【作者】 王凯；

【导师】 张华；

【作者基本信息】 东南大学 ， 通信与信息系统， 2020， 硕士

【摘要】 近些年来随着数据业务和无线应用急剧增长,传统的频谱资源日益紧缺,可见光通信（VLC,Visible Light Communication）作为射频通信的一种重要补充,受到越来越广泛的关注。室内可见光通信可以利用泛在的发光二极管（LED,Light Emitting Diode）同时实现照明与通信,具有频段与射频不冲突且无须授权、节能环保、保密性好等优点。尽管可见光通信具有潜在的巨大带宽,LED器件的非理想特性如LED的非线性、多色混合型LED的多色串扰,严重制约着可见光通信的系统性能。本文主要在传统方案的基础上,引入神经网络结构来减轻这些器件缺陷给VLC系统性能带来的影响,给VLC收发机设计提供新的解决思路。首先,本文针对LED信道有记忆非线性缺陷使得VLC系统传输性能恶化的问题,从Volterra模型和神经网络模型两个方向研究LED预失真补偿技术。首先介绍LED信道有记忆非线性的数学模型,接着从两个方向研究预失真方案对LED信道有记忆非线性进行补偿:（1）基于Volterra模型:提出基于压缩感知的稀疏Volterra预失真模型,将压缩感知理论运用于Volterra模型的关键稀疏项提取过程,以牺牲一定误比特率性能为代价,大幅降低预失真模型复杂度。（2）基于神经网络模型:提出一种适用于可见光通信的、基于直接学习结构的浅层NN-DPD（Neural Network-Digital Predistortion）模块,以较低的系统复杂度补偿LED的记忆非线性,使预失真模型将指数等级的系统复杂度降低到了平方等级。针对Volterra和神经网络模型,均通过实验验证了其对LED有记忆非线性的补偿效果。接着,本文研究基于自编码器的端到端多色可见光通信收发机。首先阐述多色VLC系统和自编码器的基本原理,讨论了多色串扰矩阵的形成原因和数学表达。接着分别针对入射角固定和入射角可变的多色信道这两种场景,给出了基于自编码器的多色VLC系统收发机方案。使用自编码器的编码器和解码器实现传输符号与颜色空间星座点之间的映射与解映射,通过光功率控制模块使LED输出满足光功率约束与动态范围约束。此外,将多色可见光信道的先验信息运用于可变入射角收发机的训练过程中,简化可变入射角的自编码器收发机训练过程。在相同噪声功率条件下,基于自编码器的多色VLC收发机相比于传统方案具有更优的误符号率性能。最后,本文研究基于自编码器的可见光DCO-OFDM（Direct Current biased Optical-Orthogonal Frequency Division Multiplexing）抗非线性技术,通过降低立方度量值来减轻VLC非线性对DCO-OFDM系统的影响。使用编码器和解码器分别对DCO-OFDM频域子载波符号星座点进行映射与解映射,以输入输出符号的均方误差和时域OFDM符号的原始立方度量的组合作为损失函数,在保证频域输入输出符号失真在可接受的范围内前提下,大幅降低DCO-OFDM时域符号的立方度量指标。通过仿真实验验证了方案的可行性,原始立方度量的互补累计函数在10^-3位置处降低了约11.2d B,降低了DCO-OFDM系统对LED线性程度的要求。更多还原

【Abstract】 In recent years,data services and wireless applications have grown rapidly,and spectrum resources have become scarcer.Visible light communication is an important supplement to radio frequency communication,and has a broad application prospect.Indoor visible light communication can use ubiquitous LED to real-ize lighting and communication at the same time.It has several advantages such as no conflict with radio frequency bandwidth,no authorization required,energy saving,environmental protection,and good con-fidentiality.In recent years,it has become a research hotspot.Although the potential bandwidth of visible light communication may be large,the poor characteristics of LED devices,such as memory nonlinearity of LEDs and multicolor crosstalk of nonlinear multicolor LEDs,severely limit the performance of visible light communication systems.This article mainly introduces the neural network structure based on the traditional scheme to mitigate the impact of these device defects on the performance of the VLC system,and provides new solutions for the design of VLC transceivers.Firstly,from the two directions of Volterra model and neural network model,the pre-distortion compen-sation technology for memory nonlinearity of LED channel is studied.The memory nonlinearity of the LED channel and the characteristic of the limited dynamic range of the LED is explained.The specific mathemat-ical model is used to characterize the nonlinearity of the LED memory.Predistortion schemes are used to suppress the memory nonlinearity of the LED channel from two directions:（1）Based on the Volterra model,first use least squares to estimate the kernel coefficients,and then explore the feasible schemes to reduce the complexity of the predistortion model system,Based on truncation and compressed sensing theory respec-tively,a better trade-off is achieved between system performance and system complexity.（2）The structure of the predistortion model based on neural network is proposed,and the parameter training process using sliding windows is explained.The predistortion model using a neural network structure reduces the exponential level of system complexity to the square level.For a nonlinear system with a memory depth of Q,the complexity of the predistortion module is O（Q²）.Simulation experiments have been carried out for different models.The simulation experiment is carried out by first using the data collected using the actual visible light com-munication platform to model the LED,and using the trained model as the ideal model for offline simulation experiments.Through experiments,the effectiveness of the pre-distortion scheme proposed in this paper on the non-linearity of LED memory is verified.Next,a type of feed-forward neural network called auto-encoder is used to build an end-to-end multicolor visible light communication transceiver.The basic principles of the multicolor VLC system and self-encoder are explained,and then the causes and mathematical expression of the multicolor crosstalk matrix are dis-cussed by mathematically modeling the multicolor VLC channel.In addition to meeting communication requirements,multicolor VLC systems also need to meet lighting constraints and dynamic range constraints.Aiming at two cases of fixed channel of visible light and variable channel of incident angle,two multicolor VLC systems based on self-encoder were designed.The encoder and decoder of the autoencoder are used to realize the mapping between transmission symbols and color space constellation points.The LED output meets the optical power constraints and dynamic range constraints through the optical power control module.Simulation experiments for these two schemes respectively verify that under the same noise power condition,the multicolor VLC transceiver based on the autoencoder has better symbol error rate performance than the traditional scheme.Finally,the DCO-OFDM system based on self-encoder was studied,from the perspective of changing the OFDM symbol waveform to reduce the linearity requirements of the LED,and to reduce the impact of nonlinear distortion on the OFDM system due to clipping.DCO-OFDM frequency-domain subcarrier symbol constellation points are mapped and demapped using encoder and decoder respectively.The combination of the mean square error of the input and output symbols and the original cubic metric of the time domain OFDM symbol is used as the loss function.Under the premise that the output symbol distortion is within an acceptable range,the cubic metric of DCO-OFDM time-domain symbols is reduced.The simulated channel environment uses the measured frequency selected LED channel.Through experimental verification,when the complementary accumulation function of RCM is 10^-3,RCM is reduced by about 11.2d B,which can decrease the linearity requirements of LED to a certain extent.更多还原