Adaptive Inference Pathway-Gated Neural Network Model for Digital Predistortion With Varying Transmission Configurations

To meet the challenge of digital predistortion (DPD) under dynamic scenarios, a structural adaptation method of neural network (NN) based on the gate mechanism is proposed. This method integrates highway network with a noise gate, achieving discrete gating network gradient backpropagation and smooth variations in the backbone network structure. Applying this method to gated dynamic NN (GDNN), the adaptive inference pathway-gated NN (AIPGNN) model is proposed. The AIPGNN is capable of adaptively activating specific finite impulse response (FIR) filter branches based on the current configuration information. In a sense, the input signal is processed only through the activated FIR filter branches, while directly passing through the inactivated FIR filter branches. This adaptive activation method allows for the training of a specialized set of FIR branches customized to the nonlinear (NL) characteristics of a particular class of configurations, while FIR branches in GDNN are required to accommodate all configurations, which results in challenging trade-offs for the FIR layer during training. Furthermore, the AIPGNN model also supports the activation of a varying number of FIR filter branches under different transmission configurations. The adaptively changed network structure enables the proposed model to adequately correct the NL behavior of the power amplifier (PA) in more complex transmission configurations, without resource wastage in simpler transmission configurations, which meets the needs of time-varying configuration scenarios. The experimental results indicate that the AIPGNN exhibits superior dynamic linearization performance and good generalization capability under varying transmission configurations.