An improved binary tree SVM and application for fault diagnosis

This paper presents an improved binary tree SVM hierarchy construction method, according to the strong influence for SVM classifier performance by binary tree hierarchy. A separability measure with weights was constructed by the average distance of samples in one class and the average distance of samples between different classes, so the class which has bigger distance from other classes and wider sample distribution within itself was first separated. The selection criteria and algorithm steps of improved binary tree SVM were proposed. Compared with recognition accuracy of standard data sets for different multi-class algorithm, the superiority of improved binary tree SVM is verified. Taken common faults of reciprocating compressor transmission mechanism as research objects, feature vectors of faults were extracted based on multifractal and singularity value decomposition, and the faults were diagnosed by the improved binary tree SVM accurately.