Abstract:

In divide-and-combine approach, multi-class support vector machines (SVMs) are divided into several binary SVMs and then the SVMs are combined to obtain multi-class classifiers. In this paper, a new probability voting strategy is presented to combine several binary SVMs. By the method, not only the unclassifiable regions existing in conventional strategy are solved, but also the decision results satisfy the probability distribution. Firstly, two most commonly combining strategy: MaxWins and FSVM are discussed, and their performances are compared through posterior probability distribution. Secondly, an estimate function for the prior probability in a binary classification problem is defined, and an adjustment function satisfying prior probability is normalized in the range of 0~1. Thirdly, a novel probability voting is improved by considering the conventional voting and the adjustment function. Finally, 5-class SVMs-based fault diagnosis models for gearbox respectively with MaxWins majority voting, FSVM and the presented strategy are tested. All the tests and data indicate that the multi-class SVMs combined by probability voting strategy has more capacity of reliability and robustness, and are suitable for fault diagnosis of gearbox.