Light Convolutional Neural Network for Digital Predistortion of Radio Frequency Power Amplifiers

Predistortion models of radio frequency (RF) power amplifier (PA), such as the generalized memory polynomial (GMP) model and artificial neural networks (ANNs) model, suffer from limited predistortion precision and high complexity. In this letter, we propose an enhanced digital predistortion (DPD) model based on a light convolutional neural network (CNN) with augmented real-valued and cross-memorized terms (ARCT). To this end, 1-D complex signals of the PA are initially mapped into 2-D real signals in the form of the ARCT matrix, which serves as the input layer. With cross-memorized terms, the matrix contains sophisticated feature information related to nonlinearity and memory effects. Then, a convolutional layer is designed utilizing macro convolutional kernels with a wide receptive field, which could reduce the number of parameters and effectively extract nonlinear feature information. Following this, a max pooling layer contributes to reducing floating-point operations (FLOPs), improving generalization capability, and preventing overfitting of the proposed model. By these means, the proposed model can significantly extract nonlinear basis functions of the PA with low computational complexity, and realize indirect learning of the DPD parameters. The experimental results, based on a 160MHz Doherty PA, indicate that the proposed model effectively decreases error vector magnitude (EVM) and adjacent channel power ratio (ACPR), compared to state-of-the-art models. In addition, the proposed model has fewer parameters and FLOPs.