The detection of bearing faults for induction motors by using vibration signals and machine learning

Abstract

Bearing faults are the most common type of faults in induction motors. Vibration monitoring is frequently employed to detect and diagnose these faults at their early stages. However, analysis of vibration signals most often requires expert knowledge and an in-depth understanding of specific tool mechanics. Recently, data-based modeling approaches coupled with machine learning algorithms gained significant attraction in the field, which can help manufacturers obtain faster and more scalable fault detection solutions. In this study, three-axis vibration signals of a healthy motor and a motor with inner-race and outer-race bearing faults are collected. Various statistical features belonging to each vibration axis are analyzed. The results show that among the statistical features extracted for each axis, peak-to-peak (p2p) and rms features of vibration signals are the most important features that can distinguish a healthy motor state from a faulty state; however further information is needed to be able to differentiate among faulty states. We have investigated various multi-axis statistical features and employed relatively two simple machine learning (ML) algorithms K-Nearest Neighbors (k-NN) and Decision Trees (DT) to obtain a model. Our results show that combined with ML models bearing faults can be distinguished with accuracies nearly up to 100 %.