You could compute the covariance that measures how much a signal is similar to another. For example in Matlab, you can do that with the function cov().

See more about this in here:

http://cnx.org/content/m10673/latest/

I would guess you must use some statistics, machine learning . Because the signals you mention, are kind of random. I don't think the signal of happines in a moment is equal to the signal of happines one hour later . But there must be a pattern somewhere .

Maybe it is as simple as differences in the mean amplitude, or at the power spectrum or stuff like that . But if it is not, you should use statistics along with that. I am not an expert at this . But that is my misinformed opinion.. Hope I am not too wrong.

Try to do FFT and multiply with conjugate of two signals.

You could compute the covariance that measures how much a signal is similar to another. For example in Matlab, you can do that with the function cov().

See more about this in here:

http://cnx.org/content/m10673/latest/

As the two signals are random variables,and the magnitude of the two signals and also the nature of variation(up mood or Down mood from the existing state etc.are not known, the signals fit in well with Stochastic Systems

Hence Tools of Stochastic Systems may be useful. Methods like Auto correlation or Cross Correlation , etc. may be thought of .

P.S.

There are a few more ways to see.

(1) Compare the frequency spectra of the two signals.

(2) Perform empirical mode decomposition of the signals and compare the components. (See : http://www.clear.rice.edu/elec301/Projects02/empiricalMode/)

Need a little more details. Are the signals one dimensional? How big is the size of the data set? Are the signals obtained from different subjects?

In my opinion you should try time-frequency analysis. Not an expert of the specific types of signals you are dealing with but I think that it may turn out to be useful to know the time involvement along with the frequency features.

Stochastic analysis would be apt, such as cross-correlation measurement etc.

You can use correlation function to compare signals directly in the time domain. You will see how the signals are similar and additionally you will see the time shift of the signals. If you want to compare the energy of the signals use coherence function. This function compares PSD (Power Spectral Density) of the functions.

Thank you all for your answers.

Actually, using fft or other frequency comparisons might not be a good idea since the signals are measures of pupil diameter and don't have frequencies.

correlation and co variance doesn't show differences because the two signals are behaving almost the same way. its just that the negative one has higher and sustained dilation compared to the positive one.

the two signals are the average for 30 participants. do you suggest that I compare each participant separately? and if this is the case what method can be used?

does any one have knowledge on fractal approach that's used in classification and shows two ranges for different features to distinguish two groups?

Fractal is a topic in Computer Graphics.Any good book on Computer Graphics deals with it

A fractal is an object or quantity that displays self-similarity, in a somewhat technical sense, on all scales. The object need not exhibit exactly the same structure at all scales, but the same "type" of structures must appear on all scales. A plot of the quantity on a log-log graph versus scale then gives a straight line, whose slope is said to be the fractal dimension. The prototypical example for a fractal is the length of a coastline measured with different length rulers. The shorter the ruler, the longer the length measured, a paradox known as the coastline paradox. Refer LINK:

http://mathworld.wolfram.com/Fractal.html

P.S.

Hi,

In time-lapse seismic their is a similar problem. They want to compare two seismic traces. They use as repeatability metrics Normalized RMS ratio (NRMS), correlation coefficient (r) and predictability (PRED):

---

Kragh, E., and Christie, P., 2001, Seismic repeatability, normalized RMS and

predictability:, in 71st Ann. Internat. Mtg Soc. of Expl. Geophys., 1656–1659.

---

In image registration, they use metrics quite similar and they use another metric based on mutual information:

---

http://www.cs.jhu.edu/~cis/cista/746/papers/mutual_info_survey.pdf

---

I guess other metrics based on entropy are similar to the one based on mutual information.

A lot of other metric can be inferred from the four metrics above.

Wavelet analysis (and time-frequency analysis by the way) is worth the complexity because it computes these metrics: if you use the Haar wavelet, you get your 1st and 2nd derivatives for example.

Regards,

Nicolas

If your signals are stationary (or at least weakly stationary over a short period), then Cross-Correlation - time series analysis can give you some very good information regarding two signals dependencies.

Coding the algorithm is also very easy to implement.

First, I think that if you provided a sample of the obtained data others can help you better. Second, Is your objective to arrive at a method to detect emotions from pupil's opening changes?. Assuming that, MMSE, or the minimum mean square error method can be used to estimate the difference between a certain waveform and the waveform of the (baseline) positive reflection or the waveform of the (baseline) negative reflection. The baseline of positive reflections can be obtained by averaging of all 30 participants when subjected to positive emotions while the baseline of negative reflections can be obtained by averaging of all 30 participants when subjected to negative emotions. If the error comparing with positive reflection baseline waveform is smaller than when comparing with the negative reflection baseline waveform it is more likely that a positive emotion occurred and vise-versa. MSE can be computed by calculating the square of difference between the two waveforms.

You may need to combine above technique with some windowing technique to ensure proper instant of comparison, i.e phase. Thanks. @AlDmour

It depends on the level of measurement of the questionnaire

Jooree,

please explain the actual data that you have for an individual patient. i am guessing you have something like two pieces of information such as maximum pupil diameter increase and time duration of increased pupil dilation?

ismat and subhajit are on the right track. you need to do a hypothesis test on your two datasets (subjects' reactions to "positive" and "negative" stimuli) and determine whether there is a statistically significant difference between the measured parameters at a suitable confidence level.

if you have bivariate data (e.g. dilation size and duration from each subject) you will need to find a suitable hypothesis test for bivariate data. if it is univariate, a two-sample student's t test might do the trick.

the vast majority of the answers that people have provided are not appropriate for your problem as far as I can tell without knowing more details. Don't use fancy analysis methods when simple methods are all you need!

Hello Joore,

If you have two different signals like digital 1 and 0, you should use PN Series to make their diifferences more noticeable. And if you use PN Series coorelation at the reciever side your data estimation will be more accurated.

If your recieved data is still not quite right, then you can try higher orders of PN Series like 16-bit , 32-bit or higher. But remember higher PN Series means slower communication speed.

I think wavelet transform technique is the best technique to find the difference between two signal because u will get the information of the two signal in diffrent levels and this information in each level will be diffrent for diffrent emotions .morover u can design a classifier that classifies the features of each level of WT in both signal .

think as Neeldrata said we can use wvelet transform technique and Furrier transform u will get good analysis

The issue with wavelet is that different mother wavelets gives you different results.