Learnable Edge-Located Activation Neural Network for Digital Predistortion of RF Power Amplifiers

In some application scenarios, radio frequency (RF) devices face strict power consumption limits, necessitating digital predistortion (DPD) models with lower complexity. To address the needs of low-complexity scenarios, a novel DPD model called learnable edge-located activation neural network (LEANN) is developed in this article. Unlike traditional neural network (NN) models that use a uniform activation function at the nodes, the core idea of the proposed LEANN model is to enhance the flexibility and interpretability of nonlinear modeling by employing learnable univariate functions as activation functions on the edges of the network. Furthermore, given the varying nonlinear characteristics of different signal components, a logarithmic regularization pruning method suitable for the LEANN model is also proposed. This method promotes a greater sparsity in the model by reducing the similarity between activation functions. Experimental results demonstrate that the proposed LEANN model achieves a lower complexity and higher performance compared to several classic linear parameter models and NN models in linearizing power amplifiers (PAs). Furthermore, the pruned LEANN (PLEANN) model further reduces the complexity without significantly decreasing the performance.