Extraction of second-order cyclostationary sources-Application to vibration analysis

This paper addresses the issue of extracting the pure second-order cyclostationary (CS2) part of a signal. This proves very useful in many situations where the CS2 part actually contains most of the information in a signal, such as in acoustics or in vibration analysis. The proposed method exploits the spectral redundancy induced by the pure CS2 part and tries to reconstruct it by combining several filtered frequency-shifted versions of the signal. It only requires the a priori knowledge of the cyclic frequencies of the sources of interest. Contrary to other blind source separation algorithms, the method does not need any knowledge about the number of sources, it does not need to set up a forgetting factor, and, it works in highly noisy contexts. The derivation of the optimal filters is described in detail. The effectiveness of the method is finally demonstrated on both simulated and real industrial examples.