Deep Learning for Effective Clipping Noise Mitigation

The dataset was generated through MATLAB-based simulations of a DCO-OFDM system employing 64 subcarriers and M-QAM modulation. The dataset comprised 10,000 pairs of transmitted and corresponding received OFDM symbols, used for training a deep learning (DL) model. Inspired by diffusion models, a similar data-generation strategy was adopted in which additive white Gaussian noise (AWGN) was iteratively introduced to the transmitted signals. For each transmitted symbol, 60 noisy received samples were collected, representing AWGN levels corresponding to signal-to-noise ratios (SNRs) ranging from 1 dB to 60 dB. These samples, obtained under non-adaptive indoor Li-Fi conditions with fixed DC bias, provide a diverse and representative dataset for modeling and learning the noise behavior in DCO-OFDM transmissions.