Wear detection of rolling element bearings using multiple-sensing technologies and mixture-model-based clustering method

Abstract

Online fault diagnostic technologies are fast emerging for detection of incipient faults on tribological components to avoid catastrophic failure. Vibration analysis has long been used to detect machine faults, but is sensitive to relatively severe conditions only. Electrostatic monitoring is a newly developed approach with the potential to detect precursor processes that indicate contact distress and wear. Recently, at the University of Southampton, both vibration and electrostatic sensors were implemented on a bearing testing rig to evaluate their effectiveness in detecting bearing faults. The results indicate that both types of sensor are sensitive to bearing deterioration shortly before complete failure. However, univariate plots of signals from both types of sensor only exhibit significant change when entering the severe wear stage. Therefore, multivariate techniques for detecting wear severity of components at different running stages need investigating. In this study, an unsupervised training method, called mixture-model-based clustering, that utilizes the expectation maximization (EM) algorithm is employed to develop further a wear detection technique. The choice and extraction of significant features from both vibration and electrostatic sensors are discussed as step one. The second step uses the clustering method to examine the behaviour of the extracted features during different running stages, and to quantify how good the sensors are at distinguishing wear severity. In the third step, a dynamic wear detection process is simulated. Clustering is applied to baseline data from a known healthy bearing and data from different wear stages to see if the data naturally group by wear condition. The result shows that the unsupervised clustering method is able not only to learn and detect wear conditions of the rolling element bearings with the developed statistical monitoring charts of occupation probability (OP) in the clusters and number of the trained clusters (NC), but also to obtain the advantage of detecting insignificant abnormalities that might be overlooked in the conventional plots.