SeqNet: Data-Driven PAPR Reduction via Sequence Classification

Abstract

To fully exploit Terahertz spectrum for the upcoming 6G communications, peak-to-average power ratio (PAPR) reduction plays an important role to enhance power efficiency and extend communication coverage. In this paper, we address the PAPR reduction with a data-driven approach and propose a PAPR reduction neural network, SeqNet, with the aid of a deep learning technique. Inspired by the well-known selected mapping (SLM) scheme, SeqNet finds a phase sequence leading to low PAPR when multiplied to a given modulated symbol block. SeqNet classifies a modulated symbol block into one of phase sequences, which are designed not to affect BER (Bit Error Rate) performance. We also propose, Split-SeqNet which splits the symbol block into multiple segments and finds a phase sequence for each. We perform comparative study with various data-driven PAPR reduction approaches and simulation result shows that SeqNet achieves better PAPR performance than conventional auto-encoder-based scheme, tone-reservation based, DFT-s-OFDM, and OFDM by about 0.2 dB, 0.4 dB, 1.6 dB, and 3.4 dB, respectively, which can be converted into 4.3%, 8.7%, 40.1% and 104.4% improvement of the communication coverage.