I am trying to analyse multivariate time series data sets. I have 6 signals for each event, representing 3 linear accelerations and 3 rotational velocities for a 40ms window. I am trying to find a way to cluster together similar events based of these 6 signals.

The method I am currently looking at is using FFT on each signal to reduce it to frequency bins, Then doing some sort of clustering algorithm on highest 3 amplitude frequencies or something along those lines.

My question is what sort of clustering algorithm should I be looking at to cluster my problem. If for example my problem has 100 events, 6 sensors, 3 frequency and amplitudes per sensor per event.

I am new to this type of signal processing so this methodology might not be feasible but I welcome any suggestions on a clustering algorithm or a completely other approach that you might think is better for my problem.


1 Answer 1


I'd start off with something simple like vector quantization (VQ). This allows you to set up a group of code words (centroids of your clusters), and simply find the code word closest to any of your data samples.

There are more complex approaches, but VQ gives you a good, simple baseline. If it doesn't give you the performance you need, then you can move to more complex approaches. The next step, after simple VQ, would be to try $K$-means clustering. These notes give a good outline.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.