A Neural Network-Based Pre-Distorter for Linearization of RF Power Amplifiers

A novel approach to the nonlinear modeling of power amplifiers and digital pre-distortion is presented. Real-valued time-delay neural network architecture is used to model both the amplifier and the digital pre-distorter using a single artificial neural network. Since the accuracy of the pre-distorter and the amplifier have no direct correlation here, the hyperparameters of the pre-distorter can be optimized according to the real-world needs of the user. A 20 W GaAs amplifier driven to its saturation is used to demonstrate the effectiveness of the method. In addition, it is also shown that without additional training, the same pre-distorter has the potential to perform well for different signals.