Dual-Block Learning: A Frequency-Based Method for Accuracy Optimization in Wideband PA Linearization

In RF communication systems, using neural network to solve the nonlinear distortion of Power Amplifier(PA) has been important and widely researched in recent years. As wideband PA brings more severe distortion to the high-frequency signal, the general neural network training method cannot achieve better performance than low-frequency situation. According to the characteristics of nonlinear regression, we propose an optimization of neural network training method for PA and DPD, which is called dual-block learning, under the mean-square error criteria. By global batch stochastic gradient descent algorithm, we make the low-frequency block(LFB) learns robust features of PA, which will smooth the output distribution. Then we use residual connection and the smoothed signal to train an added high-frequency block(HFB) to capture the high-frequency features of PA. We take PA modeling simulations on various kinds of neural networks and the normal mean-square error get improved about 3 dB generally. We also add a pre-training trick to Digital PreDistortion(DPD) system. The result shows the Adjacent Channel Power Ratio(ACPR) also outperforms the general trained models. Besides training method, we also find that the appropriate memory setting for wideband will make sense for ACPR improvement.