Automated bearing fault classification based on discrete wavelet transform method

Abstract

The prediction of the current state of the rotatory mechanical components like bearing needs to be of utmost accuracy for the efficient operation of the machine. As the bearing is normally employed at the supporting ends of the power transmission line, manual monitoring by the disengagement of the bearing is not a feasible option. Thus, there is a need for highly accurate machine learning algorithms that can successfully detect the abnormalities by capturing the running characteristics of the system. One such characteristic can be a vibration signal. In this article, various autonomous methods for condition monitoring are compared by employing machine learning and ensemble learning methods for fault classification in the bearings. Three different types of faults namely inner race fault, outer race fault, and ball fault have been considered. Discrete wavelet transform is applied to the raw signals obtained for healthy and these faulty conditions. The feature vector is extracted by decomposition and reconstruction of signals for ten different levels. The extracted feature vector is fed to three different classifiers: XGBoost, decision tree, and support vector machine. The results reveal that XGBoost outperforms the other two classifiers. The proposed manuscript signifies the capability of XGBoost classifier even if treated with readily fewer instances of raw signals.