A Study on Vibration Feature Selection for Fault Diagnostics of Rotating Machines

Abstract

Rotating machines are an important system of infrastructure such as power plants. For condition-based maintenance, the selection of reliable features is critical for fault diagnostics. Several feature selection methods are developed to avoid the curse of dimensionality and to increase the accuracy of diagnostic methods. This study proposes a new criterion, modified Fisher ratio (MFR), to select a proper set of features among available vibration feature candidates. The new MFR criterion is devised to be sensitive to outliers of vibration signals collected from degraded rotating machines. The feature pool consists of features extracted from vibration signals based on domain knowledge. To select an optimal subset of the featrues, MFR is used as an objective function of the standard genetic algorithm. A kernel support vector machine (SVM) is used to classifiy the faults of rotating machines. Case studies are conducted using vibration signals collected from the simulation model, the testbed, and real systems. It is expected that the MFR criterion can reduce the dimension of a feature subset while increasing the accuracy of the fault diagnosis of rotating machines.