Analysis method for non stationary time history signals EMD or Hilbert Huang Transform 1

Just came across this. So, if you have a time history of a varying signal the usual approach is to break it up into short sections and FFT each section (STFT), which tells you when the signal changed, and then do the same with longer sections, which can give better frequency resolution. Then you get into wavelets.

This EMD, or HHT, throws the whole Fourier approach away, it is based on envelope analysis. Fit a spline to the maxima of the signal. Fit a spline to the minima of the signal. Average the two splines together and subtract that from your signal. This residual becomes the new signal.

rinse and repeat until you are bored. This gives a whole array of the averaged splines, and one final residual. if you add them all up you get back to the original signal.



This approach has lots of problems, some of which are solved by going to EEMD, which adds noise to the original signal, then repeats the process with different noise signals, and then averages the lot together.

Also, at best, it leaves you with a cleanish signal that still needs to be analysed if you are interested in frequencies.


code for EMD, works as is in Octave


wiki page

- not very good

paper on EEMD


Cheers

Greg Locock


New here? Try reading these, they might help FAQ731-376

I've not done this type of analysis for real. Have you looked at That was supposedly the BEST for analyzing non-stationary signals, but horribly expensive computationally.

TTFN (ta ta for now)
I can do absolutely anything. I'm an expert! faq731-376 forum1529 Entire Forum list

Thanks, never heard of it. Seems a bit less dodgy than envelope analysis. To be honest computation time is the least of my worries. Here's a link for matlab, current version has it already


Cheers

Greg Locock


New here? Try reading these, they might help FAQ731-376