Low Complexity Convolutional Neural Network Digital Predistortion for Radio Frequency Power Amplifiers

The predistortion models for radio frequency (RF) power amplifiers (PAs), such as convolutional neural networks (CNNs) and deep neural networks (DNNs), has shown a significant improvement in predistortion performance. However, digital predistortion (DPD) models with neural network structures encounter challenges related to high computational complexity and prolonged training duration. In this study, we propose a DPD model utilizing a low complexity convolutional neural network (LCCNN). The proposed model simplifies the traditional CNN structure, employing an augmented real-valued cross-memorized terms (ARCT) matrix as input layer, and configuring convolutional layer to extract nonlinear features through macro convolutional kernels. Furthermore, migration learning and pruning operations are implemented to expedite model training and parameter reduction, thereby decreasing training duration and floating-point operations (FLOPs). The experiment results based on a 160 MHz Doherty PA indicate that the proposed model effectively reduces computational complexity while maintaining required predistortion performance.