3
$\begingroup$

To be correct I want to determine variance of white noise before integration (brown noise is white noise after integration). For example, I have a following model (in MATLAB) of noises acting on my sensor:

noise model

  • How can I determine variance (power) of 'white noise 2' from samples (total noise)?

    Of course noises are band-limited. I have to take into account following noise model of my sensor in my project but I got stuck on this.

\begin{align} \tilde \omega(t)&=\omega(t)+\beta (t)+\eta_v (t)\\ \dot \beta(t)&=\eta_u(t) \end{align}

  • Can I use power spectral density or Allan variance plot for this (how)?
$\endgroup$
1
$\begingroup$

If the white noise has a Gaussian probability distribution, then it should be possible to find the variance of both noise sources using a power spectral density estimate.

$S_{1}(j\omega) = {\sigma_{1}}^{2}$ and $S_{2}(j\omega) = \frac{{\sigma_{2}}^{2}}{\omega^{2}}$, hence you should be able to minimize $$ J(\sigma_{1},\sigma_{2}) = \| S_{1}(j\omega) + S_{2}(j\omega) - \hat{S}(j\omega) \| $$ where $\hat{S}(j\omega)$ is the PSD estimate of your output, in order to find $\hat{\sigma}_{1},\hat{\sigma}_{2}$.

The MATLAB script below shows how this might work in practice.

Ts = 0.1; % sampling time
Nts = 1e6; % number of samples
n = randn(2,Nts); % two Gaussian noise sequences
t = (0:length(n)-1)*Ts; % time vector

sd = 1e3.*rand(1,2); % random standard deviations for noise signals

% filter for the noise
G = sd.*[tf(1,[1 0]), tf(1)];
% zero-order hold
Gd = c2d(G,Ts,'zoh');
% time-response for G(s) by the two Gaussian white noise sources
y = lsim(Gd,n,t);
y = y(1e2:end); % remove initial transient (only want stationary response)

% compute a PSD (of stationary response)
[Pxx,f] = pwelch(y,blackmanharris(length(y)/100),[],[],1/Ts,'psd','onesided');

% scaling using logarithm provides better numerical stability when
% minimizing the function
Jfun = @(p)(log(p(1)*1./(2*pi*f(2:end)).^2 + p(2)) - log(Pxx(2:end)));

p0 = 1e3.*rand(1,2); % random initial values for curve fit
p = lsqnonlin(Jfun,p0); % minimize the cost-function Jfun

sdh = sqrt(p/(2*Ts)); % estimate of standard deviations
est_error = norm(sd - sdh);

disp('Ideal parameters')
disp(sd)
disp('Estimated parameters')
disp(sdh)
disp('Estimation error')
disp(est_error)

fa = 10*log10(p(1)*1./(2*pi*f(2:end)).^2 + p(2));
figure(2), semilogx(f,10*log10(Pxx),f(2:end),fa), grid on
xlabel('Frequency'), ylabel('Power'), legend('PSD','Fit')
$\endgroup$

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.