LSTM-Deep Neural Networks based Predistortion Linearizer for High Power Amplifiers

Abstract

Linear high power amplifiers (HPAs) are the need of current communications technology. But, almost all PAs show non-linear characteristics during amplification which are reflected in the transmitted signal in the form of distortions. Linearization is a process to suppress the effect of the nonlinear characteristic of a PA. Various methods are available to perform linearization. Predistortion (PD) linearization methods are very successful due to its simplicity in design and ease of integration with PAs. PD linearization methods observe the PA dynamic characteristics (nonlinearity) and then formulate an “inverse transfer function” to suppress this non-linearity. In the last decade, machine learning (ML) based PD linearizers are proposed and proved useful. Since then, numerous ML-PD linearizers have been developed. Shallow neural networks (NNs) based PD linearizers are successfully used to map the inverse transfer function but lack generalization performance in the presence of system conditions (IQ imbalance, DC offset). With deep learning (DL) technology, deep neural networks (DNNs) can map the complex inverse transfer function under different system conditions. This study proposes a long short-term memory (LSTM) DNN based PD linearizer for linearization of PA under different conditions. In this study, it is shown that LSTM is able to extract and exploit memory effects of PA over the perceptron. Comparative results with shallow NNs suggest reliable potential in the proposed DNN model in terms of generalization performance.