Damage identification and external effects removal for roller bearing diagnostics

In this paper we introduce a method to identify if a bearing is damaged by removing the effects of speed and load. In fact, such conditions influence vibration data during acquisitions in rotating machinery and may lead to biased results when diagnostic techniques are applied. This method combines Empirical Mode Decomposition (EMD) and Support Vector Machine classification method. The vibration signal acquired is decomposed into a finite number of Intrinsic Mode Functions (IMFs) and their energy is evaluated. These features are then used to train a particular type of SVM, namely One-Class Support Vector Machine (OCSVM), where only one class of data is known. Data acquisition is done both for a healthy bearing and for one whose rolling element presents a 450 μm damage. We consider three speeds and three different radial loads for both bearings, so nine conditions are acquired for each type of bearing overall. Feature evaluation is done using EMD and then healthy data belonging to the various conditions are taken into account to train the OCSVM. The remaining data are analysed by the classifier as test object. The real class each element belongs to is known, so the efficiency of the method can be measured by counting the errors made by the labelling procedure. These evaluations are performed by applying different kinds of SVM kernel.