Research on SVMs of small samples on rotary machine multiclass fault recognition

Abstract

The current development's major bottleneck of fault pattern recognition is the absence of fault samples, and not only the methodology itself. Most methods of mechanical fault recognition depend on the large samples of the statistical properties (such as neural network). When training limited samples, it is difficult to guarantee getting a better classification. In response to the lack of rotary mechanical diagnostic samples, this paper takes the advantages of Support Vector Machines (SVMs) in small sample classification for studying its application in small number of samples for rotary machine fault pattern recognition. For rotary machine's multi-class fault problem, we introduce three methods based on binary classifications: "one-against-all", "one-against-one", and "directed acyclic graph" SVM (DAGSVM) and then compare their performance in fault recognition. The experiments indicate that the SVMs has high adaptability for rotary machine fault diagnosis in the case of smaller number of samples and the "one-against-one" and DAG methods are more suitable for rotary machine fault diagnosis than the other.