Projection and support vector machine-based multiclass classification method

Abstract

The invention discloses a projection and support vector machine-based multiclass classification method. The method comprises the following steps of: voting to-be-classified samples on the basis of one-to-one support vector machine classifiers, and determining the type of the to-be-classified sample when the highest vote number type is unique; when and only when the highest vote type number is equal to a vote participation type number, calculating projections from vectors between the to-be-classified sample and the center of a support vector machine to vectors between the to-be-classified sample and the center of the type, eliminating a highest vote number type, the projection of which is maximum, and reducing a data recognition area; tracing among the residual highest vote number types by adoption of a one-to-one classifier voting and discriminating mechanism, and recognizing the type of the to-be-classified sample until a decision blind area disappears; and when the highest vote type number is smaller than the vote participation type number, continuing to trace among the residual highest vote number types by adoption of the one-to-one classifier voting and discriminating mechanism. According to the method disclosed by the invention, the sample classification correctness is improved, the decision time of the classifiers is shortened and the decision blind area can be completely eliminated.

Description

technical field [0001] The invention relates to a support vector machine (SVM) one-to-one multi-class classification method, in particular to a multi-class classification method based on projection and support vector machine. Background technique [0002] Support vector machine was originally designed to solve the problem of binary classification. For the classification of multiple patterns in practical applications, it mainly uses one-to-many and one-to-one multi-class classification methods. The one-to-many classification method is to train m sub-classifiers on a data set of m categories, and each sub-classifier constructs an SVM classifier with the i-th category and other categories, so that the sample points to be divided are placed on each classifier The size of the value determines the category of the sample; the one-to-one classification method is to establish m(m-1) / 2 sub-classifiers between the m categories, and determine the number of sub-classifiers to be classifi...

AI Technical Summary

Problems solved by technology

Its method to solve the decision-making blind spot is based on the method of representing the similarity between samples based on the Euclidean distance between classes. It is a weak classifier-type decision-making scheme, which affects the diagnostic accuracy; The amount of calculation is difficult to meet the real-time problem of fault identification in engineering applications

Images