PSVM Algorithm Research Based On Supersphere

Support Vector Machine (SVM) is a core technology of machine learning for classification.However, there are many classification problems which are so hard to be solved effectively in practical engineering because data collection is not timely, the sample data is incomplete, and does not cover all sample in data space, which showed low density value. This also leads to the decrease of generalization performance and robustness performance of classifier. Currently,Knowledge based Support Vector Machine can solve this problem, but complete and effective knowledge is very difficult to acquire, and not enough to describe all characteristics of the data,which makes the final classification capacity less noticeable and may even lead to negative effects.Therefore, priori knowledge from the training data was put forward so as to improve the classifier performance.Beacause any sample space can be covered by a linear combination of a limited open supersphere cover set. So, a supersphere for finitely covering the sample space can describe the distribution of sample data, and important information can be extracted for classification. Based on such understandings, this thesis described the distribution of sample data through three methods as a supersphere for finitely covering the sample space, based on which prior knowledge is constructed for improving classifier performance. Then based on three kinds of supersphere approaches new support vector machine algorithms were put forward: O-PSSVM (Proximal Ordinary-Supersphere-based Support Vector Machine), SVDD-PSSVM (Proximal SVDD-Supersphere-based Support Vector Machine),M-PSSVM (Proximal Mahalanobis-Supersphere-based Support Vector Machine).The experiments showed O-PSSVM, SVDD-PSSVM and M-PSSVM indeed could improve the generalization and robustness performance of PSVM, becaused of the introduce of ordinary supersphere, SVDD supersphere and Mahalanobis supersphere. Meanwhile, it was found that the classification performance of SVDD-PSSVM and PSVM were better than the SVDD classifier while the classification performance of M-PSSVM and PSVM were better than the Mahalanobis supersphere classifier. These indicated that SVDD and Mahalanobis supersphere are not suitable to train a classifer directly.At last, SVDD-PSSVM was applied to single-label and multi-label fault diagnosis of train air conditioning unit for better diagnosis accuracy.