Empirical mode decomposition (EMD) of speaker measurements

[Few]

While trying to teach myself a bit about wavelet analysis as an alternative to Fourier transforms when measuring loudspeaker performance, I came across "empirical mode decomposition" (EMD) and the Hilbert-Huang transform. I've done a bit of searching and trying to extract the key ideas, but it doesn't seem like the approach has really taken off with loudspeaker designers and evaluators, in spite of claims of its superiority for many applications.

Apparently Matlab provides both EMD and the H-H transform (at least if the appropriate toolboxes are available) and their website provides a more concise summary of the EMD than I've found most places:
"Another way to explain EMD is to consider a signal as a fast oscillation superimposed on a slower one. After the fast oscillation is extracted, the EMD algorithm treats the remaining slower component as the new signal and again regards it as a fast oscillation superimposed on a slower one. The algorithm continues until some exit criterion is reached. The components in EMD are referred to as intrinsic mode functions (IMF)."

For a more in-depth discussion with graphics that help convey the core ideas, this 2005 paper by Loutridis is loudspeaker-focused. It compares EMD to common approaches such as cumulative spectral decay plots ("waterfall plots") and one form of wavelet analysis.

Even though EMD was apparently introduced in the late 1990s, this is the first I've come across it. I was curious whether anyone else has any experience with either the theory or its application, and also whether anyone knows whether it was just another shiny new idea that hasn't lived up to early assertions of its superiority. It seems like the EMD might be particularly helpful to those trying to dissect a loudspeaker's impulse response to find resonances, but I say that without any hands-on experience with it.

I realize this isn't just a diy question, but diy speaker building is what brought me to the topic, and I couldn't find a better forum. I'll leave it to the moderator to relocate the thread if it belongs elsewhere.

Few

 

[dc655321]

At first glance, that sounds like it has much in common with PCA, Principle Component Analysis , and SVD, Singular Value Decomposition.

 

[Few]

Thanks for your interest! I’m trying to get myself up to speed on SVD, and I’ve read that PCA is an application of SVD, but won’t pretend to know more than I do. The somewhat odd feature of empirical mode decomposition is that it doesn’t begin with an assumed set of basis functions (like the sinusoids in Fourier transforms). It has a simple set of rules about envelopes and uses the rules to extract modes that can be transient, and have key features that vary over time (assuming the data are a sequence of values samples over time). I would say that the approach almost seems simple-minded except that it has apparently been placed on firm mathematical footing and found quite broad application. I’ve just dipped a toe in the water, which is why I thought I’d ask whether any audio junkies had experience with it and might be wiling to clear some of the fog.

Few