A comparison of vibration based bearing fault diagnostic methods

This paper presents a benchmark study in which three vibration based bearing diagnostic algorithms are compared. The three methods are a data driven approach developed by the Linz Center of Mechatronics (LCM), an physics based method of Flanders Make (FM) and an approach developed by the Center for Intelligent Maintenance Systems (IMS). Two experimental tests have been performed, an accelerated life time test to degrade a bearing and introduce an operational bearing fault and a gearbox test containing various faulty test bearings. The methods are compared based on their diagnostic performance, practical applicability, training and configuration requirements. Based on the accelerated life time test, it is concluded that the method of IMS and FM, employing bearing specific features, showed to be more sensitive for early bearing fault detection than purely statistical features used in the method of LCM. On the contrary, the method of LCM does not require specific system knowledge and is not limited to bearing monitoring only. The method is more widely applicable to fault monitoring problems. The methods of IMS and LCM seem to outperform the method of FM in the gearbox test. However, the training and testing data used by those methods were acquired for the same bearing sample and for the same bearing assembly. This could lead to a high correlation between the training and testing data and hence a misleading classification accuracy. Therefore, attention should be paid to the quality of the training data. It is concluded that the training data should comprise all relevant system variations, including e.g. remounting of the bearing, to ensure that the classification is uniquely based on bearing fault related effects. The methods of IMS and LCM require validated training data of both healthy and faulty bearing scenarios, whereas the method of FM relies on training data of healthy bearings only. In practice, the availability of training data of faulty bearings is often scarce and could make the adoption more complicated. The findings presented in this paper serve as a guideline to support the selection of an appropriate method for practical applications.