Determining the period of a discontinuous function

Question

I'm new to the field of DSP.

I'm trying to determine the period and shift of the function. I've tried using FFT, but haven't had much luck. Seems like it should be simple.

Signal (pastebin of Octave matrix):

# Created by Octave 3.2.4, Wed Dec 12 00:26:41 2012 PST <jjohnson@jjohnson>
# name: a
# type: matrix
# rows: 1
# columns: 495
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.529411764705882 0.945098039215686 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.415686274509804 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.556862745098039 0.890196078431372 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.72156862745098 0.749019607843137 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.862745098039216 0.611764705882353 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.933333333333333 0.470588235294118 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

Signal drawn with plot(a):

By hand I can tell that the shift is 75, and period is ~56. How do I get that from the results of the FFT?! Is there some other more appropriate algorithm?

FFT drawn with plot(abs(fft(a))):

Answer 1

This will give you a plot of the autocovariance up to lag 100 samples:

plot(autocov(a,100))

There you can clearly get the period of your signal.

Another approach is to explicitly get the time of each pulse:

pulses = a > 0.1;
leading_edges = diff(pulses) > 0;
times = find(leading_edges > 0);
periods = diff(times)

ans =
    55   56   56   56   56

Explanation: pulses is 1 wherever there is a pulse. leading_edges gets a 1 only on the first sample of a pulse. times gives the sample number of each leading edge.

With this result you can perform your own statistics or estimation. This second method is computationally more efficient than using transforms. Computing time is linear with number of samples.

Answer 2

My answer is based on this answer https://dsp.stackexchange.com/a/15117/3573 by https://dsp.stackexchange.com/users/80/peter-k

The following is Octave code:

y=[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.529411764705882 0.945098039215686 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.415686274509804 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.556862745098039 0.890196078431372 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.72156862745098 0.749019607843137 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.862745098039216 0.611764705882353 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.933333333333333 0.470588235294118 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0];
plot(y);
title("Signal")
print -dpng signal.png

autocorrelation=xcorr(y);
figure
plot(autocorrelation)

[pks loc] = findpeaks(autocorrelation,'DoubleSided');
hold on
plot(loc,pks,'or')
title("Autocorrelation")
print -dpng autocor.png

distance_between_peaks = diff(loc);
printf("Mean distance between peaks: %d\n",mean(diff(loc)))

The following figure is the plot of your signal y:

The following figure is the plot of the auto-correlation xcorr(y) of your signal, the peaks are marked in red:

We then take the distances between the peaks locations (with diff) and compute the mean distance, the script gives this output:

Mean distance between peaks: 55.6667

Answer 3

If you remove the DC offset before the FFT, the FFT result will show the low frequency peak a little more clearly.