Adaptive Digital Predistortion of Wireless Power Amplifiers/Transmitters Using Dynamic Real-Valued Focused Time-Delay Line Neural Networks

Neural networks (NNs) are becoming an increasingly attractive solution for power amplifier (PA) behavioral modeling, due to their excellent approximation capability. Recently, different topologies have been proposed for linearizing PAs using neural based digital predistortion, but most of the previously reported results have been simulation based and addressed the issue of linearizing static or mildly nonlinear PA models. For the first time, a realistic and experimentally validated approach towards adaptive predistortion technique, which takes advantage of the superior dynamic modeling capability of a real-valued focused time-delay neural network (RVFTDNN) for the linearization of third-generation PAs, is proposed in this paper. A comparative study of RVFTDNN and a real-valued recurrent NN has been carried out to establish RVFTDNN as an effective, robust, and easy-to-implement baseband model, which is suitable for inverse modeling of RF PAs and wireless transmitters, to be used as an effective digital predistorter. Efforts have also been made on the selection of the most efficient training algorithm during the reverse modeling of PA, based on the selected NN. The proposed model has been validated for linearizing a mildly nonlinear class AB amplifier and a strongly nonlinear Doherty PA with wideband code-division multiple access (WCDMA) signals for single- and multiple-carrier applications. The effects of memory consideration on linearization are clearly shown in the measurement results. An adjacent channel leakage ratio correction of up to 20 dB is reported due to linearization where approximately 5-dB correction is observed due to memory effect nullification for wideband multicarrier WCDMA signals.