Impact of wavelets and filter on vibration-based mechanical rub detection using Neural Networks

Abstract

The Discrete Wavelet Transform has been widely used as a mathematical tool for vibration signal analyses in Condition Monitoring Systems for the past couple of decades. But like any transformation, an effective analysis is largely dependent on the noise characteristics associated with the acquired signal, which inevitably degrades its performance. A quantitative evaluation of the impact of both filtering regimen and wavelet family on vibration analysis is presented in this paper. The classification error associated with the training of a Radial Basis Function (RBF) Neural Network is used to quantify the performance of different filtering routines and wavelet families. Faults associated with mechanical rubbing have been considered in the present research and the results indicate that a generous performance enhancement, reaching as high as 28 percent is possible when an effective filter-wavelet combination is used leading to a more reliable detection of mechanical rubbing faults in rotating machinery.