Effect of cascade methods on vibration defects detection

Within the framework of monitoring rotating machines, vibration analysis remains an effective tool for fault detection. This analysis generally consists of measuring acceleration signals from critical and judiciously chosen points of a machine with the help of piezoelectric sensors. However, relevant information concerning the machine health can be masked by disturbances such as noise. The detection reliability will then be conditioned directly by the quality of the collected signal. Signal preprocessing methods, in particular denoising methods, can significantly improve the detection quality in terms of reliability. In this paper we aim to compare two methods of denoising based on signal spectral content analysis: discrete wavelet transform and empirical mode decomposition. A first study is carried out in order to optimize specific parameters related to each of the two methods, starting from experimental data obtained on degraded bearings. In fact for each parameter, one has to define conditions which allow the best detection of periodic pulses in vibration signals thanks to indicators such as kurtosis and crest factor. The second study consists of assessing the effectiveness of each denoising method on a vibration signal measured on a failed bearing. This signal is then disturbed by various noises simulated with variable levels. This study aims to show the effectiveness of each of these two methods on the early detection of impulse defects.