BP神经网络结合变量选择方法在牛奶蛋白质含量检测中的应用北大核心CSCD

牛奶中的蛋白质含量会影响牛奶的品质,利用高光谱图像的光谱特征信息研究对牛奶蛋白质含量预测的可行性。本文提出一种基于竞争性自适应重加权算法(competitive adaptive reweighted sampling, CARS)和连续投影算法(successive projections algorithm, SPA)结合多层前馈神经网络(back propagation, BP)的预测建模方法,实验以含有不同浓度蛋白质的牛奶为对象,利用可见光/近红外高光谱成像系统共采集到5种牛奶共计250组高光谱数据,通过实验对比选择采用标准化方法对获取到的吸收光谱预处理,然后采用CARS结合SPA筛选特征波长,得到18个特征波长,建立CARS-SPA-BP模型,经过试验,CARS-SPA-BP模型的训练集决定系数和测试集决定系数R;和R;分别达到0.971和0.968,训练集均方根误差(root mean square error of calibration,RMSEC)和测试集均方根误差(root mean square error of prediction,RMSEP)达到了0.033和0.034。研究发现,采用CARS结合SPA筛选的牛奶特征波长建立的多层前馈神经网络模型,其模型预测结果与全波长建模相比并没有明显降低,因此将CARS结合SPA用于波长筛选并且结合BP神经网络基本可以完成对牛奶蛋白质含量的预测。为验证CARS-SPA-BP模型的预测能力,在相同数据环境下,使用较为传统的偏最小二乘回归(partial least squares regression, PLSR)进行建模,实验结果表明,CARS-SPA-BP相较于PLSR,R;和RMSEP均有明显提升。研究表明,CARS-SPA-BP可充分利用牛奶光谱特征信息实现较高精度的牛奶蛋白质含量检测。     

异质性和时滞作用下神经元网络的共振动力学

利用参数互异的Fitzhugh-Nagumo神经元构建了含耦合时滞的无标度神经元网络模型,通过数值模拟的方法,提出研究参数异质性和耦合时滞影响下神经元网络的共振动力学.结果发现,当耦合项中不含时滞时,适中的参数异质性能够使得神经元网络对外界弱周期信号的响应达到最优,即适中的参数异质性能够诱导神经元网络的共振响应,而且异质性诱导共振对耦合强度具有鲁棒性.更重要的是,耦合时滞对参数异质性作用下神经元网络的共振特性也有着显著性影响.当时滞约为信号周期的整数倍时,神经元网络能够周期性地出现共振现象,即适当的耦合时滞能够诱导神经元网络的多重共振,而且这种现象在异质性参数的适当范围内都能明显出现.     

一种快速自适应谱估计算法

通过对梯度格型快速自适应算法的研究,结合最大熵谱外推目前存在的缺点,将梯度格型算法应用于最大熵谱外推中,并将其与伯格算法进行了比较。     

An efficient Chebyshev-Tau spectral method for Ginzburg-Landau-Schrödinger equations

We propose an efficient time-splitting Chebyshev-Tau spectral method for the Ginzburg-Landau-Schrödinger equation with zero/nonzero far-field boundary conditions. The key technique that we apply is splitting the Ginzburg-Landau-Schrödinger equation in time into two parts, a nonlinear equation and a linear equation. The nonlinear equation is solved exactly; while the linear equation in one dimension is solved with Chebyshev-Tau spectral discretization in space and Crank-Nicolson method in time. The associated discretized system can be solved very efficiently since they can be decoupled into two systems, one for the odd coefficients, the other for the even coefficients. The associated matrices have a quasi-tridiagonal structure which allows a direction solution to be obtained. The computation cost of the method in one dimension is O(Nlog(N)) compared with that of the non-optimized one, which is O(N²). By applying the alternating direction implicit (ADI) technique, we extend this efficient method to solve the Ginzburg-Landau-Schrödinger equation both in two dimensions and in three dimensions, respectively. Numerical accuracy tests of the method in one dimension, two dimensions and three dimensions are presented. Application of the method to study the semi-classical limits of Ginzburg-Landau-Schrödinger equation in one dimension and the two-dimensional quantized vortex dynamics in the Ginzburg-Landau-Schrödinger equation are also presented.     

一种改进的谱聚类方法及其在文本分析中的应用

随着文本资源的激增,特别是网页文本的迅速增加,针对文本的挖掘分析日益受到重视。谱聚类是文本聚类分析较常用的一种新型方法。该文将非负约束引入到传统的谱聚类算法中,提出了一种基于非负约束的谱聚类方法。文中实验验证了所提出方法在中文文本聚类分析应用中的有效性。     

一种基于边界约束的流形展开方法

在流形学习的谱方法中, 流形展开被表述为优化问题. 这些优化问题的解是退化的, 即所有的样本将被嵌入到同一个点. 为了避免退化解, 谱方法对嵌入坐标人为地强加了一个单位协方差矩阵约束. 然而, 该约束往往导致流形展开的失真非常明显. 本文提出一种新的流形学习方法, 彻底抛弃了人为的单位协方差矩阵约束. 主要思路是先对流形边界进行嵌入, 然后再求流形内部的嵌入; 流形边界的嵌入位置被确定后, 流形内部样本的嵌入位置将被边界拉开, 使得它们不会都收缩到一个点上, 从而避免了退化解的出现. 将流形边界的嵌入位置作为边界条件, 求解一个线性方程组来得到内部样本的嵌入; 该线性方程组反映了尽量保持邻近样本间距离不变的要求. 流形边界的嵌入由简化流形的嵌入求出; 为此, 本文还设计了一种流形边界检测算法以及一种流形简化算法. 与目前代表性的几种流形学习方法进行了比较实验, 结果表明了本文方法的有效性, 其展开失真比谱方法明显要小.     

车载环境下基于样本熵的语音端点检测方法

在语音处理中一个关键性问题是如何准确找到语音的起止位置,目前提出许多的语音端点检测算法不能得到理想的检测结果.由于样本熵是近似熵的改进算法,提出车载环境下基于样本熵的语音端点检测方法.并采用模糊C均值聚类算法和贝叶斯信息判决算法进行样本熵特征门限估计,以及使用双门限法进行语音端点检测.在TIMIT连续语音库上的实验表明,车载噪声环境下,样本熵法和近似熵法的检测正确率均远高于谱熵法和能量谱熵法,而样本熵法相对于近似熵法具有更好的检测效果,特别是当信噪比小于等于OdB时,样本熵法的检测性能优于近似熵法近10%.因此,样本熵法在车栽智能语音领域具有很好的应用前景,能够为车载导航提供准确的语音端点检测技术.     

光谱响应差异对高精度交叉定标的影响——以FY-3A/MERSI与EOS/MODIS为研究实例

分析了MERSI与MODIS间光谱响应差异对测量表观反射比的影响。正演模拟结果显示,在极区冰雪情况下,MERSI与MODIS测量得到的表观反射比之间的相对差异是关于MODIS所测表观发射比的二次函数。MERSI的前4个通道与对应的MODIS通道之间的相对差异在0~0.8%,0~2%,0.5%~2.5%,-1.8%~-0.8%之间。将正演模拟得到的对光谱响应差异的修正关系应用到具体的SNO测量后,MERSI前3个通道的表观反射比与修正后MODIS的表观反射比之间的相对差异减小到了3%以内。MERSI的1、2通道在红外波段也有响应,这部分响应在极区冰雪条件下对表观反射率的影响在0.5%左右,基本可以忽略。     

On the use of small training sets for neural network-based characterization of mixed pixels in remotely sensed hyperspectral images

In this work, neural network-based models involved in hyperspectral image spectra separation are considered. Focus is on how to select the most highly informative samples for effectively training the neural architecture. This issue is addressed here by several new algorithms for intelligent selection of training samples: (1) a border-training algorithm (BTA) which selects training samples located in the vicinity of the hyperplanes that can optimally separate the classes; (2) a mixed-signature algorithm (MSA) which selects the most spectrally mixed pixels in the hyperspectral data as training samples; and (3) a morphological-erosion algorithm (MEA) which incorporates spatial information (via mathematical morphology concepts) to select spectrally mixed training samples located in spatially homogeneous regions. These algorithms, along with other standard techniques based on orthogonal projections and a simple Maximin-distance algorithm, are used to train a multi-layer perceptron (MLP), selected in this work as a representative neural architecture for spectral mixture analysis. Experimental results are provided using both a database of nonlinear mixed spectra with absolute ground truth and a set of real hyperspectral images, collected at different altitudes by the digital airborne imaging spectrometer (DAIS 7915) and reflective optics system imaging spectrometer (ROSIS) operating simultaneously at multiple spatial resolutions.     

Parametric polynomial spectral factorization using the sum of roots and its application to a control design problem

This paper presents an algebraic approach to polynomial spectral factorization, an important mathematical tool in signal processing and control. The approach exploits an intriguing relationship between the theory of Gröbner bases and polynomial spectral factorization which can be observed through the sum of roots, and allows us to perform polynomial spectral factorization in the presence of real parameters. It is discussed that parametric polynomial spectral factorization enables us to express quantities such as the optimal cost in terms of parameters and the sum of roots. Furthermore an optimization method over parameters is suggested that makes use of the results from parametric polynomial spectral factorization and also employs two quantifier elimination techniques. This proposed approach is demonstrated in a numerical example of a particular control problem.