基于深度学习的轨道角动量光束信息提取及应用研究

As a new mode dimension, orbital angular momentum (OAM) shows widespread applications in field of communication. Because it is does not only improving the communication capacity, but also has the potential to bring new breakthroughs in the method of communication. However, there are many problems in traditional OAM optical communication, such as difficulty in mode identification, weight information extraction, mode distortion and crosstalk, which seriously restrict the further development of OAM optical communication. By using the natural advantages of deep learning in task learning and target recognition, as well as powerful data analysis and processing capabilities, the distribution and transmission characteristics of the light field of the OAM beams will be deeply studied, and the transmission distortion and mode crosstalk law will be revealed. Based on feedforward neural network, OAM mode identification and weight information extraction technology will be developed. The feasibility of convolution neural network to realize OAM beam propagation distortion and mode crosstalk real-time compensation will be also explored. The key technology of OAM mode demodulation/demultiplexing based on deep learning will be mastered, and the OAM mode demodulation/demultiplexing communication with bit error rate less than 0.01% will be realized. The results of the research will not only resolve the key technology of OAM mode demodulation/demultiplexing communication, but also help to promote the integration of AI and OAM optical communication, and provide the key technical support for the development of new OAM optical communication.

轨道角动量（OAM）不仅能够大幅提升通信容量、频谱利用率，还有望在通信方式及全光信息交换上带来新的突破，具有重要研究意义。然而，现有OAM光通信中面临的模态识别不足、权重信息提取困难、模式畸变和串扰等基础性难点问题，严重制约了OAM光通信的进一步发展。本项目将结合深度学习技术在任务学习、目标识别以及数据处理等方面的优势，深入研究OAM光场分布和传输特性，揭示OAM光束的传输畸变、模式串扰规律；发展基于前馈神经网络的OAM光束模态识别及权重信息提取技术，探索卷积神经网络实现OAM光束传输畸变、模式串扰实时补偿的可行性；掌握基于深度学习的OAM模式解调/解复用关键技术，实现误码率低于0.01%的OAM原型光通信系统。研究结果将不仅为实现OAM光束模式信息的高效解调/解复用提供新的思路与方法，更有助于促进人工智能与OAM光通信的实际融合，为发展新型OAM光通信提供关键技术支撑。

光学轨道角动量（orbital angular momentum, OAM）模式具有无穷维希尔伯特空间，对突破传统光通信频谱资源瓶颈，实现大容量光通信和全光信息处理具有重要意义。然而，现有OAM光通信中面临的模态识别不足、权重信息提取困难、模式畸变和串扰等问题，严重制约了其进一步发展。项目针对这些问题，开展了基于深度学习的OAM光束信息提取及应用研究，并完成了以下三部分研究工作：1）利用柱透镜或叉形光栅对涡旋光束进行衍射成像来提取特征参量，并通过设计前馈或卷积神经网络来建立衍射光强特征与OAM模式的映射关系，实现了拓扑荷区间为[-25, 25]的50个OAM模态检测，以及对拓扑荷区间为[-8, 8]的混合OAM模态进行权重信息的提取和识别，解决了传统OAM模态识别方案中存在的模式检测范围有限、速度低下和识别精度低等问题。2）利用深度卷积神经网络在图片信息提取方面的能力，通过构建湍流相位与光束强度分布之间的映射关系，实现了大气湍流畸变相位信息提取及反向补偿，且补偿后的OAM模式纯度可达到97%以上，解决了OAM光束在传输过程中容易受到环境干扰发生畸变、模式串扰等问题。3）利用基于深度学习的OAM模态识别和畸变补偿技术，进一步实现了OAM模式键控和复用通信，其中补偿后的复用信道串扰从-23.15 dB降低至了-29.46 dB，大幅度提高了OAM模态信道解调速率、精度和复用通信性能。这些研究工作的开展，突破了OAM光场信息处理中的模式识别范围、速度、精度低下和畸变信息提取困难等关键科学基础问题，对OAM光通信中的模式调制、复用和畸变补偿具有重要意义，有望推动OAM调制/复用通信的进一步发展。