| T O P I C R E V I E W |
| Proust |
Posted - 01/14/2004 : 07:11:09 AM
Hello everybody,
I need to do autocorrelations of time series data. Unfortunately neighter the help function nor google were any help in doing this. Is there an autocorrelation function available in origin, or do I have to write it myself? Or can I maybe download the code from somewhere?
Thanks in advance,
Proust |
| 9 L A T E S T R E P L I E S (Newest First) |
| Mike Buess |
Posted - 08/13/2007 : 6:19:58 PM
quote:
Need to weed out noises caused by ambient factors (room T and power line fluctuations). any suggestion on that?
Depends on the 'true' signal. If only the DC signal is important you can try to smooth out the rest with Analysis> Smoothing> ? or Analysis> FFT Filter> Low Pass. If the frequency of the 'true' signal is much higher than AC you can try Analysis> FFT Filter> High Pass.
quote:
is there any difference in Correlate function between Origin 6.0 and 7.5?
No, both use fft.corr() which (as far as I know) hasn't changed since Origin 5.0 or earlier.
Mike Buess
Origin WebRing Member |
| vkatsap |
Posted - 08/13/2007 : 4:12:16 PM
Mike, thanks again.
I got your point. I ran a few test cases where Autocorrelation result is almost self-evident (DC signal, random noise, pure sinus).
Actually, I need to find correlation and autocorrelation in time series where there's strong DC component mixed up with noises caused by several factors. Need to weed out noises caused by ambient factors (room T and power line fluctuations). any suggestion on that?
also, is there any difference in Correlate function between Origin 6.0 and 7.5?
I really appreciate your help - best, Victor
quote:
quote:
I chose test time series with signal = constant, and Correlate gave me straight falling line, falling off 1.0 quickly. does that mean that constant, DC, signal isn't correlated to itself????
That's due to the finite length of the signal. When the lag is equal to the signal length there is no overlap so correlation is zero. Note that if you change the length of the signal the maximum correlation will stay at lag=0 and minimum correlation will be at lag = new length. Same thing happens with a pure sinusoid.
quote:
By popular belief, it's when Autocorrelation function drops to 1/e.
I don't think that's applicable to a DC signal. Another way to look at it is the exp time constant for a DC signal is infinite (or much longer than the signal length).
Mike Buess
Origin WebRing Member
|
| Mike Buess |
Posted - 08/10/2007 : 11:39:48 AM
quote:
I chose test time series with signal = constant, and Correlate gave me straight falling line, falling off 1.0 quickly. does that mean that constant, DC, signal isn't correlated to itself????
That's due to the finite length of the signal. When the lag is equal to the signal length there is no overlap so correlation is zero. Note that if you change the length of the signal the maximum correlation will stay at lag=0 and minimum correlation will be at lag = new length. Same thing happens with a pure sinusoid.
quote:
By popular belief, it's when Autocorrelation function drops to 1/e.
I don't think that's applicable to a DC signal. Another way to look at it is the exp time constant for a DC signal is infinite (or much longer than the signal length).
Mike Buess
Origin WebRing Member |
| vkatsap |
Posted - 08/09/2007 : 5:03:47 PM
Mike, many thanks, again.
Let me re-phrase my question: say, I have time series with a signal fluctuating and drifting. I want to find specific time when signal is not correlated to itself anymore. By popular belief, it's when Autocorrelation function drops to 1/e. I chose test time series with signal = constant, and Correlate gave me straight falling line, falling off 1.0 quickly. does that mean that constant, DC, signal isn't correlated to itself????
best wishes, Victor
quote:
quote:
So you saying that 2^n points data set > Correlate woud produce actual Autocorrelation plot?
No, because I'm not familiar with the Wiener-Khinchin theorem or how to apply it. I'm only saying that when comparing two analysis methods it's best to use the same sample data.
...Oh, you are asking if # points affects autocorrelation with Analysis> Correlate. That answer is no because correlation does not require the # points to be a power of two. In other words, Analysis> Correlate gives the true autocorrelation plot regardless of dataset size.
Mike Buess
Origin WebRing Member
Edited by - Mike Buess on 08/09/2007 10:11:18 AM
|
| Mike Buess |
Posted - 08/08/2007 : 2:05:15 PM
quote:
So you saying that 2^n points data set > Correlate woud produce actual Autocorrelation plot?
No, because I'm not familiar with the Wiener-Khinchin theorem or how to apply it. I'm only saying that when comparing two analysis methods it's best to use the same sample data.
...Oh, you are asking if # points affects autocorrelation with Analysis> Correlate. That answer is no because correlation does not require the # points to be a power of two. In other words, Analysis> Correlate gives the true autocorrelation plot regardless of dataset size.
Mike Buess
Origin WebRing Member
Edited by - Mike Buess on 08/09/2007 10:11:18 AM |
| vkatsap |
Posted - 08/08/2007 : 11:39:40 AM
Hi Mike,
thanks for getting back to me on Autocorrelation.
The point of my analysis:
By Wiener-Khinchin theorem, Autocorrelation = FFT(FFT[data]^2).
When Origin Correlation produced something I thought was strange, I tried this FFT-square up data- IFFT , with all the precautions (Normalization and Shift boxes unchecked), and I used exactly 4096 data points. Unfortunately, it also produced straight line, but diffrent from that done with Correlate function.
So you saying that 2^n points data set > Correlate woud produce actual Autocorrelation plot?
thanks in advance - Victor
quote:
Hi,
I don't understand the point of your analysis
(FFT, square Re, square Im, IFFT), but if you
used Origin's FFT tool the difference could
lie in the FFT/IFFT settings... normalization,
wrap, etc. Also, the FFT tool pads 3000 points
with zeroes to 4096 points which will contribute
to the difference...
http://www.originlab.com/www/support/resultstech.aspx?ID=570&language=English&Version=7.5
Mike Buess
Origin WebRing Member
|
| Mike Buess |
Posted - 08/07/2007 : 4:30:29 PM
Hi,
I don't understand the point of your analysis
(FFT, square Re, square Im, IFFT), but if you
used Origin's FFT tool the difference could
lie in the FFT/IFFT settings... normalization,
wrap, etc. Also, the FFT tool pads 3000 points
with zeroes to 4096 points which will contribute
to the difference...
http://www.originlab.com/www/support/resultstech.aspx?ID=570&language=English&Version=7.5
Mike Buess
Origin WebRing Member |
| vkatsap |
Posted - 08/07/2007 : 2:03:03 PM
Hi Mike,
I'm new to Origin, have v. 7.5 for evaluation.
need to do autocrrealtion on time series.
took time series sample of ~3000 size.
Did FFT forward - squared Re and Im, did FFT backward - supposedly, got Autocorerlation function 1 (Wiener-Khinchin).
Then, used just "signal" single column, and did Correlate on it. Got Autocorrelation function 2, much different from function 1.
why's the difference? which method is correct in Origin?
many thanks, Victor
quote:
Hi Proust,
fft.corr(dataset,dataset);
You didn't say which version of Origin you're using, but that LabTalk method has been available at least since Origin 5.0. You can find out about it in the LabTalk guide...
Object Reference->Alphabetical Listing of Objects->FFT->Correlation, Convolution and Deconvolution Functions
Mike Buess
Origin WebRing Member
|
| Mike Buess |
Posted - 01/14/2004 : 7:35:57 PM
Hi Proust,
fft.corr(dataset,dataset);
You didn't say which version of Origin you're using, but that LabTalk method has been available at least since Origin 5.0. You can find out about it in the LabTalk guide...
Object Reference->Alphabetical Listing of Objects->FFT->Correlation, Convolution and Deconvolution Functions
Mike Buess
Origin WebRing Member |
|
|
