the piece-wise linearization model. To the best of the authors’ knowledge, this aspect is scarcely considered in the open literature. In the proposed approach, the signal is classified into different classes based on a simple machine learning (ML) model. Each of these classes is then modeled by tailored Volterra models individually. By an extensive set of measurements, it is shown that the proposed approach can improve the PA modeling with a relatively low number of overall model coefficients. The rest of the paper is organized as follows. The proposed scheme is explained in Section II. The experimental setup is given in III and experiment results are analyzed in Section IV and then the paper is concluded in Section V. II. THE PROPOSED DPD METHOD A. ML Classification Model for Region Partition Supervised ML approaches such as k-nearest neighbors (kNN) and decision tree (DT) were considered for ML-based classification in this paper. These two are less complex, non-parametric, and do not have an assumption on the distribution of the data. Therefore, the essence of our approach lies in extracting some key features from statistics of the input signal and also the PA characteristics to build an accurate ML classifier model which partitions the input signal space into B classes (or regions) based solely on the current input sample. We denote xk and yk as the discrete-time input and output samples of PA in the baseband, respectively. Let k ∈ N bethe time index of xk where N is the set of all time sample indices. Let κ (i) ∈ N be the set of indices of the samples that are classified as belongin

FAQs

What is the main idea of the proposed piece-wise digital predistortion method?

The method partitions the input signal into regions by using an ML classifier on operational state features, then applies tailored Volterra (GMP) models per class for predistortion.

Which features are used for ML-based region partitioning?

The approach extracts statistical features from the 5G NR input signal envelope and PA characteristics, including amplitude deviation, energy, real/imaginary parts, and an energy deviation relative to the 3 dB PA compression point.

How does the proposed approach perform compared with the conventional single pruned Volterra model?

Experimental results show improved linearization with up to 3.5 dB EVM improvement for a class A PA at 28 GHz, while achieving a better performance–complexity tradeoff.

Aug 4, 2026

Upload by Noah