End-to-End Joint Optimization for PAPR Reduction and Digital Predistortion Based on Neural Network

The combination of crest factor reduction (CFR) and digital predistortion (DPD) can mitigate the average efficiency reduction of power amplifiers (PAs) due to high peak-to-average power ratio (PAPR) signals. A common CFR method is time-domain (TD) clipping, which causes irreversible signal impairment. To this end, an end-to-end (E2E) joint optimization method based on neural networks (NNs) is proposed in this letter. The E2E architecture consists of a transmitter network, a DPD model, and a PA model, enabling integrated processing of signal transmitted, transmission, and reception. The proposed method uses multiobjective joint optimization to reduce the PAPR of the TD signal through constellation point geometric shaping (GS) in the frequency domain, while simultaneously training the DPD model. While considering the interaction between PAPR reduction and DPD techniques, this approach can reduce PAPR without signal impairment and can allow them to work together to achieve high-quality signal transmission.