【作者】 肖汉光；

【导师】 蔡从中； 廖克俊；

【作者基本信息】 重庆大学 ， 凝聚态物理， 2006， 硕士

【摘要】 统计学习理论(Statistical Learning Theory,简称SLT)是由Vapnik等人提出的可应用于小样本分析的统计理论。支持向量机(Support Vector Machine,简称SVM)是基于统计学习理论的一种新的机器学习方法。它遵循结构风险最小化原则,克服了基于经验风险最小化学习方法(如贝叶斯分类器、决策树和人工神经网络等)泛化能力较差的缺点,并已在很多研究领域得到了广泛的应用。本文利用支持向量机分别根据车辆的轮廓特征、车辆运行时所产生的声波和地表震动信号特征进行车辆识别,并分析和比较了不同的特征提取和选择方法对分类准确率的影响,同时比较了支持向量机和其它分类器的分类能力。本文首次提出并应用支持向量机进行了地震预测研究。以下是本文的主要内容:(1)对目前使用的特征提取和选择方法进行了综述,介绍了遗传算法(GA)、主成分分析法(PCA)、独立成分分析法(ICA)、独立主元分析法、粒子群寻优算法(PSO)、模拟退火算法(SA)和其他常用的特征提取和选择方法的算法、及其优点和缺点。(2)介绍了几种常用的分类方法的分类原理。包括贝叶斯分类器、决策树和人工神经网络。对支持向量机的分类原理进行了详细的叙述,并介绍其算法、实现、发展和在相关领域中的应用。(3)分别利用车辆轮廓特征、声音信号和地表震动信号,结合支持向量机分类原理,对车辆类型进行了分类研究。分析了不同特征提取和特征选择方法对分类准确率的影响。同时对不同分类器的分类性能进行了比较。结果表明:支持向量机的分类准确率明显高于其他分类器;PCA降维能力优于GA,且耗时远小于GA。但是,在同种分类器下,利用PCA进行特征选择得到的测试集和独立集的分类准确率低于GA。将PCA和GA分别选出的特征向量合并后进行分类所得到的准确率均高于单独利用其中任何一种特征选择方法的(PCA或GA)。(4)针对车辆声音信号和地表震动信号的特征,提出了一种基于能谱密度的特征选择方法。该方法能重构声音和地表震动信号的特征向量。研究结果表明,采用该方法在不降低分类效果的同时,还能有效地减少特征向量的维数和提高分类准确率。(5)基于支持向量机,根据地震发生前后地表震动信号对发生在美国加利弗里亚中部的两次大地震进行了地震预测研究,分析了地震动信号的地域性,研究了地震动信号和地震发生的时间、地点和震级之间的关系。研究结果表明:震前SVM更多还原

【Abstract】 Statistical Learning Theory (SLT) proposed by Vapnik et al is a statistics theory for the analysis of a small-sample database. Based on SLT, Support Vector Machine (SVM) was proposed as a new machine learning method for classification or regression. Based on Structural Risk Minimization Rule (SMR), SVM usually achieves better generalization performance than the Empirical Risk Minimization (ERM)-based learning methods, such as Bayesian Classifier (BC), Decision Tree (DT), and Artificial Neural Network (ANN). Up to now, SVM has been widely and successfully employed in various fields.In this study, the features of vehicle profile, acoustic and seismic signals were used to recognize the types of vehicles by using SVM approach. The effect of different feature extraction and selection methods on the classification accuracy was analyzed and discussed. We also compared the generalization performance of SVM with those of other classifiers.For the first time, we proposed and applied SVM to predict earthquake.The outline of this paper is showed as below:(1) The current methods of feature selection and extraction were reviewed. The advantages and disadvantages of several algorithms were introduced, such as, Genetic Algorithm (GA), Principal Component Analysis (PCA), Independent Component Analysis (ICA), Particle Swarm Optimization (PSO), simulation annealing (SA).(2) The classification principles of popular classifiers were reviewed briefly, such as BC, DT, and ANN. We described SVM detailedly in its principle, algorithm, implementation, development and application.(3) We employed SVM and the features of vehicle profile, acoustic and seismic signals to recognize the types of vehicles, and analyzed the effect of different feature selection and extraction methods on the classification accuracy. We also compared the classification performance of different classifiers for vehicle recognition. The experimental results demonstrate that the accuracy of SVM is superior to those of other classifiers, and also reveal that PCA is more effective and faster than GA for implemention of feature dimension reduction. Under using a same classifier, the accuracies for either the training or test dataset by using PCA are higher than those of by using GA. It was found that the higher accuracy can be obtained via using the更多还原