Research On Support Vector Techniques And Their Applications

Support vector techniques can be used for solving the classification (supervised learning) and clustering (unsupervised learning) problems of machine learning. The methods are support vector machine (SVM) and support vector clustering (SVC) respectively. SVM is developed in the framework of statistical learning theory (SLT), which bases on the principle of structural risk minimization (SRM) and the VC theory. It always does well in practical applications because of its remarkable trade-off between the complexity of machines and learning ability. SVM can overcome the problems of small samples, nonlinear, over-fitting, curse of dimensionality and so on. SVC carries on and extends the idea of SVM. It use support vector technique for cluster labeling with no supervision information. It can conduct arbitrary shape datasets, and need not designate the number of clusters. Now, SVM and SVC are successfully applied for many fields such as marine diesel engine fault diagnosis and scene classification. This dissertation focuses on the research of learning methods based on support vector. The contents in this dissertation are described as follows:1.The decision tree SVM method can produce error accumulation because optimal decision tree is difficult to determine. For solving this provlem, the multi-classification method based on GA-optimized SVM decision tree is studied. The maximum separation margin between normal hyper planes of nonlinear SVM in high dimensional feature space is taken as optimization objective. Its equivalent objective function is defined as the fitness function. Roulette wheel selection operator, partially mapped crossover operator and inversion mutation operator are used. The real-valued coding genetic algorithm is used to optimize the decision tree structure. This method can decrease error accumulation because the classes with big margin can be earlier separated.2.A new classification method based on the combination of improved kernel principal component analysis (KPCA) and linear support vector machine (SVM) is proposed. In view of some principal components of KPCA are probably not suitable for classification, the principal components containing more category information are retained in the proposed method by evaluating the inter-class separability. Then soft margin linear SVM is used for classification. This method is used for scene classification, and a new method based on local Gabor features is proposed. In order to extract image features, a new local Gabor feature descriptor is proposed. Then the local Gabor feature descriptors are embedded into a bag-of-visual-words (BOVW) model. In a spatial pyramid matching framework, pyramid histograms of visual words are extracted for representing scene images. Last the classification method based on improved KPCA and linear SVM is used for classification. Numerical experiments are conducted on three commonly used scene datasets. The experimental results demonstrate the effectiveness of the method.3.A new SVC algorithm based on multi-leaf spanning tree is proposed. First union-find sets are used to construct the multi-leaf spanning tree. Compared with minimal spanning tree, multi-leaf spanning tree has more leaf nodes and more simple trunks. Then the connection check is conducted for the trunks of multi-leaf spanning tree, and an adjacency matrix is obtained. Last the depth first search algorithm is used for cluster labeling based on the adjacency matrix. In view of the superiority of nonlinear mapping, the new SVC algorithm is also expended to high dimensional feature space. Euclidean distances of sample points in high dimensional feature space are used as the weights of edges, and multi-leaf spanning tree are constructed in Hilbert space. Then cluster labeling is conducted, and good clustering results are obtained. This method is used for marine diesel engine fault diagnosis. Fault diagnosis experiments are conducted on simulation data of Kongsberg MAN B&W 5L90MC VLCC marine engine room simulator.