基于GST和多级SVMs的PQDs分类方法研究

提出一种新方法--基于广义s变换(GST)和多级支持向量机(SVMs)分类器的电能质量扰动(PQDs)分类方法．首先,引入GST提取典型PQDs特征的向量集．然后,针对电压暂降、暂升、瞬态、振荡暂态、陷波‘、尖峰、谐波、间谐波及闪变九种典型扰动构造特征集合,用于训练多级SVMs分类器．最后,使用SVMs分类器对500个PQDs测试样本进行识别,在SVMs中电能质量的N种扰动是由～一1分类器分类的．结果袁明：该方法可以有效地对PQDs检测和分类尤其分类器在训练速度和正确率方面性能突出．     

A new PQ disturbances identification method based on combining neural network with least square weighted fusion algorithm

Power quality (PQ) has been an important con-cern for utility, facility and consulting engineers in re-cent years. With the development of modern industrytechnology, especially thatof electronic and informationtechnology, the end-use equipment sensitive to PQdis-turbances are widely used in many areas. They requirenot only the conventional static PQ disturbances, suchas harmonics, voltage fluctuations, voltage deviation,frequency deviation and unbalance, but also the dy-namic PQ disturbance…     

异常检测中支持向量机最优模型选择方法

为了构建一个具有良好的学习性能和推广能力的异常检测分类器,在结构风险最小(SRM)原则下讨论了基于支持向量机(SVM)的异常检测分类器的设计准则,提出了SVM分类器模型及其参数快速选择和评估方法,并给出了异常检测分类器训练步骤.针对KDD'99网络入侵检测数据集,实验结果表明,该方法能够有效地缩短入侵检测分类模型建立时...     

一种可减少训练时间的分层并行支持向量机方法

基于支持向量的本质和并行计算方法，提出了一种新的分层并行的机器学习方法以加速支持向量机的训练过程．该方法首先按照分而治之的思想将原分类问题分成若干子问题．然后将支持向量机的训练过程分解成级联的两个层次，在每层采用并行的方法训练各个子支持向量机．各层训练集中的非支持向量被逐步筛选掉，交叉合并的规则保证问题的一致性．仿真结果表明该方法在保证分类器推广能力的同时，缩短了训练支持向量机的时间．     

基于支持向量机和遗传算法的纹理识别

为了解决尺度变化和训练样本有限给纹理识别带来的困难,提出了一种基于支持向量机和遗传算法的纹理识别新方法。该方法用小波变换各子带图像共生矩阵参数、分析窗口大小、像素均值和像素标准差等参数作为纹理特征,用多类支持向量机作为分类器。用遗传算法对纹理特征集进行了优化;用输出纠错码将二类支持向量机扩展到多类,提高了分类器的泛化能力。用包含有25类单色自然纹理的图像库进行识别试验,结果表明,该方法的识别错误率小于10%,得到了比传统的贝叶斯等方法更高的识别率和更好的推广性。     

基于支持向量机的变异语音分类研究

变异语音的训练样本有限，采用传统的分类方法进行分类，效果不够理想，而支持向量机方法在有限样本情况下可以保持很好的分类推广能力．采用支持向量机方法进行变异语音分类，提取基频和TE0基频作为变异语音分类的特征，讨论了样本预处理和参数选择等问题．提出了直接截取和DTW规正两种方法来解决语音样本特征向量长度不一致的问题．基于TE0基频特征，采用指数径向基函数(ERBF)内核，对应力(G—force)影响下的变异语音进行分类，分类正确率可达到99．2％，比传统的贝叶斯分类器和HMM分类器，分类性能分别平均提高了12．6％和6．0％．实验结果表明，采用支持向量机方法进行变异语音分类是可行的．     

基于支持向量机的变异语音分类研究

变异语音的训练样本有限,采用传统的分类方法进行分类,效果不够理想,而支持向量机方法在有限样本情况下可以保持很好的分类推广能力.采用支持向量机方法进行变异语音分类,提取基频和 TEO 基频作为变异语音分类的特征,讨论了样本预处理和参数选择等问题,提出了直接截取和 DTW 规正两种方法来解决语音样本特征向量长度不一致的问题.基于 TEO 基频特征,采用指数径向基函数(ERBF)内核,对应力(G-force)影响下的变异语音进行分类,分类正确率可达到99.2%,比传统的贝叶斯分类器和 HMM 分类器,分类性能分别平均提高了12.6%和6.0%.实验结果表明,采用支持向量机方法进行变异语音分类是可行的.     

基于SVMs的大幅遥感影像目标快速检测方法

针对高分辨率遥感影像数据幅度大、内容复杂的现状，提出了基于分层分类器的检测方法。从简化分类器的结构和缩减检测区域探讨了提高系统效率的途径。实验表明，这种新的搜索策略提高了大幅遥感影像目标检测的效率。     

Machine learning-based classification of rock discontinuity trace: SMOTE oversampling integrated with GBT ensemble learning

This paper presents a hybrid ensemble classifier combined synthetic minority oversampling technique (SMOTE), random search (RS) hyper-parameters optimization algorithm and gradient boosting tree (GBT) to achieve efficient and accurate rock trace identification. A thirteen-dimensional database consisting of basic, vector, and discontinuity features is established from image samples. All data points are classified as either “trace” or “non-trace” to divide the ultimate results into candidate trace samples. It is found that the SMOTE technology can effectively improve classification performance by recommending an optimized imbalance ratio of 1:5 to 1:4. Then, sixteen classifiers generated from four basic machine learning (ML) models are applied for performance comparison. The results reveal that the proposed RS-SMOTE-GBT classifier outperforms the other fifteen hybrid ML algorithms for both trace and non-trace classifications. Finally, discussions on feature importance, generalization ability and classification error are conducted for the proposed classifier. The experimental results indicate that more critical features affecting the trace classification are primarily from the discontinuity features. Besides, cleaning up the sedimentary pumice and reducing the area of fractured rock contribute to improving the overall classification performance. The proposed method provides a new alternative approach for the identification of 3D rock trace.     

Identification of activity stop locations in GPS trajectories by density-based clustering method combined with support vector machines

The identification of activity locations in continuous GPS trajectories is an essential preliminary step in obtaining person trip data and for activity-based transportation demand forecasting. In this research, a two-step methodology for identifying activity stop locations is proposed. In the first step, an improved density-based spatial clustering of applications with noise (DBSCAN) algorithm identifies stop points and moving points; then in the second step, the support vector machines (SVMs) method distinguishes activity stops from non-activity stops among the identified stop points. A time sequence constraint and a direction change constraint are applied as improvements to DBSCAN (yielding an improved algorithm known as C-DBSCAN). Then three major features are extracted for use in the SVMs method: stop duration, mean distance to the centroid of a cluster of points at a stop location, and the shorter of distances from current location to home and to the workplace. The proposed methodology was tested using GPS data collected from mobile phones in the Nagoya area of Japan. The C-DBSCAN algorithm achieves an accuracy of 90 % in identifying stop points in the first step, while the SVMs method is 96 % accurate in distinguishing the locations of activity stops from non-activity stops in the second step. Compared to other variants of DBSCAN used to identify activity locations from GPS trajectories, this two-step method is generally superior.

     

基于两层策略的掌纹识别新方法

提出了一种基于两层策略的掌纹识别方法。首先提出了新的定位分割方法,提取出手掌最大中心区域以及粗匹配阶段的匹配特征。然后为精确匹配提出一种有效的掌纹线特征的提取和表示方法。设计由粗到细的掌纹识别分类器,应用粗匹配算法得到符合要求的掌纹数据子集,再利用主线特征进行精确匹配。针对250幅图片做实验,精确度高达94.67%,证明了本文算法具有良好的鲁棒性和有效性。     

Research on Chinese place name recognition based on kernel classifier

SVM[1]is a new approach of pattern recognitionbased on Structural Risk Minimization which have thefollowing striking property:bode good for generalizationperformance,find a global optimal solution,complexityof the solution depends on the intricacy of the …     

基于固有时间尺度分解的信号细微特征识别

为解决无线复杂环境下同型号通信电台发送的信号识别问题,针对传统时频方法处理杂散细微特征存在不足,提出了一种基于固有时间尺度分解（intrinsic time-scale decomposition,ITD）的信号细微特征识别方法。首先通过ITD方法将稳态状态下信号分解,计算分量瞬时参数并得到信号的时频谱,然后提取频谱特征作为细微特征,最后使用SVM分类器进行模式训练以实现信号的识别。仿真结果表明,该算法能够解决传统方法的实时性和准确性差等问题,取得较好的识别效果。     

短时电能质量扰动变尺度标杆相似度识别

利用S变换模矩阵相似度识别持续时间不同的电能质量扰动信号,需要建立维数不同的标准模板.应用双线性插值法对S变换模时频矩阵进行尺度变换,对不同持续时间的同类扰动可建立统一的标杆模板,减少了标准模板数.根据最大相似度原理,通过简单的四则运算对扰动进行分类,无需添加任何分类器,分类过程简洁有效.仿真证明,该方法对噪声不敏感,能较好地解决不同持续时间的电能质量扰动分类问题.     

单相转换为三相的最佳移相措施的构想

单相转换为三相的移相电路基本上是采用RC移相电路由于电路设计和处理上的不合理,致使其在实际应用方面受到很大限制。本文结合实例说明,采用RC移相电路实行单、三转的移相措施是不可取的;同时,提出采用RLC移相电路的最佳移相措施的构想。     

基于KSVM决策树法的人脸检测与定位

传统的SVM直接在线性条件下训练SVM分类器完成人脸与非人脸的分类,分类器训练困难,计算量大且速度慢.为构造一个复杂背景下人脸检测与定位的新方法,本文用核函数把SVM推广到非线性SVM,再与二叉树相结合,可以解决多类识别问题,此即KSVM决策树人脸检测方法.在此基础上,人脸面部特征被进一步确认.本文提出了改进的四边界Prewitt边缘算子提取人眼,进而定位五官.实验结果表明该方法检测率较高,虚警率较低,定位准确.     

基于支持向量机的MSPC方法及其应用

针对传统多变量统计过程控制（MSPC）方法在故障检测、故障原因分析和故障识别中的难点，提出了多元特征提取方法与基于支持向量机（SVM）的一类分类器设计、特征选择以及多类分类器设计方法相结合的一种完整的MSPC新方法.该方法在故障检测中可去除特征满足特定分布的假设前提，并可确定多个统计量的控制限；在故障原因分析中综合考虑故障对于各个变量大小的影响以及变量变化对于故障分类的重要性，提高了关键变量选择的准确性；并且故障识别是基于SVM对故障特征分类的优良特性，避免了传统判别法中经验准则的引入.上述方法在标准仿真平台Tennessee Eastman过程上结合主元分析（PCA）进行了应用，结果显示了其优越性.     

Mandarin Digits Speech Recognition Using Support Vector Machines

A method of applying support vector machine (SVM) in speech recognition was proposed, and a speech recognition system for mandarin digits was built up by SVMs. In the system, vectors were linearly extracted from speech feature sequence to make up time-aligned input patterns for SVM, and the decisions of several 2-class SVM classifiers were employed for constructing an N-class classifier. Four kinds of SVM kernel functions were compared in the experiments of speaker-independent speech recognition of mandarin digits. And the kernel of radial basis function has the highest accurate rate of 99.33 %, which is better than that of the baseline system based on hidden Markov models (HMM) (97.08%). And the experiments also show that SVM can outperform HMM especially when the samples for learning were very limited.     

不完整数据集的MFR辐射源识别方法研究

该文提出一种基于随机森林的不完整数据集的多功能雷达（MFR）辐射源识别方法,该方法在MFR辐射源波形单元识别框架基础上,首先对参数缺失的先验知识集进行多重划分,得到多个不含缺失参数的样本子集,然后删减冗余子集并利用随机森林算法对各个子集构建弱分类器,最后根据弱分类器对识别结果贡献率的不同,进行权值设定,得到最终的识别模型。仿真实验证实了提出的MDRF-WA方法能够提高少量先验知识条件下波形单元识别的准确率和鲁棒性,降低计算成本。     

基于小波和进化网络的电能质量动态扰动自动识别

提出了一种基于小波和进化网络的电能质量动态扰动自动识别方法。首先应用Ｄａｕｂｅｃｈｉｅｓ ４小波对扰动波形进行分解,得到各个尺度上的小波系数,然后进行特征提取:采用遗传算法设计的模式特征分类器对扰动进行分类。实验表明该方法具有分类准确、网络学习速度快及收敛效果好等显著特点。