一种变分优化问题的近似解法

针对传统变分法求解困难的问题,提出一种变分优化问题的近似解法。根据最小二乘近似解法的简便性与粒子群优化算法参数少的特性,在最小二乘近似解法的求解过程中引入粒子群优化算法,并给出求解流程。数值仿真实验结果表明,该算法计算过程简单,优化效果较好。     

基于无监督学习的智能电网入侵检测

智能电网通过引入信息和通信技术服务, 带来了传统电网的技术演变, 与此同时在安全方面也带来了严重的挑战. 本文提出了一种智能电网入侵检测系统安全架构和一种基于无监督学习的新型入侵检测系统(intrusion detection system, IDS). 我们设计了区域式训练(block-training)架构, 不仅可以减轻数据中心的计算压力, 还可以对本地流量进行特征训练. 我们还提出了一种基于交叉验证的递归特征消除的差分自编码器算法(RFECV-VAE). RFECV-VAE综合了RFECV和VAE模型, 在特征选择过程使用递归特征消除交叉验证法(recursive feature elimination cross-validation, RFECV), 异常检测采用差分自编码器(variational autoencoders, VAE), 它可以对大规模高维数据进行高精度异常检测. 最后, 本文选择深度自编码器、深度自编码器高斯混合模型、单类支持向量机、隔离森林、差分自编码器作为对比算法, 采用准确率、ROC_AUC、F1_score和训练时间等指标来进行性能评估. 实验结果表明, RFECV-VAE算法结果优于其他比较算法.     

基于生成对抗网络和变分自编码器的离群点检测算法

针对传统离群点检测算法在类极度不平衡的高维数据集中难以学习离群点的分布模式,导致检测率低的问题,提出了一种生成对抗网络(generative adversarial network,GAN)与变分自编码器(variational auto-encoder,VAE)结合的GAN-VAE算法。算法首先将离群点输入VAE训练,学习离群点的分布模式;然后将VAE与GAN结合训练,生成更多潜在离群点,同时学习正常点与离群点的分类边界;最后将测试数据输入训练后的GAN-VAE,根据正常点与离群点相对密度的差异性计算每个对象的离群值,将离群值高的对象判定为离群点。在四个真实数据集上与六个离群点检测算法进行对比实验,结果表明GAN-VAE在AUC、准确率和F;值上平均提高了5.64%、5.99%和13.30%,证明GAN-VAE算法是有效可行的。     

基于条件变分自编码的密码攻击算法

使用密码猜测算法是评估用户密码强度和安全性的有效方法,提出一种基于条件变分自编码密码猜测算法PassCVAE。算法基于条件变分自编码模型,将用户个人信息作为条件特征,训练密码攻击模型。在编码器端,分别使用双向循环神经网络（GRU）和文本卷积神经网络（TextCNN）,实现对密码序列和用户个人信息的编码和特征的抽象提取;在解码器端使用两层GRU神经网络,实现对用户个人信息和密码数据隐编码的解码,生成密码序列。该算法可以有效地拟合密码数据的分布和字符组合规律,生成高质量的猜测密码数据。多组实验结果表明,提出的PassCVAE算法优于现有的主流密码猜测算法。     

基于二维变分模态分解和自适应中值滤波的图像去噪方法*

图像在采集、获取和传输过程中往往夹杂着噪声,针对几种常用方法去噪效果不理想,提出了一种新的图像去噪方法。此方法通过二维变分模态分解将图像分解为一系列不同中心频率的子模态。保留其低频模态,并对其进行自适应中值滤波处理,从而得到其去噪后的图像。实验结果表明,与其他几种常用的去噪方法相比,本文方法在滤除噪声的同时,又能较好地保留图像的边缘细节,图像也获得较好的视觉效果,此外客观评价参数也得到比较明显的改善,随着噪声强度加大去噪效果愈明显。     

基于作物生长模型和多源数据的融合技术研究进展

精确的区域作物产量估计在社会食品安全生产中起着重要作用。首先讨论了常用的2种区域作物产量估计方法,包括产量监测和产量模拟。其中,作物生长模型能够反映作物的整个生长演变过程,并能最终预报作物产量,因此在区域作物产量预报中起着重要作用,但是由于作物生长模型在输入数据、模型参数和模型结构等方面存在较大的不确定性,导致最终的模拟结果也存在较大的不确定性,尤其是应用到区域尺度时,这种不确定性使得模拟结果同真实的作物产量空间分布图存在较大的不一致性。而产量监测,尤其是利用先进的多源遥感信息,可捕捉真实的区域尺度的地面作物生长信息,但是仅为瞬时信息。因此利用数据融合算法,融合模型和数据的优点,得到更为可靠的区域产量估计结果是十分有意义的。所以在详述了当前主要的作物生长模型的基础上,重点讨论了常用的2种数据融合技术,即优化方法和顺序数据同化方法,以及目前利用这两种方法在作物生长模型中融合观测信息的部分案例。     

Computing interval-valued statistical characteristics: what is the stumbling block for reliability applications?

The application of interval-valued statistical models is often hindered by the rapid growth in imprecision that occurs when intervals are propagated through models. Is this deficiency inherent in the models? If so, what is the underlying cause of imprecision in mathematical terms? What kind of additional information can be incorporated to make the bounds tighter? The present paper gives an account of the source of this imprecision that prevents interval-valued statistical models from being widely applied. Firstly, the mathematical approach to building interval-valued models (discrete and continuous) is delineated. Secondly, a degree of imprecision is demonstrated on some simple reliability models. Thirdly, the root mathematical cause of sizeable imprecision is elucidated and, finally, a method of making the intervals tighter is described. A number of examples are given throughout the paper.     

Fast iterative algorithms for solving the minimization of curvature-related functionals in surface fairing

A number of successful variational models for processing planar images have recently been generalized to three-dimensional (3D) surface processing. With this new dimensionality, the amount of numerical computations to solve the minimization of such new 3D formulations naturally grows up dramatically. Though the need of computationally fast and efficient numerical algorithms able to process high resolution surfaces is high, much less work has been done in this area. Recently, a two-step algorithm for the fast solution of the total curvature model was introduced in Tasdizen, Whitaker, Burchard and Osher [Geometric surface processing via normal maps, ACM Trans. Graph. 22(4) (2003), pp. 1012–1033]. In this paper, we generalize and modify this algorithm to the solution of analogues of the mean curvature model of Droske and Martin Rumpf [A level set formulation for Willmore flow, Interfaces Free Bound. 6(3) (2004), pp. 361–378] and the Gaussian curvature model of Elsey and Esedo?lu [Analogue of the total variation denoising model in the context of geometry processing, Multiscale Model. Simul. 7(4) (2009), pp. 1549–1573]. Numerical experiments are shown to illustrate the good performance of the algorithms and test results.     

An iterative method for suboptimal control of a class of nonlinear time-delayed systems

This work presents an approximate solution method for the infinite-horizon nonlinear time-delay optimal control problem. A variational iteration method (VIM) is applied to design feedforward and feedback optimal controllers. By using the VIM, the original optimal control is transformed into a sequence of nonhomogeneous linear two-point boundary value problems (TPBVPs). The existence and uniqueness of the optimal control law are proved. The optimal control law obtained consists of an accurate linear feedback term and a nonlinear compensation term which is the limit of an adjoint vector sequence. The feedback term is determined by solving Riccati matrix differential equation. By using the finite-step iteration of a nonlinear compensation sequence, we can obtain a suboptimal control law. Simulation results demonstrate the validity and applicability of the VIM.