An artificial neural network for wideband pre-distortion of efficient pico-cell power amplifiers

This paper presents a Two Layer Artificial Neural Network (2LANN) model for linearization of pico-cell power amplifiers which output approximately 2W or less. Towards the goal of realizing efficient real-time pre-distortion hardware, optimization of the 2LANN model is done based on hardware considerations. The parameters of our proposed 2LANN model are obtained by measuring the wideband inputs and outputs of a Device Under Test (DUT). The said inputs and outputs are then used to train the 2LANN to exhibit system inversion of the DUT. The trained 2LANN is verified by observing its ability to linearize a 2W Class AB power amplifier with a 4-carrier WCDMA signal as its stimulus. The performance metrics for linearity are the dynamic AM-AM and AM-PM characteristics, Adjacent Channel Power Ratio (ACPR), and Error Vector Magnitude (EVM). The measured ACPR improvements due to the proposed pre-distorter are 15dB and 12dB at frequency offsets of 5MHz and 10MHz respectively. The linearized power amplifier also yields a measured EVM improvement of 2%. A comparison of our proposed model with previously published pre-distortion schemes shows the excellent linearization capability of our 2LANN.