Application of convolved blind separation based on second-order cyclic statistics in rolling element bearing feature extraction

Abstract

In condition monitoring of mechanical equipment, the collected signals are mostly mixtures of several different sources. It is very important to recover the individual effects of each source for exact feature extraction and further fault diagnosis. A blind separation method based on second-order cyclic statistics is presented for convolved cyclostationary processes such as those observed in rotating machinery. The convolutive mixtures in time domain are transformed into instantaneous mixtures in frequency domain by discrete Fourier transform (DFT). Then the aim is to find a separating matrix that simultaneously and jointly diagonalizes the set of cyclic spectral density (CSD) matrices of the observed signals. Examples of successful separation are provided on both synthetic convolutive mixed bearing signals and real data tested from a gearbox.