A multi-model approach for anomaly detection and diagnosis using vibration signals

Abstract

Continuous vibration monitoring of mechanical roller bearing parts potentially reduces machine downtime through timely prediction and diagnosis of abnormal events. Despite the progress made in the literature, challenges remain in how to assess performance related information for maintenance decision-making from large data streams. Furthermore, since roller bearings operate under various regimes (e.g., speed and load), it is not trivial to consider the effect of regime changes in the modeling in order to reduce false alarms. The paper describes a multi-model approach to monitor the condition of roller bearings under different operating regimes. Two modeling approaches for anomaly and degradation monitoring are proposed to automatically retrieve information from the data. A self-organizing map (SOM) and a support vector machines (SVM) are used comparatively for the evaluation of a bearing degradation in time (i.e., a dynamic health indicator) and for the determination of changes in the tracked features. The proposed method is validated using data from multiple bearings of the same type.