Early damage detection of roller bearings using wavelet packet decomposition, ensemble empirical mode decomposition and support vector machine

Abstract

Roller bearings are widely used in rotating machinery and one of the major reasons for machine breakdown is their failure. Vibration based condition monitoring is the most common method for extracting some important information to identify bearing defects. However, acquired acceleration signals are usually noisy, which significantly affects the results of fault diagnosis. Wavelet packet decomposition (WPD) is a powerful method utilized effectively for the denoising of the signals acquired. Furthermore, Ensemble empirical mode decomposition (EEMD) is a newly developed decomposition method to solve the mode mixing problem of empirical mode decomposition (EMD), which is a consequence of signal intermittence. In this study a combined automatic method is proposed to detect very small defects on roller bearings. WPD is applied to clean the noisy signals acquired, then informative feature vectors are extracted using the EEMD technique. Finally, the states of the bearings are examined by labeling the samples using the hyperplane constructed by the support vector machine algorithm. The data were generated by means of a test rig assembled in the labs of the Dynamics and Identification Research Group in the mechanical and aerospace engineering department, Politecnico di Torino. Various operating conditions (three shaft speeds, three external loads and a very small damage size on a roller) were considered to obtain reliable results. It is shown that the combined method proposed is able to identify the states of the bearings effectively.