语言辨识系统中最佳线性融合技术的研究

本文利用不同参数提取方法对语言辨识系统中的线性融合技术进行了研究。融合系数的获取通过三个准则进行实现,CFM准则、MSE准则和CE准则。实验系统采用了区分性高斯混合模型,利用OGI-TS多语种电话语音语料库,对决策级融合性能进行了评估。实验表明,利用决策级融合技术,选择最佳融合系数,可以很好地改善语言辨识率。     

一种可在线学习的变结构径向基函数网络及其在被动声纳目标识别中的应用

提出一种新颖的隐节点可调的变结构径向基函数,并应用进化规划最优地确定和调节变结构径向基函数网络隐层节点的数目及其核函数的中心和宽度,从而使网络具有在线学习和记忆新的目标模式的功能,并将该网络应用于被动声纳目标的识别和在线学习实验结果表明基于进化规划的变结构径向基函数网络不仅了网络的泛化能力,而且能够有效垢解决传统神经网络技术在被动声纳目标识别过程中在线学习会造成原有记忆遗忘的困难。     

一种新的基于小波包分解的EEG特征抽取与识别方法研究

为了提高脑思维任务分类精度,提出一种新的脑电特征抽取与识别方法.首先进行小波包分解,然后结合能反映脑电信号在时域与频域上的能量分布特征的小波包熵概念,从小波包库中选择最优小波包基,对各个最优基所对应的小波系数求取统计特性,然后根据不同脑思维任务下左右半脑各导联间的差异性对各个导联对求取不对称率构成分类特征向量,最后利用SVM分类器对其进行分类.实验结果表明:相对于一般的小波包分解,最优小波包基和自回归特征抽取方法,该方法对5类不同脑思维任务的所有10种不同组合任务对的平均分类预测精度可以达到95.41％～99.65％.     

基于IVIF-VIKOR的雷达辐射源信号分选识别特征性能综合评价方法

为解决雷达辐射源信号分选识别特征评价不够客观和缺乏评价依据等问题,提出了一种基于区间模糊原理、模糊交叉熵和多准则折衷法的特征评价方法. 首先通过区间模糊原理建立信噪比分级评价模型,并基于汉明距离进行寻优得出信噪比权重;其次结合信噪比权重和区间直觉模糊加权平均算子将分级模型整合成群决策矩阵,使用熵最大化法计算属性权重;最后基于多准则折衷法框架,采取模糊交叉熵实现特征方案排序. 仿真实验结果表明,所提方法能够给出与实际仿真实验相一致的分选识别特征评价排序结果,并优于逼近理想点方法,验证了所提方法的可行性和有效性.     

采用HDPHMM符号化器的语音查询样例检测方法

提出一种基于层级狄利克雷过程隐马尔科夫模型（HDPHMM）符号化器的无监督语音查询样例检测（QbE-STD）方法。该方法首先应用一个双状态层隐马尔科夫模型,其中顶层状态用于表示所发现的声学单元,底层状态用于建模顶层状态的发射概率,通过对顶层状态假设一个层级狄利克雷过程先验,获得非参贝叶斯模型HDPHMM。使用无标注语音数据对该模型进行训练,然后对测试语音和查询样例输出后验概率特征矢量,使用非负矩阵分解算法对后验概率进行优化得到新的特征,然后在此基础上,应用修正分段动态时间规整算法进行检索,构成QbE-STD系统。实验结果表明,相比于基于高斯混合模型符号化器的基线系统,本文所提出的方法性能更优,检索精度得到显著提升。      

基于DBF的汉语方言自动辨识

在汉语方言辨识中,传统的声学特征是语音信号的谱特征的参数化表示,常常包含说话人、信道、背景噪声等冗余信息,针对上述问题将深度神经网络(Deep Neural Network,DNN)引入特征提取之中,提出了与音素层面相关的深度瓶颈特征(Deep Bottleneck Feature,DBF),尝试从特征层面抑制方言冗余信息的影响.最后在实验部分对瓶颈层的位置,节点数目进行了讨论,结果显示,深度瓶颈特征相对于传统声学特征能够取得更高的识别率.     

Identification of spinal deformity classification with total curvature analysis and artificial neural network

In this paper, a multilayer feed-forward, back-propagation (MLFF/BP) artificial neural network (ANN) was implemented to identify the classification patterns of the scoliosis spinal deformity. At the first step, the simplified 3-D spine model was constructed based on the coronal and sagittal X-ray images. The features of the central axis curve of the spinal deformity patterns in 3-D space were extracted by the total curvature analysis. The discrete form of the total curvature, including the curvature and the torsion of the central axis of the simplified 3-D spine model was derived from the difference quotients. The total curvature values of 17 vertebrae from the first thoracic to the fifth lumbar spine formed a Euclidean space of 17 dimensions. The King classification model was tested on this MLFF/BP ANN identification system. The 17 total curvature values were presented to the input layer of MLFF/BP ANN. In the output layer there were five neurons representing five King classification types. A total of 37 spinal deformity patterns from scoliosis patients were selected. These 37 patterns were divided into two groups. The training group had 25 patterns and testing group had 12 patterns. The 25-pattern training group was further divided into five subsets. Based on the definition of King classification system, each subset contained all five King types. The network training was conducted on these five subsets by the hold-out method, one of cross-validation variants, and the early stop method. In each one of the five cross-validation sessions, four subsets were alternatively used for estimation learning and one subset left was used for validation learning. Final network testing was conducted with remaining 12 patterns in testing group after the MLFF/BP ANN was trained by all five subsets in training group. The performance of the neural network was evaluated by comparing between two network topologies, one with one hidden layer and another with two hidden layers. The results are shown in three tables. The first table shows network errors in estimation learning and the second table shows identification rates in validation learning. The network errors and identification rates in the last round of network training and testing are shown in the third table. Each table has a comparison for both one hidden layer and two hidden layer networks.     

最优双核复合分类算法的构造

 由于使用单一且固定的核函数,传统的核分类算法不能有效地适应复杂的数据集合,导致分类性能下降.本文提出一种基于双核复合的分类算法ODKC(Optimal Double-Kernel Combination)的构造框架,通过融合两个基本核函数的映射来构造目标核函数.研究了双核复合的三种典型方式,并把这三种复合方式纳入到统一的框架下处理.论文以核与数据的匹配性度量KTA(Kernel Target Alignment)以及分类性能验证了所提算法的有效性.     

基于改进熵特征的心律失常分类研究

心律失常类型的判断对心血管疾病的防治十分重要,针对波动散布熵(multiscale f luctuation dispersion entropy,FDE)在进行心律失常分类识别时尺度单一、不能全面反映心律失常 信息等不足,通 过改进FDE特征,提出一种基于自适应多尺度波动散布熵(adaptive multiscale fluctuati on dispersion entropy,AMFDE)的心律失常分类方法。首先在计算FDE特征前利用优化K值的变分模态分 解(variational mode decomposition,VMD)对信号进行分解,以获取预设数量的固有模态 分量(IMF),然后 提取各尺度子序列的FDE作为分类特征,并采用遗传算法(genetic algorithm, GA)对支持向量机(SVM)的惩罚 因子c和 核函数参数g进行寻优,最后通过GA-SVM模型进行模式识别。计算结 果表明, 所提方法对4类心律识 别的平均准确率达到95.3%,灵敏度达到95.4% ,特 异性达到 98.4%,相比自适应多尺度散布熵(adaptive multiscale dispersion entropy, AMDE)等其他方法优势明显,可以实现对心 电(electrocardiogram, ECG)信号的准确分类。     

Comparison of GENIE and conventional supervised classifiers formultispectral image feature extraction

The authors have developed an automated feature detection/classification system, called GENetic Imagery Exploitation (GENIE), which has been designed to generate image processing pipelines for a variety of feature detection/classification tasks. GENIE is a hybrid evolutionary algorithm that addresses the general problem of finding features of interest in multispectral remotely-sensed images. The authors describe their system in detail together with experiments involving comparisons of GENIE with several conventional supervised classification techniques, for a number of classification tasks using multispectral remotely sensed imagery     

Time-frequency and time-scale techniques for the classification ofnative and bioprosthetic heart valve sounds

The determination of diagnostic features in recorded heart sounds was investigated with Carpentier-Edwards (CE) bioprosthetic valves. Morphological features, extracted using the Choi-Williams distribution, achieved between 96 and 61% correct classification. The time-scale wavelet-transform feature set achieved 100% correct classification with native valve populations, and 87% with the CE replacements     

PCA特征提取和弹性BP神经网络的水下气泡识别

针对水下激光雷达探测得到的尾流回波信号由于非稳态造成特征提取困难、不易识别的问题,提出了基于PCA特征提取与弹性BP神经网络结合的水下气泡识别算法。首先对连续采集的回波信号进行切片预处理,然后采用PCA算法对拼接的高维样本进行主要特征提取,确定特征值个数,其次对弹性BP神经网络进行参数的选择,确定能实现最优分类的隐含层节点数、特征个数等,最后根据室内搭建的尾流探测模拟平台,实现对气泡群和干扰目标的识别。实验结果表明:在隐含节点为12,增量因子为1.15,减量因子为0.55时,选取两个特征值能对有气泡、无气泡及干扰物进行有效分类;识别率随着气泡群密度的增大提升13.4%,在低密度下的识别率随激光能量的增加平均提升6.3%,识别率随距离的增加先增大后减小,气泡群在2.2 m时的目标峰特征明显,平均识别率提升3.5%。通过与自适应附加动量BP对比,该方法在减少识别时间的同时准确率达到99.1%,证明该算法可有效运用于激光雷达舰船尾流气泡的识别。     

基于熵的前馈网络修剪算法

剪枝算法是一种通过简化网络结构来避免过拟合的有效方法之一。文章依据Shannon熵原理定义了神经网络隐层节点输出的拟熵,该熵与Shannon熵对不确定性的描述具有相同的效果,但克服了Shannon熵中无定义和零值的缺点。将交叉熵和隐节点输出拟熵作为目标函数,并采用熵周期的策略对网络参数进行寻优,通过删除合并隐层神经元达到简化网络结构的目的。仿真结果表明,此方法简单易行,对BP网络的泛化性能有较好的改善。     

基于深度特征学习的网络流量异常检测方法

针对网络流量异常检测过程中提取的流量特征准确性低、鲁棒性差导致流量攻击检测率低、误报率高等问题,该文结合堆叠降噪自编码器(SDA)和softmax,提出一种基于深度特征学习的网络流量异常检测方法。首先基于粒子群优化算法设计SDA结构两阶段寻优算法:根据流量检测准确率依次对隐藏层层数及每层节点数进行寻优,确定搜索空间中的最优SDA结构,从而提高SDA提取特征的准确性。然后采用小批量梯度下降算法对优化的SDA进行训练,通过最小化含噪数据重构向量与原始输入向量间的差异,提取具有较强鲁棒性的流量特征。最后基于提取的流量特征对softmax进行训练构建异常检测分类器,从而实现对流量攻击的高性能检测。实验结果表明:该文所提方法可根据实验数据及其分类任务动态调整SDA结构,提取的流量特征具有更高的准确性和鲁棒性,流量攻击检测率高、误报率低。     

基于弱监督学习的去噪受限玻尔兹曼机特征提取算法

针对现有特征提取方法难以实现从含有复杂背景的图像中提取有用目标特征的瓶颈问题,提出了基于弱监督学习的去噪受限玻尔兹曼机特征提取算法.首先,利用训练样本,通过无监督学习方式训练一个标准受限玻尔兹曼机模型,从而获得一个包含可视单元层和隐藏单元层的层次结构模型;然后,对可视层的每个单元引入二值转换单元,对隐藏层,根据各节点的激活值大小和激活频率将其分为两组:前景特征隐层单元和背景特征隐层单元,得到一个二元混合式去噪玻尔兹曼机的模型;最后,通过多模交互方式,利用有限数量的样本标签信息对输入样本逐像素地进行采样训练,以此来提取目标特征.实验表明,本文的特征提取算法能够有效地从复杂的干扰背景中提取目标特征,提高了目标识别精度.     

Application of non-covalent functionalized carbon nanotubes for the counter electrode of dye-sensitized solar cells

Multi-walled carbon nanotubes (MWCNTs) were functionalized noncovalently by lysozyme (LZ), cetyl pyridinium chloride (CPC), deoxycholate sodium (NaDC) and polyethylene glycol octylphenol ether (Triton X-100), respectively in this study. Four different kinds of functionalized MWCNTs were employed into dye-sensitized solar cell (DSSC) as the Pt-free counter electrode (CE). The correlation between the dispersion of MWCNTs and electrochemical active area of CE and the photovoltaic characteristic of DSSC were investigated. Among these four DSSCs, the one with Triton X-100 functionalized MWCNTs showed the best energy conversion efficiency of 2.69% which is 11.16% higher than the DSSC using pristine MWCNTs CE (2.42%), yet the efficiency is lower than the DSSC using Pt CE. While the DSSC with CPC, NaDC and LZ functionalized MWCNTs as the CE showed inferior the photovoltaic conversion efficiency than the DSSC using pristine MWCNTs CE. On analysis of the photovoltaic performance of DSSC and the dispersion of MWCNTs and electrochemical active area of CE, it is found that the high efficiency of the DSSC is associated with the good dispersion of MWCNTs and large electrochemical active area of the counter electrode material.     

基于多任务NR-DenseNet网络的航班延误预测模型

不同于目前大多数只倾向于研究单一的分类或回归任务的航班延误预测方法,该文提出一种基于多任务NR-DenseNet网络的航班延误预测模型,旨在同时实现航班延误等级分类预测与延误时间回归预测。首先,预处理相关数据;其次,建立多任务学习特征提取共享层,使用NR-DenseNet网络提取任务之间的共享参数,深度挖掘任务之间的相关特征;然后,建立多任务学习特定任务层,通过回归器与分类器分别输出特定任务的预测结果;最后,采用损失加权方法对两个任务损失函数进行优化,平衡任务间的收敛速度,提高模型泛化性。将模型应用在宁波机场数据集中,与单任务模型相比回归任务平均MSE降低了23.4%,平均MAE降低了14.2%,分类平均准确率提升了2.7%。实验结果表明,该文方法提升了分类任务的准确率降低了回归任务的误差,可以有效提升模型性能。     

基于三维特征的雷达信号脉内调制识别

针对低信噪比条件下雷达信号识别算法对噪声敏感的问题,提出了一种基于三维特征的雷达信号脉内调制识别算法。该方法通过提取信号的差分近似熵、调和平均分形盒维数和信息维数特征组成三维特征向量,使用遗传算法优化的BP神经网络分类器实现雷达信号的分类识别。仿真结果表明,所提取的三维特征在信噪比为-4~10 dB变化范围内具有较好的类内聚集度和类间分离度,可以实现对不同雷达信号进行识别,证实了该方法的有效性。     

一种创新性激光图像三维目标识别算法

针对径向基神经网络在激光图像分类识别中识别率低及训练时间长的问题,提出粗糙集与神经网络相结合的方法,将粗糙集算法简约后的样本特征作为神经网络的前置输入。首先建立不同视点的激光主动成像三维仿真图像,然后提取17个目标特征,并采用粗糙集算法选择分类的属性,从17个特征中筛选出5个影响决策的特征属性,最后选用4层径向基神经网络作为基本的网络结构,并采用在各层节点上与粗糙集相结合方法识别目标。仿真结果表明,结合粗糙集的集成神经网络方法识别正确率保持在80%以上,与未结合粗糙集的神经网络相当,但训练与识别时间缩短10倍以上。     

Wireless distributed computing in cognitive radio networks

Individual cognitive radio nodes in an ad-hoc cognitive radio network (CRN) have to perform complex data processing operations for several purposes, such as situational awareness and cognitive engine (CE) decision making. In an implementation point of view, each cognitive radio (CR) may not have the computational and power resources to perform these tasks by itself. In this paper, wireless distributed computing (WDC) is presented as a technology that enables multiple resource-constrained nodes to collaborate in computing complex tasks in a distributed manner. This approach has several benefits over the traditional approach of local computing, such as reduced energy and power consumption, reduced burden on the resources of individual nodes, and improved robustness. However, the benefits are negated by the communication overhead involved in WDC. This paper demonstrates the application of WDC to CRNs with the help of an example CE processing task. In addition, the paper analyzes the impact of the wireless environment on WDC scalability in homogeneous and heterogeneous environments. The paper also proposes a workload allocation scheme that utilizes a combination of stochastic optimization and decision-tree search approaches. The results show limitations in the scalability of WDC networks, mainly due to the communication overhead involved in sharing raw data pertaining to delegated computational tasks.