Unlike conventional topological edge states confined at a domain wall between two topologically distinct media, the recently proposed large-area topological waveguide states in three-layer heterostructures, which consist of a domain featuring Dirac points sandwiched between two domains of different topologies, have introduced the mode width degree of freedom for more flexible manipulation of electromagnetic waves. Until now, the experimental realizations of photonic large-area topological waveguide states have been exclusively based on quantum Hall and quantum valley-Hall systems. We propose a new way to create large-area topological waveguide states based on the photonic quantum spin-Hall system and observe their unique feature of pseudo-spin-momentum-locking unidirectional propagation for the first time in experiments. Moreover, due to the new effect provided by the mode width degree of freedom, the propagation of these large-area quantum spin-Hall waveguide states exhibits unusually strong robustness against defects, e.g., large voids with size reaching several unit cells, which has not been reported previously. Finally, practical applications, such as topological channel intersection and topological energy concentrator, are further demonstrated based on these novel states. Our work not only completes the last member of such states in the photonic quantum Hall, quantum valley-Hall, and quantum spin-Hall family, but also provides further opportunities for high-capacity energy transport with tunable mode width and exceptional robustness in integrated photonic devices and on-chip communications.

针对传统人脸识别方法识别性能较差，基于深度学习的方法在非限制条件下识别较为困难，人脸特征区分性弱，识别精度容易受到姿势、表情等方面影响的问题，提出了一种引入卷积块注意力模块的孪生神经网络模型结构。该结构是基于孪生神经网络（Siamese neural network）的基础框架进行改进的，在框架中引入改进的VGG11_BN模型进行特征提取。该模型是在VGG11结构的基础上引入批归一化（Batch Normalization，BN）技术，在原模型结构的基础上，提出引入CBAM混合注意力机制的特征提取网络；最后，针对目前亚洲人的人脸识别训练较少的情况，采用更加符合亚洲人脸特征的CASIA-FaceV5数据集进行识别训练。实验结果表明：本文算法在人脸识别方面的准确率达到了96.67%，并且在CAS-PEAL-R1人脸数据集上比SRGES，VGG11+siamese算法的准确率分别提升6.05%，6.7%。该算法可以在多因素影响下更好地进行人脸识别验证，具有良好的稳定性。

熔盐堆作为第四代先进核能系统之一，在安全性、经济性、防核扩散和可持续性等方面具有独特的优势。为了保障熔盐堆运行安全，需要快速、准确地识别瞬态工况，目前的瞬态识别方法主要依赖于操作员人工识别，这会引入较大的人为因素，严重影响核电安全。为了减少熔盐堆系统瞬态识别过程中引入的人为因素，提高熔盐堆运行安全，使用RELAP5-TMSR程序对美国橡树岭国家实验室建造运行的熔盐实验堆（Molten Salt Reactor Experiment，MSRE）的瞬态工况进行建模与仿真，产生数据集，基于K近邻（K-Nearest Neighbor，KNN）机器学习方法，建立了熔盐堆系统瞬态识别模型，并对识别模型在噪声下的鲁棒性进行了分析和优化。结果显示：基于KNN方法建立的熔盐堆系统瞬态识别模型在测试集上的F1分数达到99.99%；在噪声下的识别F1分数达到94.32%，具有较高的鲁棒性；进一步优化后的熔盐堆系统瞬态识别模型在噪声下的F1分数达到99.73%，能较为准确地识别MSRE的瞬态工况，满足熔盐堆系统瞬态识别需求。基于KNN方法的熔盐堆系统瞬态识别模型能够有效识别系统瞬态工况，可应用于熔盐堆智能运维，确保熔盐堆运行安全。