基于深度体素卷积神经网络的三维模型识别分类

提出一种基于深度体素卷积神经网络的三维(3D)模型识别分类算法,该算法使用体素化技术将3D多边形网格模型转化为体素矩阵,并通过深度体素卷积神经网络提取该矩阵的深层特征,以增强特征的表达能力和差异性。在ModelNet40数据集上的实验结果表明:所提算法对3D网格模型识别分类的准确率能够达到87%左右。所构建的深度体素卷积神经网络能够有效地增强3D模型的特征提取和表达能力,提高对大规模复杂3D网格模型分类识别的准确率,所提方法优于当前的主流方法。     

基于可见光谱和卷积神经网络的贝类识别方法

目前卷积神经网络（CNN）在物体种类识别方面取得突破性进展。贝类作为农业经济的重要组成部分,种类繁多,特点复杂,大多贝类存在着相似度高,各类样本分布不均衡情况,以致CNN对贝类分类的准确率偏低。针对这一情况,提出了基于可见光谱和CNN的贝类识别方法,旨在提取更有效的贝类特征,从而提高贝类分类的准确率。首先,提出了一种包含输出熵度量和正交性度量的滤波器信息度量与特征选择方法,重新初始化修剪掉的滤波器并使其正交,捕获网络激活空间中的不同方向,使神经网络模型学习到更多有用的贝类特征信息,提升模型分类准确率;其次,提出了一种包含正则化项和焦点损失项的贝类分类目标函数,通过控制各类别样本对总损失的共享权重,来减少易分类样本的权重,以使模型注意力向预测不准的样本倾斜,均衡样本分布和样本分类难度,进一步提高贝类分类的准确率。贝类图像数据集由74类贝类组成,共11 803张图像。获取原始数据集后,对数据集图像进行水平翻转、垂直翻转、随机旋转、在[0, 30°]范围内旋转、在[0, 20%]范围内缩放和移动等数据增强操作,将图像数量从11 803张增加到119 964张。整个图像数据集按8∶1∶1的比例随机分为训练集95 947张图片、验证集11 996张图片和测试集12 021张图片。在建立贝类图像数据集的基础上进行了实验验证,达到了93.38%的分类准确率,将基准网络（Resnest）的准确率提高了1.18%,相较网络SN_Net和MutualNet,准确率分别提升了4.34%和0.85% ,并且训练时长为22 320 s,将基准网络（Resnest）的训练时长缩短了960 s,训练时长分别比SN_Net和MutualNet短3 180和2 460 s。实验结果证明了该方法的有效性。     

随机静态存储器低能中子单粒子翻转效应

 建立了中子单粒子翻转可视化分析方法,对不同特征尺寸（0.13~1.50 μm）CMOS工艺商用随机静态存储器(SRAM)器件开展了反应堆中子单粒子翻转效应的实验研究,获得了SRAM器件的裂变谱中子单粒子翻转截面随特征尺寸变化的变化趋势。研究结果表明:SRAM器件的特征尺寸越小,其对低能中子导致的单粒子翻转的敏感性越高。     

基于自归一化神经网络的脉冲星候选体选择

脉冲星候选体选择是脉冲星搜寻任务中的重要步骤.为了提高脉冲星候选体选择的准确率,提出了一种基于自归一化神经网络的候选体选择方法.该方法采用自归一化神经网络、遗传算法、合成少数类过采样这三种技术提升对脉冲星候选体的筛选能力.利用自归一化神经网络的自归一化性质克服了深层神经网络训练中梯度消失和爆炸的问题,大大加快了训练速度.为了消除样本数据的冗余性,利用遗传算法对脉冲星候选体的样本特征进行选择,得到了最优特征子集.针对数据中真实脉冲星样本数极少带来的严重类不平衡性,采用合成少数类过采样技术生成脉冲星候选体样本,降低了类不平衡率.以分类精度为评价指标,在3个脉冲星候选体数据集上的实验结果表明,本文提出的方法能有效提升脉冲星候选体选择的性能.     

深亚微米SRAM质子单粒子翻转实验研究

宇航半导体器件运行在一个复杂的空间辐射环境中,质子是空间辐射环境中粒子的重要组成部分,因而质子在半导体器件中导致的辐射效应一直受到国内外的关注。利用兰州重离子加速器(Heavy Ion Research Facility In Lanzhou) 加速出的H2 分子打靶产生能量为10 MeV 的质子,研究了特征尺寸为0.5/0.35/0.15 μm体硅和绝缘体上硅(SOI) 工艺静态随机存储器(SRAM) 的质子单粒子翻转敏感性,这也是首次在该装置上开展的质子单粒子翻转实验研究。实验结果表明特征尺寸为亚微米的SOI 工艺SRAM器件对质子单粒子翻转不敏感,但随着器件特征尺寸的减小和工作电压的降低,SOI 工艺SRAM器件对质子单粒子翻转越来越敏感;特征尺寸为深亚微米的体硅工艺SRAM器件单粒子翻转截面随入射质子能量变化明显,存在发生翻转的质子能量阈值,CREME-MC模拟结果表明质子在深亚微米的体硅工艺SRAM器件中通过质子核反应导致单粒子翻转。Microelectronic devices are used in a harsh radiation environment for space missions. Among all the reliability issues concerned, proton induced single event upset (SEU) is becoming more and more noticeable for semiconductor components exposed on space. In this work, an experimental research of SEU induced by 10 MeV proton for static random access memory (SRAM) of 0.5, 0.35 and 0.15 m feature size is carried out on HeavyIon Research Facility in Lanzhou for the rst time. The experimental results show that proton induced SEUs in submicron and deep-submicron (SRAMs) are dominated by secondary ions generated by proton nuclear reaction events. The silicon-on-insulator SRAMs characters natural radiation-hardened SEU by proton. For the deep-submicron bulk-silicon technology SRAM, the proton SEU cross section is closely related to the proton energy and there is a threshold energy for the SEU occurrence by proton indirect ionization. CREME-MC simulation indicates that the SEU events in deep-submicron SRAM are induced by the proton nuclear reaction.     

八度卷积和双向门控循环单元结合的X光安检图像分类

针对主动视觉安检方法准确率低、速度慢,不适用于实时交通安检的问题,提出了八度卷积(OctConv)和注意力机制双向门控循环单元(GRU)神经网络相结合的X光安检图像分类方法。首先,利用八度卷积代替传统卷积,对输入的特征向量进行高低分频,并降低低频特征的分辨率,在有效提取X光安检图像特征的同时,减少了空间冗余。其次,通过注意力机制双向GRU,动态学习调整特征权重,提高危险品分类准确率。最后,在通用SIXRay数据集上的实验表明,对8 000幅测试样本的整体分类准确率(ACC)、特征曲线下方面积(AUC)、正类分类准确率(PRE)分别为98.73%、91.39%、85.44%,检测时间为36.80 s。相对于目前主流模型,本文方法有效提高了X光安检图像危险品分类的准确率和速度。     

基于SPOC的翻转课堂教学模式在大学物理教学中的应用

为了提高大学物理课程的教学效果,探索翻转课堂教学模式在提高教学质量方面的作用,我们利用爱课程网络平台和山东大学大学物理MOOCs课程资源,在本科生大学物理课程教学中实施了基于SPOC的翻转课堂教学模式.本文从教学设计,课程安排,考核方式等方面详细论述了翻转课堂教学模式改革的基本思路及方法.该教学模式对培养学生对大学物理课程的学习兴趣,调动学生自主学习的主观能动性及提升学生分析问题解决问题能力等方面具有积极意义,所得经验为在高校实施翻转课堂改革提供一定的借鉴.     

一种区分性互补系统构造与融合的语音识别方法

在区分性训练的框架下,提出了一种基于混淆信息加权的互补系统构造方法。首先通过统计音素对的混淆信息,利用混淆信息给音素对加以不同的惩罚权重,分别以基线系统中的3个最优识别结果作为参考,计算混淆信息加权后的音素准确率,同时以正确的标注为参考计算标准的音素准确率。然后通过同时最大化混淆信息加权后的音素准确率和最小化标准音素准确率,构建模型层互补系统,并进一步通过结合RDLT (region-dependent linear transform)特征变换过程构造特征层的互补系统。实验结果表明,与互补最小音素错误准则相比,融合模型层互补系统后识别率提高了0.76%,同时融合特征层和模型层的互补系统识别率提高了1.35%。本方法可以增大互补系统间的差异性,提高系统融合后的识别性能。      

稀疏降噪自编码结合高斯过程的近红外光谱药品鉴别方法

提出一种稀疏降噪自编码结合高斯过程的近红外光谱药品鉴别方法。首先对近红外光谱数据进行小波变换以消除基线漂移,然后用稀疏降噪自编码(SDAE)网络提取光谱特征并降维表示,最后采用高斯过程(GP)进行二分类,其中GP选用光谱混合(SM)核函数作为协方差函数,记此分类网络为wSDAG_SM。自编码网络具有很强的模型表示能力,高斯过程分类器在处理小样本数据时具有优势。wSDAG_SM网络通过稀疏降噪自编码学习得到维数更低但更有价值的特征来表示输入数据,同时将具有很好表达力的光谱混合核作为高斯过程的协方差函数,有利于更准确的光谱数据分类。以琥乙红霉素及其他药品的近红外光谱为实验数据,将该方法与经过墨西哥帽小波变换的BP神经网络(wBP)、支持向量机(wSVM), SDAE结合Logistic二分类(wSDAL)、SDAE结合采用平方指数(SE)协方差核的GP二分类(wSDAG_SE),以及未采用小波变换的SDAG_SM网络等方法进行对比。实验结果表明,对光谱数据进行墨西哥帽小波变换预处理能有效提升SDAG_SM网络的分类准确率和稳定性。wSDAG_SM方法无论从分类准确率还是分类结果稳定性方面,都优于其他分类器。     

基于中国散裂中子源的商用静态随机存取存储器中子单粒子效应实验研究

利用中国散裂中子源反角白光中子束线开展13款商用静态随机存取存储器的中子单粒子效应实验.研究了测试图形、特征尺寸和版图工艺差异对单粒子效应的影响.结果表明测试图形对器件的单粒子翻转截面影响不大,但对部分器件的多单元翻转占比有较大的影响;特征尺寸对器件单粒子翻转截面的影响没有明显的规律,但对多单元翻转的影响规律明显,多单元翻转占比和最大位数都随着特征尺寸的降低而增大;器件版图工艺差异对器件的单粒子翻转截面和多单元翻转占比都有较大的影响.此外,通过与高原辐照实验结果对比,发现在反角白光中子源获得的多单元翻转占比小于高原辐照实验的结果,其原因是反角白光中子源实验中,中子的最高能量和高能成分占比偏小,且中子束流只有垂直入射.因此,利用反角白光中子源评估器件的大气中子单粒子效应时可能会低估多单元翻转情况.本文的结果可为研究者利用反角白光中子源开展相关研究提供参考.     

Point-wise measurements of MRS volume selection performance are insensitive to magnetic susceptibility effects of phantom materials

The purpose was to analyse magnetic susceptibility effects on accuracy of point-wise measurements of signal profiles in the assessment of MRS volume selection performance. An existing phantom design consisting of a sphere with a movable signal source was used for the investigation. The influence from the phantom on magnetic field homogeneity was measured with phase sensitive 1H imaging and 31P spectroscopy on a 1.5 T whole body MR system. The susceptibility effects for such a phantom design can be separated in 1/ A variation in the background magnetic field, which is caused by the stationary structures and has a significant influence on spatial accuracy. 2/ A magnetic field distortion, which is caused by the movable signal source and has very little influence on accuracy. The spatial inaccuracy due to susceptibility effects in this phantom, was 0.03 mm for positions of the signal source covering a 40-mm VOI. Susceptibility effects from the movable signal source were substantial but had very little influence on spatial accuracy. Still, improvements of this phantom design are possible. Point-wise measurements using a phantom with a movable signal source is inherently insensitive to susceptibility effects from the signal source and permits accurate signal profile measurements of high spatial (sub-mm) resolution.     

单点金刚石超精密车削快点火靶丸微孔精度控制北大核心CSCD

研究了单点金刚石超精密车削技术(SPDT)加工靶丸微孔中的精度控制方法,建立了靶丸微孔加工误差的仿真模型,并理论分析了不同误差因素对微孔尺寸误差的影响规律;根据误差分析结果提出了基于刀具阶梯进给运动方式的微孔精度控制方法,用以控制靶丸微孔精度;在单点金刚石超精密车床上进行了辉光放电聚合物(GDP)靶丸微孔的车削实验,实验结果表明:采用该精度控制方法,靶丸微孔尺寸误差和圆度误差分别降低了70.7%和87.5%,实验结果表明了所提出方法的有效性。     

Coils for ingestible capsules: Near-field magnetic induction link

Monolayer- and compact-multilayer-stacked ingestible TX coils are investigated for ingestible capsule systems. The inductive link through the human body is modeled. The efficiency of the near-field magnetic induction link budget is evaluated in the air, in the homogeneous human body and in the three-layer human body. The variations of the position and the orientation of the TX capsule coil are taken into account to evaluate the coupling response between the TX ingestible and RX on-body coils.     

康普顿照相中的背景噪声分析北大核心CSCD

康普顿照相是用于压缩靶丸成像、诊断内爆压缩对称性的一种有效技术手段。如何屏蔽成像过程中的背景噪声是康普顿照相技术的重点问题,同时也是一个难点问题。分析了靶环境以及各种背景信号的来源,并提出了相应的屏蔽建议。通过分析可以看到,在中子产额小于1013sr-1、压缩靶丸中心温度小于5 ke V时,通过使用相应的屏蔽措施并采用合适的诊断措施和记录设备,可以把背景噪声降至109sr-1以下,使康普顿照相中的面密度测量精度达到5%左右。     

再入飞行器等离子体预测与气体组分相关性

对一类大钝头再入飞行器,理论分析电子密度预测与气体组分的相关性,使用经验证的数值模拟方法研究化学反应模型对等离子体预测的影响.研究发现:马赫数是影响等离子体预测与气体组分相关性的重要参数,马赫数越高,不同气体组分模型所得电子密度差异越大.气体组分模型对等离子体预测的影响在驻点强压缩区域和身部位置基本一致.对于大钝头再入飞行器,高度H=60 km,马赫数大于23时应该采用11组分化学反应模型.     

不同部位矿质元素与红外光谱数据融合对美味牛肝菌产地溯源研究

野生食用菌产地溯源研究中,采用单一有机成分或矿质元素指纹存在一定局限性。利用不同指纹分析技术的互补性与协同性,将不同部位与类型的化学信息进行融合,探讨此方法对野生食用菌产地溯源的可行性,以期为野生食用菌溯源提供新的思路与科学依据。通过测定云南7个产地、124个美味牛肝菌(菌柄、菌盖)中15种矿质元素的含量,以及子实体傅里叶变换红外光谱(FTIR)。标准正态变换(SNV)、二阶导数(2D)等算法对原始光谱进行预处理。基于低级与中级数据融合策略,将预处理后的FTIR光谱与菌柄、菌盖矿质元素数据进行融合,结合支持向量机(SVM)分别建立菌柄、菌盖、FTIR、低级数据融合(菌柄+菌盖,菌柄+菌盖+FTIR)与中级数据融合(菌柄+菌盖+FTIR)判别模型;分析比较模型参数,确定快速甄别美味牛肝菌产地的可靠方法。结果显示：(1)菌盖中Cd,Cr,Cu,Li,Mg,Na,P和Zn元素平均含量高于菌柄,Ba,Ca,Co,Ni,Rb,Sr和V元素在菌柄中平均含量高于菌盖。美味牛肝菌中人体必需矿质元素Ca,Cu,Mg,P和Zn平均含量远高于小麦、水稻干品和新鲜蔬菜,与动物干制品含量相似;(2)FTIR光谱数据最佳预处理方法为3D+SNV,其Q2和R2Y分别为76.64%,88.91%;(3)菌柄、菌盖、FTIR、低级数据融合与中级数据融合SVM模型,c值分别为8 192,4 096,1.414 2,11.313 7,1和0.7071 1,菌柄和菌盖模型c值较大,表明采用单一菌柄或菌盖矿质元素含量数据,SVM训练存在过拟合风险,判别效果较差;(4)FTIR、低级数据融合和中级数据融合SVM模型,样品分类错误总数分别为7,9,7和0,中级数据融合(菌柄+菌盖+FTIR)模型样品分类正确率最高。表明基于中级融合策略将不同部位矿物元素和子实体FTIR光谱数据融合,可作为野生食用菌产地溯源的一种有效方法。     

An improved penalty immersed boundary method for fluid–flexible body interaction

An improved penalty immersed boundary (pIB) method has been proposed for simulation of fluid–flexible body interaction problems. In the proposed method, the fluid motion is defined on the Eulerian domain, while the solid motion is described by the Lagrangian variables. To account for the interaction, the flexible body is assumed to be composed of two parts: massive material points and massless material points, which are assumed to be linked closely by a stiff spring with damping. The massive material points are subjected to the elastic force of solid deformation but do not interact with the fluid directly, while the massless material points interact with the fluid by moving with the local fluid velocity. The flow solver and the solid solver are coupled in this framework and are developed separately by different methods. The fractional step method is adopted to solve the incompressible fluid motion on a staggered Cartesian grid, while the finite element method is developed to simulate the solid motion using an unstructured triangular mesh. The interaction force is just the restoring force of the stiff spring with damping, and is spread from the Lagrangian coordinates to the Eulerian grids by a smoothed approximation of the Dirac delta function. In the numerical simulations, we first validate the solid solver by using a vibrating circular ring in vacuum, and a second-order spatial accuracy is observed. Then both two- and three-dimensional simulations of fluid–flexible body interaction are carried out, including a circular disk in a linear shear flow, an elastic circular disk moving through a constricted channel, a spherical capsule in a linear shear flow, and a windsock in a uniform flow. The spatial accuracy is shown to be between first-order and second-order for both the fluid velocities and the solid positions. Comparisons between the numerical results and the theoretical solutions are also presented.     

Compensation for time-dependent radiation-drive asymmetries in inertial-fusion capsules

An approach is presented to design inertial-fusion capsules compensated for time-dependent radiation-drive asymmetries. This approach uses in depth variable doping of the capsule ablator, i.e., the addition of small amounts of material to tailor the opacity. Simulations show that an inertial-fusion capsule, using a beryllium ablator variably doped with gold, can be designed to compensate for a constant P(2) radiation asymmetry as high as 20% and still produce nominal yield (80% of a symmetrically driven capsule). In contrast, without variable doping the P(2) asymmetry must be less than 2% to obtain nominal yield. Similarly encouraging results are obtained for modes P(1), P(4), and P(6). Simulations also demonstrate that variable doping can compensate for nearly arbitrary time-dependent radiation-drive asymmetries by varying the polar dependence of the doping fraction with depth.     

毛头鬼伞的红外光谱研究

利用傅里叶变换红外光谱法(FTIR)研究了人工栽培与野生的毛头鬼伞,及其液化前后的菌盖样品。毛头鬼伞的红外光谱主要由蛋白质、多糖的特征吸收峰组成。人工菌与野生菌的光谱差异表现为蛋白质和多糖的特征峰强度比,野生菌的多糖特征吸收峰要比人工菌的吸收峰明显。菌盖液化黑色物与液化前的菌盖光谱相比,多糖特征峰明显减弱,表明菌褶液化现象主要是多糖的变化。红外光谱结果为毛头鬼伞的进一步的开发研究提供了参考信息。     

An Improvement to Conformer-Based Model for High-Accuracy Speech Feature Extraction and Learning

Owing to the loss of effective information and incomplete feature extraction caused by the convolution and pooling operations in a convolution subsampling network, the accuracy and speed of current speech processing architectures based on the conformer model are influenced because the shallow features of speech signals are not completely extracted. To solve these problems, in this study, we researched a method that used a capsule network to improve the accuracy of feature extraction in a conformer-based model, and then, we proposed a new end-to-end model architecture for speech recognition. First, to improve the accuracy of speech feature extraction, a capsule network with a dynamic routing mechanism was introduced into the conformer model; thus, the structural information in speech was preserved, and it was input to the conformer blocks via sequestered vectors; the learning ability of the conformed-based model was significantly enhanced using dynamic weight updating. Second, a residual network was added to the capsule blocks, thus, the mapping ability of our model was improved and the training difficulty was reduced. Furthermore, the bi-transformer model was adopted in the decoding network to promote the consistency of the hypotheses in different directions through bidirectional modeling. Finally, the effectiveness and robustness of the proposed model were verified against different types of recognition models by performing multiple sets of experiments. The experimental results demonstrated that our speech recognition model achieved a lower word error rate without a language model because of the higher accuracy of speech feature extraction and learning using our model architecture with a capsule network. Furthermore, our model architecture benefited from the advantage of the capsule network and the conformer encoder, and also has potential for other speech-related applications.