Parametric-margin projection twin support matrix machine and its application in fault diagnosis of roller bearing

Abstract

Support matrix machine (SMM) is an innovative classifier that accepts matrix as inputs to make full use of the structure information between matrices. However, SMM aims at constructing two parallel hyperplanes to segment different types of samples, which makes the model inflexible and sensitive to the impact of noise, resulting in poor performance on complex data classification. Given this consideration, a novel parametric-margin projection twin support matrix machine (PPTSMM) is proposed in this paper. PPTSMM introduces the additional regularization term, which is the parametric-margin between the projected centers of the two classes, and the separability of projected classes based on the parametric-margin rather than unit distance. Meanwhile, the slack vector is employed for reformulating the within-class least square loss in PPTSMM, which address the time-consuming matrix inverse operation to reduce the complexity of calculations. Two kinds of roller bearing fault signals are used to analysis the performance of PPTSMM, and the analysis results indicate that PPTSMM is effective to reduce the impact of noise and computation time in fault diagnosis of roller bearing.