An Analgorithm Research On Multi-Class Classification Based On Support Vector Machine And Its Application In Fault Identification Of Rolling Bearing

In 1990's,a new algorithm of machine learning named support vector machine was brought forward based on statistical learning theory. It attracted much attention for its theory and generalization ability, and solved some problems in the field of machine learning. However, support vector machine can only be used for classification between two classes. It is very important to use support vector machine for multi-classification for there are always multiple classes needed to be classified in projects. Many scholars at home and abroad had assumed large amount of studying in this field and designed various multiclass classification algorithm based on supporting vector machine. Based on this, they achieved the purpose of multiclass classification based on support vector machine. However, these algorithms have some defects respectively and much problem has to be solved further. Generally, the algorithm research on multiclass classification based on support vector machine is in a stage probing unceasingly. All research jobs focus on the following problems mainly. First, which common ground do they have on algorithm structure among the available multiclass classifiers based on support vector machine? Can we divide them into some categories? Second, how to design the algorithm structure of the multiclass classifier for improving its generalization ability? What effect does the kernel function has to the space distribution of the samples of multiple classes? Can we estimate the space distribution of the samples of multiple classes directly in the feature space? Can we design new algorithm of multiclass classification based on support vector machine and let it has good generalization ability?For all the above problems, the following research jobs are done. Firstly, the analysis job of several algorithms such as one against rest, one against one, multiclass classifier based on binary tree and DAGSVM is done in this paper. They are divided into two classes according to their structures. Secondly, the experiment of rolling bearings fault diagnosis is done and the five kinds of rolling bearing vibration signals are recorded. Finally, the five sample sets are gained by feature extraction using wavelet packet. Thirdly, some problems related to classifiability between two classes are researched. The classifiability between two classes is measured in high dimensional feature space by using kernel function. Fourthly, what effects different kernel functions and different parameters have to the classifiability between two classes in high dimensional feature space is analyzed and compared. Based on this, a conclusion is made. Fifthly, a new method of designing the binary tree is put forward in this paper and they are proved good in the experiment of rolling bearing fault identificaiton. Sixthly, two new algorithms of multiclass classification based on support vector machine are suggested and they performance well in the rolling bearing fault identification.The method of designing the binary tree of the multiclass classifier based on support vector machine , new algorithms of multiclass classification based on support vector machine, comparation of classifiability between any two classes and its connection with the kernel function are discussed mainly in this paper. The five working conditions of the rolling bearing are recognized by appling the new methods to the algorithms of multiclass classification based on support vector machine. Generally, the algorithm research of multiclass classification based on support vector machine focuses on designing better classifier using simple methods, and it is still to be solved for many problems of the research facing us.The project is the study content of the Shanxi Province Natural Science Foundation 'Statistical learning theory and support vector machine and itsapplication in fault diagnosis of machine'.Foundation program number:2006011056...