Low Computational Complexity Digital Predistortion Based on Convolutional Neural Network for Wideband Power Amplifiers

Abstract

The convolutional neural network (CNN) based power amplifier (PA) model has been proven to reduce the model complexity significantly. However, due to the calculation mode of the convolutional structure, the application of the CNN-based predistortion model still faces the problem of high computational complexity. In this letter, we use one lightweight CNN to propose a modeling method of the predistorter with low computational complexity for the wideband PA. This method first decomposes the traditional two-dimensional convolution kernels into two kinds of one-dimensional convolution kernels, to create the predesigned filter layer. These two kinds of convolution kernels are used to successively construct the nonlinear terms and the cross basis function terms required by the digital predistortion (DPD) model, respectively. Then, the unnecessary connections of the fully connected structure are removed using the pruning method based on amplitudes, to further reduce the complexity. Experimental results based on 100 MHz Doherty PA show that this predistortion model can significantly reduce the computational complexity, while ensuring that the linearization effects do not deteriorate.