Fault Recognition with Labeled Multi-category Support Vector Machine

Abstract

Support vector machine is intrinsically a binary classifier providing no theoretically formulated procedure for multi-category classification. Several methods have been developed to extend it to multi-category problems. Combining strengths of them, an improved "labeled multi-category support vector machine" is proposed. The proposed method explicitly labels samples and performs multi-category classification with only a single support vector machine classifier. Labeling samples leads to the sample number disparity between positive and negative classes. The techniques of setting different cost parameters for different classes are employed to enhance the algorithm's performance. Generalization error bound estimates are theoretically derived by the new technique of maximal discrepancy. Experiments with a benchmark dataset show that the algorithm can accurately classify multi-category data. Rotor mechanical fault recognition applications confirm that the algorithm can efficiently perform multi-category fault detection and identification.