深度学习辅助的宽带可见光通信动态非线性信道均衡研究

For improving the capacity of the visible light communication (VLC) systems based on intensity modulated direct detection (IM/DD), it is necessary to solve the transmission distortions introduced by memory nonlinearity and multipath interference. Deep learning (DL) provides a new approach to deal with the distortion compensations for the received signal, and DL-based nonlinear channel equalizer (NCE) becomes a research hotspot. Although the DL-based NCE has the broad prospect, its effectiveness relies on network topologies and hypothetical spaces essentially. In addition, it faces a lot of challenges such as complexity and training difficulty. Therefore, in this project, theory and key techniques for DL-based NCE which combines mathematical model solving and example-based learning will be studied. Specifically, NCE model structure optimization and model identification will be investigated in the situation of memory nonlinearity and multipath environment, which are foundation and performance guarantee for the DL-based network topologies. On the other hand, the relationship between low-dimensional structure and depth model for high-dimensional datas will be explored, and an equivalent deep neural network for the mathematical model solving will be established, which are the core technology for DL-based NCE. The final research achievements are expected to provide solid theoretical foundation and technical supports for the application of DL in the physical layer of VLC systems.

为提高以强度调制直接检测为基础架构的宽带可见光通信（VLC）系统容量，记忆非线性和多径引入的传输损伤是必须要解决的关键问题。深度学习（DL）为解决接收信号失真补偿困难开辟新的出路，形成DL非线性信道均衡这一研究热点。尽管DL非线性信道均衡前景广阔，但其有效性却本质地依赖网络拓扑和假设空间，同时又面临复杂度和训练耗时等诸多挑战。为此，课题基于模型求解与样本学习相结合的新思路，研究宽带VLC系统DL非线性信道均衡理论与关键技术：研究复杂动态非线性场景中均衡器模型结构优化新方法、完善观测干扰下模型辨识理论，它们是DL均衡器网络拓扑的构建基础；发掘高维数据的低维结构与深度模型之间的内在联系，研究模型求解问题的深度神经网络等效方法，它是DL均衡器行之有效的核心关键。本项目最终研究成果有望为DL在VLC通信物理层中的应用打下扎实的理论基础和提供关键技术储备。

本项目面向强度调制直接检测（IM/DD）架构的宽带可见光通信（VLC）动态非线性场景，通过模型求解与深度学习（DL）相结合，研究高效、绿色的宽带VLC非线性信道均衡技术。主要研究内容如下：（1）在发光二极管（LED）深度限幅场景中，提出一种基于稀疏重构的非线性削波噪声抑制方案，实现直流偏置正交频分复用（DCO-OFDM）系统的信号与削波噪声的分离，在保持发送端削波增益的同时提升接收端符号解调性能。（2）在IM/DD深度记忆高阶非线性场景中，面向Volterra均衡器复杂度高难以实用的缺陷，提出一种模型结构优化方法，可降低模型结构复杂度，减小参数计算量，并同时提供优异的均衡精度。（3）针对观测干扰下模型求解问题的不适定性和不稳定性，挖掘均衡模型内核的稀疏特性，提出具有实时实现、数值稳定和快速收敛的系数提取算法，推导出模型求解的理论均方误差（MSE），并深入分析观测干扰下的系统性能。（4）针对DL均衡网络拓扑部署的盲目性，融合VLC物理层信号处理模型到DL网络拓扑结构设计中，提出模型驱动深度学习的均衡器设计方案，克服DL网络拓扑和假设空间的选取困难，降低网络复杂度和训练耗时，并提升网络收敛速率和计算精度。以上研究成果可应用于高速数传、绿色VLC通信等。综上所述，本项目达到了预期目标，取得了丰硕的研究成果，为高效、绿色可见光通信的应用提供重要的理论依据和强有力的技术储备。