Reflectivity measurement technology of special high reflective mirrors and uncertainty analysis of measurement results

In order to meet the reflectivity measurement requirements of any incident angle at different points of the large size special high mirror, a rotating cavity ring-down spectroscopy high reflectivity measurement system was built, in which the rotation function of the resonant cavity was set, and the lifting and parallel travelling mechanism of the measured mirror was added. Furthermore, the uncertainty of the measurement results was analyzed and calculated. The results showed that the reflectivity of a high reflective mirror measured by the system was 99.979 5%, the measurement accuracy reached the order of 10-6, and the combined standard uncertainty of reflectivity measurement was 0.002 8%. Collectively, these results provide a detection guarantee for the maintenance of the large size special high mirror, and provide ideas and methods for the uncertainty analysis of measurement results of similar equipment parameters.     

基于特征融合的无监督跨模态哈希

已有的无监督跨模态哈希（UCMH）方法主要关注构造相似矩阵和约束公共表征空间的结构,忽略了2个重要问题：一是它们为不同模态的数据提取独立的表征用以检索,没有考虑不同模态之间的信息互补;二是预提取特征的结构信息不完全适用于跨模态检索任务,可能会造成一些错误信息的迁移。针对第一个问题,提出一种多模态表征融合结构,通过对不同模态的嵌入特征进行融合,从而有效地综合来自不同模态的信息,提高哈希码的表达能力,同时引入跨模态生成机制,解决检索数据模态缺失的问题;针对第二个问题,提出一种相似矩阵动态调整策略,在训练过程中用学到的模态嵌入自适应地逐步优化相似矩阵,减轻预提取特征对原始数据集的偏见,使其更适应跨模态检索,并有效避免过拟合问题。基于常用数据集Flickr25k和NUS-WIDE进行实验,结果表明,通过该方法构建的模型在Flickr25k数据集上3种哈希位长检索的平均精度均值较DGCPN模型分别提高1.43%、1.82%和1.52%,在NUS-WIDE数据集上分别提高3.72%、3.77%和1.99%,验证了所提方法的有效性。     

Intrinsically Flexible and Aging Resistant Fluorene-Based Rod-Coil Copolymer for Bendable Deep-Blue PLEDs

In the field of flexible light-emitting display, goal-oriented intelligent molecular design is used to control various behaviors of molecules, which provides potential for the development of flexible light-emitting conjugated polymers (LCPs). The introduction of non-conjugated units into polymer molecules is a key prerequisite for realizing the intrinsic flexibility, but its easy interchain slip will also lead to the formation of interchain excited states, which is detrimental to the efficiency of light-emitting diodes. Herein, two kinds of fluorene-based rod-coil copolymer with stable deep blue emission characteristics is presented and with Commission Internationale de L'Eclairage (CIE) coordinates of (0.18, 0.14) and (0.15, 0.09), respectively. Surprisingly, the copolymer films show efficient blue emission even at 100% tension. Meanwhile, the rod-coil copolymer possesses better aging resistance compared to rigid π-conjugated counterparts. Finally, both rigid and flexible light-emitting diodes based on rod-coil copolymer exhibit stable deep blue emission, and the G2-based PLED with CIE coordinates of (0.16, 0.08), which approach National Television System Committee standard blue specification. These results confirm the validity of rod-coil copolymer design strategy in constructing inherently flexible polymers with deep blue emission, which have great application potential in flexible PLEDs.     

预训练语言模型实体匹配的可解释性

实体匹配可以判断两个数据集中的记录是否指向同一现实世界实体,对于大数据集成、社交网络分析、网络语义数据管理等任务不可或缺.作为在自然语言处理、计算机视觉中取得大量成功的深度学习技术,预训练语言模型在实体识别任务上也取得了优于传统方法的效果,引起了大量研究人员的关注.然而,基于预训练语言模型的实体匹配技术效果不稳定、匹配结果不可解释,给这一技术在大数据集成中的应用带来了很大的不确定性.同时,现有的实体匹配模型解释方法主要面向机器学习方法进行模型无关的解释,在预训练语言模型上的适用性存在缺陷.因此,以Ditto、JointBERT等BERT类实体匹配模型为例,提出3种面向预训练语言模型实体匹配技术的模型解释方法来解决这个问题:(1)针对序列化操作中关系数据属性序的敏感性,对于错分样本,利用数据集元特征和属性相似度实现属性序反事实生成;(2)作为传统属性重要性衡量的补充,通过预训练语言模型注意力机制权重来衡量并可视化模型处理数据时的关联性;(3)基于序列化后的句子向量,使用k近邻搜索技术召回与错分样本相似的可解释性优良的样本,增强低置信度的预训练语言模型预测结果.在真实公开数据集上的实验结果...     

面向样本扩充的新型风格迁移网络研究

全监督语义分割网络在训练时需要耗费大量的人力与时间成本来标注样本。所以减少人工标注样本的时间,同时提升语义分割效果,对于深度学习网络的快速部署和应用推广具有重要意义。提出一种基于改进图像风格迁移网络（CycleGAN-AD）的样本扩充方法。以CycleGAN为基础,在生成器中引入注意力机制并将深度残差网络改为密集连接卷积网络。利用计算机批量产生自带标签的模拟样本,使用CycleGAN-AD网络将模拟样本风格迁移成为真实样本风格（标签不变）,并用于扩充训练样本。对石墨电极的钢印字符进行语义分割的实验结果表明,采用CycleGAN-AD网络进行样本扩充后,其分割效果得到显著提升,MIoU值最高升至0.826 0。可见,提出的样本扩充方法有希望在显著减少人工标注工作量的同时,获得高质量的训练样本。     

Fixed-Time Stabilization of a Class of Strict-Feedback Nonlinear Systems via Dynamic Gain Feedback Control

This paper presents a novel fixed-time stabilization control (FSC) method for a class of strict-feedback nonlinear systems involving unmodelled system dynamics. The key feature of the proposed method is the design of two dynamic parameters. Specifically, a set of auxiliary variables is first introduced through state transformation. These variables combine the original system states and the two introduced dynamic parameters, facilitating the closed-loop system stability analyses. Then, the two dynamic parameters are delicately designed by utilizing the Lyapunov method, ensuring that all the closed-loop system states are globally fixed-time stable. Compared with existing results, the “explosion of complexity” problem of backstepping control is avoided. Moreover, the two designed dynamic parameters are dependent on system states rather than a time-varying function, thus the proposed controller is still valid beyond the given fixed-time convergence instant. The effectiveness of the proposed method is demonstrated through two practical systems.      

基于ATLG混合模型的股票价格预测

股市是金融市场的重要组成部分,对股票价格预测有着重要的意义.同时,深度学习具有强大的数据处理能力,可以解决金融时间序列的复杂性所带来的问题.对此,本文提出一种结合自注意力机制的混合神经网络模型(ATLG).该模型由长短期记忆网络(LSTM)、门控递归单元(GRU)、自注意力机制构建而成,用于对股票价格的预测.实验结果表明:(1)与LSTM、GRU、RNN-LSTM、RNN-GRU等模型相比, ATLG模型的准确率更高;(2)引入自注意力机制使模型更能聚焦于重要时间点的股票特征信息;(3)通过对比,双层神经网络起到的效果更为明显.(4)通过MACD (moving average convergence and divergence)指标进行回测检验,获得了53%的收益,高于同期沪深300的收益.结果证明了该模型在股票价格预测中的有效性和实用性.     

融合语音、脑电和人脸表情的多模态情绪识别

本文提出了一种多模态情绪识别方法, 该方法融合语音、脑电及人脸的情绪识别结果来从多个角度综合判断人的情绪, 有效地解决了过去研究中准确率低、模型鲁棒性差的问题. 对于语音信号, 本文设计了一个轻量级全卷积神经网络, 该网络能够很好地学习语音情绪特征且在轻量级方面拥有绝对的优势. 对于脑电信号, 本文提出了一个树状LSTM模型, 可以全面学习每个阶段的情绪特征. 对于人脸信号, 本文使用GhostNet进行特征学习, 并改进了GhostNet的结构使其性能大幅提升. 此外, 我们设计了一个最优权重分布算法来搜寻各模态识别结果的可信度来进行决策级融合, 从而得到更全面、更准确的结果. 上述方法在EMO-DB与CK+数据集上分别达到了94.36%与98.27%的准确率, 且提出的融合方法在MAHNOB-HCI数据库的唤醒效价两个维度上分别得到了90.25%与89.33%的准确率. 我们的实验结果表明, 与使用单一模态以及传统的融合方式进行情绪识别相比, 本文提出的多模态情绪识别方法有效地提高了识别准确率.     

基于自注意力机制模拟实体信息的实体关系抽取

在信息抽取领域,从非结构化文本中抽取实体关系是一项基础且重要的任务,且面临实体重叠和模型误差累积等挑战.本文以关系为导向,提出一种改进的实体关系联合抽取方法.该方法将实体关系抽取任务分为关系抽取与实体抽取两个子任务.在关系抽取任务上采用自注意力机制关注词与词之间的重要程度从而模拟实体信息,并使用平均池化来表征整个句子信息;在实体抽取任务上结合关系信息使用条件随机场识别该关系下的实体对.本模型不仅能够利用存在关系必定存在实体对的思想解决实体对重叠问题,还能够在训练过程中利用数据集中已知的关系使实体抽取模块不依赖于关系抽取模块的结果来训练,从而在训练阶段避免误差累积.最后,在WebNLG和NYT公开数据集上验证了该模型的有效性.