Abstract: The ear ly rub-impact fault detection of an engine rotor system in noise circumstances is researched. Anoptimum par ameter sear ching method that can improve the ICA algor ithm is proposed and applied to the separationbetween vibra-signal and noise. After that, the separated vibra-signal is analyzed on the wavelet packet decomposition to obtain the early r ub-impact fault features. Simulation results show that the modified ICA algor ithm is effective on denoise of vibr a-signal, and wavelet packet decomposit ion may provide richer ear ly rub-impact fault information than wavelet decomposition. The proposed approach may be a desired solution to fault detection and diagnosis ofearly rub-impact in engineering practice.

Key words: rotor, early rub impact, ICA, blind separation, wavelet packet, fault detection