Scalable Behavioral Models and Predistorters for Broad Band Power Amplifiers

Abstract

With the ever-changing improvements in the realm of telecommunication, power amplifiers (PAs), being an indispensable component, are required to adhere to very high demands. The technology behind manufacturing power amplifiers has also improved over time but they are still prone to suffer from nonlinearities. Since amplification requires the PAs to be driven at high voltage levels, it is inevitable for them to exhibit such nonlinear behavior. This thesis was a pursuit to find a neural network (NN) based digital predistorter (DPD) to rectify the power amplifier’s nonlinear behavior. The proposed model was scalable and obliges with changing average power levels, varied bandwidth as well as heterogeneous carrier configurations of signals. The proposed neural network was assessed for behavioral modeling and showed that it is capable of accurately mimicking the memory as well as static nonlinearities of the device under test (DUT) with an average normalized mean square error (NMSE) of -29.67dB. The proposed DPD NN model was investigated for robustness with respect to the signal’s characteristics, such that the offline model does not require signal dependent updates. The signal with the highest memory effect intensity (MEI) was then proposed for the model's initial training and was found to be linearizing all the rest of the various configurations and reaching ACLR values up to -55dBc. The proposed DPD has been tested on 20MHz long-term evolution (LTE) as well as 40MHz, 30MHz, 20MHz and 10MHz new radio (NR) signals with various carrier configurations and power levels and has been observed to be meeting the 5G NR ACLR requirements. Furthermore, the proposed DPD was also trained on reduced sampling rate data to accommodate for limited hardware capabilities. It proved to be still scalable and provided satisfying linearization performance with an average ACLR of -49.29dBc.