External condition removal in bearing diagnostics through EMD and One-Class SVM

The removal of the running conditions influencing data acquisitions in rotating machinery is a very important task because it could avoid some misunderstandings when diagnostic techniques are applied. This paper introduces a new parameter that could be able to identify damage in a rotating element of a roller bearing removing the effect of speed and external load. The parameter proposed in this paper is evaluated through Empirical Mode Decomposition (EMD). Our algorithm proposes firstly the decomposition of the acceleration vibration signals into a finite number of Intrinsic Mode Functions (IMFs) and then the evaluation of the energy for each one of these. Data are acquired both for a healthy bearing and for one with a 450 μm large indentation on a rolling element. Three different speeds and three radial loads are monitored for both cases, so nine conditions can be evaluated for each type of bearing overall. The parameters obtained, namely energy evaluated for a certain number of IMFs, are then used to train a One-Class Support Vector Machine (OCSVM). Healthy data belonging to the nine different conditions are taken into account and OCSVM is trained while other acquisitions are given to the classifier as test object. Since the real class membership is known, we consider how many errors the labelling produces. We compare these results with those obtained by considering a wavelet decomposition. Energies are evaluated for each level of decomposition and the previous approach is then applied to these parameters.