Is it possible to calculate the FFT of a signal that was not sampled uniformly? I have an MPU6050 connected to an Arduino uno. I write the samples at the default frequency, approximately 20 milliseconds, but it's not always constant (see the image). Also see this related question. enter image description here

This is my first time I approach to the FFT, and I have a civil engineer background, so is my first time with signal analysis, never studied Signal Theory before.

My interest for the FFT is to define the best low pass filter (example apply a butterworth but I don’t know how to choose the filter order and cutoff frequency.

Thanks for your time, thanks for you patience. I tried to solve this with the code below, but I think is only rubbish.


MatFilt= M(cleaning:n,:);
A= MatFilt(2:m,1);
meanrate= mean(C);
f= 1000/meanrate;
T= sum(MatFilt);    
TemporalSample= MatFilt(:,1);
Axs=Ax*19.6/32768; %I converted the data in m/s^2
SignalLength= (MatFilt(m,1)-MatFilt(1,1))/1000;
title('Signal Corrupted with Zero-Mean Random Noise')
xlabel('time (milliseconds)');
NFFT = 2^nextpow2(m); % Next power of 2 from length of y
Y = fft(Axs,NFFT)/m;
f = f/2*linspace(0,1,NFFT/2+1);

enter image description here


1 Answer 1


You are essentially referring to nonuniform signal sampling in which data samples are not acquired at exact periods but have same random jitter deviations from their exact uniform timing positions.

Now under some suitable conditions (such as the jitter length being less than a period), there are very effective algorithms which can perfectly convert the nonuniformly sampled data into their equivalent uniform samples. (i.e., given the values of nonuniform samples they give a new set of values which correspond to the perfect uniform samples).

After this nonuiform to uniform conversion, you can then apply the usual FFT, which assumes that the data to be transformed was uniformly sampled.

However, before such a DSP attempt, I would suggest you understand the sampling and read out mechanism of the Arduino Uno and MPU 6050 6DOF sensor combination. That sensor can be read in either a single data read mode or in a burst buffer data read mode. In this second mode, I think it will provide a set of buffered data which are sampled by the MPU6050 unit itself as uniform as possible, compared to a sigle data read mode mastered by the Arduino.

From your data, there seems to be a drift of a few milliseconds per sample. I assume a 20 milli second of sampling period is quite achievable with less than 1 ms of timing accuracy using the Arduino in a proper configuration. So may be your main problem is on the software configuration?

EDIT: Try the following MATLAB code as is: NOTE: The following code MAY result in WORSE results than available before processing... The success highly depends on the application.

% ----------------------------
N=100;       % Converts a block of length N nonuniform samples to uniform
K=N;        % 
M = N*K;    % 
Tsam = 0.020; % Sampling period in seconds,

% tp represents the nonuniform sampling times
tp = [0:Tsam:(K-1)*Tsam] + Tsam*(rand(1,K)-0.5);
xnunds = sin(2*pi*1*tp);        % A simulated nonuniformly sampled data...

% -----------------------------------
A=zeros(K,K);       % initial matrix to be inverted
for p=1:K
    for q=1:K


AI=inv(A);          % Inverse matrix is here

% -----------------------------
xint=zeros(1,K);    % interpolate to uniform samples

for n=1:N:M           


    for m=1:K       % per nonuniform sample loop


        for q=1:K   % calculation of PSI for each m

            PSI = PSI + AI(q,m) * sinc( ((n-1)*Tsam/N - tp(q))/Tsam);


        ACS=ACS+xnunds(m)*PSI;  % acumulation of multiplicant


    xint( (n-1)/N + 1)=ACS;    % result for each n is assigned

hold on
title('nonuniform samples imposed on true samples');

hold on;
title('interpolated uniform samples imposed on true uniform samples)
  • $\begingroup$ I am going to implement on Matlab a lagrange interpolation, there are better algorithm to solve the problem? $\endgroup$ Feb 21, 2017 at 8:51
  • $\begingroup$ Lagrange interpolation is nice, especially for polynomial signals. Simulate the problem and see if it's accuracy is sufficient. Otherwise for the general bandlimited signals there are more specific algorithms such as of Yen's. $\endgroup$
    – Fat32
    Feb 21, 2017 at 9:20
  • $\begingroup$ i tried to implement "interp" and "resampling" function on Matlab. Now i am stucked with another problem :) Thanks for your time <3 link $\endgroup$ Feb 22, 2017 at 22:11
  • $\begingroup$ Ok. I provide you a simple algorithm that converts nonuiform smaples to uniform ones, provided that the original signal was strictly bandlimited. I hope you know what a bandlimied signal is. Take the code as is and try to fit into your application. It requires a matrix inverison of leng NxN and really is on the order of O(N^3) computational cost. So it's not efficient anyweher. However It may improve your FFT computation... $\endgroup$
    – Fat32
    Feb 22, 2017 at 22:32
  • $\begingroup$ I am studying the theory behind, in the specific DFT, DTFT, the sinc function and a litle bit of signal filtering. I am trying to undestrand ACS and PSI, but is not clear :( $\endgroup$ Mar 16, 2017 at 17:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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