Enhancing Deep Learning Semantics:The Diffusion Sampling and Label-Driven Co-Attention Approach

The advent of self-attention mechanisms within Transformer models has significantly propelled the advancement of deep learning algorithms,yielding outstanding achievements across diverse domains.Nonetheless,self-attention mechanisms falter when applied to datasets with intricate semantic content and extensive dependency structures.In response,this paper introduces a Diffusion Sampling and Label-Driven Co-attention Neural Network(DSLD),which adopts a diffusion sampling method to capture more comprehensive semantic information of the data.Additionally,themodel leverages the joint correlation information of labels and data to introduce the computation of text representation,correcting semantic representationbiases in thedata,andincreasing the accuracyof semantic representation.Ultimately,the model computes the corresponding classification results by synthesizing these rich data semantic representations.Experiments on seven benchmark datasets show that our proposed model achieves competitive results compared to state-of-the-art methods.

Shear mechanical properties and fracturing responses of layered rough jointed rock-like materials

This study aims to investigate mechanical properties and failure mechanisms of layered rock with rough joint surfaces under direct shear loading.Cubic layered samples with dimensions of 100 mm×100 mm×100 mm were casted using rock-like materials,with anisotropic angle(α)and joint roughness coefficient(JRC)ranging from 15°to 75°and 2-20,respectively.The direct shear tests were conducted under the application of initial normal stress(σ_(n)) ranging from 1-4 MPa.The test results indicate significant differences in mechanical properties,acoustic emission(AE)responses,maximum principal strain fields,and ultimate failure modes of layered samples under different test conditions.The peak stress increases with the increasingαand achieves a maximum value atα=60°or 75°.As σ_(n) increases,the peak stress shows an increasing trend,with correlation coefficients R² ranging from 0.918 to 0.995 for the linear least squares fitting.As JRC increases from 2-4 to 18-20,the cohesion increases by 86.32%whenα=15°,while the cohesion decreases by 27.93%whenα=75°.The differences in roughness characteristics of shear failure surface induced byαresult in anisotropic post-peak AE responses,which is characterized by active AE signals whenαis small and quiet AE signals for a largeα.For a given JRC=6-8 andσ_(n)=1 MPa,asαincreases,the accumulative AE counts increase by 224.31%(αincreased from 15°to 60°),and then decrease by 14.68%(αincreased from 60°to 75°).The shear failure surface is formed along the weak interlayer whenα=15°and penetrates the layered matrix whenα=60°.Whenα=15°,as σ_(n) increases,the adjacent weak interlayer induces a change in the direction of tensile cracks propagation,resulting in a stepped pattern of cracks distribution.The increase in JRC intensifies roughness characteristics of shear failure surface for a smallα,however,it is not pronounced for a largeα.The findings will contribute to a better understanding of the mechanical responses and failure mechanisms of the layered rocks subjected to shear loads.

Effective data sampling strategies and boundary condition constraints of physics-informed neural networks for identifying material properties in solid mechanics

Material identification is critical for understanding the relationship between mechanical properties and the associated mechanical functions.However,material identification is a challenging task,especially when the characteristic of the material is highly nonlinear in nature,as is common in biological tissue.In this work,we identify unknown material properties in continuum solid mechanics via physics-informed neural networks(PINNs).To improve the accuracy and efficiency of PINNs,we develop efficient strategies to nonuniformly sample observational data.We also investigate different approaches to enforce Dirichlet-type boundary conditions(BCs)as soft or hard constraints.Finally,we apply the proposed methods to a diverse set of time-dependent and time-independent solid mechanic examples that span linear elastic and hyperelastic material space.The estimated material parameters achieve relative errors of less than 1%.As such,this work is relevant to diverse applications,including optimizing structural integrity and developing novel materials.

Inversion of river-bottom sediment parameters using mechanically sampled specimens and subbottom profiling data

The study of river dynamics requires knowledge of physical parameters, such as porosity, permeability, and wave propagation velocity, of river-bottom sediments. To do so, sediment properties are determined on mechanically sampled specimens and from subbottom profiling. However, mechanical sampling introduces disturbances that affect test results, with the exception of grain-size distribution. In this study, we perform inversion of acoustic data using the grain-size distribution of mechanically sampled specimens and the relation between porosity and permeability from the Kozeny-Carman equation as prior information. The wave reflection coefficient of the water-silt interface is extracted from the raw subbottom profile. Based on the effective density fluid model, we combine the Kozeny-Carman equation and the wave reflection coefficient. We use experimental data from two Yellow River reservoirs to obtain the wave velocity and density of multiple sections and their spatial variations, and find that the inversion and testing results are in good agreement.

Influence analysis of complex crack geometric parameters on mechanical properties of soft rock

Soft rocks, such as coal, are afected by sedimentary efects, and the surrounding rock mass of underground coal mines is generally soft and rich in joints and cracks. A clear and deep understanding of the relationship between crack geometric parameters and rock mechanics properties in cracked rock is greatly important to the design of engineering rock mass struc‑tures. In this study, computed tomography (CT) scanning was used to extract the internal crack network of coal specimens. Based on the crack size and dominant crack number, the parameters of crack area, volume, length, width, and angle were statistically analyzed by diferent sampling thresholds. In addition, the Pearson correlation coefcients between the crack parameters and uniaxial compression rock mechanics properties (uniaxial compressive strength UCS, elasticity modulus E) were calculated to quantitatively analyze the impact of each parameter. Furthermore, a method based on Pearson coefcients was used to grade the correlation between crack geometric parameters and rock mechanical properties to determine threshold values. The results indicated that the UCS and E of the specimens changed with the varied internal crack structures of the specimens, the crack parameters of area, volume, length and width all showed negative correlations with UCS and E, and the dominant crack played an important role both in weakening strength and stifness. The crack parameters of the angle are all positively correlated with the UCS and E. More crack statistics can signifcantly improve the correlation between the parameters of the crack angle and the rock mechanics properties, and the statistics of the geometric parameters of at least 16 cracks or the area larger than 5 mm2 are suggested for the analysis of complex cracked rock masses or physical reproduction using 3D printing. The results are validated and further analyzed with triaxial tests. The fndings of this study have important reference value for future research regarding the accurate and efcient selection of a few cracks with a signifcant infuence on the rock mechanical properties of surrounding rock mass structures in coal engineering.

Research on the Encapsulation Device for Lunar Samples

The encapsulation of lunar samples is a core research area in the third phase of the Chinese Lunar Exploration Program.The seal assembly,opening and closing mechanism(OCM),and locking mechanism are the core components of the encapsulation device of the lunar samples,and the requirements of a tight seal,lightweight,and low power make the design of these core components difficult.In this study,a combined sealing assembly,OCM,and locking mechanism were investigated for the device.The sealing architecture consists of rubber and an Ag-In alloy,and a theory was built to analyze the seal.Experiments of the electroplate Au coating on the knife-edge revealed that the hermetic seal can be significantly improved.The driving principle for coaxial double-helical pairs was investigated and used to design the OCM.Moreover,a locking mechanism was created using an electric initiating explosive device with orifice damping.By optimizing the design,the output parameters were adjusted to meet the requirements of the lunar explorer.The experimental results showed that the helium leak rate of the test pieces were not more than 5×10^(-11) Pa·m^(3)·s^(-1),the minimum power of the OCM was 0.3 W,and the total weight of the principle prototype was 2.9 kg.The explosive driven locking mechanism has low impact.This investigation solved the difficulties in achieving tight seal,light weight,and low power for the lunar explorer,and the results can also be used to explore other extraterrestrial objects in the future.

Distributed model predictive control based on adaptive sampling mechanism

In this work,an adaptive sampling control strategy for distributed predictive control is proposed.According to the proposed method,the sampling rate of each subsystem of the accused object is determined based on the periodic detection of its dynamic behavior and calculations made using a correlation function.Then,the optimal sampling interval within the period is obtained and sent to the corresponding sub-prediction controller,and the sampling interval of the controller is changed accordingly before the next sampling period begins.In the next control period,the adaptive sampling mechanism recalculates the sampling rate of each subsystem’s measurable output variable according to both the abovementioned method and the change in the dynamic behavior of the entire system,and this process is repeated.Such an adaptive sampling interval selection based on an autocorrelation function that measures dynamic behavior can dynamically optimize the selection of sampling rate according to the real-time change in the dynamic behavior of the controlled object.It can also accurately capture dynamic changes,meaning that each sub-prediction controller can more accurately calculate the optimal control quantity at the next moment,significantly improving the performance of distributed model predictive control(DMPC).A comparison demonstrates that the proposed adaptive sampling DMPC algorithm has better tracking performance than the traditional DMPC algorithm.

Few-shot object detection based on positive-sample improvement

Traditional object detectors based on deep learning rely on plenty of labeled samples,which are expensive to obtain.Few-shot object detection(FSOD)attempts to solve this problem,learning detection objects from a few labeled samples,but the performance is often unsatisfactory due to the scarcity of samples.We believe that the main reasons that restrict the performance of few-shot detectors are:(1)the positive samples is scarce,and(2)the quality of positive samples is low.Therefore,we put forward a novel few-shot object detector based on YOLOv4,starting from both improving the quantity and quality of positive samples.First,we design a hybrid multivariate positive sample augmentation(HMPSA)module to amplify the quantity of positive samples and increase positive sample diversity while suppressing negative samples.Then,we design a selective non-local fusion attention(SNFA)module to help the detector better learn the target features and improve the feature quality of positive samples.Finally,we optimize the loss function to make it more suitable for the task of FSOD.Experimental results on PASCAL VOC and MS COCO demonstrate that our designed few-shot object detector has competitive performance with other state-of-the-art detectors.

基于Electra预训练模型并融合依存关系的中文事件检测模型

事件检测是信息提取领域的一个重要研究方向。现存的事件检测模型受到语言模型训练目标的限制,只能被动地获取词与词之间的依赖关系,使得模型在训练的过程中过多地关注与训练目标不相关的成分,从而导致检测结果错误。以往的研究表明,充分理解上下文信息对于基于深度学习的事件检测技术至关重要。因此,在Electra预训练模型的基础上,引入KVMN网络来捕捉单词之间的依赖关系,以增强单词的语义特征,并采用了一种门控机制来加权这些特征。然后,为了解决中文事件检测中模型识别错误决策的问题,在输入中加入负样本,对不同样本加入不同程度的噪声,使模型学习更好的嵌入表示,有效提高了模型对未知样本的泛化能力。最后,在公共数据集LEVEN上的实验结果表明,该方法优于现有方法,取得了93.43%的F1值。

基于广度-深度采样和图卷积网络的谣言检测方法

现有谣言检测方法存在早期数据丢失、特征利用不充分问题,为此提出新的检测方法.为了充分挖掘事件的早期传播特征,提出广度采样方法并构建与事件对应的传播序列,利用Transformer挖掘长距离评论间的语义相关性并构建事件的传播序列特征.为了有效挖掘事件的传播结构特征,提出基于路径长度的深度采样方法,构建事件对应的信息传播子图和信息聚合子图,利用图卷积网络在挖掘图结构特征方面的优势,获得与事件对应的传播结构特征.将事件对应的传播序列特征表示与传播结构特征表示进行拼接,得到事件对应的最终特征表示.在公开数据集Weibo2016和CED上开展所提方法的有效性验证实验.结果表明,所提方法普遍优于现有典型方法.与基线方法相比,所提方法的准确率和F1值均有显著提升,所提方法在谣言检测领域的有效性得到验证.

结合注意力和多路径融合的实时肺结节检测算法

现有单阶段目标检测算法在肺结节检测中结节检出不敏感,卷积神经网络(CNN)在特征提取时多次上采样导致微小结节特征提取困难、检测效果差,并且现存肺结节检测算法模型复杂,不利于实际应用部署落地。针对上述问题,提出一种结合注意力机制和多路径融合的实时肺结节检测算法,并在此基础上改进上采样算法,提升肺部结节的检测精度和模型推理速度,且模型的权重小容易部署。首先,在特征提取的主干网络部分融合通道和空间的混合注意力机制;其次,改进采样算法,提高生成特征图的质量;最后在加强特征提取网络部分,在不同路径之间建立通道,实现深层和浅层特征的融合,将不同尺度的语义和位置信息融合。在LUNA16数据集的实验结果表明,相较于原始YOLOv5s算法,所提算法的精确率、敏感度和平均精度分别提升9.5、6.9和8.7个百分点,帧率达到131.6 frame/s,模型权重文件仅有14.2 MB,表明了所提算法可以实时检测肺结节,并且精度远高于YOLOv3和YOLOv8等现有单阶段检测算法。

“双曲线”型煤样承载力学特性试验研究

煤炭地下气化结束后,两气化炉间形成类“双曲线”形煤柱,支撑覆岩,保障气化区域安全稳定。为研究类“双曲线”形煤柱承载力学特性,基于声发射监测系统和XTDIC三维全场应变测量系统,开展了不同侧向拱高(h=0,3,7,10,13,17 mm)的6组“双曲线”形煤样单轴压缩试验,分析了h对煤样峰值载荷、变形破坏及声发射特征的影响,揭示了其承载破坏机制。结果表明:①“双曲线”形煤样可分为矩形结构(主要承载体)和侧向拱结构,其承载破坏机制与其受力形式、侧向拱结构有关;随着h增大,煤样承载能力降低,与h=0煤样相比,峰值载荷分别降低了7.66%,13.56%,26.83%,35.28%,62.75%。②随着h增大,煤样整体受力形式由以受压为主向受压–受弯曲转变,中部区域产生应力集中而形成薄弱区,对应的水平位移场向中部迁移,最终汇集于中部边缘处;而垂直位移场由水平条带状向倾斜条带状转变,最终集中于煤样侧向拱结构上端。③在轴向载荷作用下,煤样侧向拱结构对其矩形结构中部区域产生等效作用力,加之煤样非均质性影响,加剧了薄弱区损伤程度,该作用随着h增大而增强,煤样承受载荷未超过其抗拉强度即产生剪切破坏,其破坏模式由拉–剪混合破坏向剪切破坏转变,均伴随着不同程度的剥落和局部弹射破坏。④煤样声发射累计计数–时间曲线演化可分为3种类型,当h为0和3 mm时,分为“上凸”式增长、相对快速增长、快速增长、“突变”式增长4个阶段,其演化特征与常规煤岩试样一致;当h为7 mm和10 mm时,分为相对快速增长、快速增长、“突变”式增长3个阶段;当h为13 mm和17 mm时,分为快速增长和“突变”式增长2个阶段;峰后阶段均呈“突变”式增长,而峰前阶段增长形式不一致是由煤样裂纹稳定扩展和中部区域持续损伤共同导致的。

基于注意力与密集重参数化的目标检测算法

针对目标检测任务中背景复杂、目标尺寸差异大等因素导致目标检测结果较差的问题,本文提出基于注意力和密集重参数化的目标检测算法。首先,基于CSP-DarkNet提出高效的特征提取网络,主要包括密集重参数化模块和CASA模块2个设计。前者利用密集连接保留浅层特征,又通过重参数化结构降低网络复杂度;后者CASA模块用于获取需要的目标信息。其次,特征融合在特征金字塔(FPN)和路径聚合网络(PAN)的基础上,引入内容感知特征重组(CARAFE)进行上采样,有效解决了邻近插值法等未能捕捉丰富语义信息的问题;提出更高效的C3-G模块,获取丰富的梯度信息,增强模型表达能力和感知能力;同时,引入深度可分离卷积提升运算效率。最后,检测输出采用在更大范围上跨领域正负样本匹配策略扩充正样本数量,提升检测效果。该算法在MS COCO和PASCAL VOC数据集上的mAP@0.5分别达到了57.5%和83.0%,充分说明了本文算法的先进性。

基于混合特征提取的流数据概念漂移处理方法

大数据时代,越来越多的数据以数据流的形式产生,由于其具有快速、无限、不稳定及动态变化等特性,使得概念漂移成为流数据挖掘中一个重要但困难的问题.目前多数概念漂移处理方法存在信息提取能力有限且未充分考虑流数据的时序特性等问题.针对这些问题,提出一种基于混合特征提取的流数据概念漂移处理方法(concept drift processing method of streaming data based on mixed feature extraction,MFECD).该方法首先采用不同尺度的卷积核对数据进行建模以构建拼接特征,采用门控机制将浅层输入和拼接特征融合,作为不同网络层次输入进行自适应集成,以获得能够兼顾细节信息和语义信息的数据特性.在此基础上,采用注意力机制和相似度计算评估流数据不同时刻的重要性,以增强数据流关键位点的时序特性.实验结果表明,该方法能有效提取流数据中包含的复杂数据特征和时序特征,提高了数据流中概念漂移的处理能力.

基于双分支点流语义先验的路面病害分割模型

针对基于深度学习的真实路面病害图像识别算法主要面临的复杂道路背景与病害前景比例不同、病害尺度小等导致的类别严重不平衡、路面病害与道路的几何结构特征对比不明显导致其不易识别等问题,本文提出一种基于双分支语义先验网络,用于指导自注意力骨干特征网络挖掘背景与病害前景的复杂关系,运用高效自注意力机制和互协方差自注意力机制分别对二维空间和特征通道进行语义特征提取,并引入语义局部增强模块提高局部特征聚合能力。本文提出了一种新的稀疏主体点流模块,并与传统特征金字塔网络相结合,进一步缓解路面病害的类别不平衡问题;构建了一个真实场景的道路病害分割数据集,并在该数据集和公开数据集上与多个基线模型进行对比实验,实验结果验证了本模型的有效性。

基于改进YOLOv5的黑色素瘤图像自动诊断

为解决现有黑色素瘤智能诊断模型中存在的对毛发遮挡目标识别精度不足、样本不均以及轻量化程度不够的问题,提出一种改进的YOLOv5模型。首先,基于改进的C3结构和自注意力机制设计CS_Neck结构,从而有效区分黑色素瘤和毛发的相关特征;其次,提出一种二次筛选难样本挖掘方法,利用焦点损失函数降低简单样本权重,引入损失秩排序(loss rank mining,LRM)思想降低简单样本数量;最后,设计轻量级骨干网络,提出使用改进的RepVGG结构替换普通卷积提取特征,提高推理速度,并引入宽度乘子降低参数量和权重,实现模型轻量化。基于ISIC2019数据集的实验结果表明,所提算法的权重和参数量仅为7.9 MB和4.0×10^(6),精度达到92.9%。所提算法有效提升了精度且实现了轻量化,可以满足高效诊断黑色素瘤的要求。

特异小样本工业产品表面缺陷检测方法研究

基于机器视觉的工业产品表面缺陷检测设备和系统大量应用在工业制造领域,目前其难点在于工业检测数据的采集,由于训练样本缺失导致深度学习网络模型无法有效训练.为解决上述问题,首先,提出一种基于不规则掩码的伤痕样本生成算法,改善了钢板表面缺陷检测任务中特异小样本数据集正负样本不均衡的情况;然后,在YOLOv8主干网络引入MHSA多头自注意力,提高对钢板表面缺陷的关注度;最后,使用SIoU替换原损失函数,增强网络模型的定位能力,提高检测的准确性.基于热轧钢板表面缺陷检测问题的实验结果表明,该方法能够有效解决特异小样本工业探伤的具体问题.

训练样本标签误差对高光谱影像分类影响

在影像尤其是高光谱影像分类中,用于学习训练的标签质量对分类成效影响并未得到充分重视。为此,文章基于PyTorch框架,利用Indian Pines高光谱数据集,探讨了在RF、BP、CNN和SSConvNeXt模型下,光谱特征相似度较高的地物在不同比例人为误标注情况时对分类结果的影响。分析结果认为:同样错误标注情况下,SSConvNeXt和CNN相较RF、BP模型体现出20%以上的分类精度优势;在无人为错误标注、10个错误噪声标签、错误标签占比15%和25%时,SSConvNeXt和CNN模型的分类精度都在96%以上,体现了模型的容错性和稳定性;在相对传统的RF和BP模型中,错误标签对分类影响较大且离散。最后重点分析了SSConvNeXt模型在分类方面的机制优势。该研究可从训练样本角度为遥感影像分类精度问题给予一定的方法选择和定量分析依据。

动静组合加载下裂隙试样力学特性和破坏模式

目的为了深入了解裂隙岩石在同时受到地应力和动力扰动时的工程灾变机制。方法采用改进的分离式霍普金森压杆系统,对预制裂隙类岩石试样进行了动静组合加载试验。研究了应变率、裂隙倾角和预应力对试样动态力学特性和能耗特征的影响,并对试样的破坏模式进行了分析。结果当应变率从61.82s^(-1)增加到195.57s^(-1),裂隙试样的动态增长因子不断增大,具有显著的应变率效应;动态弹性模量变化规律不明显,应变率效应不显著。随着入射能的增加,裂隙试样的单位耗散能密度不断增加。当裂隙倾角由0°变化到90°,试样的动态增长因子增不断增加,动态弹性模量和单位耗散能密度不断降低。当施加预应力后,30°、45°和60°裂隙试样的动态增长因子增长均有所增加;45°裂隙试样动态弹性模量有所增加,30°和60°裂隙试样的变化规律不明显;45°和60°裂隙试样单位耗散能密度有所降低,30°裂隙试样的变化规律不明显。裂隙试样的破坏模式主要包括张拉破坏、剪切破坏和张拉-剪切复合破坏3种,施加预应力对试样破坏模式的影响不明显。结论0°裂隙对工程岩体的承载能力削弱程度最小,90°裂隙有利于提高破岩效率。预应力在一定程度上会提高工程岩体的承载能力。研究结果可以为预测和评估岩石工程中的风险,制定合理的加固措施提供参考。

马尔科夫链关联下电动汽车机械变速器可靠性评估方法设计

可靠性评估理论应用到电动汽车机械式变速器齿轮系的设计中,当前方法多是根据电动汽车动力性要求,在保证零件结构强度和刚度可靠使用的条件下,以变速器安全和功率使用区间为目标函数,建立了汽车机械式变速器多目标可靠性模型,但是这种方式忽略了目标之间的整体关联性。为此,设计一种电动汽车机械变速器可靠性评估方法。对电动汽车机械变速器各个部件展开运行模式分析,确定影响其可靠性的主要因素。根据运行模式分析的结果,使用状态枚举和蒙特卡洛抽样方法得到各个故障模式的马尔科夫链,并进行时域离散化处理,根据时域离散化处理得到的状态样本,构建模拟状态转移矩阵,可以计算出电动汽车机械变速器的可靠性指标,用于评估自动变速器的可靠性能。实验结果表明,采用所提方法可以获取高精度和高效率的电动汽车机械变速器可靠性评估结果。