基于对称式对比学习的齿轮箱无监督故障诊断方法

针对跨工况下无监故障诊断特征提取难、模型泛化性弱的问题,提出一种基于对称式对比学习策略的齿轮箱无监督故障诊断方法。首先,利用原始信号构建正负样本集,通过加噪声、序列倒转等数据增强后,分别输入两个结构相同的卷积神经网络提取高维特征;其次,度量正负样本的相似程度进行编码学习数据的隐藏表示,通过对称式自监督对比学习优化正负样本的对比估计损失函数,从而有效利用样本自身标签信息,提升网络从无标签样本中学习判别特征的能力;最后,在齿轮箱数据集上对所提方法开展试验验证,通过聚类准确率、分类系数和划分熵进行综合评估。结果表明,所提方法聚类精度可达98%以上,相比其他方法,呈现了更强的聚类能力和泛化性能。     

基于深度学习的表面缺陷检测技术研究进展

随着计算机视觉技术的快速发展,基于深度学习的表面缺陷检测技术实现了爆发式的应用,并逐步成为了主流发展方向。基于深度学习的缺陷检测技术可以近似为计算机视觉任务中的分类、检测、分割等任务,其主要目的是找出物体表面缺陷的类别和所在位置,相较于传统图像处理方法,深度学习在特征提取能力和环境适应能力上优势明显。以缺陷数据标签类型为依据,对近年来基于深度学习的表面缺陷检测技术进行梳理划分,总结目前技术的优点与不足,重点阐述了监督学习下的三种缺陷检测方法。探讨了表面缺陷检测技术面临的小样本以及不平衡样本等关键问题：对于小样本问题目前有结构优化、数据增广、迁移学习等解决方法;针对不平衡样本问题,介绍了近年来热点的无监督、弱监督与半监督学习模型。随后介绍了常用的工业表面缺陷数据集并展现了近年来提出的算法在NEU数据集上的应用效果。最后对进一步的研究工作提出展望,希望能给缺陷检测研究提供有意义的参考。     

基于加权在线样本更新的目标长时跟踪方法

针对在长时跟踪过程中因丢失视野导致目标跟踪失败的问题,提出了基于加权在线样本更新的目标长时跟踪方法。 首 先,使用 ResNet50 网络提取目标深度特征并增强初始帧样本优化目标模型,提高初始帧样本权重影响;然后,利用目标模型对 测试帧样本进行分类,并采用置信度分值加权在线学习样本以增强样本质量,提升模型的分类效果;其次,使用置信度分值判别 目标状态并跟踪定位目标,目标丢失时使用时空约束搜索在丢失处自适应扩展区域并随机搜索目标,同时利用在线学习快速优 化目标模型,增强其对目标的搜索能力;最后,针对搜索过程设计一种自适应阈值判别方法,充分利用图像背景信息,将目标丢 失时背景置信度分值作为判别阈值,降低搜索过程中相似背景的影响以准确找回目标。 使用 LTB50 数据集进行实验验证,成 功率和跟踪 F-score 分别为 66. 1% 和 64. 4% ,优于其他方法;在四足移动机器人平台上进行真实场景实验,目标完全遮挡和视野 外两种情况下成功率分别为 87. 8% 和 85. 8% ,证明了方法的有效性。     

基于边界辅助判别的滚动轴承故障特征增强及诊断方法

滚动轴承作为机械设备重要部件,对保障设备安全稳定运行具有重要意义。针对实际诊断中的滚动轴承故障数据不平衡问题,提出了一种基于边界辅助判别的辅助分类生成对抗网络模型(BD-ACGAN)。首先,设计了一种可用于提取故障样本边界细节特征的边界辅助判别器,以引导生成器生成更真实的样本,并采用该生成样本解决了数据不平衡的问题;其次,采用了自适应权重损失模块,动态调整了损失权重,使该模型更加关注重要的特征信息,从而提高了该模型的生成质量和特征表达能力;利用生成样本和真实样本数据对BD-ACGAN模型进行了增强训练,提高了该模型的泛化能力和诊断能力;最后,进行了消融实验及对照实验,对BD-ACGAN模型的特征增强能力和诊断效果进行了验证,分别采用美国凯斯西储大学和西安交通大学滚动轴承数据集对模型进行了实验验证。研究结果表明：该BD-ACGAN模型能够有效利用故障样本的边界特征解决数据不平衡问题,并且故障诊断精确度为98.79%,优于其他对照模型,为滚动轴承故障诊断提供了一种新的方法。     

面向高光谱影像分类的深度流形重构置信网络

鉴于传统深度学习方法只提取了高光谱图像中的深度抽象信息,而未能充分揭示样本之间的局部几何结构关系,限制了分类性能的提升,本文提出了一种新的特征提取网络——深度流形重构置信网络。该网络首先通过深度置信网络提取深度抽象特征,为进一步增强抽象特征的鉴别能力,在图嵌入框架下通过样本数据的邻域点和各邻域的同类近邻重构点来构建类内图和类间图,并在低维空间中分离类间近邻点与其重构点的同时压缩类内近邻点和相应的重构点,实现提取深度鉴别特征,以改善不同类数据的可分性,进而提升地物分类精度。在KSC和MUUFL Gulfport高光谱数据集上的实验结果表明,本文算法的总体分类精度分别达到了94.71%和86.38%。相比较其他算法,本文算法有效提升了地物分类能力,更有利于实际应用。     

深度学习框架下数控机床运动误差溯因方法

为了快速准确地识别数控机床误差因素,提出了一种基于深度学习框架的运动误差智能溯因方法。首先,建立数控机床圆运动误差轨迹基准圆模型,生成理论样本解决深度学习大数据训练问题;然后,基于深度学习目标检测框架,提出数控机床运动误差溯因模型,构建深度卷积网络层自动提取圆运动轨迹特征,改进候选区域生成机制实现轨迹判别及区域生成;最后,模型的区域识别层准确识别误差轨迹类型,通过圆运动轨迹与误差源的映射关系实现对运动误差快速准确智能溯因。实验表明,本文方法可行且适应性好,识别溯源准确,正确率超过96%。     

联合训练生成对抗网络的半监督分类方法

深度神经网络需要大量数据进行监督训练学习,而实际应用中往往难以获取大量标签数据。半监督学习可以减小深度网络对标签数据的依赖,基于半监督学习的生成对抗网络可以提升分类效果,但仍存在训练不稳定的问题。为进一步提高网络的分类精度并解决网络训练不稳定的问题,本文提出一种基于联合训练生成对抗网络的半监督分类方法,通过两个判别器的联合训练来消除单个判别器的分布误差,同时选取无标签数据中置信度高的样本来扩充标签数据集,提高半监督分类精度并提升网络模型的泛化能力。在CIFAR-10和SVHN数据集上的实验结果表明,本文方法在不同数量的标签数据下都获得更好的分类精度。当标签数量为2 000时,在CIFAR-10数据集上分类精度可达80.36%;当标签数量为10时,相比于现有的半监督方法,分类精度提升了约5%。在一定程度上解决了GAN网络在小样本条件下的过拟合问题。     

面向高光谱图像分类的半监督双流网络

为了提取高光谱图像中的深度鉴别特征,往往需要大量标记样本,但是高光谱图像样本标定困难,基于高光谱图像的“图谱合一”特性提出一种基于深度-流形学习的半监督双流网络。该网络用卷积网络和神经网络分别提取少量标记样本以及大量无标记样本中的空-谱联合特征,然后分别构建基于监督图和非监督图的流形重构图模型,以挖掘其中的本征流形结构。在此基础上设计了基于均方误差和流形学习的联合损失函数,以协同度量流形边界和空-谱概率残差,实现双流网络的一体化反馈和优化,进而实现地物分类。在WHU-Hi龙口和黑河高光谱数据集上实验的总体分类精度分别达到97.53%和96.79%,有效提升了地物分类能力。     

多模态特征融合与多任务学习的特种视频分类

特种视频(本文特指暴力视频)的智能分类技术有助于实现网络信息内容安全的智能监控。针对现有特种视频多模态特征融合时未考虑语义一致性等问题,本文提出了一种基于音视频多模态特征融合与多任务学习的特种视频识别方法。首先,提取特种视频的表观信息和运动信息随时空变化的视觉语义特征及音频信息语义特征;然后,构建具有语义保持的共享特征子空间,以实现音视频多种模态特征的融合;最后,提出基于音视频特征的语义一致性度量和特种视频分类的多任务学习特种视频分类理论框架,设计了对应的损失函数,实现了端到端的特种视频智能识别。实验结果表明,本文提出的算法在Violent Flow和MediaEval VSD 2015两个数据集上平均精度分别为97.97%和39.76%,优于已有研究。结果证明了该算法的有效性,有助于提升特种视频监控的智能化水平。     

CNN-Transformer结合对比学习的高光谱与LiDAR数据协同分类

针对高光谱图像（hyperspectral images, HSI）与LiDAR数据多模态分类任务中的跨模态信息表达和特征对齐等问题,提出一种基于对比学习CNN-Transformer高光谱和LiDAR数据协同分类网络（Contrastive Learning based CNN-Transformer Network, CLCT-Net）。CLCT-Net通过由ConvNeXt V2 Block构成的共有特征提取模块,获得不同模态间的共性特征,解决异构传感器数据之间语义对齐的问题。构建了包含空间-通道分支和光谱上下文分支的双分支HSI编码器,以及结合频域自注意力机制的LiDAR编码器,以获取更丰富的特征表示。利用集成对比学习进行分类,进一步提升多模态数据协同分类的精度。在Houston 2013和Trento数据集上的实验结果表明,相较于其他高光谱图像和LiDAR数据分类模型,本文所提模型获得了更高的地物分类精度,分别达到了92.01%和98.90%,实现了跨模态数据特征的深度挖掘和协同提取。     

Minimal Realization of Linear Graph Models for Multi-physics Systems

An engineering system may consist of several different types of components,belonging to such physical"domains"as mechanical,electrical,fluid,and thermal.It is t...     

INTEGRAL EQUATION METHOD＇S APPLICATION IN HOLE-EDGE STRESS OF COMPOSITE MATERIAL PLATE WITH DIFFERENT SHAPED HOLES

The strength of composite plate with different hole-shapes is always one of the most important but complicated issues in the application of the composite material. The holes will lead to mutations and discontinuity to the structure. So the hole-edge stress concentration is always a serious phenomenon. And the phenomenon makes the structure strength decrease very quickly to form dangerous weak points. Most partial damage begins from these weak points. According to the complex variable functions theory, the accurate boundary condition of composite plate with different hole-shapes is founded by conformal mapping method to settle the boundary condition problem of complex hole-shapes. Composite plate with commonly hole-shapes in engineering is studied by several complex variable stress fimction. The boundary integral equations are founded based on exact boundary conditions. Then the exact hole-edge stress analytic solution of composite plate with rectangle holes and wing manholes is resolved. Both of offset axis loadings and its influences on the stress concentration coefficient of the hole-edge are discussed. And comparisons of different loads along various offset axis on the hole-edge stress distribution of orthotropic plate with rectangle hole or wing manhole are made. It can be concluded that hole-edge with continuous variable curvatures might help to decrease the stress concentration coefficient; and smaller angle of outer load and fiber can decrease the stress peak value.     

Non-monotonic material contrast in scanning ion and scanning electron images

30 keV Ga⁺ focused ion beam induced secondary electron (iSE) imaging was used to determine the relative contrast between several materials. The iSE signal compared from C, Si, Al, Ti, Cr, Ni, Cu, Mo, Ag, and W metal layers does not decrease with an increase in target atomic number Z₂, and shows a non-monotonic relationship between contrast and Z₂. The non-monotonic relationship is attributed to periodic fluctuations of the stopping power and sputter yield inherent to the ion–solid interactions. In addition, material contrast from electron-induced secondary electron (eSE) and backscattered electron (BSE) images using scanning electron microscopy (SEM) also shows non-monotonic contrast as a function of Z₂, following the periodic behavior of the stopping power for electron–solid interactions. A comparison of the iSE and eSE results shows similar relative contrast between the metal layers, and not complementary contrast as conventionally understood. These similarities in the contrast behavior can be attributed to similarities in the periodic and non-monotonic function defined by incident particle–solid interaction theory.     

Application of Feature Extraction through Convolution Neural Networks and SVM Classifier for Robust Grading of Apples

This paper proposes a novel grading method of apples,in an automated grading device that uses convolutional neural networks to extract the size,color,texture,an...     

分布动态载荷识别的抗噪处理

针对正交多项式频域法在用多种响应对矩形薄板进行载荷识别中抗噪性较差的问题,综合运用平均法、矩阵预处理和奇异值截断法等方法对之进行改善,并引入空间映射的思想,将该方法的应用范围拓展为复杂的模型.利用仿真算例,证实了该方法具有较好的抗噪性.     

基于MELTAC-N平台的核电厂安全级DCS环网研究与测试

针对工程实践中环网通讯相关问题的处理缺乏理论基础及国产化安全级DCS平台的开发缺乏成熟经验借鉴问题,对基于MELTAC-N平台核电厂安全级DCS环网的软硬件实现进行了研究。提出了安全级DCS环网双环网冗余设计、光切换开关设计等硬件设计方法,以及以RPR协议为基础,采用全数据收发策略的软件设计方法。在CPR1000安全级DCS平台上对安全级DCS环网的可靠性及实时性进行了评价,并进行了容错能力、响应时间及响应时间稳定性测试验证实验。结果表明,基于MELTAC-N平台安全级DCS环网软硬件设计具有较好的容错能力及响应时间稳定性。     

Short-Term Relay Quality Prediction Algorithm Based on Long and Short-Term Memory

The fraction defective of semi-finished products is predicted to optimize the process of relay production lines, by which production quality and productivity ar...     

End-to-End Multiview Gesture Recognition for Autonomous Car Parking System

The use of hand gestures can be the most intuitive human-machine interaction medium.The early approaches for hand gesture recognition used device-based methods....     

The laboratory simulation of abrasive wear

Abrasive wear has long been recognised as one of the most potentially serious tribological problems facing the operators of many types of plant and machinery; several industrial surveys have indicated that wear by abrasion can be responsible for more than 50% of unscheduled machine and plant stoppages. Locating the operating point of a tribological contact in an appropriate operational ‚map’︁ can provide a useful guide to the likely nature and origins of the surface degradation experienced in use, though care must be exercised in choosing the most suitable parameters for the axes of the plot. Laboratory testing of materials and simulations of machine contacts are carried out for a number of purposes; at one level for the very practical aims of ranking candidate materials or surface hardening treatments in order of their wear resistance, or in an attempt to predict wear lives under field conditions. More fundamentally, tests may be aimed at elucidating the essential physical mechanisms of surface damage and loss, with the longer term aim of building an analytical and predictive model of the wear process itself. In many cases, component surface damage is brought about by the ingress of hard, particulate matter into machine bearing or sealing clearances. These may be running dry although, more usually, a lubricant or service fluid is present at the interface. A number of standardised wear test geometries and procedures have been established for both two- and three-body wear situations, and these are briefly described. Although abrasive wear is often modelled as following an ‚Archard’︁ equation (i.e. a linear increase in material loss with both load and time, and an inverse dependence on specimen hardness) both industrial experience and laboratory tests of particularly lubricated contacts show that this is not always the case: increasing the hardness differential in an abrasively contaminated lubricated pair may not always reduce the rate of damage to the harder surface.     

ANGULAR DISTORTION ALLOWABLE VALUE AND ASSESS METHOD OF WELD ON COKE TOWER

A graph as the new engineering method for estimate the safety of bulging deformation of coke tower is proposed. Through stresses analysis of circumferential weld of coke tower and comparing the stresses produced by pressure with heat stress of steady state, residual stress, bending stress produced by both itself weight and wind loads, it showed that the stresses produced by pressure on the angle distortion are the main factor of equivalent stress of the combined stress. After comparing four kinds of stress controlling conditions, the relation to stress with depth of angular distortion, grade of curvature of angular distortion and half of region of angular distortion has been inferred. Graph of deformation allowable value of coke tower for different condition by angular distortion and half of region of angular distortion has been plotted. The five steps for its engineering use have been explained. The lighter the grade of curvature is, the larger of bulge allowance, may be, and the bigger of depth of angular distortion may pose too. For the coke tower with a popular structure of Dg 5 400 mm×28 mm, the result by graph is nearly more than the result of two formulas formed by other research, the error is less than 7.0%. But, the graph can be easily applied to different size of angular distortion.