Investigation on the use of Artificial Neural Network Equalizer in Indoor Visible Light Communication Systems

In this paper, we investigate a non-line-of-sight visible light communication system with the artificial neural network (ANN)-based equalizer that uses the machine learning algorithm Levenberg-Marquardt (LM). We investigate the system performance in terms of the bit error rate for 2-, 4-, 8-, 16-, 32-of pulse amplitude modulation (PAM) scheme using an ANN-based equalizer with 4, 5, 10, 17, and 20 hidden neurons that are optimized. The signal to noise ratio (SNR) penalties are below 10 dB at a bit error rate of $10^{-4}$, which is below the 7% forward error correction limit of $3.8 \times 10^{-3}$. We also compare the LM algorithm over Broyden-Fletcher-Goldfarb-Shanno) quasi-newton, resilient backpropagation, and gradient descent backpropagation. LM offers the best result with a 7 dB SNR penalty at a BER of $2\times 10^{-4}$. Lastly, a 1 Mbit/s 4-PAM lin with an ANN-based equalizer with 5 hidden neurons is demonstrated over transmission distances of 1, 3, and 6 m is performed, with the lowest SNR penalty of 0.5 dB for the 1 m link.