Nonlinearity Mitigation For Visible Light Communication Systems Based On Deep Learning

With the rapid development of various wireless applications,there is more and more demand for frequency spectrum.The scarcity of frequency resources hinders the development of the wireless network to meet the need and scenarios of the future wireless applications.Visible light communication(VLC)is a promising wireless transmission technology due to its rich spectrum resources,low implementation cost and high transmission rate.Nonlinearity widely exists in the elements.In addition,the DC-biased optical orthogonal frequency-division multiplexing(DCO-OFDM)modulation commonly adopted in VLC has the problem of high peak-to-average power ratio(PAPR),which makes it easier to cause nonlinear distortion.However,current research on nonlinearity mainly focuses on reducing the PAPR of general OFDM signals to alleviate nonlinear distortion.The nonlinearity reduction measures for VLC has not been intensively investigated.To deal with these problems,this thesis investigates the nonlinearity mitigation problem for VLC by exploiting deep learning techniques,including following two aspects.Firstly,the thesis proposes an LSTM network based linearization method for VLC system.The main idea is to design an LSTM network based pre-distortion module so that the overall response of the concatenated pre-distortion module and the nonlinear VLC transmission module exhibits a good linearity.On the one hand,the nonlinearity of VLC transmission module is compensated by the LSTM network.On the other hand,the memory effect of the VLC transmission is neutralized by adopting multiple cascading LSTM blocks.The proposed pre-distortion network implements the signal linearization functionality at the transmission side and reduces the computational complexity of the equalization at the receiver side.The pre-distortion network can be trained with only transmitter and can be applied in general VLC systems.Simulation results show that the proposed scheme can achieve better symbol error rate performance compared with the traditional non-linearity mitigation measures.Next,a nonlinearity mitigation method with end-to-end training is proposed.It involves the visible light propagation channel response in the network training process.In the end-to-end training process,effective signal pre-distortion for input signals can be trained without the need of pre-determined ideal system response as previous method.Meanwhile,the compensation for the memory and nonlinearity effects can automatically be learned.This thesis investigates the optimization of the network structure and loss function.Comparisons in terms of the LSTM-based pre-distortion method and the autoencoder based nonlinearity mitigation method show that the proposed algorithm is effective for the suppression of the nonlinearity of VLC systems under non-ideal channel conditions.Comparison in terms of network training,information feedback,time complexity,implementation difficulty and non-linearity mitigation capability,the proposed scheme is more effective and practical.It can adapt to the VLC applications with flexible lighting and communication range configurations.