Dirichlet过程混合模型的聚类算法

有限混合模型进行高维数据聚类分析时需预先估计聚类个数,因而聚类的准确性和泛化性受到影响.通过建立Dirichlet过程无限混合模型对高维数据开展聚类分析,采用Dirichlet过程的Urn模型分析出模型中各参数的后验分布,利用Gibbs采样MCMC方法估计出模型中各参数及数据中潜在的聚类数.在五维的仿真数据集和IRIS测试数据集上的聚类结果表明:经过200次Gibbs采样MCMC过程,该算法能够正确地估计出数据中潜在的聚类数.单次Gibbs采样MCMC过程的平均占用时间分别为0.185 0s和0.145 5s,其时间复杂度和数据的样本个数N有关,为O(N).     

AFM research on Fe-based nanocrystal crystallization mechanism

The cross-section pattern of Fe-based alloy ribbon (Fe_73.5Cu₁Nb₃Si_13.5B₉) annealed at different temperatures was investigated by AFM (atomic force microscope), and the effect mechanism of Nb and Cu in Fe-based alloy ribbon annealing was analyzed with XRD diffraction crystal analysis technique and other research results. New concepts of encapsulated grain, Nb vacancy cluster, Nb-B atom cluster and so on were proposed and used to describe the formation mechanism of α-Fe (Si) nanocrystal. Finally, a three-phase (separation phase, encapsulated phase and nanocrystalline phase) interconnected structure model in Fe-based nanocrystalline alloy was established. Supported by the Natural Science Foundation of Zhejiang Province (Grant No. Y405021), Zhejiang Provincial Science and Technology Key Project (Grant No. 2006C21109) and Key Project of Science and Technology Research of China Ministry of Education (Grant No. 204059)     

An LMI-based unified approach to reduced-order controller design

The design of reduced-order controllers is considered for stabilization control, covariance upper bound control, linear quadratic regulator, ∞ control, H∞ control, positive real control problems and robust H2 control, robust ∞ control and robust H∞ control problems for generalized linear plants without any additional assumptions. An upper bound of the order is obtained with which the (robust) controllers can guarantee stability constraints and satisfy the same design objectives as the so-called "full-order" controllers. A unified linear-matrix-inequality ( LMI) based approach to reduced-order (robust) controller design for all the above-mentioned problems is provided. It us shown that each of these problems is solvable if and only if two uncoupled LMIs with an LMI-type constraint have a pair of positive definite solutions, one of which has a lower dimension than that given by Skelton and iwasaki (1995). All desired reduced-order (robust) controllers are parameterized via the solutions of LMIs Moreov     

Durability of asphalt mixture in different corrosion solution

The corrosion to asphalt mixture under different kinds of corrosion solution, such as pH=2 solution, pH=12 solution, pH = 12 solution and 10% Na₂SO₄ solution, was studied. The performance attenuation of asphalt mixture was analyzed under the normal environment and the freeze-thaw environment, and the analysis was given on the sensitivity of the test results to the evaluation index. The experimental results show that the performance of asphalt mixture is attenuated faster under the acidic solution, alkaline solution and sulfate solution. Corrosion factor K _c, freeze-thaw corrosion factor K _f, and freeze-thaw effect factor K _fc are proposed to evaluate asphalt mixture resistance to corrosion in different kinds of corrosion solution. The values of K _c and K _fc decrease with the increasing of corrosion time. The change rule of K _f show that the rate of corrosion is decreased by the action of freeze-thaw in acidic solution and in alkaline solution, but is increased by the action of freeze-thaw in sulfate solution. The microscopic analysis indicates that acid solution reacts with aggregate of asphalt mixture, alkaline solution reacts with asphalt cement of asphalt mixture, the surface tension of sulfate solution and crystallization of sulfate are the main reasons which weak the performance of asphalt mixture.     

K-均值聚类算法的MapReduce模型实现

针对日益严峻的大数据处理时间长、执行速率低等问题,通过深入分析,提出了一种提高大规模数据聚类效率的方法。以K-均值聚类算法为原型,利用Map Reduce模型在大规模数据处理方面的优势,对原有算法进行并行化改进,设计出一种基于Hadoop分布式云平台的K-均值聚类Map Reduce模型。应用此模型,对淘宝用户仿真数据进行聚类试验,试验结果表明,对K-均值聚类算法的Map Reduce模型实现后,性能优于原算法性能,缩短了聚类时间,提高了聚类效率,特别适于对海量数据进行聚类处理。     

融合AP和GMM的说话人识别方法研究

针对在说话人识别过程中经典的高斯混合模型(Gaussian Mixture Model,GMM)阶数的确定具有很大随意性的问题,提出采用吸引子传播聚类方法(AP聚类)自动获取GMM的阶数,进而实现说话人识别的方法.首先,采用Mel频率倒谱系数法(MFCC)与差分倒谱相结合的方法,提取语音特征参数;其次,采用吸引子传播聚类方法(AP聚类)对语音特征参数进行聚类处理,从而自动获得GMM的阶数;在此基础上进行GMM模型的训练;最后,采用训练好的GMM模型对Timit标准语音库以及自制网络志愿者语音库进行说话人识别测试实验.实验结果为：使用了AP聚类算法获取GMM阶数的情况下,对Timit标准语音库的测试结果为100%;在自制网络志愿者语音库中,训练样本为168个,其中潮汕话样本10个,湖南话样本10个,测试样本为42个,测试结果为97.6%.实验结果表明,引入AP聚类自动获取GMM的阶数,可以显著提高说话人识别的精度和效率.     

土石坝有限元分析中土体卸载时计算方法分析

对糯扎渡心墙堆石坝进行了三维有限元计算,对比分析两种常用于邓肯E-B模型的加卸载准则以及回弹模量取值对土石坝应力变形的影响。结果表明:大坝施工过程中,基于加载函数的加卸载准则(简称准则1)判断的卸载回弹区域一直较大;准则1下数值计算所得的坝体位移较小,仅为采用基于应力水平s与偏应力(σ_1-σ_3)的加卸载准则(简称准则2)时计算位移的一半左右,加卸载准则的选取对坝体应力影响较小;监测数据表明糯扎渡大坝采用准则2时计算位移更为准确,在准则2下回弹模量取值对大坝应力变形的影响主要表现为满蓄期坝体水流向位移随回弹模量的增大而减小。     

Simulation calculation of solubility of insoluble compound MmAa in complex system

A simulation calculation model for the solubility of insoluble compound M _m A _a in complex system was established. According to coordination equilibrium principle, relevant dissociation reaction, complexation reaction, self-complexation reaction and protonation reaction during insoluble compound dissolving were considered and then the mass balance equations about solubility calculation were obtained. In the case analysis, the solubility of silver chloride in ammonia system was obtained by simulation calculation, and curved surface charts of thermodynamic equilibrium about the total concentration of silver ions, pH and concentration of ammonia ions were drawn correspondingly. The results show that under the conditions of room temperature and 6 mol/L ammonia concentration, the calculated solubility value of silver chloride (34 g/L) is close to the actual value (31 g/L), demonstrating that this model is suitable for solubility calculation of insoluble compound M _m A _a in the complex system.     

基于高斯混合模型的轨迹模仿学习表征参数优化

针对高斯混合模型(Gaussian mixture model,GMM)参数选取效率较低的问题,提出了一种在基于GMM的轨迹模仿学习表征中综合求解GMM参数估计的方法.该方法基于多中心聚类算法中的最大最小距离算法改进kmeans算法,得到最优初始聚类中心,并基于贝叶斯信息准则(Bayesian information criterion,BIC)通过遗传算法优化求解,同时获取GMM的4个重要参数.该方法通过提高划分初始数据集的效率,在优化初始聚类中心基础上确定混合模型个数,有效地避免了因为初值敏感而导致的局部极值问题.通过多组仿真实验验证了该方法的有效性.     

Prediction of water activities in multicomponent solutions of metal chlorides from the thermodynamic properties of corresponding binary systems

Ｉｎｈｙｄｒｏｍｅｔａｌｌｕｒｇｙ ,ｅｌｅｃｔｒｏｌｙｔｅｓｏｌｕｔｉｏｎｉｓｏｆｔｅｎｄｉｓ ｃｕｓｓｅｄ.Ｔｈｅａｃｔｉｖｉｔｙｏｆｗａｔｅｒｉｓｖｅｒｙｉｍｐｏｒｔａｎｔｔｏｓｔｕｄｙｔｈｅｓｏｌｕｔｉｏｎ ,ｗｈｉｃｈｃａｎｂｅｅｍｐｌｏｙｅｄｔｏｃａｌｃｕｌａｔｅｍａｎｙｔｈｅｒｍｏｄｙｎａｍｉｃａｌｐｒｏｐｅｒｔｉｅｓｏｆｉｔ.Ｆｏｒｅｘａｍｐｌｅ ,ｔｈｅａｃｔｉｖｉ ｔｉｅｓｏｆｔｈｅｓｏｌｕｔｅｓｃａｎｂｅｏｂｔａｉｎｅｄｆｒｏｍｉｔｂｙｕｓｉｎｇｃｒｏｓｓ ｄｉｆｆｅｒｅｎｔｉａｔｉ…     

Linearization of T-S fuzzy systems and robust H_∞ control

Takagi-Sugeno (T-S) fuzzy model is difficult to be linearized because of membership functions included. So, novel T-S fuzzy state transformation and T-S fuzzy feedback are proposed for the linearization of T-S fuzzy system. The novel T-S fuzzy state transformation is the fuzzy combination of local linear transformation which transforms local linear models in the T-S fuzzy model into the local linear controllable canonical models. The fuzzy combination of local linear controllable canonical model gives controllable canonical T-S fuzzy model and then nonlinear feedback is obtained easily. After the linearization of T-S fuzzy model, a robust H _∞ controller with the robustness of sliding model control (SMC) is designed. As a result, controlled T-S fuzzy system shows the performance of H _∞ control and the robustness of SMC.     

无线传感器网络中基于时空分集的分布式能量有效成簇算法

为了延长无线传感器网络的生存时间,需要设计适合无线传感器网络特点的能量有效的协议。成簇算法能够减少无线传感器网络的能量消耗,对增强网络的可扩展性和延长网络的生存时间有着重要的作用。但现有算法对多路衰退的链路并不十分能量有效,针对这些算法的不足,提出了基于时空的分布式能量有效成簇算法,每个簇由一个簇首节点和一个协作节点相互合作运用时空编码实现数据传输的时空分集,以减少网络能耗和均衡节点能量。模拟实验结果显示,与现有重要成簇算法相比,该算法能够提供更长的网络生存时间。     

Effects of mixture inhomogeneity and combustion temperature on soot surface activity and soot formation in diesel engines

The soot surface growth plays significant role on the soot mass accumulation,which starts with H(hydrogen)atom abstraction forming activated soot surface sites,and is followed by the acetylene addition process.In this study,the effect of the mixture inhomogeneity and combustion temperature on the soot surface activity and soot formation was investigated by developing a new multi-step phenomenological(MSP)soot model of diesel engines.A new detailed soot surface growth mechanism was proposed by correlation analysis of combustion parameters with soot formation.The inhomogeneity coefficient of soot surface activityαCH and the specific rate of soot surface growth R CH were derived to highlight the effect of inhomogeneity of mixture and combustion temperature on soot formation.The predicted diesel engine-out soot agreed well with experimental findings in wide ranges of combustion conditions.In the case of lower engine load with single fuel injection and higher EGR(exhaust gas recirculation)rate,it had quiet homogeneous mixtures before ignition when the combustion temperature dominated the soot surface activity.At medium engine load with multi-pulse fuel injections,it got mixture slightly stratified before ignition and revealed that the mixture inhomogeneity became more dominated on soot surface activity than the combustion temperature.An increased soot surface activity led to increased soot emission.Under the full engine loads with single fuel injection but quite high boost pressure over 0.4 MPa,it led to the combustion conditions of higher mixture density and higher mixture heat capacity,which benefits the mixture homogeneity.The decay rate of soot surface activity became lower due to the decreased combustion temperature and the soot surface activity decreased due to improved mixture homogeneity.In addition,the lowered intake oxygen concentration due to usage of EGR played a role to lower the specific rate of soot surface growth R CH,but to increase the soot surface activityαCH.     

基于深度学习编码模型的图像分类方法

针对矢量量化编码的量化误差严重,而稀疏编码只是一种浅层学习模型,容易导致视觉词典对图像特征缺乏选择性的问题,提出了一种基于深度学习特征编码模型的图像分类方法。首先,采用深度学习网络无监督的受限玻尔兹曼机（RBM）代替传统的K-Means聚类及稀疏编码等方法对SIFT特征库进行编码学习,生成视觉词典;其次,对RBM编码添加正则化项分解组合每个特征的稀疏表示,使得生成的视觉单词兼具稀疏性和选择性;然后,利用训练数据的类别标签信息有监督地自上而下对得到的初始视觉词典进行微调,得到图像深度学习表示向量,以此训练SVM分类器并完成图像分类。实验结果表明,本文方法能有效克服传统矢量量化编码及稀疏编码等方法的缺点,有效地提升图像分类性能。     

中国便利店的照明能耗预测模型与主观评价

便利店销售额在传统零售业中增幅最高,照明在所调研样本总室内能耗中占比为43%。基于全国257家便利店的数据,探讨便利店的照明特征。采用多重线性回归方法建立照明能耗预测模型,采用因子分析方法探讨客观参数与主观评价的联系。研究表明,对便利店,考虑照明节能时需同步关注其主观评价的变化;进行照明设计时需因地制宜;照明能耗预测模型为ln y=0.479 ln x₁+ 0.248ln x₂-0.217（y为照明功率密度,x₁为单位面积照明造价,x₂为典型货架位置水平照度平均值）;顾客对照明充足度的“敏感性”比照明均匀度高;提高照度平均值有助于提高主观评价。     

基于Markov随机场的Student's t混合模型的脑MR图像分割

为解决在使用期望最大化(EM)算法求解混合模型前需要额外的计算问题,提出一种新的基于Markov随机场的Student's t混合模型,该模型能直接利用简单有效的EM算法求解。试验结果表明,该方法能有效克服噪声对图像分割的影响,获得较好的分割结果。     

一种基于高斯混合模型的无监督粗糙聚类方法

针对数据统计分布的随机性和复杂性,从统计聚类的角度出发,采用高斯混合模型来描述整个数据的概率密度函数,提出了一种基于高斯混合模型的粗糙聚类分析方法.该方法首先利用粗糙集理论的不可区分关系性质以及生成的逻辑规则来设定EM算法的初始近似参数,然后通过Expectation-M axim ization(EM)算法估计各分量概率密度分布的最大似然参数集,最后通过密度分布概率大小来确定类别的归属.与传统的k-m eans聚类算法的试验结果比较表明,该方法是有效的,并且具有较高的聚类精度,用规则集来描述聚类的结果具有可解释性和合理性.     

中国区域性大气加权平均温度线性模型精度评估

以欧洲中期天气预报中心的再分析资料ERA5为参考数据,评估由探空数据建立的中国区域88个单站大气加权平均温度(Tm)与地表气温(Ts)线性关系模型的精度.各站Tm-Ts线性模型计算的Tm(计算值)与ERA5气压层数据积分所得的Tm(参考值)间偏差均方根值(RMSE)为1.8～5.5 K.不同站模型计算值与参考值间存在-...     

聚类分析法在水质监测断面水质分析中的应用

为了在不影响污水治理的前提下减少监测断面的设定,利用SPSS中聚类分析的方法对滏阳河河段上八个监测断面的监测资料,选取COD、氨氮等五个水质指标监测数据进行优化分类,并对监测断面的水质变化趋势做了简要的分析。结果表明:Ⅰ类监测断面(南留旺、东武仕水库、木鼻、张庄桥)水质较好,Ⅱ类监测断面(沁河苏曹站、曲周南干渠)和Ⅲ类监测断面(苏里、莲花口)的水质污染较为严重,而这些监测断面均位于邯郸市区内或者是经过市区以后的一段距离内。因此邯郸市区内的重污染企业仍是水污染的主要污染源。聚类分析法可以很好地将多个水质断面进行归类分析,并对河水污染治理起到指导作用。