Blind extraction of a cyclostationary signal using reduced-rank cyclic regression—A unifying approach

This paper addresses the issue of the blind extraction of a second-order cyclostationary source drowned by an unknown number of interferences and additive noise. It first reviews two recently developed methods based respectively on a subspace decomposition of the observed signals via their cyclic statistics and on multiple cyclic regression (MCR). It then proposes a unifying and refined approach using reduced-rank cyclic regression (RRCR) which combines the respective advantages of the two previous methods and suppresses their drawbacks. It also reveals that unlike the classical MCR technique, the power of the additive noise at the output of RRCR does not depend neither on the number of frequency shifts used in the regression nor on the number of available measured signals. This property is verified by means of simulations where the behaviour of all the methods with respect to many parameters is compared. RRCR is finally applied to the diagnostics of bearings and gears where it is shown to achieve a very good extraction of fault signatures.