基于随机重构图结构的图神经网络分类算法

目的:为了解决在特征聚合过程中的确定性传播所导致的节点相似性破坏和节点对邻域依赖性高的问题,构建基于随机重构图结构的图神经网络分类算法。方法:首先,随机特征变换根据学习的权重值对随机保留的部分节点特征进行增强,生成随机特征。然后,利用生成的特征计算融合系数,对原始图和k近邻图进行自适应融合,重构出随机图结构。最后,将提取的多支浅层特征加入到重构图结构的卷积层,使模型随着层数的加深可得到浅层信息的补充。此外,对联合优化分类损失和自监督学习损失,保持节点相似性和平滑性。结果:与其他节点分类方法在Cora、Citeseer和Pubmed数据集上进行半监督实验和全监督实验结果对比,本文的算法精度提高了0.9%～2.3%。结论:基于随机重构图结构的网络分类算法在节点分类任务中取得较好的性能。     

矩阵输入的多层前向神经网络学习算法

单隐层前向神经网络的学习能力是有限的.特别地,作为分类器,单隐层前向神经网络对于图像的复杂信息和不同图像之间的细节信息很难学习和处理.文章借鉴深度神经网络的思想,将单隐层矩阵输入的神经网络拓展到多隐层神经网络,并采用传统的反向传播算法对其训练并给出学习算法.通过多个数据库的实验对比,结果显示所提出的算法具有良好的效果.     

密集连接神经网络在远距离水声目标探测中的性能分析

由于水声目标辐射噪声的低信噪比特性,探测远距离水声目标具有一定挑战。为提升远距离水声目标探测的准确率,文章提出一种基于密集连接神经网络和自注意力机制的方法。该方法提取信号的梅尔倒谱系数作为特征,在密集连接神经网络头部添加自注意力模块以捕获关键信息,经过多个密集块后输出探测结果。在实测数据集上进行实验,分析了自注意力机制添加与否、输入特征不同、接收端深度不同时模型的性能变化。应用在未来几天的数据测试模型的任务中,探测范围在小于10 km时,探测准确率为93.3%,探测范围扩大至20 km时,探测准确率为90.34%。实验结果表明,模型在信噪比不小于-6 dB时实现了水声目标探测,在增加更多的低信噪比样本后,仍具有一定探测能力,且其性能优于其他模型。此外,训练集包含多种信噪比条件下的数据时,探测性能会有进一步提升。     

基于深度迁移学习的微型细粒度图像分类

现有的细粒度分类模型不仅利用图像的类别标签,还使用大量人工标注的额外信息。为解决该问题,本文提出一种深度迁移学习模型,将大规模有标签细粒度数据集上学习到的图像特征有效地迁移至微型细粒度数据集中。首先,通过衔接域定量计算域间任务的关联度。然后,根据关联度选择适合目标域的迁移特征。最后,使用细粒度数据集视图类标签进行辅助学习,通过联合学习所有属性来获取更多的特征表示。实验表明,本文方法不仅可以获得较高精度,而且能够有效减少模型训练时间,同时也验证了进行域间特征迁移可以加速网络学习与优化这一结论。     

图模型核学习模式识别算法

本文提出一类核学习结构的算法,该算法是基于一种称为核协方差的度量方法。此度量方法允许在由特征空间中处理映射的高斯变量。并且能学习离散和连续变量的模型。本文还用离散和连续变量测试了该方法的计算性能,并给出如何在线性时间内完成相关统计的计算。     

一种融合图结构的阅读理解答案预测方法

机器阅读理解(简称阅读理解)是自然语言处理领域的一项重要任务.目前基于深度学习的阅读理解模型在一些数据集上(如SQuAD)取得了很好的性能,但是这些模型不能有效整合句法结构、句子间长距离语义信息,对问题和篇章理解得不够充分.为解决以上问题提出一种新方法,该方法利用图结构对句子内部的句法结构与句子间的语义关系进行建模,然后与基于注意力机制的表示进行融合,最终形成对篇章与问题更好的理解与表示.同时,该方法还引入问题分类任务,与问答任务共同形成多任务学习框架,进一步优化了问题与文本的表示,获得了更好的答案预测准确率.相关数据集上的实验结果表明,所提方法取得了比基线模型更好的效果.     

旅游场景下的基于深度学习的文本方面级细粒度情感分类

方面级细粒度情感分类是指针对文本数据,分析其在指定方面的情感极性.由于获取到的评论样本往往涉及不同的方面,导致各个方面的情感极性不平衡.为了减少不平衡数据对模型训练的影响,本文提出了一种新的数据平衡方法——批处理平衡方法(BB),用来平衡多标签多类别数据.同时,由于评论文本蕴含多个方面,传统模型结构往往每次只能预测一个...     

基于深度学习的垃圾图像分类模型的研究

在现实生活中,垃圾形态、拍照角度、光线、背景等都存在差异,要精准的识别垃圾分类,需要加入更多的数据样本.传统的目标检测以手工特征检测算法为主,识别效果不够好,准确率不高,计算量特别大,运行速度很慢,可能会产生多个正确的识别结果.为了快速推进垃圾分类地有效进行垃圾分类,运用改进YOLOv3架构,构建基于深度学习技术的图像...     

基于深度学习的水体色度分类算法研究

运用深度学习技术进行非接触、快速水体色度检测与分类,采用无人机采集水体图像,运用色度仪对标定的图像完成分类,建立数据集。采用图像归一化处理减少环境因素对分类结果的影响,设计多特征的分步边缘检测算法,检测水域图像边缘,剔除无关像素。对VGG 16、GoogleNet-V3和ResNet 18卷积神经网络进行水体色度分类模型构建与训练,后筛选Inception结构和残差结构为基本构建单元,设计专门用于水体色度分类的WCNet 15与WCNet 21神经网络模型。在训练集上训练参数并利用验证集完成对2个模型的准确率的比较,筛选准确率高的WCNet 21模型作为最终水体色度分类模型。WCNet21模型的最优准确率可达97.8%,满足水体色度分类需求,可应用到具体的水体色度分类工作当中。     

基于迁移学习的钢金相组织分类与识别方法的研究

金相检验是分析钢内部组织的常用方法,其中检验图像由人工判别,容易受到主观因素的影响而造成结果的不确定.近年来,深度学习(Deep learning,DL)方法中的卷积神经网络(Convolutional neural networks,CNN)能从原始图像中学习复杂的特征,在图像分类与识别领域得到了广泛的应用.CNN建模需要大量的训练样本才能达到较好的泛化能力,材料科学与工程领域针对具体问题的数据集往往较小,不能满足CNN建模的条件,制约了DL在材料领域的应用.本研究基于lmageNet数据集预训练VGG19模型,对火力发电机组耐热钢金相检验图像进行识别,采用冻结全部卷积层权值和微调部分卷积层权值两种迁移学习方法,可以克服金相图像数据集较小的问题,实现小样本数据集的深度学习建模,两种方法的准确率分别为92.5％和94.2％.微调方式的迁移学习CNN模型具有较快的收敛速度、较高的训练精度与泛化能力,能够对金相组织图像进行较为准确的分类与识别,是一种智能的钢金相组织识别方法,也是自动化分类与识别钢金相组织的一种新方法.     

A model for labour pooling in batch manufacturing

Cross-training workers to perform multi-skilled jobs is one of the modern trends in job design. As companies engage in downsizing, the remaining workforce is expected to do more and different tasks. This paper presents a formal definition and a practical solution for optimizing the size and cost of the pool of multi-skilled workers for production units operated under batch manufacturing. The pool size is optimized through a search procedure applied separately to just-in-time (JIT) and level production plans, which are derived from the stones heuristic. The method allows direct calculation of the cost savings from labour pooling. This paper was inspired by consulting in the food industry, where implementation of these results has significantly reduced labour costs.     

A production-recycling-inventory model with learning effect

This paper develops an integrated production-recycling system over a finite time horizon. Here, the dynamic demand is satisfied by production and recycling. The used units are bought back and then either recycled or disposed of which are not repairable. The used units are collected continuously from the customers. Recycling products can be used as new products which are sold again. The rate of production and disposal are assumed to be function of time. The setup cost is reduced over time due to “Learning curve” effect. The optimum results are presented both in tabular form and graphically.     

A context-aware citation recommendation model with BERT and graph convolutional networks

Scientometrics - With the tremendous growth in the number of scientific papers being published, searching for references while writing a scientific paper is a time-consuming process. A technique...     

A model of polymer composite structure with intermediate layer

In real structures isothermal planes may be distorted, and it is necessary to take this into account while carrying out the calculations of thermal conductivity λ_k of a material. The distortion depends on thermal conductivities of constituent particles and the matrix. With computer speeds already available, we can calculate temperature fields of the structural elements and to evaluate the thermal conductivity of the material. A structural model of polymeric compositions with solid particles with a shell and with microspheres is proposed. It is possible to evaluate the influence of a particle distribution order and particle spacing on the properties of compositions using structural elements. The model is suitable for the estimation with the use of corresponding coefficients of general regularities in compositions which are different by physical nature and yet are represented by analogous mathematical expressions (Fourier, Fick and Ohm laws). Solutions of integral equations for the structural element provide relative magnitudes and are, therefore, simple to use when evaluating other physical parameters of a composition. The application of the model is illustrated by examples.     

A finite element model for the analysis of flexibly connected steel frames

A finite element model for the non-linear analysis of flexibly connected steel frames is proposed. The element is regarded as an assemblage of individual parts whose behaviour is described in terms of generalized stresses and strains. Displacements and plastic strain fields are independently modelled. The former are a function of 13 nodal parameters and constitute a typical application of a hierarchical procedure. The latter are modelled over the cross section, along the element and in the six springs located at the ends of the beam. Examples to show the importance of considering the non-linear behaviour of the connections in the frame analysis are proposed.     

Heat pipe with combined capillary-porous structure

The authors present results of an experimental investigation of the heat-transfer characteristics of low-temperature heat pipes with longitudinal capillary grooves coated with a layer of porous metal.Translated from Inzhenerno-Fizicheskii Zhurnal, Vol. 52, No. 4, pp. 588–592, April, 1987.     

A general steady state mathematical model for fin-and-tube heat exchanger based on graph theory

Fin-and-tube heat exchangers are widely used in air conditioners, chillers, etc. A lot of factors, including arrangement of refrigerant circuits, configure specification of fins and tubes, and operating conditions, have significant influence on the performance of fin-and-tube heat exchangers. For the purpose of fast design of high performance heat exchangers, a simulator reflecting the influence of these factors is necessary. In this paper, a general steady state mathematic model based on the graph theory is presented. With the help of the directed graph and graph-based traversal methods (Breadth-first search and Depth-first search), this model is capable to describe any flexible refrigerant circuit arrangement, and quantify the refrigerant distribution in the refrigerant circuit and heat conduction through fins. An alternative iteration method is also developed to solve the conservation equations, which can shorten the simulating time effectively. The model is verified with the experimental results, and the maximum error is within ±10.0%. A simulator based on this model has been used for designing practical fin-and-tube heat exchangers.     

A combined thermo-mechanical model for Tribological Surface Transformations

Some of the French railroad network’s rails have been affected for the last about twenty years or so by an undesirable physical phenomenon known as Tribological Surface Transformations (TSTs), which are irreversible, quasi-surface solid–solid phase transformations. Although the physical causes of TSTs have not yet been clearly established, it seems likely that the mechanical loads combined with the thermal effects of the wheel/rail contacts may be largely responsible. The present study is based on the assumption that the combined thermo-mechanical constraints may generate solid–solid phase transformations. The thermo-mechanical model developed on these lines was based on previous studies on TRansformation Induced Plasticity (TRIP), which are extended here to account for TSTs. The ability of the model to describe TSTs initiation and development is also assessed by performing a 2-D finite element analysis.     

具有三维连通网络结构的多孔SiC/C材料的电磁损耗特征

通过有机泡沫浸渍/高温炭化和热压固化/高温炭化两种工艺分别制备了具有宏观三维连通网络结构的SiC/C泡沫和显微三维连通多孔结构的SiC/C泡沫同质压制块。使用谐振腔微扰法对比测试了2450MHz频率下两种材料的电磁参数随电导率的变化。结果表明:随着电导率的增加,SiC/C泡沫及其同质压制块的介电常数实部εr′均逐渐增加;电损耗tgδe均先增加,达到最大值后逐渐减小;SiC/C泡沫的磁损耗tgδm不断增加,而其同质压制块的tgδm值则先快速增加,达到最大值后缓慢下降。当二者具有相同有效电导率时,SiC/C泡沫的εr′值比其同质压制块约小1/2,tgδe值至少大2倍,而压制块的tgδm值超过SiC/C泡沫4倍多。SiC/C泡沫及其同质压制块是非磁性的,它们的磁损耗均由其特殊结构与电磁场相互作用产生的,是一种非本征磁损耗。     

A learning curve model for an equivalent number of units

Conventional learning curve models are able to deal only with past data that includes an integer number of cycles and time per cycle. However, in real-life situations, the data collected are of a different nature: periodical information, which includes total work time, in-process inventory and completed units, for example. In such circumstances it would be wrong to consider just the completed units, disregarding the in-process inventory. The concept of Equivalent Number of Units (ENU) has been introduced to permit one to sum up all the work performed, and express it in a manner that allows one to use the learning curve model for such cases. Also, time per unit may not be given for single units and the ENU produced per reported period may not be an integer. In order to solve the above problems, a simple procedure was developed for periodical non-integer data. The end result of the procedure is the product learning curve parameters.