I'm working on detecting fault impulses in a vibration signal. I have a vibration signal of 25.6 kHz where I first perform band pass filtering (7000 Hz, 9000 Hz) and then perform Empirical Mode Decomposition (EEMD).

I then want to select the IMF where the kurtosis value is maximised (I'm looking for fault impulses in a gearbox).

See the attached image with the IMFs.

Question:

If I select IMF 5 and perform a FFT I see that the signals are in the 2000 Hz range. How is this possible, given that I already band passed at (7000 Hz, 9000 Hz)? Is it incorrect to use this IMF for further processing then? Thanks!

This paper does something similar.

Article Faulty bearing signal recovery from large noise using a hybr...