Neural network based frequency adaptive digital predistortion of RF power amplifiers

Abstract

Linearization of power amplifiers (PAs) is a big challenge in high-dimensional radio frequency (RF) designs, and to tackle this drawback we propose an adaptive strategy with the combination of neural networks (NNs) and band-pass filters for input signals with different frequencies that results in reduced computational costs. The proposed linearization approach is based on utilization of NN for modeling the PA and band-pass filters for contributing to frequency adaptability without feedback loop. Thus, even if the frequency of the input signal changes, the system may still produce linear output. The proposed model consists of sub-digital predistortion (DPD) blocks where each sub-DPD block generates DPD coefficients only for the specified frequency range. Thanks to sub-DPD blocks without feedback, the computational load of the model is reduced and computation time is saved. To validate the proposed model, the PA is first characterized using the neural network. Then, the frequency of the input signal is determined via band-pass filtering. Based on this frequency information, the corresponding NN-based sub-DPD block is activated to linearize the PA’s nonlinear behavior. For the presented PA that is operating from 1.7 GHz to 2 GHz, four different input signal frequencies values as 1.7 GHz, 1.9 GHz, 2.1 GHz, 2.4 GHz respectively are carried out. The achieved results prove that the proposed model provides improved PA modeling and nonlinear compensation compared to the other methods. The 1-dB compression point of the PA is measured as–6.88 dBm without DPD, 4.49 dBm with look-up table-based DPD, and 7 dBm with NN-based DPD.