Decision tree and feature selection by using genetic wrapper for fault diagnosis of rotating machinery

Abstract

In this work, an expert system methodology, based on the decision tree, is proposed to exploit the information provided by the vibration signal time indicators. A study was carried out on a set of six indicators expressed in three physical quantities, namely acceleration, velocity and displacement, to extract expert rules in order to identify the origin of the fault and diagnose the machine. The reduction of the high number of attributes was performed by genetic wrapper approach with classifiers C4.5. This approach permitted us to reduce and select the best input characteristics to build the decision-tree system. These attributes are as follows: kurtosis and crest factor expressed by the acceleration and displacement, and the RMS and peak expressed by the velocity and displacement. The methodology proposed in this paper has been validated on an experimental test bench with five types of operating states (the good condition of the machine, the misalignment, the bearing defects, the unbalance and the combination between the bearings and the unbalance).