Comparison on Zhang Neural Network and Gradient Neural Network for
219506089_基于自组织映射的卷积神经网络架构研究
第10期2023年5月无线互联科技Wireless Internet TechnologyNo.10May,2023作者简介:赵义爱(1989 ),女,河南郑州人,硕士;研究方向:物联网㊂基于自组织映射的卷积神经网络架构研究赵义爱(郑州工业应用技术学院,河南郑州451100)摘要:辅助驾驶和自动驾驶技术将对人类的生活方式带来巨大影响,而交通标志识别技术则是其中至关重要的一环㊂为了进一步完善交通标志识别理论,文章提出了一种融合自组织映射的卷积神经网络架构㊂自组织映射能将图像样本量化至拓扑空间中,从而对微小的变化提供降维和不变性处理㊂该方法与卷积神经网络相结合,能充分利用卷积神经网络对平移㊁旋转㊁缩放和形变的部分不变性能,从而提高该架构的效率和准确度㊂经过初步测试,本系统在200个测试样本中表现出98.5%的准确率,取得了显著的成效㊂关键词:卷积神经网络;交通标志识别;自组织映射;深度学习中图分类号:TP183㊀㊀文献标志码:A 0㊀引言㊀㊀辅助驾驶和自动驾驶技术需要快速准确地从图像中检测交通标志㊂交通标志识别主要依赖于图像视觉信息,如标志的形状㊁大小和颜色等[1]㊂然而,传统的识别算法在实时测试中面临着照明强度㊁摄像头角度㊁障碍物等问题㊂此外,实现多目标检测也是目前技术面临的困难之一[2]㊂虽然深度学习作为一种机器学习方法很早就被提出,但近年来随着计算机硬件和神经网络架构的不断发展,深度学习才开始进入发展黄金时期[3]㊂卷积神经网络(Convolutional Neural Networks,CNN)是一种重要的深度学习架构,可以实现对人脸或其他图像的学习和识别[4],在人脸识别[5]㊁自动驾驶汽车[6]和智能医疗[7]等领域得到广泛应用㊂当采用CNN 模型时,并不需要太多的数据预处理任务,它主要是通过卷积完成图像特征的提取,同时不会丢失重要信息,在降维的同时也保留了与特征相关的信息[8-9]㊂因此,本研究系统性地研究了基于自组织映射(Self -organizing Map,SOM)的CNN 网络架构来实现交通标志识别㊂主要包括3个方面:对交通标志数据集进行分析和预处理㊁介绍CNN 在图像识别中的应用以及提出SOM -CNN 架构并进行数据集地训练和测试㊂实验结果表明,本研究工作具有一定的发展潜力㊂1㊀针对交通标志识别的CNN 架构1.1㊀CNN 的典型结构㊀㊀典型的CNN 由多个层组成,主要包含卷积层和池化层,如图1所示[10]㊂其中,卷积层包含有多个面,每个面都具有固定的特征检测器,与前一层的局部窗口做卷积㊂池化层在卷积层后面,用于进行局部平均和下采样操作,减少数据量,同时保留重要特征㊂在交通标志识别任务中,CNN 已被证明可以有效地识别交通标志的形状㊁颜色㊁符号等㊂该模型的成功应用主要是因为其卷积层可以学习低级特征,例如边缘和角点等,然后在池化层中进一步提取和减少特征,最终在全连接层中组合这些特征,形成对交通标志的分类和识别㊂此外,通过使用反向传播梯度下降法进行训练,可以进一步优化网络中的连接策略,从而减少网络中的权重数量,提高模型的精度和效率㊂1.2㊀SOM -CNN㊀㊀为了进一步提高CNN 架构的效率和精度,本文引入了SOM 来优化整个系统,提出了SOM -CNN 架构,如图2所示㊂该方法能将图像样本投影到量化的低维空间的SOM 上,实现局部图像采样和部分光照不变性技术的使用㊂具体研究工作包括:(1)对于集中训练的图像,在整个图像上设置固定大小的窗口(例如5ˑ5),并在每一步提取局部图像样本,其中每步中窗口移动4个像素点;(2)SOM 的3个维度可以被认为是3个特征,在图1㊀典型的CNN架构前一阶段的向量上进行训练时,SOM将25维输入向量量化为125个拓扑值;(3)在训练集和测试集中的所有图像上都会出现与(1)中相同的窗口㊂局部图像样本在每个步骤都通过SOM,从而在SOM创建的输出空间中创建新㊀㊀的训练和测试集㊂此时,每个输入图像由3个映射表示,每个映射对应于SOM中的维度㊂这些映射的大小等于输入图像的大小除以步长;(4)采用新创建的训练集训练CNN网络㊂图2㊀引入SOM的CNN架构2㊀实验设计和结果2.1㊀实验设计㊀㊀本实验采用了比利时交通标志数据集(BelgiumTraffic Signs Dataset),该数据集包括警示标志㊁优先通行标志㊁禁止通行标志㊁强制通行标志㊁停车路牌㊁指定通行标志等6大类,训练和测试数据文件夹包含62个子文件夹,所有图像的格式均为ppm㊂因此,本实验的任务是将给定图像分类为表示交通标志面板的62个类别之一㊂实验环境搭建基于Python的Tensorflow㊂Python是一种常用的编程语言,在机器学习中广泛使用㊂本实验使用的Python模块包含一个名为scikit-learntool的模块,该模块集成了大量用于监督和非监督问题的机器学习算法㊂Tensorflow则是一个多用途开源库,可以在Python,C++,Java,Scala,R等多种编程语言中使用,并可以在Unix,Windows,iOS和Android等平台上运行㊂Keras是TensorFlow的官方高级API,用于提供开发接口,其模型制作简单,支持卷积神经网络和递归神经网络以及两者的组合,支持任意连接方案(包括多输入和多输出训练)㊂2.2㊀实验结果㊀㊀经过多组实验,本文利用SOM-CNN模型对交通标志进行分类识别㊂每次实验分别具有200张训练图像,200张测试图像,并且训练和测试集之间没有重叠㊂为了比较训练和执行时间,本实验使用了NVIDIA GeForce RTX2060和3080进行对比实验,如表1所示㊂表1㊀SOM-CNN的训练时间和分类时间硬件平台训练时间/min分类时间/ms2060平均14平均0.213080平均29平均0.35实验表明,该模型对交通标志识别效果良好㊂在多次实验中,平均每200张测试图像种有3张被错误分类,其正确率约为98.5%,表现出很好的分类效果㊂3㊀结语㊀㊀交通标志检测通常基于机器学习方法,而深度神经网络的出现进一步提升了其分类精度㊂深度神经网络基于许多简单互连的神经元,可以从大量数据中提取有意义的特征以解决复杂的分类问题㊂为了进一步提高交通标志识别的准确率,文章结合现有理论开发了SOM-CNN架构,并对交通标志数据集进行了分类,取得了良好的效果㊂尽管其准确率与主流方法相比尚有差距,但其高效性使其具有一定的发展潜力㊂未来,研究人员可以通过改进结构来进一步提高SOM-CNN模型的效率和精度,并扩展模型以检测其他对象,如行人㊁动物和其他复杂障碍物,以期在实际应用中取得更好的效果㊂参考文献[1]SAADNA Y,BEHLOUL A.An overview of traffic sign detection and classification methods[J]. International Journal of Multimedia Information Retrieval,2017(6):193-210.[2]HE Z,NAN F,LI X,et al.Traffic sign recognition by combining global and local features based on semi‐supervised classification[J].IET Intelligent Transport Systems,2020(5):323-330.[3]LECUN Y,BENGIO Y,HINTON G.Deep learning [J].Nature,2015(521):436-444.[4]ALZUBAIDI L,ZHANG J,HUMAIDI A J,et al. Review of deep learning:concepts,CNN architectures, challenges,applications,future directions[J].Journal of Big Data,2021(8):71-74.[5]KASAR M M,BHATTACHARYYA D,KIM T H. Face recognition using neural network:a review[J]. International Journal of Security and Its Applications,2016(3):81-100.[6]ALAM A,PRAVEEN S.A review of automatic driving system by recognizing road signs using digital image processing[J].Journal of Informatics Electrical and Electronics Engineering(JIEEE),2021(2):1-9.[7]SINGH S P,WANG L,GUPTA S,et al.3D deep learning on medical images:a review[J].Sensors,2020 (18):5097.[8]LI Z,LIU F,YANG W,et al.A survey of convolutional neural networks:analysis,applications, and prospects[J].IEEE Transactions on Neural Networks and Learning Systems,2021(4):1-21. [9]HINTON G E.To recognize shapes,first learn to generate images[J].Progress in Brain Research,2007 (165):535-547.[10]HAJI S H,ABDULAZEEZ A parison of optimization techniques based on gradient descent algorithm:a review[J].PalArch s Journal of Archaeology of Egypt/Egyptology,2021(4):2715-2743.(编辑㊀王雪芬)Research on convolutional neural network architecture based on self-organizing mapZhao Yi aiZhengzhou University of Industrial Technology Zhengzhou451100 ChinaAbstract Assisted driving and automatic driving technologies are poised to greatly impact human lifestyle with traffic sign recognition technology representing a pivotal aspect thereof.To refine the theoretical underpinnings of traffic sign recognition a convolutional neural network framework combined with self-organizing map algorithm is proposed.And self-organizing map algorithm can quantize image samples into topological space thus providing dimensionality reduction and invariance processing for small changes in image samples.This method combined with the partial invariant performance of convolution neural network for translation rotation scaling and deformation improves the efficiency and accuracy of this architecture.The experimental results show that the accuracy of proposed system is 98.5%in200test samples representing a highly promising outcome.Key words convolution neural network traffic sign recognition self-organizing map deep learning。
图神经网络综述
第47卷第4期Vol.47No.4计算机工程Computer Engineering2021年4月April 2021图神经网络综述王健宗,孔令炜,黄章成,肖京(平安科技(深圳)有限公司联邦学习技术部,广东深圳518063)摘要:随着互联网和计算机信息技术的不断发展,图神经网络已成为人工智能和大数据处理领域的重要研究方向。
图神经网络可对相邻节点间的信息进行有效传播和聚合,并将深度学习理念应用于非欧几里德空间的数据处理中。
简述图计算、图数据库、知识图谱、图神经网络等图结构的相关研究进展,从频域和空间域角度分析与比较基于不同信息聚合方式的图神经网络结构,重点讨论图神经网络与深度学习技术相结合的研究领域,总结归纳图神经网络在动作检测、图系统、文本和图像处理任务中的具体应用,并对图神经网络未来的发展方向进行展望。
关键词:图神经网络;图结构;图计算;深度学习;频域;空间域开放科学(资源服务)标志码(OSID ):中文引用格式:王健宗,孔令炜,黄章成,等.图神经网络综述[J ].计算机工程,2021,47(4):1-12.英文引用格式:WANG Jianzong ,KONG Lingwei ,HUANG Zhangcheng ,et al.Survey of graph neural network [J ].Computer Engineering ,2021,47(4):1-12.Survey of Graph Neural NetworkWANG Jianzong ,KONG Lingwei ,HUANG Zhangcheng ,XIAO Jing(Federated Learning Technology Department ,Ping An Technology (Shenzhen )Co.,Ltd.,Shenzhen ,Guangdong 518063,China )【Abstract 】With the continuous development of the computer and Internet technologies ,graph neural network has become an important research area in artificial intelligence and big data.Graph neural network can effectively transmit and aggregate information between neighboring nodes ,and applies the concept of deep learning to the data processing of non-Euclidean space.This paper briefly introduces the research progress of graph computing ,graph database ,knowledge graph ,graph neural network and other graph-based techniques.It also analyses and compares graph neural network structures based on different information aggregation modes in the spectral and spatial domain.Then the paper discusses research fields that combine graph neural network with deep learning ,and summarizes the specific applications of graph neural networks in action detection ,graph systems ,text and image processing tasks.Finally ,it prospects the future development research directions of graph neural networks.【Key words 】graph neural network ;graph structure ;graph computing ;deep learning ;spectral domain ;spatial domain DOI :10.19678/j.issn.1000-3428.00583820概述近年来,深度学习技术逐渐成为人工智能领域的研究热点和主流发展方向,主要应用于高维特征规则分布的非欧几里德数据处理中,并且在图像处理、语音识别和语义理解[1]等领域取得了显著成果。
研究生英语论文写作
(2) All the experiments draw on the observations made during the First Global GARP Experiment (FGGE) of 1979, when an intensive international effort acquired as complete a set of global observations as was ever made before or since, in each set of experiments, forecasts were made drawing on all of the observing systems available----satellite temperatures, satellite cloud-track winds, winds and temperatures from commercial aircraft, surface observations (land and ship), and balloon observations. Then new sets of forecasts drew on this maximum system with one or more observing systems deleted from it. -----―Weather Satellites Coming of Age‖, Science, Vol.229. 1. Complexity Written language is relatively more complex than spoken language. Written texts are lexically dense compared to spoken languagethey have proportionately more lexical words than grammatical words. Writ-ten texts are shorter and have longer, more complex words and phrases. They have more noun-based phrases, more nominalizations, and more lexical variation. Written language is grammatically more complex than spoken language. It has more subordinate clauses, more ―that/to‖ complement clauses, more long sequences of prepositional phrases, more attributive adjectives and more passives than spoken language. 1.2.2 Formality Academic writing is relatively formal. Formal writing doesn‘t use contractions, or colloquialisms and slang. In general this means that in an essay you should avoid colloquial words and expressions.
模拟ai英文面试题目及答案
模拟ai英文面试题目及答案模拟AI英文面试题目及答案1. 题目: What is the difference between a neural network anda deep learning model?答案: A neural network is a set of algorithms modeled loosely after the human brain that are designed to recognize patterns. A deep learning model is a neural network with multiple layers, allowing it to learn more complex patterns and features from data.2. 题目: Explain the concept of 'overfitting' in machine learning.答案: Overfitting occurs when a machine learning model learns the training data too well, including its noise and outliers, resulting in poor generalization to new, unseen data.3. 题目: What is the role of a 'bias' in an AI model?答案: Bias in an AI model refers to the systematic errors introduced by the model during the learning process. It can be due to the choice of model, the training data, or the algorithm's assumptions, and it can lead to unfair or inaccurate predictions.4. 题目: Describe the importance of data preprocessing in AI.答案: Data preprocessing is crucial in AI as it involves cleaning, transforming, and reducing the data to a suitableformat for the model to learn effectively. Proper preprocessing can significantly improve the performance of AI models by ensuring that the input data is relevant, accurate, and free from noise.5. 题目: How does reinforcement learning differ from supervised learning?答案: Reinforcement learning is a type of machine learning where an agent learns to make decisions by performing actions in an environment to maximize a reward signal. It differs from supervised learning, where the model learns from labeled data to predict outcomes based on input features.6. 题目: What is the purpose of a 'convolutional neural network' (CNN)?答案: A convolutional neural network (CNN) is a type of deep learning model that is particularly effective for processing data with a grid-like topology, such as images. CNNs use convolutional layers to automatically and adaptively learn spatial hierarchies of features from input images.7. 题目: Explain the concept of 'feature extraction' in AI.答案: Feature extraction in AI is the process of identifying and extracting relevant pieces of information from the raw data. It is a crucial step in many machine learning algorithms, as it helps to reduce the dimensionality of the data and to focus on the most informative aspects that can be used to make predictions or classifications.8. 题目: What is the significance of 'gradient descent' in training AI models?答案: Gradient descent is an optimization algorithm used to minimize a function by iteratively moving in the direction of steepest descent as defined by the negative of the gradient. In the context of AI, it is used to minimize the loss function of a model, thus refining the model's parameters to improve its accuracy.9. 题目: How does 'transfer learning' work in AI?答案: Transfer learning is a technique where a pre-trained model is used as the starting point for learning a new task. It leverages the knowledge gained from one problem to improve performance on a different but related problem, reducing the need for large amounts of labeled data and computational resources.10. 题目: What is the role of 'regularization' in preventing overfitting?答案: Regularization is a technique used to prevent overfitting by adding a penalty term to the loss function, which discourages overly complex models. It helps to control the model's capacity, forcing it to generalize better to new data by not fitting too closely to the training data.。
联合训练生成对抗网络的半监督分类方法
光学 精密工程Optics and Precision Engineering第 29 卷 第 5 期2021年5月Vol. 29 No. 5May 2021文章编号 1004-924X( 2021)05-1127-09联合训练生成对抗网络的半监督分类方法徐哲,耿杰*,蒋雯,张卓,曾庆捷(西北工业大学电子信息学院,西安710072)摘要: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,旦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°关键词:生成对抗网络;半监督学习;图像分类;深度学习中图分类号:TP391文献标识码:Adoi :10. 37188/OPE. 20212905.1127Co -training generative adversarial networks forsemi -supervised classification methodXU Zhe , GENG Jie * , JIANG Wen , ZHANG Zhuo , ZENG Qing -jie(School of E lectronics and Information , Northwestern Polytechnical University , Xian 710072, China )* Corresponding author , E -mail : gengjie@nwpu. edu. cnAbstract : Deep neural networks require a large amount of data for supervised learning ; however , it is dif ficult to obtain enough labeled data in practical applications. Semi -supervised learning can train deep neuralnetworks with limited samples. Semi -supervised generative adversarial networks can yield superior classifi cation performance ; however , they are unstable during training in classical networks. To further improve the classification accuracy and solve the problem of training instability for networks , we propose a semi -su pervised classification model called co -training generative adversarial networks ( CT -GAN ) for image clas sification. In the proposed model , co -training of two discriminators is applied to eliminate the distribution error of a single discriminator and unlabeled samples with higher confidence are selected to expand thetraining set , which can be utilized for semi -supervised classification and enhance the generalization of deep networks. Experimental results on the CIFAR -10 dataset and the SVHN dataset showed that the pro posed method achieved better classification accuracies with different numbers of labeled data. The classifi cation accuracy was 80. 36% with 2000 labeled data on the CIFAR -10 dataset , whereas it improved by收稿日期:2020-11-04;修订日期:2021-01-04.基金项目:装备预研领域基金资助项目(No. 61400010304);国家自然科学基金资助项目(No. 61901376)1128光学精密工程第29卷about5%compared with the existing semi-supervised method with10labeled data.To a certain extent, the problem of GAN overfitting under a few sample conditions is solved.Key words:generative adversarial networks;semi-supervised learning;image classification;deep learning1引言图像分类作为计算机视觉领域最基础的任务之一,主要通过提取原始图像的特征并根据特征学习进行分类[11o传统的特征提取方法主要是对图像的颜色、纹理、局部特征等图像表层特征进行处理实现的,例如尺度不变特征变换法[21,方向梯度法[31以及局部二值法[41等。
aigc检测相似度的原理
aigc检测相似度的原理AIGC, or Artificial Intelligence Graphic Comparison, is a technology that is used to detect and measure the similarity between two or more images. The principle behind AIGC is based on the use of advanced algorithms and machine learning techniques to analyze and compare the visual features of images. This technology has variousapplications in different fields, including image recognition, content-based image retrieval, and copyright protection.One of the key principles behind AIGC is the use of deep learning algorithms, such as convolutional neural networks (CNNs), to extract and analyze the visual features of images. CNNs are a type of artificial neural networkthat is particularly well-suited for image recognition tasks, as they are able to automatically learn and extract hierarchical features from raw pixel data. By using CNNs, AIGC is able to effectively capture and analyze the visual content of images, allowing for accurate and reliablesimilarity measurements.Another important principle behind AIGC is the use of similarity metrics, such as cosine similarity or Euclidean distance, to quantitatively measure the similarity between images. These metrics allow AIGC to compare the visual features of images and determine how similar or dissimilar they are. By using these metrics, AIGC is able to provide a numerical measure of similarity, which can be used for various applications, such as image retrieval or copyright infringement detection.In addition to deep learning algorithms and similarity metrics, AIGC also relies on large-scale image datasets for training and evaluation. By using these datasets, AIGC is able to learn and extract a wide range of visual features, allowing for robust and accurate similarity measurements. Furthermore, the use of large-scale datasets also enables AIGC to generalize well to new and unseen images, making it a powerful tool for various real-world applications.From a practical perspective, AIGC works by firstextracting visual features from the input images using deep learning algorithms. These features are then compared using similarity metrics to determine the degree of similarity between the images. The result is a quantitative measure of similarity, which can be used for various purposes, such as image search, content-based recommendation systems, or copyright protection.Overall, the principle behind AIGC is based on the use of advanced algorithms, machine learning techniques, and large-scale image datasets to analyze and compare thevisual features of images. By leveraging these principles, AIGC is able to provide accurate and reliable similarity measurements, making it a valuable tool for various applications in different fields.。
基于多尺度各向异性高斯核主方向角度变化的角点检测
2021年第05期(总第221期)《蓿窟画蓿基于多尺度各向异性高斯核主方向角度变化的角点检测王天赋,任劫,章为川,晁凯(西安工程大学大学电子信息学院,陕西西安710048)摘要:角点检测在图像分析和计算机视觉领域有着及其重要的作用。
各向异性高斯方向导数(AG D Ds)可以很好的提取 不同方向的图像局部灰度变化信息,并且具有很强的噪声鲁棒性,在对不同尺度下的轮康信息和各向异性高斯方向导数 进行研究后,将两者相结合提出了一种基于多尺度各向异性高斯核主方向角度变化的角点检测算法。
该算法可以有效 的降低角点检测的误检率。
实验采用了两幅标注了真实角点测试图,在无噪声和加入不同等级噪声的情况下,对该算法 和其它三种经典的角点检测算法从角点检测的能力和角点定位的精度进行对比。
实验表明,研究所得的算法对于角点 的检测具有更好的噪声稳健性和更低的误检率。
关键词:角点检测;图像轮摩;各向异性高斯核;多尺度;鲁棒性中图分类号:TP391.41 文献标识码:B文章编号=2096-9759(2021)05-0032-04A muHkeale coma-detection v ia ai^ledifi»iHice ofpnnc^le directions of a nis»tropk Gaussian daivativesW ang T ianfu,Ren Jie9Zhang W eichuan,Chao kai(School o f Electronics and Inform ation,X i;an Polytechnic University,X i^an710048, China)A bstract:Comer detection plays an important role in image analysis and computer vision.Anisotropic Gaussian directionalderivatives(AGDDs)can extract the local gray change inform ation o f images in different directions,and has strong noise ro-bustness.A fter studying the contour inform ation and anisotropic Gaussian directional derivatives in different scales,a comer detection a lg o ritto based on multi-scale anisottopic Gaussian kernel principal direction angle change is proposed.The algor-ithm can effectively reduce the false detection rate o f comer detection.Two real comer test images are used in the experiment.In the case o f no noise and adding different levels o f noise,the comer detection ability and comer positioning accuracy o f the algorithm are compared w ith other three classical comer detection algorithms.Experimental results show that the proposed al圓gorithm has better noise robustness and lower false detection rate for comer detection.Key words:comer detection;image contours;anisotropic Gaussian kernel;m ulti-scale;robustness〇引言角点中包含了图像中重要的结构信息,是图像的基本特 征。
东北大学本科毕业设计论文《基于支持向量机算法的电网故障诊断方法研究》
ABSTRACT
With electricity demand growth and technology progress, power grid has become larger and more complex. Due to the formation of large power grids, the quality of electricity supply and electric security improves, also, resources complementary has been strengthened. Once fault occurs, however, it will spread to a wider area with a faster speed. For these merits, this study focuses on the fault diagnosis for power network based on support vector machine. By analyzing relative literatures and building a simulation model, this thesis finishes the analyzing of fault waveforms and harmonic distribution, and studies fault characteristics from the perspective of signal synthesis. To extract fault features submerged in original fault data, this thesis deeply studies the fuzzy processing method, the value detection of instantaneous current and the common fault feature extraction method based on wavelet singular entropy. For the error-prone of instantaneous current detection, fuzzing set ideas is drew to optimize the training samples and by modifying diagnostic strategies, the shortcoming is overcame. To reduce the elapsed time of the common fault feature extraction method based on wavelet singular entropy, a new fault feature combination is proposed by comparing the method with instantaneous current detection. This new combination can inspect faults rapidly when current has a sharp rise such as no- load line closing serious short circuit and improve the diagnostic accuracy when fault current rise is more gentle by taking advantage of wavelet transform which has a wealth of information. Under the condition that the fault features are extracted entirely, artifirt vector machine are used to diagnose power network faults. On one hand, a comparison of the two methods and a study on kernels, multi-class classification methods and SVM training algorithms are carried out. On the other hand, for a figurative expression of the diagnostic results, two dimensions are constructed from the training samples and a twodimensional optimal hyperplane is established by analyzing simulation system structure and data characteristics. Finally, by analyzing the spatial distribution of sample points, the three-dimensional optimal hyperplane is explored. -III-
基于深度预测的单目SLAM绝对尺度估计
2021年6月计算机工程与设计June2021第42卷第6期COMPUTER ENGINEERING AND DESIGN Vol.42No.6基于深度预测的单目SLAM绝对尺度估计张建博,袁亮+,何丽,冉腾,唐鼎新(新疆大学机械工程学院,新疆乌鲁木齐830047)摘要:针对单目同时定位与地图构建(simultan-eous localization and mapping,SLAM)技术存在的尺度不确定性问题,提出一种结合深度预测网络来估计绝对尺度的单目SLAM算法。
利用MonoDepth网络对单目图像进行深度预测,与从单目图像中提取的ORB特征点进行深度值的数据关联,通过设定深度阈值的方法剔除具有不可靠深度值的特征点,恢复单目的绝对尺度,根据特征点的真实深度信息,通过光束法平差优化位姿图,校正尺度漂移,减少累积误差&通过室外KIT-TI数据集进行对比实验,其结果表明,该方法能够获得更高的定位精度&关键词:同时定位与地图构建;深度预测网络;尺度漂移;绝对尺度估计;数据关联中图法分类号:TP242文献标识号:A文章编号:1000-7024(2021)061749-07doi:10.16208/j.issnl000-7024.2021.06.033Absolute scale estimation of monocular SLAM based on depth prediction ZHANG Jian-bo,YUAN Liang+,HE Li&RAN Teng&TANG Ding-xin(School of Mechanicd Engineering,Xinjiang University&Urumqi830047,China)Abstract:In view of the scale uncertainty of simultaneous localization and mapping(SLAM)technology&an approach based on depth prediction network was proposed to estimate the absolute scale of SLAM system.The MonoDepthconvolutionalneural network was used to predict the depth of monocular images&and the ORB feature points extracted from monocular images were associated with the data of the depth values.The feature points with unreliable depth values were removed by se t ing the distance threshold&andtheabsolutescaleofmonocularimageswasrecovered.Accordingtotherealdepthinformationofthefeature points&posegraphwasoptimizedthroughthebundleadjustmentmethod&inwhichwaydriftofscalewascorrectedandthecu-mulative error was reduced.Through the comparison experiment on outdoor KITTI data set&the results show that the proposed methodcane f ectivelyimprovethepositioningaccuracy.Key words:SLAM;depth prediction network;scale drift;absolute scale estimation;data association0引言由于单目相机具有成本低、适用范围广和校准过程简单等优势,使得单目视觉同时定位与地图构建(simultaneous localization and mapping,SLAM)成为机器人在未知环境中自主定位的一个重要研究方向。
资本结构论文参考文献范例
资本结构论文参考文献一、资本结构论文期刊参考文献[1].管理者认知异质性、资本结构与税收效应.《系统工程理论与实践》.被中信所《中国科技期刊引证报告》收录ISTIC.被EI收录EI.被北京大学《中文核心期刊要目总览》收录PKU.被南京大学《核心期刊目录》收录CSSCI.2013年9期.陈克兢.李延喜.冯宝军.张敏.[2].基于效用的公司证券定价与资本结构选择.《系统工程理论与实践》.被中信所《中国科技期刊引证报告》收录ISTIC.被EI收录EI.被北京大学《中文核心期刊要目总览》收录PKU.被南京大学《核心期刊目录》收录CSSCI.2014年1期.王晓林.杨招军.[3].我国上市公司资本结构影响因素的实证分析.《厦门大学学报《会计研究》.被北京大学《中文核心期刊要目总览》收录PKU.被南京大学《核心期刊目录》收录CSSCI.2002年3期.吕长江.王克敏.[5].文化对资本结构影响实证研究:以重商文化为例.《管理科学》.被中信所《中国科技期刊引证报告》收录ISTIC.被北京大学《中文核心期刊要目总览》收录PKU.被南京大学《核心期刊目录》收录CSSCI.2014年4期.魏汉泽.[6].资本结构影响因素和双向效应动态模型来自中国上市公司面板数据的证据.《会计研究》.被北京大学《中文核心期刊要目总览》收录PKU.被南京大学《核心期刊目录》收录CSSCI.2004年2期.肖作平.[7].中国上市公司资本结构行业间差异实证研究.《系统工程理论与实践》.被中信所《中国科技期刊引证报告》收录ISTIC.被EI收录EI.被北京大学《中文核心期刊要目总览》收录PKU.被南京大学《核心期刊目录》收录CSSCI.2004年5期.郭鹏飞.杨朝军.孙培源.[9].法律渊源、关系异象与资本结构决策机理:基于17个国家地区制造业的面板证据.《管理工程学报》.被中信所《中国科技期刊引证报告》收录ISTIC.被北京大学《中文核心期刊要目总览》收录PKU.被南京大学《核心期刊目录》收录CSSCI.2014年4期.吴泽福.二、资本结构论文参考文献学位论文类[1].基于动态调整模型的中国上市公司资本结构研究.被引次数:27作者:严浩军.技术经济及管理同济大学2006(学位年度)[2].高层管理团队异质性、资本结构决策效果与企业绩效.作者:李晓颖.会计学重庆大学2014(学位年度)[3].中国上市公司资本结构研究.被引次数:56作者:李义超.农业经济管理浙江大学2001(学位年度)[4].河南省民营上市公司资本结构问题研究.作者:王怀川.金融学河南大学2014(学位年度)[5].基于动态调整模型的公司资本结构影响因素分析——以创业板为例. 作者:朱维芳.金融学浙江工商大学2014(学位年度)[6].人力资本结构与区域经济发展研究.被引次数:34作者:李福柱.区域经济学东北师范大学2006(学位年度)[7].财政政策对我国上市公司动态资本结构影响的研究.被引次数:1作者:刘程蕾.财务学厦门大学2014(学位年度)[8].我国房地产上市公司资本结构的区间优化——基于不同规模的公司分类分析.被引次数:1作者:刘倩.技术经济及管理重庆大学2013(学位年度)[9].中国上市公司资本结构研究.被引次数:27作者:童勇.会计学复旦大学2006(学位年度)[10].我国医药制药行业上市公司资本结构与企业价值的关系研究.被引次数:3作者:王梦云.金融学河南大学2013(学位年度)三、相关资本结构论文外文参考文献[1]ResourceFlexibilityandCapitalStructure.JiriChodJianerZhou《Managementscience:JournaloftheInstituteofManagementSciences》,被EI 收录EI.被SCI收录SCI.20143[2]Capitalstructureoptimizationforbuildoperatetransferprojectsusi ngastochasticandmultiobjectiveapproach. SungminYunSeungHeonHanHyoungkwanKim《CanadianJournalofCivilEngineering》,被EI收录EI.被SCI收录SCI.20095[3]Capitalstructureoptimizationforbuild–operate–transferprojectsusingastochasticandmultiobjectiveapproach. SungminYunSeungHeonHanHyoungkwanKimJongHoOck 《CanadianJournalofCivilEngineering》,被EI收录EI.被SCI收录SCI.20095[4]Capitalstructureoptimizationforbuildoperatetransferprojectsusi ngastochasticandmultiobjectiveapproach. SungminYunSeungHeonHanHyoungkwanKim 《CanadianJournalofCivilEngineering》,被EI收录EI.被SCI收录SCI.20095[5]Intellectualcapitalandfinancingdecisions:EvidencefromtheU.S.pa tentdata.Liu,Q.Wong,K.P.《Managementscience:JournaloftheInstituteofManagementSciences》,被EI 收录EI.被SCI收录SCI.201110[6]Measuringcapitalmarketefficiency:Globalandlocalcorrelationsstr ucture.Kristoufek,L.Vosvrda,M.《Physica,A.Statisticalmechanicsanditsapplications》,被EI收录EI.被SCI收录SCI.20131[7]Acomparisonofneuralnetworkandmultipleregressionanalysisinmodel ingcapitalstructure.HsiaoTienPao《ExpertSystemswithApplication》,被EI收录EI.被SCI收录SCI.20083[8]Anempiricalstudyoftheimpactoffirmresourcesonalliancegovernance structures.DanielleD.DunneShanthiGopalakrishnanJoanneL.Scillitoe 《JournalofEngineeringandTechnologyManagement》,被EI收录EI.被SCI收录SCI.20093[9]FinethreedimensionalPwavevelocitystructurebeneaththecapitalreg ionanddeepenvironmentforthenucleationofstrongearthquakes. HUANGJinliZHAODapeng《Chinesesciencebulletin》,被EI收录EI.被SCI收录SCI.20056[10]HowDoIndustryPeersRespondtoControlThreats?. HenriServaesAneTamayo《Managementscience:JournaloftheInstituteofManagementSciences》,被EI 收录EI.被SCI收录SCI.20142四、资本结构论文专著参考文献[1]货币政策、企业性质和资本结构.范蕊,2012中国会计学会2012年学术年会[2]资本结构与企业绩效的互动关系研究基于创业板上市公司的实证检验. 陈德萍.曾智海,2013中国对外经济贸易会计学会2013年学术年会[3]我国商业银行资本结构的经营绩效分析.马巾英,2013第19届中国财务学年会[4]基于结构方程建模的资本结构概念模型设计以中国中小板上市公司为例.辛琳,2013第19届中国财务学年会[5]知识产权保护水平对高科技企业资本结构影响研究基于信息不对称的效应分析.李莉.顾春霞.李奇旋,2013第八届中国管理学年会——中国管理的国际化与本土化[6]建筑业企业资本结构、业务蔓延与盈利绩效研究以土木工程上市公司为例.张静晓.李慧.周天华.郭慧婷,20132013中国工程管理论坛[7]控制权防守、利益转移对资本结构敏感性的交互影响.韩亮亮.徐业坤,2011第四届全国比较管理研讨会[8]基于行业环境条件下的上市公司创新战略与资本结构研究.于晓红.汪伟,2011第十届中国实证会计国际研讨会[9]企业持有水平与资本结构动态调整.何威风.熊回,2013第19届中国财务学年会[10]制造业上市公司证券投资与资本结构相关性研究.林丽萍.朱文健,2013第八届中国管理学年会——中国管理的国际化与本土化。
recurrent neural network based language model
PPL 93.7 85.1 80.0 73.5
Simple experiment: 4M words from Switchboard corpus Feedforward networks used here are slightly different than what Bengio & Schwenk use
i
(1) (2)
xi (t)uji sj (t)vkj
j
yk (t) = g
(3)
where f (z ) is sigmoid activation function: f (z ) = and g (z ) is softmax function: g (zm ) = e zm zk ke (5)
8 / 24
1 1 + e−z
(4)
Overview Introduction Model description ASR Results Extensions MT Results Comparison Main outcomes Future work
Comparison of models
Model KN 5gram feedforward NN recurrent NN 4xRNN + KN5
4 / 24
Overview Introduction Model description ASR Results Extensions MT Results Comparison Main outcomes Future work
Model description - feedforward NN
WER static 24.5 24.1 23.3
WER dynamic 22.8
adaptive
Parallel analysis for lightweight network incident detection using nonlinearadaptive systemsRuo Ando,Yoshiyasu TakefujiNational Institute of Information and Communication Technology,4-2-1Nukui-Kitamachi,Koganei,Tokyo184-8795Japanruo@nict.go.jpKeio University,Graduate School of Media and Governance,Endo5322Fujisawa252-8520JapanAbstractThe rapid increasing of security incidents imposes a great burden on Internet users and system administrators. In this paper we discuss a parallel analysis for lightweight network incident detection using nonlinear adaptive sys-tems.We run AID(anomaly intrusion detection)and MID (misuse intrusion detection)systems in parallel.Two de-tectors generate binary output misuse={Y ES/NO} and anomaly={Y ES/NO}.Then,we can determine whether we need to perform network or security operation. We apply clustering algorithm for AID and classification al-gorithm for MID.The nonlinear adaptive system is trained for running MID and AID in parallel.Proposed parallel system is more lightweight and simple to operate even if the number of incident patterns is increased.Experimental results in the case where false positive is frequently caused show that our method is functional with a recognition rate of attacks less than10%,whilefinding the anomaly status. Also,performance evaluation show that proposed system can work with reasonable CPU utilization compared with conventional serial search based system.1Introduction1.1Anomaly and misuse detectionIDS(Intrusion detection system)is kind of alarm de-ployed on computer system and network to detect activity called misuse,that is something unauthorized action such as leaking or compromising.Recent increasing of the number of attacks against computer systems is rapid enough to pare off the effectiveness of human response.More effective,automated and intelligent detection method is researched in manyfields to take some measures for the unseen incidents. Generally,researches are objective to construct a system treating attacks with automatic response.Among attacks, network intrusion includes far-flung activity to threaten the stability or assurance of information stored in the system linked to network.NIDS(Network Intrusion detection sys-tem)works on network to analyze the packets that passing through thefirewall.NIDS monitors traffic of internal and external malicious activity so that administrator of system can take corrective countermeasure.Almost traditional IDS is using signature-based detection methods that collate pat-terns in packet data,and comparing the patterns to dataset of signatures afforded manually by experts.Because signa-ture database is now maintaining by system administrator, these methods is not able tofind unknown types of intrusion unless the new signature is updated to the database.Intru-sion detection techniques are generally classified into two categories:anomaly detection and misuse detection.Mis-use detection is performed by looking for the behavior of a known exploit scenario,which is usually described by a specific sequence or data.Current signature based methods is classified in misuse detection but lacks the scalability to extract features from attacks observed to detect even deriva-tives of conventional incidents by itself.Misuse learning algorithms on labeled data generally cannot detect new in-trusion as it is.In addition,processing labeled data in order tofind variation is usually so expensive.On the other hand, anomaly detection is performed by the inspection for the state that deviates from the baseline or normal state defined before.Profiling algorithms for AID on unlabeled data is frequently causing false positive because the audit data can be very large.And the output of this method is inclined to depend much on the numbers and features of data to train.1.2Increase of IDS signaturesIt is supposed that a large number of service and system will be connected to the internet.And the number of ex-posed security holes,flaws and vulnerabilities is increasing rapidly still now.In addition,crackersfind out the security holes,flaws and exploits code from vast number of program code.Concerning external attacks,it may be almost impos-sible to generate unknown signature of all exploit no matter how much the data mining techniques is progressed in the view of feasibility to training data of vast information of all systems and services connected to the internet.To specify the unknown exploit,we should eventually use the pattern matching method according to the information from trust-worthy sources such as mailing list managed by experts.On the other hand,the signatures of current intrusion detection system is increasing and its managing is becoming so com-plicated that administrators is required to spend much time to learn how to handle rules and maintain databases.Cur-rent IDS checks all internal and external packets and logs part of them according to signature rule set.With the com-plexity of managing signatures,there is matter of concern that increase in the number of signatures unnecessary im-pose the great burden to the system and worse,the improper setting of signature rule set drop the packets of coming at-tacks.1.3Tradeoffs between clustering and clas-sificationThere are two major data mining techniques applied for intrusion detection,clustering or classification.Clustering is the automated,unsupervised process that allows one to group together data into similar characteristics.Classifi-cation is the method to learn to assign data to predefined classes.These two methods are applied for two detec-tion styles as we mentioned before,anomaly detection and misuse detection.Anomaly detection uses clustering algo-rithms because the behavior tofind is unlabeled,with no external information sources.Misuse detection adapts clas-sification algorithm because the activity to analyze requires that detector know how classes are defined.The tradeoff about the accuracy and range of detection exists between clustering and classification.Classification deal with pre-defined data,so it affords detection of weaker signal and figure out accurate recognition.But in some cases,it may be biased by incorrect data to train and it is not able to de-tect new type of attacks in the sense that the attack does not belong to any category defined before.Clustering is not dis-torted by previous knowledge,but therefore needs stronger signal to discover.At the same time it can deal with unla-beled attacks because the training doesn’t specify what the detection system is trying tofind while clustering go too far to perceive the activity that is not included incident affair. 2Proposed systemFigure1shows the comparison of signature-based in-spection and proposed parallel analysis.In signature based IDS such as Snort[24],packet payload is matched after preprocess of IP header.If some signature is matched, the packet is inspected in detail and classified according to database..Then,network administrator can distinguish whether the packet is generated by anomaly event or ma-licious behavior.In this method,as the number of signa-ture is increasing,the system is more heavy and compli-cated to operate.Besides,we cannot know if we should treat the detected event by network or security operation until database transaction is completed.Lower side of Fig-ure1show the proposed parallel analysis method.After preprocess,we run anomaly detector and misuse detector in parallel.Both modules is quantitative analyzer.By set-ting proper threshold,two detectors generate binary output misuse={Y ES/NO}or anomaly={Y ES/NO}.In the case of anomaly=Y ES and misuse=NO,we start countermeasure by network operation.In the case of anomaly=Y ES and misuse=NO,we start counter-measure by security operation.That is,anomaly∧misuse=1→work.operation anomaly∨misuse=1→do.security.operation.By AND or OR operation of binary output from two de-tectors,we can address the tradeoff between clustering and classification.Details about this solution is discussed in sec-tion4.Instead of qualitative specification of attack packet, proposed parallel system is more lightweight and simple to operate.Also,according to numerical output or perfor-mance evaluation in section5,our system is more robust to attack for IDS itself which is serious for network adminis-tration.3Related workOne of the most classic researches of intrusion detec-tion was[1]in1980.Since then,many data mining tech-niques have been proposed for more accurate detection as one activefield[2][3].Anomaly detection to model nor-mal behavior is implemented in the statistical techniques [4][5][6]and machine learning to recognize anomalous user and program[7][8][9].Among anomaly detection tech-niques,clustering is popular and well studied method in this field[10].There could be categorized anomaly detection into two types:profiling-emphasize[11][12][13]and real-time detection aiming type.Misuse detection to discoverparison of serial matching based system and proposed parallel analysis system.After packet preprocess.we run anomaly and misuse detection in parallel.exploitation that is recognized by a specific pattern or se-quence of the events data observed is also performed in the expert systems[14].In[15],discussion is mainly based on how to classify the predefined data.In this paper,we adapt two-stage processing based on neural networks.Early work on the applying neural network for intrusion detection was due to Henning et al[16].Ghosh and Schwartzbard have proposed the application of neural for anomaly and misuse detection.They use artificial neural networks for anomaly detection in order to detect unseen behavior and for misuse detection in order to detect variations of known attacks.In recent years,there are various techniques in applying neural networks.The advantages of neural networks are pointed in the potential to process the limited,incomplete and nonlin-ear data sources by Cannaby[17].Recently,applying SVM (support vector machine)for IDS is well researched.In [18],SVM is applied for anomaly detection.Katherine A Heller et al.applied SVM for anomalous registry accesses in[19].DJ.Wang et al.discuss the efficient optimal probe deployment of distributed IDS in[20].In[21],distributed IDS in MANET is discussed.4Training nonlinear adaptive systems4.1Two stage matrix generationClustering algorithms are applied for learning”nor-mal”or”usual”state of computer network.To output anomaly=YES and misuse=NO in parallel,we need to ap-pend misuse dataset to clustered dataset.Figure2show the two stage signature generation for parallel analysis.Let data[i][j]data for AID and data[k][h]for MID where i,k are the number of class(kinds of aggregated data)and j,h are time(j=h).First,traffic data[i][j]is processed by clustering module.Then,generated matrix and misuse data[k][h]is processed in classification module.In detection,distance of AID and recognition rate of MID is calculated in parallel.4.2Neural networksTo make proposed system lightweight,we apply two neural networks:maximum neural network and functional link network.Two algorithm have been developed for faster convergence and lightweight recognition.Figure2.Two stage matrix generation.Let data[i][j]data for AID and data[k][h]for MID where i,k are the number of class(kinds of aggregated data)and j,h are time(j=h).After clustering,we append misuse dataset[k][h]in classification module.4.2.1Maximum neural networkMaximum neural network is one of the algorithms for unsu-pervised clustering.This paper use a two-dimensional max-imum neural network that has M clusters N neuronsfigur-ing out M*N processing elements.In this algorithm,only one neuron willfire in each cluster and the output of other neurons in the same cluster become zero so that a stable state will always represent a valid solution[22].In discrim-ination analysis,Mahalanobis distance is the useful way of calculating similarity in units of standard deviation from the group mean.In this measurement,thefigured out circum-scribing ellipse formed around the cluster defined the one standard deviation boundary of that group.4.2.2Functional link networkTo assign instances to predefined classes,we use the lightweight classification algorithm called functional link neural network.A standard neural network typically con-tains on lots of simple computational elements or nodes ar-ranged in on or more stages between input and output.The inputs to a node are linearly weighted before the sum before calculating sum by some nonlinear function,which gives to the network its nonlinear approximation ability.If nonlinear processing rules are adapted,the learning rate is often slow and local minimum may case problems.The functional link model eliminates all layers between input and output by us-ing single step of processing is one way to avoid the nonlin-ear learning.The simplicity of the network,the time it takes to complete a model trained by prediction is so fast mea-sured in milliseconds.Another benefit of functional link neural network isflexibility when the learning time is based on the numerous processing elements necessary for com-puting[23].It is known that functional link model could be faster than the other multi-layer learning algorithm in the problem where the convergence is predicted to some extent. 5System outputs and measuring perfor-mance5.1System outputsAs we discussed before,network trouble such as rapid increase of SYN or ICMP packet is not easy to be distin-guished from status caused by malicious activities.Admin-istrators usually inspect many kinds of packetsflow com-prehensively and compare the current status with his ex-perience.We show this kind of event(network trouble by misoperation)is good example for validating the effective-ness of proposed method.Table1shows the dataset for network trouble(rapid un-usual packet increase)by the misoperation.Figure3shows the output of AID,the mahalanobis distance of clusters as profilefigured out by maximum neural network.FigureTable 1.Data set for Network trouble by misoperation.1the number of inward packets2the number of outward packets3the number of inward bytes4the number of outward bytes5the number of SYN packets6the number of port connection7the number of host inward connection8the number of host outward connection9the number of PING packets10the number of ICMP unreachpackets Figure3.Mahalanobis distance of the states occured by networktrouble.Figure4.Recognition rate of the state in point99of Figure3.This indicates that the state isnot occured by malicious action but unusual.4show the output of MID in recognizing network trou-ble.Clustering is able to detect the large number of attacks because the unsupervised learning and discrimination does not require predefined data.However,clustering algorithms can onlyfind just unusual or abnormal event compared with the profiles from system monitored.Thus,we need to run classification module in parallel.To proceed to specify the events alerted,we should process weaker signal.Classifica-tion algorithm can handle weaker signal tofigure out accu-rate recognition.Figure4show the recognition rate of the state in point99of Figure3.This indicates that anomaly is not occurred by malicious action but misoperation.5.2Performance evaluationNetwork administrators are recently concerned about the cost by deploying software loaded on their services.Some-times it is discussed that impact on CPU and disk utilization of utilizing IDS is considerably high.In this section,we compare the performance of proposal method with IDS with database transaction such as Snort and PostgreSQL[25].We pick up an attack called stick attack.This attack creates DoS (Denial of Service)attacks with packet data that generates numerous false alarms.The purpose of this technique is to trigger IDS events on aflooding level.Consequently IDS become disable to handle aflood of alerts owing to CPU usage and disk space being full of logsfiles.AndfinallyFigure5.CPU idle time of snort and proposedsystem.IDS drop packets or fail-on allowing malicious access to the network monitored.The effectiveness of the stick attack de-pends on how much CPU utilization or disk space of system where IDS resides could be wasted by the performance of tools and the attacker’s available bandwidth.Figure5shows that CPU idle time both of snort using text signature matching and proposed system using quanti-tative analysis.Under stick attack,proposal method is more functional in the sense that it works with less CPU utiliza-tion and returns to normal status faster.In Figure6,comparison is shown between the CPU utilization of snort applying PostgreSQL database transac-tion and proposed system with two neural networks.From performance measurementfigured out,proposal method is faster and more reasonable about CPU utilization.6Conclusion and further workThe rapid increasing of security incidents imposes a great burden on Internet users and system administrators. In this paper we discuss the parallel analysis for lightweight network incident detection using nonlinear adaptive sys-tems.In proposed system,we run anomaly detector and misuse detector in parallel.Both modules is quantitative analyzer.By setting proper threshold,two detectors gener-ates binary output misuse={Y ES/NO}or anomaly= {Y ES/NO}.In the case of anomaly=Y ES and misuse=NO,we start countermeasure by networkopera-Figure6.CPU idle time of snort with post-greSQL and proposed system.tion.In the case of anomaly=Y ES and misuse=NO, we start countermeasure by security operation.That is, anomaly∧misuse=1→work.operation anomaly∨misuse=1→do.security.operation.By{AND/OR}operation of binary output of two detec-tors,we can address the tradeoff between clustering and classification.We apply clustering algorithm for AID and classification algorithm for MID.Lightweight nonlinear adaptive system(maximum and functional link neural net-works)is trained for generating matrix for AID and MID. Proposed parallel system is more lightweight and simple to operate even if the number of incident patterns is increased. Experimental results in the case where false positive is fre-quently caused show that the proposal system is functional with a recognition rate of attacks less than10%,whilefind-ing the anomaly status.Also,performance evaluation show that for stick attack,this method can work with reasonable CPU utilization compared with serial signature matching based system.References[1]James P.Anderson,”Computer security threat moni-toring and surveillance”,Technical report,James P.Anderson Co.,Fort Washington,PA,1980.[2]Wenke Lee,Salvatore J.Stolfo,Kui W.Mok,”A DataMining Framework adaptive intrusion detection”,in Proceedings of the7th USENIX Security Symposium, 1998.[3]David Marchette,”A Statistical Method for ProfilingNetwork Traffic”,in Proceedings offirst USENIX Workshop on Intrusion Detection and Network Moni-toring,1999.[4]David Wagner and Drew,”Intrusion detection viaStatistic Analysis”,IEEE Symposium on Security and Privacy,2001.[5]Marina Thottan,Chuanyi Ji,”Proactive Anomaly De-tection Using Distributed Intelligent Agents”,IEEE Network,Special Issue on Network Management,vol.12,1998[6]Ghosh,Anup,K.Wanken,James and Charron,Frank.”Detecting Anomalous and Unknown Intru-sions Against Programs”,in Proceedings of the14th IEEE Annual Computer Security Applications Con-ference,1998.[7]Leonid Portnoy,Eleazar Eskin and Salvatore J.Stolfo,”Intrusion Detection with unlabeled data using clus-tering”,in Proceedings of ACM CSS Workshop on Data Mining Applied to Security(DMSA-2001), Philadelphia,PA,2001.[8]Makoto Iguchi and Shigeki Goto,”Detecting Ma-licious Activities through Port Profiling”,IEICE TRANS.INF.V ol.E82-D,1999.[9]Anup K.Ghosh,Aaron Schwartzbart,Michael Schatz,”Learning Program Behavior Profiles for Intrusion Detection”,in Proceedings1st USENIX Workshop on Intrusion Detection and Network Monitoring,1999.[10]W.Lee,S.Stolfo,P.Chan,E.Eskin,W.Fan,M.Miller,S.Hershkop,and J.Zhang,”Real time data mining-based intrusion detection”,in Proceedings of Second DARPA Information Survivability Conference and Exposition,2001.[11]James Cannady,”icial Neural Networks for MisuseDetection”,in Proceedings of the1998National In-formation Systems Security Conference(NISSC’98), Arlington,V A,1998.[12]Fawcett,T.and F.Provost,”Combining data miningand machine learning for effective user Profiling”,in Proceedings of the Second International Conference on Knowledge Discovery and data mining(KDD-96), 1996.[13]S.T.Eckmann,G.Vigna,and R.A.Kemmerer,”STATL:An Attack Language for state-based intru-sion detection”,Journal of Computer Security,vol.10, no.1/2,2002.[14]Ulf Lindqvist and Erland Jonsson,”How to Systemat-ically Classify Computer Security Intrusions”,In Pro-ceedings of the1997IEEE Symposium on Security& Privacy,Oakland,California,May4-7,1997. [15]S.W.Shieh,Virgil D.Gligor,”A Pattern-OrientedIntrusion-Detection Model and Its Applications”, IEEE Symposium on Security and Privacy1991p327-342,1991.[16]K.Fox,R.Henning,J.Reed,and R.Simonian,”Aneural network approach towards intrusion detection”, Tech.Rep.,Harris Corporation,1990.[17]Wenjie Hu,Yihua Liao and V.Rao Vemuri,”Ro-bust Support Vector Machines for Anomaly Detec-tion in Computer Security”,The2003International Conference on Machine Learning and Applications (ICMLA’03),Los Angeles,California,2003. [18]Katherine A Heller,Krysta M Svore,Angelos D.Keromytis,and Salvatore J.Stolfo,”One Class Sup-port Vector Machines for Detecting Anomalous Win-dow Registry Accesses”,3rd IEEE Conference Data Mining Workshop on Data Mining for Computer Se-curity,Florida,2003[19]J.Wang,N.Nitta,and H.Seki,”An efficient methodfor optimal probe deployment of distributed IDS,”The 2004Symposium on Cryptography and Information Security,2004.[20]Ricardo Puttini,Jean-Marc Percher,L.Me and Rafaelde Sousa,”A Fully Distributed IDS for MANET”,in Proceedings of the9th IEEE Symposium on Comput-ers and Communications(ISCC’2004),2004[21]James Cannady,”Artificial Neural Networks forMisuse Detection”,in Proceedings of the1998 National Information Systems Security Conference (NISSC’98),1998[22]Takefuji Y.-Lee K.C.-Aiso H,”An artificial maxi-mum neural network:a winner-take-all neuron model forcing the state of the system in a solution do-main”,Biological Cybernetics,67:Springer-Verlag Ltd,London,1992.[23]Y.H.Pao,and Y.Takefuji,”Functional-link net com-puting:theory,system architecture and functionali-ties”,IEEE Computer,25,5,1992。
基于知识蒸馏的脉冲神经网络强化学习方法
北京大学学报(自然科学版) 第59卷第5期 2023年9月Acta Scientiarum Naturalium Universitatis Pekinensis, Vol. 59, No. 5 (Sept. 2023)doi: 10.13209/j.0479-8023.2023.065基于知识蒸馏的脉冲神经网络强化学习方法张领曹健†张袁冯硕王源北京大学软件与微电子学院, 北京 102600; †通信作者,E-mail:**************摘要提出一种基于知识蒸馏的脉冲神经网络(SNN)强化学习方法SDN。
该方法利用STBP梯度下降法, 实现深度神经网络(DNN)向SNN强化学习任务的知识蒸馏。
实验结果表明, 与传统的SNN强化学习和DNN强化学习方法相比, 该方法可以更快地收敛, 能获得比DNN参数量更小的SNN强化学习模型。
将SDN部署到神经形态学芯片上, 证明其功耗比DNN低, 是高性能的SNN强化学习方法, 可以加速SNN强化学习的收敛。
关键词脉冲神经网络; 强化学习; 知识蒸馏Reinforcement Learning of Spiking Neural NetworkBased on Knowledge DistillationZHANG Ling, CAO Jian†, ZHANG Yuan, FENG Shuo, WANG Yuan School of Software and Microelectronics, Peking University, Beijing 102600; †Correspondingauthor,E-mail:**************Abstract We propose the reinforcement learning method of Spike Distillation Network (SDN), which uses STBP gradient descent method to realize the knowledge distillation from Deep Neural Network (DNN) to Spiking Neural Network (SNN) reinforcement learning tasks. Experiment results show that SDN converges faster than traditional SNN reinforcement learning and DNN reinforcement learning methods, and can obtain a SNN reinforcement learning model with smaller parameters than DNN. SDN is deployed to the neuromorphology chip, and the power consumption is lower than DNN, proving that SDN is a new and high-performance SNN reinforcement learning method and can accelerate the convergence of SNN reinforcement learning.Key words spiking neural network (SNN); reinforcement learning; knowledge distillation大脑是动物最重要的器官。
基于1DCNN鄄LSTM_和迁移学习的短期电力负荷预测
第41卷 第1期吉林大学学报(信息科学版)Vol.41 No.12023年1月Journal of Jilin University (Information Science Edition)Jan.2023文章编号:1671⁃5896(2023)01⁃0124⁃07基于1DCNN⁃LSTM 和迁移学习的短期电力负荷预测收稿日期:2022⁃03⁃16作者简介:姜建国(1966 ),男,新疆奇台人,东北石油大学教授,硕士生导师,主要从事智能电网及电气自动化㊁深度学习负荷预测研究,(Tel)86⁃137****3588(E⁃mail)jjgnepu@㊂姜建国,万成德,陈 鹏,郭晓丽,佟麟阁(东北石油大学电气信息工程学院,黑龙江大庆163318)摘要:针对在短期电力负荷预测中,当某区域电力负荷数据较少时,负荷预测精度较差的问题,提出一种基于1DCNN⁃LSTM(1D Convolutional Neural⁃Long Short⁃Term Memory Networks)和参数迁移的短期负荷预测方法,并采用1DCNN⁃LSTM 结合迁移学习针对性提高预测精度㊂使用美国某地区的实际负荷数据进行仿真分析,实验结果表明,该方法能有效提升区域电力负荷数据缺失时负荷预测的精度㊂关键词:负荷预测;迁移学习;一维卷积神经网络;长短期记忆网络中图分类号:TM715文献标志码:AShort⁃Term Power Load Prediction Based on 1DCNN⁃LSTM and Transfer LearningJIANG Jianguo,WAN Chengde,CHEN Peng,GUO Xiaoli,TONG Linge(School of Electrical and Information Engineering,Northeast Petroleum University,Daqing 163318,China)Abstract :In the short⁃term power load forecasting,when the power load data is sufficient,the accuracy of load forecasting is usually high,but when the data is missing or the data quantity is small,the accuracy of load forecasting is often poor.Therefore,when the power load data in a certain region is small,the load prediction accuracy is difficult to meet the prediction accuracy requirements.A short⁃term load prediction method based on 1DCNN⁃LSTM (1D Convolutional Neural⁃Long Short⁃Term Memory Networks )and parameter transfer is proposed.1DCNN⁃LSTM combined with transfer learning is used to solve the problem of low prediction accuracy.The actual load data of a certain area in the United States are used for simulation analysis.Experimental results show that this method can effectively improve the accuracy of load prediction when regional power load data is missing.Key words :load forecasting;transfer learning;one⁃dimensional convolutional neural networks;long and short⁃term memory networks 0 引 言在对某地区未来短期电力负荷采用神经网络预测方法进行预测时[1],人们通常选择将预测日的日期信息和所在地的气象数据作为影响因素输入,并可得到较好的预测结果㊂然而在实验过程中,人们发现电力负荷数据的充足与否决定了短期负荷预测精度的高低,并且在对某些区域进行负荷预测时,由于不可控因素的影响,电力负荷历史数据出现数据缺失或存储不足的现象,导致经典深度学习预测方法无法实现高精度的短期负荷预测㊂为克服数据样本少所引起的精度低的问题,有人给出了基于迁移学习的估计方式㊂通过迁移相关领域中已经进行过训练的模型参数,帮助目标领域的模型训练,最后形成目标领域的预测模型㊂Hu 等[2]针对新建风电场历史负荷数据严重不足的情况,采用深度学习和迁移学习相结合的方法构建预测模型并成功解决了短期风速预测问题㊂Ma 等[3]则针对时间分辨率增大后伴随的数据稀疏性问题,采用迁移学习对问题进行了针对性解决㊂栗然等[4]提出一种新的神经网络,即长短期记忆神经网络(LSTM:Long Short⁃Term Memory),使用该网络可有效提取电力负荷数据时序特征㊂刘月峰等[5]则将卷积神经网络和长短期记忆网络相结合,利用卷积神经提取特征,长短期记忆网络处理时序,解决了在长短期记忆网络中当输入时序过长所导致预测精度下降的问题㊂刘婉兵等[6]研究在城市中,如何进行有效预测不同类型的电力负荷㊂李俊卿等[7]将CNN (Convolutional Neural Networks),LSTM,CNN⁃LSTM (Convolutional Neural⁃Long Short⁃Term Memory Networks)进行对比,实验结果表明CNN⁃LSTM 模型训练速度和拟合效果均优于其他两种网络㊂笔者在上述研究基础上,针对预测区域电力负荷数据较少,无法使用经典预测方法进行预测的问题,提出一种基于一维卷积神经⁃长短期记忆网络(1DCNN⁃LSTM:1D Convolutional Neural⁃Long Short⁃Term Memory Networks)和迁移学习相结合的短期预测方法,既挖掘了负荷数据的隐藏特征,又保留了负荷数据作为时序的特性,将负荷数据和使用气象数据㊁日期数据进行卷积处理,以更好地提取数据新特征,增强输入数据与输出数据的相关性[8],并有效利用长短期记忆网络的特性,在长时间上有效避免梯度消失和爆炸㊂1 理论基础1.1 1DCNN 网络卷积神经网络是一种在深度学习领域中被广泛应用的深度前馈神经网络㊂其结构近似于多层感知器,并可较好地解决预测精度低的问题,但训练所需的参数规模过大,预测速度慢㊂卷积神经网络在此基础上进行了改进㊂采用稀疏连接的方式链接相邻的两层神经元,同一神经元共享信息,并对样本进行池化运算㊂通过上述操作,使数据在学习过程中不仅可以充分地将原有特征转化为抽样特征,还可提高网络的稀疏性与预测效率㊂卷积神经网络已成功地应用于二维图像处理,同时也可将相同的思想用于处理一维数据[9]㊂一维卷积神经网络(1DCNN)也是由输入层㊁隐藏层和输出层组成,隐藏层包括一维卷积层和一维池化层㊂一维卷积与池化运算如图1所示㊂图1 一维卷积与池化运算Fig.1 One⁃dimensional convolution and pooling operations 当输入信号为x i (i ∈(1,n ))时,卷积层的计算公式为x l i =f ∑k j =i x l -1i w l j +b ()l j , i =1,2, ,n -r +1, k =i +r -1,(1)最大池化运算如下p l i (j )=max (j -1)w +1≤t ≤jw {q l -1i (t )},(2)其中f 为激活函数运算,x l i 为第l 层第i 个节点的输出;x l -1j 为第l -1层j 第个节点输出;r 为卷积核步长;j 的取值范围为i ~k ,表示稀疏链接;w l j 为x l -1j 对应的权值;b l j 为x l -1j 对应的偏置;q l -1i (t )为第l -1个特521第1期姜建国,等:基于1DCNN⁃LSTM 和迁移学习的短期电力负荷预测征矩阵㊂1.2 LSTM 网络LSTM 由存储信息的若干存储单元组成,每个存储单元包括了输入门㊁遗忘门和输出门,以此实现信息的保护和控制,LSTM 单元的基本构成如图2所示[10]㊂其中h t -1为LSTM 在前一时刻的输出,x t 为输入,h t 为当前单元输出,c t 和c t -1分别为存储单元当前状态和上一时刻的状态㊂图2 LSTM 单元结构图Fig.2 LSTM unit structure diagram LSTM 网络传输公式为i t =σ(w i [~h t -1,x t ]+b i ),(3)~c t =tanh(w c [~h t -1,x t ]+b c ),(4)c t =f t c t -1+i t ~c t ,(5)o t =σ(w o [~h t -1,x t ]+b o ),(6)h t =o t tanh(c t ),(7)其中~h 为t 时刻隐藏单元向量的下标,h 为t 时刻新记忆单元的下标,w 为权重矩阵,b 为偏置,f 为sigmoid函数㊂1.3 迁移学习迁移学习主要有两个域的概念,源域和目标域㊂域一般用D ={X ,P (X )}表示,包含两部分:特征图3 迁移学习基本示意图Fig.3 Basic schematic diagram of transfer learning 空间X ={x x i ∈X ,i =1,2, ,n }和边缘分布P (x )㊂在给定一个域D ={X ,P (X )}的情况下,一个任务可表现为T ={Y ,f (X )},包含两部分:标签控件Y 和目标预测函数f (X ),f (X )也可以看作条件概率P (y x )㊂给出源域的数据D s 和T s ,目标域的数据D t 和T t ,迁移学习旨在使用源域D s 和T s 中的知识去改进对目标域的预测函数f (X ),其中D s ≠D t ,T s ≠T t ㊂其基本示例如图3所示㊂2 基于1DCNN⁃LSTM 和参数迁移的短期负荷预测预测模型建立过程主要由以下步骤组成㊂1)获取数据㊂初始选择某地电力负荷数据作为实验数据,在数据集中加入最高㊁最低㊁平均温度和负荷数据以及相关日期信息等构成测试样本㊂2)数据处理㊂初始数据集中除了含有电力负荷值外,还有最高㊁最低㊁平均温度等温度数据以及日期信息的相关数据,在进行模型训练前,需要对数据进行清洗预处理,排除异常数据,将数据所提供的信息量化,并划分训练集和测试集㊂3)确认源域与目的域,针对目的域历史数据缺失时的负荷预测困难的情况,需要在源域所包含的众多区域之间选取与目的域匹配度最高的区域作为迁移对象㊂图4 1DCNN⁃LSTM 网络模型Fig.4 1DCNN⁃LSTM network model 4)搭建1DCNN⁃LSTM 网络模型,对模型进行训练,网络模型如图4所示㊂首先将处理后的数据输入1DCNN 网络中,由于输入数据为一维时间序列,因此使用一维卷积层㊂在卷积层中,卷积步长设为1,使用Relu 函数作为激活函数㊂5)参数迁移,构建新网络,进行负荷预测㊂将训练完毕的1DCNN⁃LSTM 模型参数全部迁移至目标域预测模型作为初始参数,固定1DCNN⁃LSTM 模型的前621吉林大学学报(信息科学版)第41卷l 层,利用已知样本数据微调其剩余参数,完成最终训练[11]㊂6)模型评价㊂将测试集数据输入到训练好的预测模型中,以平均绝对百分比误差(MAPE:Mean Absolute Percentage Error)和均方根误差(RMSE:Root Mean Square Error)作为准确度评价指标对模型进行评价[8],M MAPE =1n ∑n i =1a i -b i a i ×100%,(8)R RMSE =1n ∑ni =1(a i -b i )2,(9)其中a i 和b i 分别为预测时刻i 的预测值和真实值,n 为样本总数㊂3 算例分析对源域NH(New Hampshire)的数据集,按照8∶2的比例将其划分训练集和测试集,为给目的域提供一个高精度的预测模型进行模型参数迁移,在源域展开各种模型预测对比实验,将1DCNN⁃LSTM 与1DCNN㊁LSTM㊁RNN(Recurrent Neural Network)㊁1DCNN⁃RNN(1D Convolutional Neural⁃Recurrent Neural Network)模型进行预测精度对比,其中1DCNN⁃LSTM 模型参数设置如表3所示㊂表1 模型参数设置Tab.1 Model parameter settings 超参数Num_epochs Batch_size AdamConvld Convld_1Convld_2Convld_3值60640.01326412864超参数Convld_4LSTM LSTM_1LSTM_2Input_featuersOutput_class 值6432641282424 表1中,设置的训练次数为60,一次训练中的最大迭代次数为64,学习率0.01,CNN 网络第1层Conv1D 卷积核数目依次设为32㊁64㊁128,第2层ConvlD 卷积核数目依次为64㊁64,LSTM 网络部分共设置3层LSTM 网络层,各层神经元数目依次为32㊁64㊁128,输入24(特征为9),输出24(特征为1)㊂图5 多种模型预测效果对比图Fig.5 Comparison of prediction effects of various models 预测结果对比如图5所示,为直观分析其预测精度,采用MAPE 和RMSE 进行精度评估,评估结果如表2所示㊂由表2可知,预测效果最优的模型为1DCNN⁃LSTM,其MAPE 为4.35%,RMSE 为71.56,负荷预测精度为95.65%,各项参数指标均为最优㊂分析其原因:1DCNN⁃LSTM 神经网络预测模型,结合了一维卷积神经网络在特征提取上的优势和LSTM 神经网络的长时记忆功能,在处理时间序列较多时能有效防止梯度消失和爆炸㊂因此在源域选择使用1DCNN⁃LSTM模型作为预测模型㊂表2 预测模型评估指标Tab.2 Evaluation indicators of prediction model 评估指标预测模型1DCNN LSTM RNN 1DCNN⁃LSTM 1DCNN⁃RNN MAPE /%11.7812.24 6.47 4.3511.65RMSE 171.93187.58102.5771.56191.087预测精度/%88.2287.7693.5395.6588.35 笔者使用美国9个州的电力负荷数据,9个州分别为:内布拉斯加州(NE:Nebraska)㊁康涅狄格州(CT:Connecticut)㊁麻萨诸塞州(MASS:Massachusetts)㊁缅因州(ME:Maine)㊁新罕布什尔州(NH:New721第1期姜建国,等:基于1DCNN⁃LSTM 和迁移学习的短期电力负荷预测Hampshire)㊁罗得岛州(RI:Rhode lsland)㊁得克萨斯州(TX:Texas)㊁佛蒙特州(VT:Vermont)㊁南达科他州(SD:South Dakota)㊂电力负荷曲线如图6所示㊂其中ME 为目的域(仅有7天电力负荷数据),其他图6 电力负荷曲线图Fig.6 Power load curve diagram 8个州为源域(数据充足),为选取与ME 相似性最高的区域作为实验的源域,以ME 为基准,计算其欧氏距离d (x ,y )=∑(x i -y i )()2,(10)结果如表3所示㊂欧氏距离表示每个时间点上,两个区域的电力负荷差的绝对值,所以欧氏距离之和越小,即两个区域的负荷点对应的差值越小,两个区域的相似度越高㊂从表3中可看出,在以ME 为基准的情况下,NH 与ME 之间的欧几里得距离最小,故而笔者将以ME 为目的域,NH 为源域,开展基于预测模型参数迁移的短期负荷预测实验㊂表3 欧氏距离表 Tab.3 Euclidean distance table MWh 地区CT MA NE NH RI SD TX VT欧氏距离38305580985822770448276777557598206303692560 在确定源域预测模型后,首先将迁移源域模型作为目标域初始模型,其次使用目标域原始数据对模型进行微调,最后使用目标域没有训练过的数据作为测试集对微调后的模型进行精度验证㊂采用迁移学习冻结网络,如图7所示㊂图7 基于参数迁移的1DCNN⁃LSTM 预测模型示意图Fig.7 1DCNN⁃LSTM prediction model based on parameter migration 为取得最优预测效果,对冻结层数进行调整,对比以下预测模型(P1~P5)㊂P1:建立1DCNN⁃LSTM 预测网络,将ME 数据集最后一天作为输入,输出即为预测结果㊂P2:迁移源域建立的1DCNN⁃LSTM 预测模型,不冻结参数层,利用目的域前6d 的电力负荷数据对预测模型进行微调㊂P3:同P2,冻结至倒数第2层参数㊂P4:同P2,冻结至倒数第3层参数㊂P5:同P3,冻结至倒数第4层参数㊂实验结果如图8所示㊂预测效果评估如表4所示㊂由图8和表4可知,P4的预测精度最高,为95.91%,即冻结至倒数第3层网络,并用目的域数据对821吉林大学学报(信息科学版)第41卷倒数两层网络进行微调㊂该类模型可以在原模型的基础上,融合目的域数据的基本特征,其预测效果可以更好贴合目的域㊂图8 迁移模型预测结果Fig.8 Migration model prediction results 表4 预测效果评估表Tab.4 Evaluation table of prediction effect 类别MAPE /%RMSE 预测精度/%P110.53157.1289.47P247.43671.0353.57P3 5.3494.2994.66P4 4.0964.8395.91P59.70146.2390.30综上所述,得益于源域预模型所提供的参数初始值以及倒数两层参数学习到关于目的域的负荷特征,笔者方法有效提升了目的域数据短缺时负荷预测的精度和模型的稳定性㊂4 结 语笔者提出了一种基于1DCNN⁃LSTM 和迁移学习的短期负荷预测方法㊂该方法可有效解决目的域由于电力负荷数据缺失所导致短期电力负荷预测困难的问题,1DCNN⁃LSTM 预测模型既可以有效提取负荷特征,又可以有效避免在长时间维度上发生梯度消失和爆炸㊂实验结果表明,本模型均比1DCNN㊁RNN㊁LSTM㊁1DCNN⁃RNN 模型的预测效果优秀,预测精度最高可达95.91%㊂由于在数据处理阶段,仅对数据集中所涵盖的温度㊁日期信息进行处理,而实际上电力负荷与湿度㊁节假日及极端天气等多种因素有关,后续将搜集更多相关影响因素数据,进一步提高模型的预测准确度㊂参考文献:[1]王激华,仇钧,方云辉,等.基于深度长短期记忆神经网络的短期负荷预测[J].广东电力,2020,33(8):62⁃68.WANG J H,QIU J,FANG Y H,et al.Short⁃Term Load Forecasting Based on Deep Long Short⁃Term Memory Neural Network [J].Guangdong Electric Power,2020,33(8):62⁃68.[2]HU Q H,ZHANG R J,ZHOU Y C.Transfer Learning for Short⁃Term Wind Speed Prediction with Deep Neural Networks [J].Renewable Energy,2016,85:83⁃95.[3]MA J,CHENG J C P,LIN C Q,et al.Improving Air Quality Prediction Accuracy at Larger Temporal Resolutions Using DeepLearning and Transfer Learning Techniques [J].Atmospheric Environment,2019,214:116885.[4]栗然,孙帆,丁星,等.考虑多能时空耦合的用户级综合能源系统超短期负荷预测方法[J].电网技术,2020,44(11):4121⁃4134.LI R,SUN F,DING X,et al.Ultra⁃Short⁃Term Load Forecasting Method for User⁃Level Integrated Energy System ConsideringSpatio⁃Temporal Coupling of Multi⁃Energy [J].Power Grid Technology,2020,44(11):4121⁃4134.[5]刘月峰,杨宇慧.基于CNN⁃LSTM 的短期电力负荷预测研究[J].科技创新与应用,2020(1):84⁃85.LIU Y F,YANG Y H.Research on Short⁃Term Power Load Forecasting Based on CNN⁃LSTM [J].Science &Technology Innovation and Application,2020(1):84⁃85.[6]刘婉兵,李妍,杜明秋,等.城市负荷空间分布的聚类群簇分析[J].电力系统自动化,2019,43(5):59⁃65,82.LIU W B,LI Y,DU M Q,et al.Clustering Analysis of Urban Load Spatial Distribution [J].Automation of Electric Power Systems,2019,43(5):59⁃65,82.[7]李俊卿,陈雅婷.LSTM⁃CNN 网络在同步电机励磁绕组匝间短路故障预警中的应用[J].华北电力大学学报(自然科学版),2020,47(4):61⁃70.LI J Q,CHEN Y T.Application of LSTM⁃CNN Network in Early Warning of Inter⁃Turn Short Circuit Fault of SynchronousMotor Excitation Winding [J].Journal of North China Electric Power University (Natural Science Edition),2020,47(4):61⁃70.921第1期姜建国,等:基于1DCNN⁃LSTM 和迁移学习的短期电力负荷预测031吉林大学学报(信息科学版)第41卷[8]张立峰,刘旭.基于CNN⁃GRU神经网络的短期负荷预测[J].电力科学与工程,2020,36(11):53⁃57. ZHANG L F,LIU X.Short⁃Term Load Prediction Based on CNN⁃GRU Neural Network[J].Electric Power Science and Engineering,2020,36(11):53⁃57.[9]KUANG D,BANG X.Predicting Kinetic Triplets Using a1D Convolutional Neural Network[J].Thermochimica Acta,2018, 669:8⁃15.[10]张勇,陶一凡,巩敦卫.迁移学习引导的变源域长短时记忆网络建筑负荷预测[J].控制与决策,2021,36(10): 2328⁃2338.ZHANG Y,TAO Y F,GONG D W.Migration Guided Learning as the Source Domain Length Memory Network Building Load Prediction[J].Control and Decision,2021,36(10):2328⁃2338.[11]蔡秋娜,苏炳洪,闫斌杰,等.基于参数迁移的节假日短期负荷预测方法[J].电气自动化,2020,42(4):59⁃60,98. CAI Q N,SU B H,YAN B J,et al.Short⁃Term Holiday Load Forecasting Method Based on Parameter Migration[J].Electric Automation,2020,42(4):59⁃60,98.(责任编辑:刘俏亮)。
基于Transformer模型的手写数学公式语法树解码器
北京大学学报(自然科学版) 第59卷 第6期 2023年11月Acta Scientiarum Naturalium Universitatis Pekinensis, Vol. 59, No. 6 (Nov. 2023)doi: 10.13209/j.0479-8023.2023.085基于Transformer模型的手写数学公式语法树解码器周伯瀚曹健†王源北京大学软件与微电子学院, 北京 102600; †通信作者摘要目前对数学公式进行树结构解码的方法大多基于循环神经网络的结构, 训练效率低, 训练过程复杂,基于此问题, 提出一种基于Transformer结构的手写数学公式识别模型, 可以直接对公式的语法树进行解码。
在手写公式识别任务多个数据集上的实验结果表明, 所提出的Transformer树解码方法都取得超越Trans-former序列解码方法的性能, 并展现出超越循环神经网络树解码方法的潜力。
关键词手写数学公式识别; Transformer; 树解码器; 图表理解A Transformer-based Syntax Tree Decoder for HandwrittenMathematical Expression RecognitionZHOU Bohan, CAO Jian†, WANG YuanSchool of Software and Microelectronics, Peking University, Beijing 102600; † Corresponding author,Abstract Most of the existing tree-structured decoding methods of handwritten mathematical expression recog-nition are based on the recurrent neural networks, which have low training efficiency and complicated training process. In order to prove this problem, the authors propose a handwritten mathematical expression recognition model based on Transformer structure, which can decode the syntax tree of expressions directly. Experimental results show that the proposed tree-structured decoding method achieves better performance than the string decoding methods base on Transformer on several datasets of handwritten formula recognition tasks, and show the potential to surpass recurrent neural network tree decoding methods.Key words handwritten mathematical expression recognition; Transformer; tree decoder; document comprehension作为光学文字识别(optical character recognition, OCR)的一部分, 手写公式识别在文档理解、图表理解、AI作业批改和AI解题等领域发挥着重要作用。
资本结构论文参考文献范例
资本结构论文参考文献一、资本结构论文期刊参考文献[1].管理者认知异质性、资本结构与税收效应.《系统工程理论与实践》.被中信所《中国科技期刊引证报告》收录ISTIC.被EI收录EI.被北京大学《中文核心期刊要目总览》收录PKU.被南京大学《核心期刊目录》收录CSSCI.2013年9期.陈克兢.李延喜.冯宝军.张敏.[2].基于效用的公司证券定价与资本结构选择.《系统工程理论与实践》.被中信所《中国科技期刊引证报告》收录ISTIC.被EI收录EI.被北京大学《中文核心期刊要目总览》收录PKU.被南京大学《核心期刊目录》收录CSSCI.2014年1期.王晓林.杨招军.[3].我国上市公司资本结构影响因素的实证分析.《厦门大学学报《会计研究》.被北京大学《中文核心期刊要目总览》收录PKU.被南京大学《核心期刊目录》收录CSSCI.2002年3期.吕长江.王克敏.[5].文化对资本结构影响实证研究:以重商文化为例.《管理科学》.被中信所《中国科技期刊引证报告》收录ISTIC.被北京大学《中文核心期刊要目总览》收录PKU.被南京大学《核心期刊目录》收录CSSCI.2014年4期.魏汉泽.[6].资本结构影响因素和双向效应动态模型来自中国上市公司面板数据的证据.《会计研究》.被北京大学《中文核心期刊要目总览》收录PKU.被南京大学《核心期刊目录》收录CSSCI.2004年2期.肖作平.[7].中国上市公司资本结构行业间差异实证研究.《系统工程理论与实践》.被中信所《中国科技期刊引证报告》收录ISTIC.被EI收录EI.被北京大学《中文核心期刊要目总览》收录PKU.被南京大学《核心期刊目录》收录CSSCI.2004年5期.郭鹏飞.杨朝军.孙培源.[9].法律渊源、关系异象与资本结构决策机理:基于17个国家地区制造业的面板证据.《管理工程学报》.被中信所《中国科技期刊引证报告》收录ISTIC.被北京大学《中文核心期刊要目总览》收录PKU.被南京大学《核心期刊目录》收录CSSCI.2014年4期.吴泽福.二、资本结构论文参考文献学位论文类[1].基于动态调整模型的中国上市公司资本结构研究.被引次数:27作者:严浩军.技术经济及管理同济大学2006(学位年度)[2].高层管理团队异质性、资本结构决策效果与企业绩效.作者:李晓颖.会计学重庆大学2014(学位年度)[3].中国上市公司资本结构研究.被引次数:56作者:李义超.农业经济管理浙江大学2001(学位年度)[4].河南省民营上市公司资本结构问题研究.作者:王怀川.金融学河南大学2014(学位年度)[5].基于动态调整模型的公司资本结构影响因素分析——以创业板为例. 作者:朱维芳.金融学浙江工商大学2014(学位年度)[6].人力资本结构与区域经济发展研究.被引次数:34作者:李福柱.区域经济学东北师范大学2006(学位年度)[7].财政政策对我国上市公司动态资本结构影响的研究.被引次数:1作者:刘程蕾.财务学厦门大学2014(学位年度)[8].我国房地产上市公司资本结构的区间优化——基于不同规模的公司分类分析.被引次数:1作者:刘倩.技术经济及管理重庆大学2013(学位年度)[9].中国上市公司资本结构研究.被引次数:27作者:童勇.会计学复旦大学2006(学位年度)[10].我国医药制药行业上市公司资本结构与企业价值的关系研究.被引次数:3作者:王梦云.金融学河南大学2013(学位年度)三、相关资本结构论文外文参考文献[1]ResourceFlexibilityandCapitalStructure.JiriChodJianerZhou《Managementscience:JournaloftheInstituteofManagementSciences》,被EI 收录EI.被SCI收录SCI.20143[2]Capitalstructureoptimizationforbuildoperatetransferprojectsusi ngastochasticandmultiobjectiveapproach. SungminYunSeungHeonHanHyoungkwanKim《CanadianJournalofCivilEngineering》,被EI收录EI.被SCI收录SCI.20095[3]Capitalstructureoptimizationforbuild–operate–transferprojectsusingastochasticandmultiobjectiveapproach. SungminYunSeungHeonHanHyoungkwanKimJongHoOck 《CanadianJournalofCivilEngineering》,被EI收录EI.被SCI收录SCI.20095[4]Capitalstructureoptimizationforbuildoperatetransferprojectsusi ngastochasticandmultiobjectiveapproach. SungminYunSeungHeonHanHyoungkwanKim 《CanadianJournalofCivilEngineering》,被EI收录EI.被SCI收录SCI.20095[5]Intellectualcapitalandfinancingdecisions:EvidencefromtheU.S.pa tentdata.Liu,Q.Wong,K.P.《Managementscience:JournaloftheInstituteofManagementSciences》,被EI 收录EI.被SCI收录SCI.201110[6]Measuringcapitalmarketefficiency:Globalandlocalcorrelationsstr ucture.Kristoufek,L.Vosvrda,M.《Physica,A.Statisticalmechanicsanditsapplications》,被EI收录EI.被SCI收录SCI.20131[7]Acomparisonofneuralnetworkandmultipleregressionanalysisinmodel ingcapitalstructure.HsiaoTienPao《ExpertSystemswithApplication》,被EI收录EI.被SCI收录SCI.20083[8]Anempiricalstudyoftheimpactoffirmresourcesonalliancegovernance structures.DanielleD.DunneShanthiGopalakrishnanJoanneL.Scillitoe 《JournalofEngineeringandTechnologyManagement》,被EI收录EI.被SCI收录SCI.20093[9]FinethreedimensionalPwavevelocitystructurebeneaththecapitalreg ionanddeepenvironmentforthenucleationofstrongearthquakes. HUANGJinliZHAODapeng《Chinesesciencebulletin》,被EI收录EI.被SCI收录SCI.20056[10]HowDoIndustryPeersRespondtoControlThreats?. HenriServaesAneTamayo《Managementscience:JournaloftheInstituteofManagementSciences》,被EI 收录EI.被SCI收录SCI.20142四、资本结构论文专著参考文献[1]货币政策、企业性质和资本结构.范蕊,2012中国会计学会2012年学术年会[2]资本结构与企业绩效的互动关系研究基于创业板上市公司的实证检验. 陈德萍.曾智海,2013中国对外经济贸易会计学会2013年学术年会[3]我国商业银行资本结构的经营绩效分析.马巾英,2013第19届中国财务学年会[4]基于结构方程建模的资本结构概念模型设计以中国中小板上市公司为例.辛琳,2013第19届中国财务学年会[5]知识产权保护水平对高科技企业资本结构影响研究基于信息不对称的效应分析.李莉.顾春霞.李奇旋,2013第八届中国管理学年会——中国管理的国际化与本土化[6]建筑业企业资本结构、业务蔓延与盈利绩效研究以土木工程上市公司为例.张静晓.李慧.周天华.郭慧婷,20132013中国工程管理论坛[7]控制权防守、利益转移对资本结构敏感性的交互影响.韩亮亮.徐业坤,2011第四届全国比较管理研讨会[8]基于行业环境条件下的上市公司创新战略与资本结构研究.于晓红.汪伟,2011第十届中国实证会计国际研讨会[9]企业持有水平与资本结构动态调整.何威风.熊回,2013第19届中国财务学年会[10]制造业上市公司证券投资与资本结构相关性研究.林丽萍.朱文健,2013第八届中国管理学年会——中国管理的国际化与本土化。
牛顿方法的用导一致性
牛顿方法的用导一致性张雨浓;陈宇曦;付森波;肖林【摘要】导数及其使用是一种重要的数学工具与方法,在数值方法研究中起着十分重要的作用,很多经典数值计算方法理论中都涉及了对导数的运用.作为常用的数值方法,牛顿方法大量利用了求导的思想.通过归纳牛顿迭代法的显性用导特性和牛顿插值多项式的隐性用导特性,得到牛顿方法的用导一致性.%Derivatives and their use as an important mathematical tool and method plays an important role in the research of numerical methods. In addition,many classical numerical methods involve the use of derivatives. Being commonly-used.Newton methods exploit the idea of using derivatives extensively. By summing up the explicit use of derivatives in Newton iteration method and the implicit use of derivatives in Newton interpolation polynomial, the consistency of using derivatives in Newton methods is thus illustrated.【期刊名称】《甘肃科学学报》【年(卷),期】2012(024)002【总页数】4页(P5-8)【关键词】牛顿迭代法;牛顿插值多项式;用导一致性【作者】张雨浓;陈宇曦;付森波;肖林【作者单位】中山大学信息科学与技术学院,广东广州510006;中山大学信息科学与技术学院,广东广州510006;中山大学信息科学与技术学院,广东广州510006;中山大学信息科学与技术学院,广东广州510006【正文语种】中文【中图分类】O241.4随着当今科学技术的快速发展,计算机在自然科学和社会科学的各个领域发挥着举足轻重的作用,导数及其使用作为数值方法研究的一种工具和方法日渐被研究者所重视.以下从牛顿迭代法和牛顿插值多项式两个方面探讨牛顿方法所体现的用导一致性现象.17世纪英国数学家牛顿开展了对牛顿迭代法的研究工作:他基于二项式定理,推导出牛顿迭代法的基本表达形式.18世纪英国数学家辛普森从微积分的角度给出了牛顿迭代法的现代表达形式[1].由于每次迭代都需要求出和利用函数导数值,因此求导和用导成为牛顿迭代法的一个重要特征.定理1 设f∈C2[a,b],且存在数p∈ [a,b],满足f(p)=0.如果f′(p)≠0,则存在一个数δ>0,对任意初始近似值p0∈[p-δ,p+δ],使得由如下迭代定义的序列pk(k=1,2,…)收敛到p[1-8]:pk =g(pk-1)=pk-1-f(pk-1)/f′(pk-1). (1)设在几何上y=f(x)是一曲线[2-6],如图1所示.它与x轴的交点的横坐标p为方程的根,假设已给出一个近似根p0,我们用曲线在点(p0,f(p0))处的切线逼近该曲线.令p1是该切线与x轴交点的横坐标,由图1可见,p1对根的近似比p0的效果要好.由此可以写出切线L的表达式为整理上式可得用以上方法,利用一系列的切线与x轴的交点来逐步逼近曲线y=f (x)与x轴的交点,即可得到p2,p3,…,pk,…,最终使序列{pk}收敛到p,它正是局部线性化与迭代法结合产生的方法.对比式(1)和式(2)可以发现,它们在表达形式上是一致的.可见,式(1)就是牛顿迭代法对函数几何模型迭代求解的一个函数表达式.它是建立在对函数图形求解切线斜率的基础上,得出比初始选取点更接近方程根的一系列点,在精度要求范围内经过有限次迭代求出函数根的近似值.牛顿迭代法中还要求近似根p0的选取必须充分靠近真正根p,这可以通过泰勒多项式进行分析.即用泰勒级数展开,可得[1,2,4]用x=p代入式(3),并利用f (p)=0,可得如果p0足够逼近p,则式(4)中的最后一项比前两项的和要小得多,因此可以将最后一项(非线性项)忽略,这样便得到牛顿迭代公式比较发现,式(5)的结果与式(2)一样,它们都要求近似根的选取充分靠近真实根.因此从泰勒级数展开式的角度再次体现牛顿迭代式对一阶导数的适当应用,证明了牛顿迭代法几何上的斜率分析法的正确性和有效性.通过上述对泰勒级数展开式的分析,体现了牛顿迭代法的求导和用导特性.由式(1)可见,牛顿迭代法中每次迭代都需要求一次导数,因此需要一种与牛顿法的收敛速度差不多快的方法且不用计算f′(x).我们可以用函数值的差商近似导数,那么每次的迭代只需计算两次f(x),这就是割线法的思想.可以证明它在单根上的收敛阶R≈1.618 033 989,与收敛阶为2的牛顿迭代法差不多快[1-8]. 设曲线f(x)的图形如图2中所示[2,5,6],经过点(p,0),即p为函数f (x)的一个根.(p0,f(p0))和(p1,f(p1))是靠近点(p,0)的两个已知初始点.定义p2为经过两个初始点的直线与x轴的交点的横坐标,则p2比p0或p1更接近p.由式(6)整理可得到因此,根据以上两点的迭代公式可得到一般式为这个迭代表达式被称为求解非线性方程的割线法公式.假设割线法中所选取的两个点(p0,f(p0))和(p1,f(p1))无限逼近,则理论而言,过这两个点得到的割线将无限逼近于函数在该处的切线.因此牛顿迭代法其实可以看作割线法的一种极限情况.而在这个极限情况下,从割线法的角度来看,牛顿迭代法同样体现了牛顿方法的求导和用导特征.以下给出n维非线性方程组的牛顿迭代法求解方法,并同时观察和分析牛顿方法的求导和用导思想在高维非线性方程组求解时的运用.由p2、p1和p0可得到以下方程:设考虑的非线性方程组问题为[4-6,8]其中函数f1,f2,…,fn在某解p的邻域D内具有二阶连续偏导数,列向量x定义为x = [x1,x2,…,xn]T,则该非线性方程组的牛顿迭代式为[4-6,8]其中pk 代表第k次迭代向量值(k=1,2,3,…),雅可比矩阵J满足Jij =∂fi/∂xj,i,j∈ {1,2,…,n}.对比可见式(8)与式(1)基本上是一致的.在保证函数可微的前提下,牛顿方法从一维非线性方程求解扩展到n维非线性方程组求解,其中的一维求导和用导对应地变成了求取和利用一阶偏导数而已(最终形成具有明显的牛顿方法求导和用导特征的雅可比矩阵J,从形式上看,其相当于一维牛顿迭代式(1)中的导数f′). 插值多项式是插值的一种重要方法.常见的插值多项式有拉格朗日多项式、牛顿多项式和切比雪夫多项式等.牛顿多项式因其在增加节点时具有“承袭性”[1-8],避免了多项式的重新构建,因此更为简单和适用,具有明显更大的优势,从而成为一种更为实用的插值多项式数值算法.定理2 设目标函数f在n+1个节点x0,x1,…,xn 上的值f(x0),f(x1),…,f(xn)为已知,存在惟一的至多n次的多项式Nn(x),具有插值性质f(xi)=Nn(xi),其中i=0,1,2,…,n.该多项式的牛顿形式为[2,5,6]式(9)即被称为基于点x0,x1,…,xn 的牛顿插值多项式.推论1 设Nn(x)是上述定理中给出的牛顿插值多项式[2,5],并用来逼近目标函数f(x),即如果f∈Cn+1[a,b],则对每个x∈ [a,b],对应地存在(a,b)内的数c=c(x),使得误差项形如从式(11)能够看出,在对函数f(x)逼近的误差项的给出中,涉及到了函数高阶导数,即使用了目标函数f(x)在c=c(x)处的n+1阶导数f(n+1)(c),因此牛顿方法的求导和用导思想在估计牛顿插值多项式的误差项时也得到了有效的使用.我们引入如下差商的概念,并同时观察牛顿插值多项式对导数的运用情况.定义1 设目标函数f关于xi的零阶差商f[xi]为f 在xi 的值[4,5,9,10]:由零阶差商出发可以归纳地定义各阶差商,其中f关于xi+1 与xi 的一阶差商记作f[xi+1,xi].且有以下关系:一般而言,设f 关于xi+k,…,xi+1,xi 的k 阶差商为根据上述牛顿插值多项式和差商的定义,可以求得式(9)中的系数如下:即牛顿插值多项式的系数可以用如下一般式确定:即牛顿插值多项式每一项的系数ck(k=0,1,2,…,n)等于目标函数f(x)对应阶的差商.另外,通过对差商与牛顿插值多项式的余项分析还可以得到以下关系.定理3 设目标函数f在[a,b]上存在n+1阶导数,且x0,x1,…,xn ∈ [a,b],则存在c∈ [a,b]满足[4]从定义1和定理3可见:导数,具体就是差商的一种带权极限,所以差商本身就包含了求导和用导的思想.牛顿迭代法是建立在对切线求斜率(导数)的基础上的,且在迭代过程中需要显式地计算出来.假设割线法中所选取的两个点无限逼近,则过这两个点得到的割线将无限逼近于函数在该处的切线(即导数),因此牛顿迭代法其实可以看作是割线法的一种极限情况,而这个极限过程体现了牛顿类方法的求导和用导特征.当牛顿迭代法推广高维非线性方程组求解时,在表达式的整体形式上保持了一维非线性方程求解的特征,而对函数求偏导数形成的雅可比矩阵则对应了一维非线性方程求解的函数导数项,其中雅可比矩阵正体现了牛顿方法在求解非线性方程组中的求导和用导特征.牛顿插值多项式项数随节点增加而“承袭性”地增加,因此不需要重构该插值多项式.这一方便性远远高于拉格朗日插值多项式.牛顿插值多项式虽然看上去与函数的求导没什么关系,但经过仔细观察和推敲便可发现,其中使用的差商与导数是有密切关系的.导数其实就是差商的一种带权极限,所以在求解牛顿插值多项式的系数过程中处处隐性地用到了导数.另外,由差商与导数的关系,可以将牛顿插值多项式的误差项以差商的形式写为:其进一步体现了求导和用导思想在估计和求解牛顿插值多项式误差余项过程中的具体运用.可见,不论是牛顿迭代法求解非线性方程(组)还是牛顿插值多项式的构造,在它们的计算和扩展中均使用到了求导和用导思想.通过以上对牛顿迭代法和牛顿插值多项式的分析可见,牛顿方法广泛运用求导和用导思想,比如函数的几何分析、函数的泰勒展开式、插值函数系数和误差项的求取等等.在这些问题中,导数的运用使得牛顿法在同类数值算法中占据优势.当今各个科学领域的研究中,人们已经离不开对求导和用导思想和方法的运用,比如文献[11]关于一阶微分公式的研究,以及文献[12-17]关于神经网络和神经动力学的研究.显然,倘若这些研究失去了导数的辅助,将使其陷入困境甚至无法进行.此外,牛顿方法的用导一致性现象也说明,在数值方法的研究中,要善于利用求导工具进行数值计算方法的设计和优化;在研究新的数值算法时可从多方面加强对求导和用导思想方法的运用.【相关文献】[1]McNAMEE J M.Numerical Methods for Roots of Polynomials[M].London:Elsevier,2007.[2]John H M,Kurtis D F.数值方法(MATLAB版)[M].周璐,陈渝,钱方,等译.北京:电子工业出版社,2009.[3]李庆扬,峰杉白,关治.数值计算原理[M].北京:清华大学出版社,2000.[4]Jaan K.Numerical Methods in Engineering with Matlab[M].Cambridge:Cambridge University Press,2005.[5]李建良,蒋勇,汪光先.计算机数值方法[M].南京:东南大学出版社,2000.[6]Gerald Recktenwald(USA).数值方法和 MATLAB实现与应用[M].伍卫国,万群,张辉,等译.北京:机械工业出版社,2009.[7]关治,陆金甫.数值方法[M].北京:清华大学出版社,2006.[8]冯康.数值计算方法[M].北京:科学出版社,1981.[9]《现代应用数学手册》编委会.现代应用数学手册:计算与数值分析卷[M].北京:清华大学出版社,2005.[10]Nicholas J H.Accuracy and Stability of Numerical Algorithms[M].SecondEdition.Washington:Society for Industrial and Applied Mathematics,2002.[11]张雨浓,郭东生,徐思洪,等.未知目标函数之一阶数值微分公式验证与实践[J].甘肃科学学报,2009,21(1):13-18.[12]张雨浓,陈扬文,易称福,等.Hermite正交基前向神经网络的权值直接确定法[J].甘肃科学学报,2008,20(1):82-86.[13]Zhang Yunong,Li Zhan,Li plex-valued Zhang Neural Network for Online Complex-valued Time-varying Matrix Inversion[J].Applied Mathematics and Computation,2011,217(24):10 066-10 073.[14]Yunong Zhang,Yiwen Yang,Ning Tan,etal.Zhang Neural Network Solving forTime-varying Full-rank Matrix Moore-Penrose Inverse[J].Computing,2011,92(2):97-121.[15]Zhang Yunong,Xiao Lin,Ruan Gongqin,etal.Continuous and Discrete Time Zhang Dynamics for Time-varying 4th Root Finding[J].Numerical Algorithms,2011,57(1):35-51.[16]Yunong Zhang,Yiwen Yang,Gongqin Ruan.Performance A-nalysis of Gradient Neural Network Exploited for Online Time-varying Quadratic Minimization and Equality-constrained Quadratic Programming[J].Neurocomputing,2011,74(10):1 710-1 719.[17]Zhang Yunong,Yi Chenfu,Guo Dongsheng,parison on Zhang Neural Dynamics and Gradient-based Neural Dynamics for Online Solution of Nonlinear Time-varying Equation[J].Neural Computing and Applications,2011,20(1):1-7.。
输电线路覆冰预测技术研究综述
输电线路覆冰预测技术研究综述【摘要】架空输电线路的覆冰轻则引起线路过荷载以及导线舞动等危害,重则致使断线、倒塔,以致电网瘫痪。
由此可见,架空输电线路覆冰严重地影响了输电网的可靠性,给电力系统的安全稳定运行带来了巨大的威胁。
由于输电线路覆冰主要受气象因子的影响,因此从气象因子方面讨论输电线路覆冰规律以及建立大气候背景下的输电线路覆冰预测模型具有十分重要的意义。
【关键词】输电线路覆冰预测技术气象因子1 前言当严重的冰灾持续来袭时,输电线路不可避免的会出现覆冰。
与其他事故相比,冰p按照覆冰的形成条件,可以把导线覆冰分为雨凇、雾凇、混合淞、白霜、雪和雾[6],其性质和形成条件如表1所示。
一般过冷却水滴越小,越容易形成雾凇;过冷却水滴较大,在海拔较低的地区,则容易形成雨凇。
在我国,雨凇多见于湖南、粤北、赣南、湖北、河南及皖南等丘陵地区,而雾凇多见于云贵高原或海拔在1000米以上的山区,尤其海拔在2000-3000米的山区最多。
3 导线覆冰的影响因素分析导线覆冰受不同因素的影响,其厚度、密度和单位长度覆冰量等都会存在差异。
而影响导线覆冰的因素很多,主要有气象条件、地形条件、海拔高程、凝结高度、导线悬挂高度、导线直径、水滴直径、风速风向、负荷电流及电场强度等[2]。
重点探讨气象条件、地形条件对输电线路覆冰的影响。
(1)气象条件对覆冰的影响:导线覆冰主要是由气象条件决定的,是受温度、湿度、风、冷暖空气对流以阳光照射等因素影响的综合物理现象。
其中影响导线覆冰的主要因素有 4 种:空气温度、空气中液态水含量、空气中或云中过冷却水滴直径、以及风速风向。
这4种因素的不同组合确定了导线的覆冰类型。
雨凇覆冰形成时,通常温度较高,一般在-5—0℃之间,水滴直径大,一般在10-40μm之间:雾凇覆冰形成时,温度较低,在-8℃以下,一般在-15到-10℃之间,水滴直径在1-20μm之间:混合凇则通常介于雨凇和雾凇之间,混合凇覆冰时的温度范围为-9到-3℃,水滴直径在5-35μm之间。
简单英文自我介绍
简单英文自我介绍In addition ,I like swimming ,on one hand, ming can exercise my body: on the other hand ,I enjoy the feeling in the can make me relax,and reduce the pressure of learning. What’s more,and the mostimportant,it can help me lose weight ne who want to swim I think we can go to together.That's all, thank you!2.简单英文自我介绍篇2I am a third year master major in automation at Shanghai Jiao Tong University, P. R. Ch ina. With tremendous interest in Industrial Engineering, I am writing to apply for acceptan ce into your Ph.D. graduate program. Education background In 1995, I entered the Nanjing University of Science & Technology (NUST) -- widely co nsidered one of the China’s best engineering schools. During the following undergraduate study, my academic records kept distinguished among the whole department. I was granted First Class Prize every semester, and my overall GPA(89.5/100) ranked No.1 among 113 students. In 1999, I got the privilege to enter the graduate program waived of the admiss ion test. I selected the Shanghai Jiao Tong University to continue my study for its best re putation on Combinatorial Optimization and Network Scheduling where my research interes t lies. At the period of my graduate study, my overall GPA(3.77/4.0) ranked top 5% in the depa rtment. In the second semester, I became teacher assistant that is given to talented and m atured students only. This year, I won the Acer Scholarship as the one and only candidate in my department, which is the ultimate accolade for distinguished students endowed by my university. Presently, I am preparing my graduation thesis and trying for the honor of Excellent Graduation Thesis. Research experience and academic activity When a sophomore, I joined the Association of AI Enthusiast and began to narrow down my interest for my future research. In 1997, I participated in simulation tool development for the scheduling system in Prof. Wang’s lab. With the tool of OpenGL and Matlab, I d esigned a simulation program for transportation scheduling system. It is now widely used by different research groups in NUST. In 1998, I assumed and fulfilled asewage analysis & dispose project for Nanjing sewage treatment plant. This was my first practice to conv ert a laboratory idea to a commercial product. In 1999, I joined the distinguished Professor Yu-Geng Xis research group aiming at Netw ork flow problem solving and Heuristic algorithm research. Soon I was engaged in the Fu Dan Gene Database Design. My duty was to pick up the useful information among differe nt kinds of gene matching format. Through the comparison and analysis for many heuristi c algorithms, I introduced an improved evolutionary algorithm -- Multi-population GeneticAlgorithm. By dividing a whole population into several sub-populations, this improved algorithm can effectively prevent GA fromlocal convergence and promote various evolutionary orientations. It proved more efficiently than SGA in experiments, too. In the second semester, I joined the workshop-scheduling research in Shanghai Heavy Duty Tyre plant. The scheduling was designed for the rubber-making process that covered not only discrete but also continuous circumstances. To make a balance point between optimization quality and time cost, I proposed a Dynamic Layered Scheduling method based on hybrid Petri Nets. The practical application showed that the average makespan was shortened by a large scale. I also publicized two papers in core journals withthis idea. Recently, I am doing research in the Composite Predict of the Electrical Power system assisted with the technology of Data Mining for Bao Steel. I try to combine the Decision Tree with Receding Optimization to provide a new solution for the Composite Predictive Problem. This project is now under construction.Besides, In July 20xx, I got the opportunity to give a lecture in English in Asia Control Conference (ASCC) which is one of the top-level conferences among the world in the area of control and automation. In my senior year, I met Prof. Xiao-Song Lin, a visiting professor of mathematics from University of California-Riverside, I learned graph theory from him for my network research. These experiences all rapidly expanded my knowledge of English and the understanding of western culture.I hope to study in depthIn retrospect, I find myself standing on a solid basis in boththeory and experience, which has prepared me for the Ph.D. program. My future research interests include: Network Scheduling Problem, Heuristic Algorithm research (especially in GA and Neural network), Supply chain network research, Hybrid system performance analysis with Petri nets and Data Mining.3.简单英文自我介绍篇3I spend most of my time on study,i have passed CET4/6 . and i have acquired basic knowledge of my major during my school time.In July 20xx, I began work for a small private company as atechnical support engineer in QingDao use I'm capable of more responsibilities, so I decided to change my job.And in August 20xx,I left QingDao to BeiJing and worked for aforeign enterprise as a automation software test use I want to change my working environment, I'd like to find a job which is more challenging. Morover Motorola is a global company, so I feel I can gain the most from working in this kind of company ennvironment. That is the reason why I come here to compete for this position.I think I'm a good team player and I'm a person of great honesty to others. Also I am able to work under great pressure.That’s all. Thank you for giving me the chance.。
- 1、下载文档前请自行甄别文档内容的完整性,平台不提供额外的编辑、内容补充、找答案等附加服务。
- 2、"仅部分预览"的文档,不可在线预览部分如存在完整性等问题,可反馈申请退款(可完整预览的文档不适用该条件!)。
- 3、如文档侵犯您的权益,请联系客服反馈,我们会尽快为您处理(人工客服工作时间:9:00-18:30)。
I. I NTRODUCTION The problem of linear matrix equations solving (including matrix-inverse problems as a sub-topic) is considered to be a very fundamental problem widely encountered in science and engineering. It could usually be an essential part of many solutions; e.g., in control system design [1][2] and image-processing [3]. In view of these, we consider in this paper the following general problem formulation of linear matrix equation: AXB = C , where coefficient matrices A ∈ Rm×m , B ∈ Rn×n and C ∈ Rm×n , while X ∈ Rm×n is the unknown matrix to be found. Evidently, when B = C = I and m = n, the problem reduces to the matrix-inversion problem. There are two general types of solutions to the problem of linear matrix equations. One is the numerical algorithms per-
formed on digital computers (i.e., on our today’s computers). Usually, such numerical algorithms are of serial-processing nature and may not be efficient enough for large-scale online or real-time applications. Being the second general type of solution, many parallel-processing computational methods have been developed, analyzed, and implemented on specific architectures [3]-[13]. The dynamic-system approach is one of such important parallel-processing methods for solving linear matrix equations. Recently, because of the in-depth research in neural networks, numerous dynamic and analog solvers based on recurrent neural networks (RNN) have been developed and investigated [3][6][10]-[13]. The neural dynamic approach is thus now regarded as a powerful alternative to online computation of matrix problems because of its parallel distributed nature and convenience of hardware implementation [10][14]. Different from gradient neural networks for constant problems solving [2][3][6][10][15]-[17], a special kind of recurrent neural networks has recently been proposed by Zhang et al [10]-[12] for real-time solution of time-varying problems solving. In other words, in our context of AXB = C , coefficient matrices A, B and C could be A(t), B (t) and C (t), time-varying ones. The design method of Zhang neural network is completely different from that of gradient neural networks. In this paper, we generalize such a design method to solving online the time-varying linear matrix equation, A(t)X (t)B (t) = C (t) over time t ∈ [0, +∞). Theoretical and simulation results both demonstrate the efficacy of the proposed ZNN neural approach. To the best of our knowledge, there is little work dealing with such a timevarying problem in the literature at present stage, except some preliminary results presented in [11]-[13]. The main contributions of the paper are thus as follows. 1) In our paper, we propose a special kind of recurrent neural network to solve the time-varying linear matrix equation in real-time. As far as we know (with 10year research experience on neural networks), there are almost no other papers working on these time-varying
1
Yunongct— For solving online the linear matrix equation AXB = C with time-varying coefficients, this paper presents a special kind of recurrent neural networks by using a design method recently proposed by Zhang et al. Compared with gradient neural networks (abbreviated as GNN, or termed as gradient-based neural networks), the resultant Zhang neural network (termed as such and abbreviated as ZNN hereafter for presentation convenience) is designed based on a matrixvalued error function, instead of a scalar-valued error function. Zhang neural network is deliberately developed in the way that its trajectory could be guaranteed to globally exponentially converge to the time-varying theoretical solution of given linear matrix equation. In addition, Zhang neural network is described by an implicit dynamics, instead of an explicit dynamics usually describing recurrent neural networks. Convergence results of Zhang neural network are presented to show the neuralnetwork performance. In comparison, we develop and simulate the gradient neural network as well, which is exploited to solve online the time-varying linear matrix equation. Computersimulation results substantiate the theoretical efficacy and superior performance of Zhang neural network for the online solution of time-varying linear matrix equation, especially when using a power-sigmoid activation function. Index Terms— Recurrent neural network; Gradient neural network; Time-varying linear matrix equation; Matrix-valued error function; Implicit dynamics.