CNN direct equalization in OFDM-VLC systems: evaluations in a numerical model based on experimental characterizations

An investigation on the orthogonal frequency division multiplexing (OFDM) equalization using deep learning architectures for a multipath single-input single-output visible light communication (VLC) channel is presented in this work. Convolution neural networks (CNN) architectures are applied in a direct OFDM mapped symbols equalization, without channel estimation, interpolation nor element-wised division, denominated CNN-Direct Equalization (CNN-DE). The performance analysis of the proposed equalizer is evaluated by considering the mean square error, bit error rate (BER), and error vector magnitude, over different signal-to-noise ratio (SNR) scenarios. Simulation results show that the proposed CNN-DE outperforms the least-square channel estimation (LS) for lower values of SNR (lower than 10 dB), which validates the CNN-DE application for noisy channels. The CNN-DE performs similarly as LS-based equalization, in terms of BER, when the LED non-linear effects and a more realistic VLC channel are taken into consideration.