I have a time series, which is produced from a sensor. This sensor detect sound in the real World. My aim is to separate sounds with different frequencies from this time series.
With which technique can I do this job?
$\begingroup$
$\endgroup$
1
-
1$\begingroup$ I think that the so called wavelet analysis is the thing that could help you, but this is a pretty complex topic... $\endgroup$– DatamineRCommented Jan 10, 2014 at 12:06
Add a comment
|
3 Answers
$\begingroup$
$\endgroup$
If your time series appears to be stationary and highly periodic, it should be worthwhile to analyse its characteristics in the frequency domain. Even if your purpose is not prediction, this post on time-series-analysis-and-order-prediction-r may get you started on how to use neural networks for signal decomposition in the frequency domain, using the most flexible software. For non-stationary time series, search more on wavelets after checking out this post on Wavelet Spectrogram Non-Stationary Financial Time Series analysis using R .
answered Jan 10, 2014 at 18:56
$\begingroup$
$\endgroup$
A standard means of separating sounds from multiple sources is called blind source separation (BSS), on such method which is commonly used is called Independent Component Analysis (ICA) .
One reported real world application is the separation of seizure activity from background activity in EEG (the brains electrical signal).
$\begingroup$
$\endgroup$
this recent article present a fairly comprehensive review of methods for multi pitch analysis (whis is method to separate sounds with different frequencies) http://www.disit.dinfo.unifi.it/articoli/articoloIEEEfinal.pdf
This is a good bibliographical starting point.
I have personnaly some promising work in progress on a pure spectral method (without a priori knowledge of tones or music) but I published nothing so far.
