2
$\begingroup$

I need to compute the DC offset, amplitude, and phase of a sine wave and I would like some help validating that my current technique is correct and get some tips to improve my technique. I am a DSP noob.

I have a signal that is sampled at 100 kHz and I am reasonably sure of the signal's frequency. To find the DC offset I'm just computing a single bin DFT (Goertzel) at 0 Hz.

My problem comes when I need to compute the phase and amplitude at the signal's frequency (2400 Hz). As you can see from the graph it seems to get close to the actual signal but it looks like the amplitude and phase are just slightly off.

Reconstruction of a noiseless sinewave

The performance also seems to degrade as noise is introduced to the signal. I say seems because I'm judging the performace of this algorithm based purely on visual inspection alone.

Reconstruction of a noisy sinewave

Question 1 Am I on the right track as far as measuring phase, amplitude? Are my results reasonable?

Question 2 How can I improve the accuracy of my phase and amplitude reconstruction?

Bonus Question 3 Why do I need to divide the result from Goertzel's algorithm by N when computing the DC offset?

Bonus Question 4 Why is the result of Goertzel's algorithm seemingly shifted by 90 degrees? When I do the reconstruction I used a cosine wave instead of a sine wave because it looked better, which to me is an unsatisfying answer. I still don't understand all that's happening in Goertzel's algorithm.


Matlab code to generate the first graph:

close all
clear all

signal = [145,113,87,67,54,46,44,43,44,49,61,79,102,131,159,195,232,273,305,339,374,406,434,460,479,493,503,507,505,499,486,467,446,419,388,355,319,284,251,215,179,147,115,91,70,55,47,44,43,44,49,63,78,100,127,159,191,232,264,302,337,371,403,431,457,478,492,503,508,505,502,488,470,447,419,392,358,323,287,254,217,183,150,119,91,70,55,46,44,44,44,48,59,76,99,127,158,191,230,265,300,335,369,399,430,457,476,491,502,505,505,499,489,471,447,424,392,361,325,291,255,220,183,151,121,95,75,55,47,44,43,44,48,59,74,95,124,156,191,227,264,296,332,367,399,428,453,475,491,502,507,505,499,488,471,451,425,396,363,328,291,259,223,188,154,124,99,72,57,47,44,42,43,48,57,73,95,121,151,189,223,262,294,329,365,397,425,451,471,489,499,505,507,502,489,473,453,427,398,366,332,296,262,223,190,158,127,99,78,57,48,44,43,44,47,55,72,92,119,153,187,222,255,291,328,363,395,424,447,471,488,502];

fs = 100e3;
f  = 2400;
N  = length(signal);
t  = 0:1/fs:N/fs-1/fs;

s = [goertzel(signal, 0) / N, goertzel(signal, f)];

dc    = s(1);
amp   = abs(s(2));
phase = angle(s(2));
reconstruction = amp * cos(2*pi*f*t + phase) + dc;

h = figure(1);
hold on
title('Signal Reconstruction')
xlabel('Time (ms)')
ylabel('ADC Counts')
grid on
plot(t*1000, signal, 'b');
plot(t*1000, reconstruction, 'r');
ylim = get(gca, 'ylim');
xlim = get(gca, 'xlim');
text(xlim(1), ylim(2) - 15, [num2str(amp) ' cos(2\pi * ' num2str(f) 't + ' num2str(phase * 180/pi) ') + ' num2str(dc)]);
legend('Signal', 'Reconstruction')
hold off

Matlab code to generate the second graph:

close all
clear all

noisy = [2395,2376,2365,2351,2364,2359,2367,2349,2335,2348,2351,2374,2367,2376,2374,2378,2399,2414,2431,2430,2431,2446,2447,2475,2479,2487,2477,2472,2479,2486,2495,2487,2479,2470,2460,2463,2461,2463,2440,2423,2415,2399,2414,2396,2383,2365,2351,2359,2359,2379,2367,2365,2365,2367,2383,2399,2418,2415,2414,2428,2439,2463,2468,2479,2476,2477,2493,2495,2511,2495,2492,2492,2487,2495,2492,2487,2471,2455,2462,2447,2455,2438,2423,2399,2383,2396,2383,2391,2377,2364,2367,2367,2383,2383,2396,2383,2383,2415,2428,2447,2447,2447,2447,2460,2486,2494,2508,2495,2495,2505,2508,2526,2511,2508,2495,2487,2495,2487,2494,2472,2455,2447,2438,2443,2431,2420,2398,2383,2383,2383,2399,2383,2382,2380,2383,2408,2415,2430,2423,2423,2431,2444,2472,2478,2486,2479,2485,2503,2511,2527,2511,2511,2511,2508,2525,2511,2511,2495,2479,2479,2479,2487,2473,2460,2441,2423,2430,2415,2415,2398,2383,2383,2380,2399,2399,2408,2399,2399,2428,2438,2463,2458,2461,2463,2472,2495,2506,2511,2511,2508,2511,2523,2543,2537,2527,2511,2510,2511,2511,2511,2495,2487,2477,2463,2473,2463,2459,2431,2415,2415,2413,2423,2414,2408,2399,2398,2419,2428,2441,2431,2438,2447,2447,2479,2487,2495,2492,2493,2511,2511,2541,2536];

fs = 100e3;
f  = 2400;
N  = length(noisy);
t  = 0:1/fs:N/fs-1/fs;

s = [goertzel(noisy, 0) / N, goertzel(noisy, f)];

dc    = s(1);
amp   = abs(s(2));
phase = angle(s(2));
reconstruction = amp * cos(2*pi*f*t + phase) + dc;

h = figure(1);
hold on
title('Noisy Signal Reconstruction')
xlabel('Time (ms)')
ylabel('ADC Counts')
grid on
plot(t*1000, noisy, 'b');
plot(t*1000, reconstruction, 'r');
ylim = get(gca, 'ylim');
xlim = get(gca, 'xlim');
text(xlim(1), ylim(2) - 15, [num2str(amp) ' cos(2\pi * ' num2str(f) 't + ' num2str(phase * 180/pi) ') + ' num2str(dc)]);
legend('Signal', 'Reconstruction')
hold off

Goertzel's algorithm at any frequency:

function y = goertzel(data, f)
N  = length(data);
w  = 2 * pi * f / N;

s0 = 0;
s1 = 0;
s2 = 0;

for i=1:1:N
  s2 = s1;
  s1 = s0;
  s0 = data(i) + 2*cos(w) * s1 - s2;
end

y = (s0 - s1 * exp(-j*w)) * (exp(-j*w*(N-1)))
$\endgroup$
  • 1
    $\begingroup$ If your samples are duplicated due to a 'quirk', shouldn't you remove the duplicates before doing the analysis? Also, when you say the sine wave is distorted, what sine wave are you talking about? Please clarify. $\endgroup$ – MBaz Jun 30 '16 at 22:56
  • $\begingroup$ Your data goes from about 44 to 508, for an amplitude of (508-44)/2 = 232, and a DC offset of (232+44)=276. And what do you think the DC value will compute to when you have about 6.5 cycles of a sine wave? Suggest you use a single bin DFT tuned to the frequency of the input, eg: dsp.stackexchange.com/questions/3301/… Or use a sliding DFT, or other methods. $\endgroup$ – user14819 Jul 1 '16 at 5:02
  • $\begingroup$ Appreciate your efforts in detailing your question but it's still unclear what exactly you are asking... A hardware solution? A formula correction? A method implementation? A concept validation? A software bug to solve? please state (after isolating it as much as possbile) what your exact problem is ? $\endgroup$ – Fat32 Jul 1 '16 at 11:40
  • $\begingroup$ @Fat32 I've had a sleep on this problem and the muddy waters have calmed but are still muddy. Hope my edits have made the waters clearer for every else as well. Let me know if the question still needs some work. $\endgroup$ – thndrwrks Jul 1 '16 at 19:21
1
$\begingroup$

In response to Question 1, I suspect that Goertzel algorithm , is like DFT itself, suffers from leakage and this might affects your estimated parameters, especially in your case, where your signal length (window) is too short (only about 250 samples). This is DFT of your signal:

enter image description here

Leakage is so obvious here.

So I suggest to do something else instead of Goertzel or DFT to estimate your parameter, like "sum of sines" curve fitting. In MATLAB simply use cftool, to estimate your Phase, Amplitude and DC. I've tried to estimate the your parameters using some of two sines (one with zero frequency) and got:

General model Sin2:
       f(x) =  a1*sin(b1*x+c1) + a2*sin(b2*x+c2)
Coefficients (with 95% confidence bounds):
       a1 =       270.2  (-80.04, 620.5)
       b1 =        4.79  (-848.3, 857.9)
       c1 =        1.48  (-12.69, 15.65)
       a2 =       235.7  (235.1, 236.3)
       b2 =   1.53e+004  (1.529e+004, 1.53e+004)
       c2 =       3.687  (3.682, 3.693)

Try this parameters, you've got the exact signal (blue dots are your samples, red curve is fitted sine):

enter image description here

If your are seeking an efficient method and if you know your frequencies (well it seems you know one is DC and the other is 2400), you can use much lighter computations using a linear regression method which is well explained in this post:

https://math.stackexchange.com/questions/1876900/sinusoidal-fitting-when-frequencies-are-known-however-some-samples-are-missing

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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