Mask-Based PAPR Reduction Scheme With Deep Learning for OFDM Systems

High peak-to-average ratio (PAPR) is a major drawback in orthogonal frequency division multiplexing (OFDM) systems. Despite that non-linear transformation techniques achieve low PAPR, it meanwhile degrades the bit error rate (BER) performance. To tackle this issue, deep learning based tightly coupled architectures have been proposed, but their significant computational overheads make them impractical for real-world deployments. This paper explores the ability of separated architecture to efficiently reduce PAPR while maintaining a good BER performance. Specifically, we analyze the distribution of OFDM signals to find an interesting phenomenon that the transmitter can achieve a large PAPR reduction (about 2.4 dB) as long as masking a small percentage (0.55%) of high-PAPR symbols. Driven by this finding, we propose a mask-based scheme including a PAPR reduction module at the transmitter and a mask-location assisted signal reconstruction module at the receiver. To capture the mask-location information (MLI) at the receiver, we outline two ways tailored for scenarios with different communication quality and computing capabilities, respectively. Simulation results reveal that our proposed scheme can achieve satisfactory BER performance with substantially reduced PAPR against conventional schemes. Moreover, the robustness of our scheme is demonstrated across various channels, including an additional white Gaussian noise channel, the Extended Pedestrian A channel and the Extended Vehicular A channel.