Deep Unfolding ADMMnet for PAPR Reduction in Mixed-Numerology Systems

In this letter, a deep unfolding network is proposed for peak-to-average power ratio (PAPR) reduction in mixed-numerology systems. The guard band between adjacent subbands with distinct numerologies is reserved for the peak cancellation signals (PCS) injection and then the power loss minimization problem with PAPR constraint is formulated. The alternating-direction multiplier method (ADMM) approach is introduced and unfolded into a multilayer structure. In order to achieve fewer iterations corresponding to fewer unfolded layers, the penalty factor is set as the trainable parameter in both cross-layer and intra-layer. In addition, the cancellation ratio is designed as the soft-constraint so that it can reach a suboptimal solution when PAPR constraint is too strict to attain any feasible set. These parameters can be optimized by minimizing a labeled error loss function offline and refined by an unlabeled metric loss function online. Simulation results demonstrate that the proposed deep unfolding network achieves better PAPR performance and less power loss than the conventional ADMM approach with comparable computation complexity.