Anomaly detection of the tapered roller bearings with statistical data-driven approaches

Chen, S.L., Wang, L., Wood, R.J.K., Callan, R. and Powrie, H.E.G. (2010) Anomaly detection of the tapered roller bearings with statistical data-driven approaches. Insight, 52 (8), 428-436. (doi:10.1784/insi.2010.52.8.428).

Abstract

Current bearing monitoring generally relies on univariate sensing techniques and the development in automatic bearing fault detection is far from satisfactory. This study has developed an innovative anomaly detection strategy for bearing fault detection based on multi-sensing variables and a series of statistical data-driven approaches. The major advantage of using multi-sensors is that it enables the capture of abnormal signals due to different contact physics, for example phase transformation or surface cracking that may not be detected by one type of sensor.

Experiments have been carried out on a bench top bearing test-rig under both normal and seeded defect conditions. Sensors measuring electrostatic charge, vibration and debris are used in this study. The signals under normal bearing conditions are used to train Gaussian mixture models (GMM) to establish a preliminary normal model for anomaly detection. The preliminary model is firstly adapted by a novel ‘data cleaning’ method that identifies and removes unexpected anomalies in the training data. The purpose of data adaptation is to eliminate fault masking by the anomalies during training. Seeded defect bearing tests provide signals containing data from a bearing under ‘faulty’ or failure conditions. Hotelling's T-squared statistics of the test data are used to detect anomalies in this study. The results from a series of bearing tests have demonstrated that the application of the GMM adaptation techniques based entropy score and distance can help to optimise the trend of the T-squared statistic, minimise noise and extract/enhance valuable abnormal trends. Furthermore, to achieve automatic detection with the T-squared statistic, a novel threshold set-up approach based on the combination of the GMM and extreme value theory (EVT) has been developed in this study. The threshold set by the new approach can reduce the false alarm rate by 20% compared to the thresholds set by other conventional thresholds. Details of the developed approaches as well as a comparison with some conventional approaches are presented in this paper

This record has no associated files available for download.

Contributors

Author: L. Wang

Author: R.J.K. Wood

Download statistics