Automatic diagnosis of diabetic retinopathy using vision transformer based on wide-field optical coherence tomography angiography

Diabetic retinopathy(DR)is one of the major causes of visual impairment in adults with diabetes.Optical coherence tomography angiography(OCTA)is nowadays widely used as the golden criterion for diagnosing DR.Recently,wide-field OCTA(WF-OCTA)provided more abundant information including that of the peripheral retinal degenerative changes and it can contribute in accurately diagnosing DR.The need for an automatic DR diagnostic system based on WF-OCTA pictures attracts more and more attention due to the large diabetic population and the prevalence of retinopathy cases.In this study,automatic diagnosis of DR using vision transformer was performed using WF-OCTA images(12 mm×12 mm single-scan)centered on the fovea as the dataset.WF-OCTA images were automatically classified into four classes:No DR,mild nonproliferative diabetic retinopathy(NPDR),moderate to severe NPDR,and proliferative diabetic retinopathy(PDR).The proposed method for detecting DR on the test set achieves accuracy of 99.55%,sensitivity of 99.49%,and specificity of 99.57%.The accuracy of the method for DR staging reaches up to 99.20%,which has been proven to be higher than that attained by classical convolutional neural network models.Results show that the automatic diagnosis of DR based on vision transformer and WF-OCTA pictures is more effective for detecting and staging DR.

基于残差U-Net和自注意力Transformer编码器的磁场预测方法

利用有限元方法对几何结构复杂的电机和变压器进行磁场分析,存在仿真时间长且无法复用的问题。因此,该文提出一种基于残差U-Net和自注意力Transformer编码器的磁场预测方法。首先建立永磁同步电机(PMSM)和非晶合金变压器(AMT)有限元模型,得到深度学习训练所需的数据集;然后将Transformer模块与U-Net模型结合,并引入短残差机制建立ResUnet-Transformer模型,通过预测图像的像素实现磁场预测;最后通过Targeted Dropout算法和动态学习率调整策略对模型进行优化,解决拟合问题并提高预测精度。计算实例证明,ResUnet-Transformer模型在PMSM和AMT数据集上测试集的平均绝对百分比误差(MAPE)均小于1%,且仅需500组样本。该文提出的磁场预测方法能减少实际工况和多工况下精细模拟和拓扑优化的时间和资源消耗,亦是虚拟传感器乃至数字孪生的关键实现方法之一。

Equivalent Doping Transformation Method for Predicting Breakdown Voltage and Peak Field at Breakdown of Epitaxial-Diffused Punch-Through Junction

Based on a new semi empirical analytical method, namely equivalent doping transformation, the breakdown voltage and the peak field of the epitaxial diffused punch through junction have been obtained. The basic principle of this method is introduced and a set of breakdown voltage and peak field plots are provided for the optimum design of the low voltage power devices. It shows that the analytical results coincide with the previous numerical simulation well.

基于Transformer的多尺度遥感语义分割网络

为了提升遥感图像语义分割效果,本文针对分割目标类间方差小、类内方差大的特点,从全局上下文信息和多尺度语义特征2个关键点提出一种基于Transformer的多尺度遥感语义分割网络(muliti-scale Transformer network,MSTNet)。其由编码器和解码器2个部分组成,编码器包含基于Transformer改进的视觉注意网络(visual attention network,VAN)主干和基于空洞空间金字塔池化(atrous spatial pyramid pooling, ASPP)结构改进的多尺度语义特征提取模块(multi-scale semantic feature extraction module, MSFEM)。解码器采用轻量级多层感知器(multi-layer perception,MLP)配合编码器设计,充分分析所提取的包含全局上下文信息和多尺度表示的语义特征。MSTNet在2个高分辨率遥感语义分割数据集ISPRS Potsdam和LoveDA上进行验证,平均交并比(mIoU)分别达到79.50%和54.12%,平均F1-score(m F1)分别达到87.46%和69.34%,实验结果验证了本文所提方法有效提升了遥感图像语义分割的效果。

基于MTF-Swin Transformer的风机齿轮箱故障诊断

针对风机齿轮箱实际工况复杂多变及含有强噪声,传统故障诊断方法对风机齿轮箱故障诊断识别准确率较低的问题,文章提出了MTF-Swin Transformer风机齿轮箱故障诊断模型。首先,采用马尔科夫变迁场(MTF)图形编码方法将原始一维振动时序信号转化为具有关联时间信息的二维特征图谱;然后,将特征图谱作为Swin Transformer模型的输入,基于自注意力机制进行自动特征提取;最后,实现对不同故障类型的分类。仿真结果表明,该方法对齿轮箱故障诊断准确率达到了99.48%,证明了该方法的有效性和优越性。

COUPLED NUMERICAL SIMULATION ON COLD ROLLER'S TEMPERATURE FIFLD-PHASE TRANSFORMATION - STRESS FIELD DURING ITS QUENCHING PSOCESS

Complicated changns occur inside the steel parts during quenching process. The abruptly changed boundary conditions make the temperature field,micro - structure and stress field change dramatically in very short time, and these variables take a contact interactions in the whole process. In this paper, a three dimensional non - linear mathematical model for queeching process has been founded and the numerical simulation on temperature field,microstructre and stress field has been realized.In the FEM analysis, the incremental iteration method is used to deal with such complicated nonlinear as boundary nonlinear, physical property nonlinear,transformation nonlinear etc.The effect of stress on transformation kinetics has been considered in the calculation of microstructure. In the stress field anal- ysis,a thermo- elasto - plastic model has been founded, which considers such factors as transforma- tion strain,transformation plastic strain, themal strain and the effect of temperature and transforma- tion on mechanical propertier etc. The transient temperature field, microstructure distribution and stress field of the roller on any time can be displayed vividly,and the cooling curve and the changes of stress on any position can also be given.

CALCULATION OF 3-D TRANSIENT TEMPERATURE FIELD DURING LASER TRANSFORMATION HARDENING PROCESS

A three-dimensional transient heat transfer model for laser transformation hardening process has been developed in this paper. The finite size of the laser treated sample, the surface heat loss of the sample, the latent heat of phase transformation and the temperature dependence of thermal properties of materials were considered. The heat source was considered as a moving Gaussian heat flux with a constant velocity. Three-dimension unequally spatial grid explicit finite difference equations, alternating direction implicit finite difference equations and implicit finite difference equations were deduced respectively. Three programs to calculate the temperature field were developed using Fortran language. The transient temperature fields of C22, 42CrMo, C60 steel samples during laser transformation hardening process were calculated using these programs, and the widths and depths of laser transformation hardening zones were also predicted. C22, 42CrMo, C60 steel samples were treated by CO_2 laser,the widths and depths of laser transformation hardening zones of these samples were also measured experimentally. The calculated widths and depths of laser transformation hardening zones are in good agreement with the experimental results.

Calculation of Temperature Field During Laser Transformation Hardening of Cylindrical Bodies

A Two-dimensional transient heat transfer model for laser transformation hardening process of cylindrical bodies has been developed. The surface heat loss of the sample, the latent heat of phase transformation, and the temperature dependence of thermal properties of material were considered. The laser heat source was considered as a moving Gaussian heat flux with a constant velocity. Finite element method was used to calculate the transient temperature field. A program to calculate the temperature field was developed using FORTRAN language. The transient temperature field, the width and depth of laser transformation hardening zone of 42CrMo steel sample during laser transformation hardening process was calculated. The widths and depths of laser transformation hardening zones of 42CrMo steel samples were also measured experimentally. The calculated widths and depths of laser transformation hardening zones are in good agreement with the experimental results.

A-E equation of potential field transformations in the wavenumber domain and its application

A shift sampling theory established by author （1997a） is a generalization of Fourier transform computation theory. Based on this theory, I develop an Algorithm-Error （A-E） equation of potential field transformations in the wavenumber domain, which not only gives a more flexible algorithm of potential field transformations, but also reveals the law of error of potential field transformations in the wavenumber domain. The DFT0η η（0.5, 0.5） reduction-to-pole （RTP） technique derived from the A-E equation significantly improves the resolution and accuracy of RTP anomalies at low magnetic latitudes, including the magnetic equator. The law （origin, form mechanism, and essential properties） of the edge oscillation revealed by the A-E equation points out theoretically a way of improving the effect of existing padding methods in high-pass transformations in the wavenumber domain.

Mathematical model of wave transformation over radial sand ridge field on continental shelf of South Yellow Sea

According to a deformed mild-slope equation derived by Guang-wen Hong and an enhanced numerical method, a wave refraction-diffraction nonlinear mathematical model that takes tidal level change and the high-order bathymetry factor into account has been developed. The deformed mild-slope equation is used to eliminate the restriction of wave length on calculation steps. Using the hard disk to record data during the calculation process, the enhanced numerical method can save computer memory space to a certain extent, so that a large-scale sea area can be calculated with high-resolution grids. This model was applied to wave field integral calculation over a radial sand ridge field in the South Yellow Sea. The results demonstrate some features of the wave field： （1） the wave-height contour lines are arc-shaped near the shore; （2） waves break many times when they propagate toward the shore; （3） wave field characteristics on the northern and southern sides of Huangshayang are different; and （4） the characteristics of wave distribution match the terrain features. The application of this model in the region of the radial sand ridge field suggests that it is a feasible way to analyze wave refraction-diffraction effects under natural sea conditions.

Semi-analytical investigation of heat transfer in a porous convective radiative moving longitudinal fin exposed to magnetic field in the presence of a shape-dependent trihybrid nanofluid

The thermal examination of a non-integer-ordered mobile fin with a magnetism in the presence of a trihybrid nanofluid(Fe_3O_4-Au-Zn-blood) is carried out. Three types of nanoparticles, each having a different shape, are considered. These shapes include spherical(Fe_3O_4), cylindrical(Au), and platelet(Zn) configurations. The combination approach is utilized to evaluate the physical and thermal characteristics of the trihybrid and hybrid nanofluids, excluding the thermal conductivity and dynamic viscosity. These two properties are inferred by means of the interpolation method based on the volume fraction of nanoparticles. The governing equation is transformed into a dimensionless form, and the Adomian decomposition Sumudu transform method(ADSTM) is adopted to solve the conundrum of a moving fin immersed in a trihybrid nanofluid. The obtained results agree well with those numerical simulation results, indicating that this research is reliable. The influence of diverse factors on the thermal overview for varying noninteger values of γ is analyzed and presented in graphical representations. Furthermore, the fluctuations in the heat transfer concerning the pertinent parameters are studied. The results show that the heat flux in the presence of the combination of spherical, cylindrical, and platelet nanoparticles is higher than that in the presence of the combination of only spherical and cylindrical nanoparticles. The temperature at the fin tip increases by 0.705 759% when the value of the Peclet number increases by 400%, while decreases by 11.825 13% when the value of the Hartman number increases by 400%.

CALCULATION OF RESIDUAL STRESS FIELD IN LASER TRANSFORMATION HARDENED MoCu NODULAR IRON

In this paper,a thermal elastic-plastic 2-D finite element model of stress generation during laser transformation hardening process was developed. In this model, the mechanical properties of the material, Young's modules E, Poisson's ratio v, yield limit s, and thermal expansion coefficient α, are all change with temperature. The equivalent expansion method was used to deal with the problem with phase transformation. Based on this model, a program to calculate the residual stress field was developed using FORTRAN language. The residual stress fields in CO2 laser transformation hardened MoCu nodular iron were calculated. The calculated results showed that in the transformation hardened zone, the residual stress state is compress,whereas adjacent to this zone the residual stress state is tensile, and there is a tensile stress peak close to the transformation hardened zone. To verify this model, the residual stress fields were measured using X-ray diffraction method. The calculated results of residual stress fields are in good agreement with the experimental results.

Calculation of Transient Stress Field During Laser Transformation Hardening Process

A thermal elastic-plastic Two-D finite element model to calculate the transient stress field during laser transformation hardening process was developed in this paper. The mechanical properties of material, Young’s module E, Poisson’s ratio v, yield limit s2 and thermal expansion coefficient a are all considered to change with temperature. The equivalent expansion method was used to deal with the problem with phase transformation. A program to calculate the transient stress field was developed using FORTRAN language. The transient and residual stress fields during CO2 laser transformation hardening process of MoCu nodular iron were calculated. The residual stress fields were measured using X-ray diffraction method. The calculated residual stress field of laser transformation hardened MoCu nodular iron is in good agreement with the experimental result.

Modeling electric field of power metal-oxide-semiconductor field-effect transistor with dielectric trench based on Schwarz–Christoffel transformation

A power metal-oxide-semiconductor field-effect transistor(MOSFET) with dielectric trench is investigated to enhance the reversed blocking capability. The dielectric trench with a low permittivity to reduce the electric field at reversed blocking state has been studied. To analyze the electric field, the drift region is segmented into four regions, where the conformal mapping method based on Schwarz–Christoffel transformation has been applied. According to the analysis, the improvement in the electric field for using the low permittivity trench is mainly due to the two electric field peaks generated in the drift region around this dielectric trench. The analytical results of the electric field and the potential models are in good agreement with the simulation results.

Three-dimensional transformation of magnetization direction and magnetic field component at low latitudes based on vertical relationship

The transformation of the magnetization direction and the magnetic fi eld component is one of the important methods in magnetic data processing and transformation,which can be conducted in both wavenumber and spatial domains.The transformation method in the wavenumber domain has simpler processing expression and higher processing effi ciency than in the spatial domain;however,they are unstable at low latitude.In this paper,the conclusion that the sum is 0 of two vertical magnetic fi eld components(magnetization inclinations are also perpendicular)in 2D is used for the 3D transformation of the magnetization direction and the magnetic field component.In addition,the transformation method at low latitudes based on vertical relationship(VMT)is proposed,which is an iterative algorithm that converts the transformation of the magnetization direction and the magnetic field component at the low latitude into the high latitude.This method restrains the instability of transformation of constant and variable magnetization direction and magnetic fi eld components in low latitudes.The accuracy,stability,and practicality are verifi ed from synthetic models and real data.

Multi-phase field simulation of multi-grain peritectic transition in multiple phase transformation

Taking Fe-C binary alloy as an example,based on the multi-phase field model,the nucleation and growth ofδphase,peritectic reaction,peritectic transformation,and the growth of subsequent austenite are simulated.Effects of the nucleation site of austenite on the peritectic reaction rate and the starting time of the peritectic transformation were studied.The simulation results show that theγphase,as a shell,surrounds theδphase and grows rapidly when the peritectic reaction occurs between the dendriticδgrains,and a layer ofγphase shell is formed aroundδphase after the peritectic reaction.After theδphase is surrounded byγphase completely,the membrane shell separates the L phase from theδphase,so that the phase transfers from peritectic reaction to peritectic transformation.During the peritectic transformation,since the solute diffusion coefficient of the liquid phase is much greater than that of the solid phase,the average growth rate of austenite in the liquid phase is visibly higher than that of theδphase.The peritectic reaction rate is related to the curvature of the nucleation site of theγphase on theδphase grains.The peritectic reaction rate at the large curvatures is faster than that at small curvatures.

Effect of high magnetic field on phase transformation temperature and microstructure of Fe-based Alloys

The effect of a high magnetic field up to 30T on phase transformation temperature and microstructure of Fe-based alloys has been reviewed. A high magnetic field accelerates ferrite transformation, changes the morphology of the transformed microstructures and increases the A3 and A1 temperature. In a magnetic field of 30T, the A1 temperature increases by about 37.1℃ for Fe-0.8C, the A3 temperature for pure Fe increases by about 33.1℃. The measured transformation temperature data are not consistent with calculation results using Weiss molecular field theory. Ferrite grains are elongated and aligned along the direction of magnetic field in Fe-0.4C and Fe-0.6C alloys by ferrite transformation, but elongated and aligned structure was not found in pure Fe, Fe-0.05C alloy and Fe-1.5Mn-0.11C-0.1V alloy.

Phase Transformation and Enhancing Electron Field Emission Properties in Microcrystalline Diamond Films Induced by Cu Ion Implantation and Rapid Annealing

Cu ion implantation and subsequent rapid annealing at 500℃ in N2 result in low surface resistivity of 1.611 ohm/sq with high mobility of 290 cm2 V-1S-1 for microcrystalline diamond （MCD） films. Its electrical field emission behavior can be turned on at Eo = 2.6 V/μm, attaining a current density of 19.5μA/cm2 at an applied field of 3.5 V/#m. Field emission scanning electron microscopy combined with Raman and x-ray photoelectron mi- croscopy reveal that the formation of Cu nanoparticles in MCD films can catalytically convert the less conducting disorder/a-C phases into graphitic phases and can provoke the formation of nanographite in the films, forming conduction channels for electron transportation.

基于CNN与Transformer相融合的心跳分类算法

心电图(Electrocardiogram,ECG)对于心血管疾病的诊断有着重要的作用,精确的心跳分类有助于后续的疾病治疗。但原始的心电图信号存在的大量噪声会影响计算机的判断,而去噪可能会导致信号特征丢失。为此,提出了利用格拉姆角场(Gramian Angular Field,GAF)将一维信号转化为图像,以保证信息完整性,然后利用计算机视觉处理技术分析心跳信号。此外,针对视觉Transformer(Vision Transformer,ViT)在局部特征捕获能力受限的问题,提出了在ViT模型内部和外部引入带残差连接的卷积神经网络(Convolutional Neural Network,CNN),使整体网络能够同时关注图像的局部特征和全局特征。同时,通过将ViT中单一的自注意块改进为双支路通道空间注意力结构,提升了模型的整体分类性能。实验结果表明,所提出的方法有效提高了网络的性能。

基于FPGA的卷积神经网络和视觉Transformer通用加速器

针对计算机视觉领域中基于现场可编程逻辑门阵列(FPGA)的传统卷积神经网(CNN)络加速器不适配视觉Transformer网络的问题,该文提出一种面向卷积神经网络和Transformer的通用FPGA加速器。首先,根据卷积和注意力机制的计算特征,提出一种面向FPGA的通用计算映射方法;其次,提出一种非线性与归一化加速单元,为计算机视觉神经网络模型中的多种非线性和归一化操作提供加速支持;然后,在Xilinx XCVU37P FPGA上实现了加速器设计。实验结果表明,所提出的非线性与归一化加速单元在提高吞吐量的同时仅造成很小的精度损失,ResNet-50和ViT-B/16在所提FPGA加速器上的性能分别达到了589.94 GOPS和564.76 GOPS。与GPU实现相比,能效比分别提高了5.19倍和7.17倍;与其他基于FPGA的大规模加速器设计相比,能效比有明显提高,同时计算效率较对比FPGA加速器提高了8.02%~177.53%。