This paper presents the linearization of radio frequency power amplifier using a deep neural network (DNN) for distortion compensation. The linearization results are shown for heterojunction bipolar transistor based RF power amplifier. The proposed DNN digital predistortion (DPD) uses Swish or Sigmoid-weighted linear unit activation function instead of sigmoid and rectified linear units (ReLU) activation function to avoid the gradient vanishing and dead neurons problem. A comparative study of different activation functions for real-valued focused time delay neural network has been carried out. The DNN-DPD achieves 3dB more linearity improvement over sigmoid and 2dB over ReLU activation function respectively for a 55 MHz 256 quadrature amplitude signal (QAM) signal. These results support that the conventional activation function can be substituted with the Swish based activation function for power amplifiers with strong memory effects.
Paper
Full text
Swish Activation Based Deep Neural Network Predistorter for RF-PA
Semantic Scholar · Engineering · 2019
Abstract
This paper presents the linearization of radio frequency power amplifier using a deep neural network (DNN) for distortion compensation. The linearization results are shown for heterojunction bipolar transistor based RF power amplifier. The proposed DNN digital predistortion (DPD) uses Swish or Sigmoid-weighted linear unit activation function instead of sigmoid and rectified linear units (ReLU) activation function to avoid the gradient vanishing and dead neurons problem. A comparative study of different activation functions for real-valued focused time delay neural network has been carried out. The DNN-DPD achieves 3dB more linearity improvement over sigmoid and 2dB over ReLU activation function respectively for a 55 MHz 256 quadrature amplitude signal (QAM) signal. These results support that the conventional activation function can be substituted with the Swish based activation function for power amplifiers with strong memory effects.