Power Transformer No-Load Loss Prediction with FEM Modeling and Building Factor Optimization

Estimation of power transformer no-load loss is a critical issue in the design of distribution transformers. Any deviation in estimation of the core losses during the design stage can lead to a financial penalty for the transformer manufacturer. In this paper an effective and novel method is proposed to determine all components of the iron core losses applying a combination of the empirical and numerical techniques. In this method at the first stage all computable components of the core losses are calculated, using Finite Element Method (FEM) modeling and analysis of the transformer iron core. This method takes into account magnetic sheets anisotropy, joint losses and stacking holes. Next, a Quadratic Programming (QP) optimization technique is employed to estimate the incomputable components of the core losses. This method provides a chance for improvement of the core loss estimation over the time when more measured data become available. The optimization process handles the singular deviations caused by different manufacturing machineries and labor during the transformer manufacturing and overhaul process. Therefore, application of this method enables different companies to obtain different results for the same designs and materials employed, using their historical data. Effectiveness of this method is verified by inspection of 54 full size distribution transformer measurement data.

结合坐标Transformer的轻量级人体姿态估计算法

针对现有的大多数自底向上人体姿态估计算法存在模型规模大、计算成本高及对边缘设备不友好等问题,提出了一种基于YOLOv5s6-Pose的轻量级多人姿态估计网络模型YOLOv5s6-Pose-CT。该模型在颈部网络中引入空间和通道重建卷积,以减少空间和通道维度上的特征冗余。同时,提出了一种坐标Transformer嵌入于主干网络中,使模型专注于长距离依赖和拥有高效的局部特征提取能力。其次,通过使用无偏特征位置对齐来解决多尺度融合过程中出现的特征错位问题。最后,使用损失函数MPDIoU对边界框的回归损失重新定义。在COCO 2017数据集上的实验结果表明,本文优化的网络模型与主流的轻量级网络EfficientHRNet-H1模型相比,在保持相同精度的同时,参数量和计算量分别减少16.2%和66.1%。相比于基准模型YOLOv5s6-Pose,参数量减少11.2%,计算量降低5.8%,平均检测精度和平均召回率分别提升2.5%和2.6%。

结合Transformer与生成对抗网络的水下图像增强算法

由于水下环境的多样性和光在水中受到的散射及选择性吸收作用,采集到的水下图像通常会产生严重的质量退化问题,如颜色偏差、清晰度低和亮度低等,为解决以上问题,本文提出了一种基于Transformer和生成对抗网络的水下图像增强算法。以生成对抗网络为基础架构,结合编码解码结构、基于空间自注意力机制的全局特征建模Transformer模块和通道级多尺度特征融合Transformer模块构建了TGAN(generative adversarial network with transformer)网络增强模型,重点关注水下图像衰减更严重的颜色通道和空间区域,有效增强了图像细节并解决了颜色偏差问题。此外,设计了一种结合RGB和LAB颜色空间的多项损失函数,约束网络增强模型的对抗训练。实验结果表明,与CLAHE(contrast limited adaptive histogram equalization)、UDCP(underwater dark channel prior)、UWCNN(underwater based on convolutional neural network)、FUnIE-GAN(fast underwater image enhancement for improved visual perception)等典型水下图像增强算法相比,所提算法增强后的水下图像在清晰度、细节纹理和色彩表现等方面都有所提升,客观评价指标如峰值信噪比、结构相似性和水下图像质量度量的平均值分别提升了5.8%、1.8%和3.6%,有效地提升了水下图像的视觉感知效果。

基于改进双流视觉Transformer的行为识别模型

针对现有行为识别方法中抗背景干扰能力差和准确率低等问题,提出了一种改进的双流视觉Transformer行为识别模型。该模型采用分段采样的方法来增加模型对长时序列数据的处理能力;在网络头部嵌入无参数的注意力模块,在降低动作背景干扰的同时,增强了模型的特征表示能力;在网络尾部嵌入时间注意力模块,通过融合时域高语义信息来充分提取时序特征。文中提出了一种新的联合损失函数,旨在增大类间差异并减少类内差异;采用决策融合层以充分利用光流与RGB流特征。针对上述改进模型,在基准数据集UCF101和HMDB51上进行消融及对比实验,消融实验结果验证了所提方法的有效性,对比实验结果表明,所提方法相比时间分段网络在两个数据集上的准确率分别提高了3.48%和7.76%,优于目前的主流算法,具有较好的识别效果。

基于Swin-Transformer与生成对抗网络的地震随机噪声压制方法

目前深度学习类地震数据去噪方法大多基于卷积神经网络,而此类方法受限于卷积核的局部操作,缺少对地震数据全局特征的分析,因而降低了去噪效果。另外,以L 1,L 2损失函数为指标的网络模型容易出现过度平滑效应,产生虚假同相轴以及虚高的峰值信噪比(PSNR)与结构相似性(SSIM)值。为此,提出一种基于Swin-Transformer(Swin-T)和生成对抗网络的去噪方法(ST-GAN)。该方法以Swin-Transformer作为生成对抗网络中的生成网络对地震数据去噪,判别网络基于卷积神经网络。Transformer的自注意力机制是全局操作,可以有效提取地震数据的全局特征,并能与卷积神经网络的局部操作互补,提升网络模型的特征提取能力。而对抗损失则提升了网络模型的细节恢复能力,有效避免因过度平滑效应产生的同相轴假象。将该方法应用于地震数据去噪,并与现有去噪方法进行对比分析,实验结果表明,该方法具有更加优异的特征提取能力,能在有效压制随机噪声的同时,恢复和保留更多的细节信息,提高了地震信号的信噪比。

多尺度Transformer的在线更新无锚框工件跟踪方法研究

针对工业场景目标工件跟踪任务精度低、失败率高的问题,提出了多尺度Transformer在线更新的工件跟踪算法。首先,采用Transformer特征金字塔结构,融合多层次特征信息,以实现鲁棒的对目标表观建模;其次,使用Transformer模块对高级语义信息进行特征融合,使得网络模型专注于目标工件本身;然后,提出了基于排序的交并化(IoU)损失函数优化策略,有效地抑制干扰物对跟踪器影响;最后,设计一种在线更新策略更新目标模板,增强网络的鲁棒性。实验结果表明,在VOT-2018上准确率和失败率分别比基准跟踪器提高3.8%和4.1%,且能保持53 fps的实时跟踪速度;在LaSOT数据集上精度与成功率别为0.578和0.573,均优于基准跟踪器。通过CCD工业相机采集视频序列验证算法可以准确且鲁棒的跟踪目标工件。

基于Transformer改进的Faster RCNN在复杂环境下的车辆检测

在监控视角中目标车辆较小、遮挡较为严重,导致检测精度低。通过探讨卷积神经网络和Transformer模型的互相借鉴和联系,并结合损失函数等常规改进,提出了新的Faster RCNN模型。通过借鉴Transformer模型的思想,对原有的特征提取网络进行了改进,将原block比例3∶4∶6∶3改为3∶3∶27∶3、卷积核由3×3改为7×7,增大其感受野,能够更好捕捉图像中的全局特征,使用DW卷积来减少参数量并略微提高性能,使用Channel shuffle解决通道间信息不交流的问题。将原先交并比IoU改为CIoU,与改进后的特征提取网络结合,进一步提高小目标和遮挡目标的检测效果。在UA-DETRAC数据集上,改进后的模型在mAP@0.5:0.95方面比原算法提高了20.20%,并在大、中、小目标下分别提高了15.8%、23%和45.8%,相较于其他模型,如YO⁃LOv7、YOLOv5和Cascade RCNN,mAP@0.5:0.95分别提高了3.3%、5%和6.69%。

A Swin Transformer and Residualnetwork Combined Model for Breast Cancer Disease Multi-Classification Using Histopathological Images

Breast cancer has become a killer of women's health nowadays.In order to exploit the potential representational capabilities of the models more comprehensively,we propose a multi-model fusion strategy.Specifically,we combine two differently structured deep learning models,ResNet101 and Swin Transformer(SwinT),with the addition of the Convolutional Block Attention Module(CBAM)attention mechanism,which makes full use of SwinT's global context information modeling ability and ResNet101's local feature extraction ability,and additionally the cross entropy loss function is replaced by the focus loss function to solve the problem of unbalanced allocation of breast cancer data sets.The multi-classification recognition accuracies of the proposed fusion model under 40X,100X,200X and 400X BreakHis datasets are 97.50%,96.60%,96.30 and 96.10%,respectively.Compared with a single SwinT model and ResNet 101 model,the fusion model has higher accuracy and better generalization ability,which provides a more effective method for screening,diagnosis and pathological classification of female breast cancer.

基于双交叉注意力Transformer网络的小样本图像语义分割

小样本图像语义分割只用少量样本就能分割出新类别。针对现有方法中语义信息挖掘不充分的问题,本文提出一种基于双交叉注意力网络的小样本图像语义分割方法。该方法采用Transformer结构,利用双交叉注意力模块同时从通道和空间维度上学习多尺度查询特征和支持特征的远程依赖性。首先,本文提出通道交叉注意力模块,并结合位置交叉注意力模块构成双交叉注意力模块。其中,通道交叉注意力模块用于学习查询和支持特征之间的通道语义相互关系,位置交叉注意力模块用来捕获查询和支持特征之间的远程上下文相关性。然后,通过多个双交叉注意力模块能够为查询图像提供包含丰富语义信息的多尺度交互特征。最后,本文引入辅助监督损失,并通过上采样和残差连接将多尺度交互特征连接至解码器以得到准确的新类分割结果。本文方法在数据集PASCAL-5i上的mIoU达到了69.9%(1-shot)和72.4%(5-shot),在数据集COCO-20i上的mIoU达到了48.9%(1-shot)和54.6%(5-shot)。与主流方法相比,本文方法的分割性能达到了最先进的水平。

融合Swin Transformer与UNet的云检测架构

目前在遥感图像云检测的领域中,对于薄云和碎云以及不同下垫面的云检测精度还有待进一步提高,且存在较多误判漏判和数据不均衡等问题。针对这些问题,提出了一个结合Transformer和卷积思想的云检测网络架构Cloud TransUnet。首先,使用Transformer模块替换掉UNet编码器阶段的卷积模块,利用Transformer的全局注意力机制提取出更多的细节特征信息,减少误判和漏判的发生;同时,添加边缘检测模块,加强对边缘信息的提取,提高检测的精度;最后,使用中值频率平衡对损失函数进行改进以处理数据不平衡的问题。实验结果表明,Cloud TransUnet对薄云和碎云的检测、不同下垫面的云检测都具有良好的表现,与卷积语义分割网络UNet、SegNet、ResUnet和Swin_Unet相比,Cloud TransUnet在检测速度略微提升的同时,其MIoU、总体精度、精确率等指标都得到了一定的提升。

改进视觉Transformer的人脸识别方法

目前大多数人脸识别方法依赖于卷积神经网络,通过级联的形式构建多层处理单元,利用卷积操作融合局部特征,忽略了人脸全局语义信息,缺乏对人脸重点区域的关注度。针对上述问题,提出一种基于改进视觉Transformer的人脸识别方法,引入Shuffle Transformer作为特征提取骨干网络,通过自注意力机制以及Shuffle操作捕捉特征图全局信息,建立特征点之间的长距离依赖关系,提高模型的特征感知能力;同时,结合ArcFace损失函数和中心损失函数的特点,设计融合损失作为目标函数,利用类内约束扩大角度间隔,提高特征空间的辨别性。该方法在LFW、CALFW、CPLFW、AgeDB-30和CFP五个具有挑战性的基准测试人脸数据集上分别取得了99.83%、95.87%、90.05%、98.05%、97.23%的平均准确率,能够有效提升人脸特征提取能力,识别效果优于同等规模卷积神经网络。

基于Transformer的暂态稳定评估模型的可解释性分析与模型更新研究

深度学习算法在电力系统暂态稳定性评估问题上有着优秀的表现,然而模型评估结果的不可知性与决策过程的不可控性阻碍了其在实际中进一步的应用。构建了基于Transformer编码器的暂态稳定评估模型,尝试通过模型对于特征量的注意力权重,解释和分析模型所关注和学习到的规则。在此基础上,结合可解释性结果提出了一种利用物理信息指导模型优化的模型更新方法。从模型的损失函数出发,通过微调的方式修正模型对特征量的注意力权重分布,加强对于样本失稳模式的挖掘。在微调模型的过程中,引入注意力引导函数提高对特定失稳模式关键机组的关注度,以减少对于特定失稳模式样本的误判,进一步提高整体的预测精度。在IEEE39节点系统和华东电网系统的仿真均验证了该方法的有效性。

融合Transformer和CNN的手掌静脉识别网络

针对手掌静脉特征提取识别精度不高问题,提出了掌静脉识别网络PVCodeNet。该网络设计了改进的BasicBlock和Transformer Encoder模块结合并运用扩大决策边界的损失函数AAM-Loss(additive angular margin loss)。该网络首次将Transformer Encoder模块成功用于掌静脉图像全局特征提取,改进的BasicBlock使用深度超参数化卷积Do-Conv取代传统卷积Conv进行特征提取使提取的特征更加具有区分性,该模块还加入规一化的注意力机制NAM模块,通过应用权重稀疏性惩罚项抑制不显著性特征的权值来提取图像在通道和空间域上重要的细节特征。在手掌关键点定位、ROI提取、图像增强方面作了详细描述,在特征向量维度、AAM-Loss参数设置方面做了详细实验,在PolyU数据库和自建库SEPAD-PV数据库上进行消融实验测试,EER均达到了0,成功实现了最高识别率的突破。为了验证该网络的泛化性能,还在具有相似纹理特征的掌纹数据库Tongji和指静脉数据库SDUMLA上进行验证,EER远远优于其他主流算法,充分证明了提出算法的优越性。

DT-Net:Joint Dual-Input Transformer and CNN for Retinal Vessel Segmentation

Retinal vessel segmentation in fundus images plays an essential role in the screening,diagnosis,and treatment of many diseases.The acquired fundus images generally have the following problems:uneven illumination,high noise,and complex structure.It makes vessel segmentation very challenging.Previous methods of retinal vascular segmentation mainly use convolutional neural networks on U Network(U-Net)models,and they have many limitations and shortcomings,such as the loss of microvascular details at the end of the vessels.We address the limitations of convolution by introducing the transformer into retinal vessel segmentation.Therefore,we propose a hybrid method for retinal vessel segmentation based on modulated deformable convolution and the transformer,named DT-Net.Firstly,multi-scale image features are extracted by deformable convolution and multi-head selfattention(MHSA).Secondly,image information is recovered,and vessel morphology is refined by the proposed transformer decoder block.Finally,the local prediction results are obtained by the side output layer.The accuracy of the vessel segmentation is improved by the hybrid loss function.Experimental results show that our method obtains good segmentation performance on Specificity(SP),Sensitivity(SE),Accuracy(ACC),Curve(AUC),and F1-score on three publicly available fundus datasets such as DRIVE,STARE,and CHASE_DB1.

A Discrete State Event Driven Simulation based Losses Analysis for Multi-terminal Megawatt Power Electronic Transformer

At present,power electronic transformers(PETs)have been widely used in power systems.With the increase of PET capacity to the megawatt level.the problem of increased losses need to be taken seriously.As an important indicator of power electronic device designing,losses have always been the focus of attention.At present,the losses are generally measured through experiments,but it takes a lot of time and is difficult to quantitatively analyze the internal distribution of PET losses.To solve the above problems,this article first qualitatively analyzes the losses of power electronic devices and proposes a loss calculation method based on pure simulation.This method uses the Discrete State Event Driven(DSED)modeling method to solve the problem of slow simulation speed of large-capacity power electronic devices and uses a loss calculation method that considers the operating conditions of the device to improve the calculation accuracy.For the PET prototype in this article,a losses model of the PET is established.The comparison of experimental and simulation results verifies the feasibility of the losses model.Then the losses composition of PET was analyzed to provide reference opinions for actual operation.It can help pre-analyze the losses distribution of PET,thereby providing a potential method for improving system efficiency.

可形变Transformer辅助的胸部X光影像疾病诊断模型

针对胸部X光影像中的灰雾现象、病变区域重叠等问题,提出可形变Transformer辅助的胸部X光影像疾病诊断模型.将扩展后的ResNet50作为特征提取网络,添加压缩型双注意力模块,增强病变区域与非病变区域之间的特征差异,降低冗余信息的干扰,提高图像数据的特征提取效果;通过可形变Transformer解码器内部的交叉注意力模块,引入类别表征作为先验知识,引导影像特征进一步融合,提高不同疾病在影像区域重叠情况下的特征区分度;将解码器的输出传入分类器中以获得最终的诊断结果.压缩型双注意力模块和可形变Transformer均起到降低模型计算复杂度的作用,引入非对称损失函数可以更好地解决正负样本不均衡.利用所提模型在公开数据集ChestX-Ray14和CheXpert上进行多组实验,在2个数据集上的受试者操作的特征曲线下面积值(AUC)分别达到0.8398和0.9061,表明该模型在胸部X光影像的疾病诊断方面具有正确性和有效性.

Unifying Convolution and Transformer Decoder for Textile Fiber Identification

At present,convolutional neural networks(CNNs)and transformers surpass humans in many situations(such as face recognition and object classification),but do not work well in identifying fibers in textile surface images.Hence,this paper proposes an architecture named FiberCT which takes advantages of the feature extraction capability of CNNs and the long-range modeling capability of transformer decoders to adaptively extract multiple types of fiber features.Firstly,the convolution module extracts fiber features from the input textile surface images.Secondly,these features are sent into the transformer decoder module where label embeddings are compared with the features of each type of fibers through multi-head cross-attention and the desired features are pooled adaptively.Finally,an asymmetric loss further purifies the extracted fiber representations.Experiments show that FiberCT can more effectively extract the representations of various types of fibers and improve fiber identification accuracy than state-of-the-art multi-label classification approaches.

跨级可变形Transformer编解码视网膜图像分割算法

眼底视网膜血管图像分割对青光眼、糖尿病等疾病的预防和诊断具有重要意义。针对视网膜血管图像边缘分割模糊、微细血管漏缺和模型感受野不足等问题,提出了一种跨级可变形Transformer编解码(Cross-stage Deformable Transformer Encoding and Decoding Net,CTED-Net)视网膜图像分割算法。在特征提取网络中融入通道像素增强模块和跨级融合骨干,实现对视网膜血管全局特征的粗提取;在网络编码部分加入可变形自适应编码Transformer模块(Deformable Adaptive Coding Transformer Module,DACT),通过可变形编码的方式增大模型感受野;在编解码结构底层加入深层语义门控注意模块,实现对视网膜血管深层特征的充分学习,以改善血管图像边缘分割模糊的问题。在模型训练阶段采用加权交叉焦点损失函数,克服视网膜血管图像样本不平衡的问题。在公共数据集DRIVE和STARE上进行仿真实验,所提算法灵敏度、特异性、准确率和AUC指标在DRIVE上达到84.25%、98.17%、96.46%和98.70%,在STARE上达到80.22%、98.64%、96.71%和98.78%。通过与其他先进算法对比分析可以看出,所提算法分割效果可靠且整体性能先进。

Maximum Inner Product Method for Extracting the Path-Loss Parameters in Primary Transformer Substations

In order to deploy short-range wireless communication devices in the primary transformer substation, a Maximum Inner Product(MIP) Method is proposed to extract the path-loss parameters in 110 kV and 220 kV primary transformer substations. The maximum inner product of the testing data is calculated to find out the loss index n and the standard deviation σ, and then the pathloss models can be set up. By comparing the MIP with Minimum Mean Square estimation(MMSE) and Cumulative Sum(CUSUM), MIP can match the measured values best. In order to apply the MIP path-loss model, under the initial signal to noise ratio(SNR) at 5 dB and 10 dB, a ZigBee simulation system is constructed to validate the situation that bit error rate(BER) varies with distance. And the ZigBee devices with 5 units are tested in a 220 kV primary transformer substation. The result of the test proves that the path-loss model is accurate.

LSTFormer:基于Swin Transformer的轻量化语义分割网络

针对现有基于Transformer的语义分割网络普遍存在计算复杂度高的问题,提出了一种基于Swin Transformer的轻量化语义分割网络。该网络通过Swin Transformer获取多个尺度的特征图;采用全感知模块和改进的级联融合模块跨层融合不同尺度的特征图,减小不同层级特征图的语义差距;引入单个Swin Transformer block对初分割特征映射进行优化,通过移动窗口自注意力机制提升网络对不同像素点进行分类的能力;训练阶段加入Dice损失函数和交叉熵损失函数,提高网络的分割性能和收敛速度。实验结果表明,LSTFormer在数据集ADE20K和Cityscapes上mIoU分别达到49.47%和81.47%,相较于SETR和Swin-UPerNet等同类网络,LSTFormer在保持相当分割精度的同时具有更低的参数量和计算量。