Neural Network-Based Linear Companding Scheme for PAPR Reduction in OFDM Systems

The high Peak-To-Average Power Ratio (PAPR) of Orthogonal Frequency Division Multiplexing (OFDM) systems severely limits signal performance. This paper proposes a Neural Network-Driven Linear Companding (NNLC) scheme that generates parameterized companding functions through neural networks, effectively balancing the trade-off between PAPR suppression and signal distortion. First, we construct a linear companding architecture: maintaining linear transmission in low-amplitude regions to reduce quantization noise, implementing dynamic compression via adjustable slopes in medium-high amplitude regions. Second, we design a feature-driven parameter generation network that extracts real-time signal statistics and maps them to optimal segment slope parameters through a fully connected network. We introduce time-frequency domain joint regularization constraints to ensure the companding function satisfies strict monotonicity and hardware implementability. Compared with state-of-the-art traditional schemes like Amplitude Limiting Companding(ALC) and segmented linear approaches using traditional optimizing methods, proposed solution demonstrates significant PAPR suppression gains while maintaining lower signal distortion. The scheme can be implemented on Field-Programmable Gate Arrayss (FPGAs) via Look-Up Table (LUT) mechanisms with lower computational complexity than conventional methods, providing an efficient dynamic PAPR control solution for 5G/6G systems.