Performance Study of Model-Driven Based Neural Network for VLC Channel Impairments Compensation

Inspired by the model-solving procedure of Volterra equalizer, an efficient nonlinear post equalization (NPE) is proposed to mitigate the channel nonlinearity in visible light communications (VLCs). Our insight is to employ the Volterra feature as spatial input, and then deploy the convolutional neural network to extract the ambiguity and implicit of nonlinearity feature. After that, we use the long-short term memory network for predicting the original transmitted signal from the received ones. Simulation results show that the proposed NPE can converge to the desired target loss with a comforting speed at the training phase and can provide the well-compensation for the overall nonlinearity at the testing phase. Moreover, compared with the conventional equalizer, the proposed one can achieve an excellent recovery accuracy and bit error rate performance, showing the validity for channel nonlinearity compensation in VLC system.