I'm a noob on this subject, so please be extra clear :)

With a system of equations: $$\dot\omega_x = \alpha_s\omega_y - \epsilon\omega_s\omega_y + \epsilon\omega_s\eta + Q_x$$ $$\dot\omega_y = -\frac{\alpha_s}{1-\epsilon}\omega_x -\frac{\epsilon\gamma}{1-\epsilon}\omega_y +\frac{\epsilon\gamma}{1-\epsilon}\eta -\frac{\epsilon\beta}{1-\epsilon}e_m +\frac{Q_y}{1-\epsilon}$$ $$\dot\eta = \gamma\omega_y - \gamma\eta +\beta e_m$$ $$y(t)=\omega_x + n$$ where $Q_x$ and $Q_y$ are plant momentum disturbances, with $$\phi_{Q_x}(w) = \phi_{Q_y}(w)= \frac{0.05}{(\frac{w}{10^5})+1}$$ and n is output noise $$\phi_n(w) = 0.625$$ I have been asked to design a discrete time kalman filter, and given the guidance in form of general steps. It is as the first of these steps I fail. $$$$

"Start to convert all the disturbances into white noise. You need to perform Spectral factorization. And define the noise with intensities $R_1$, $R_2$ and $R_{12}$" $$$$It has previously been seen that $$R=[\matrix{R_1& R_{12}\\ R^T_{12}& R_2}].$$

I have gone over the litterature and can't seem to understand what I should do. What I have tried is:

Spectral factorization

We wish to find $G_v(w)$ such that $\phi_{Q_x}(w) = |G_v(jw)|^2$, which I have found to be $G_v(jw)=\frac{\sqrt{5}\cdot10^4}{10^5+jw}$. From what I have read we would have noise $v=[v_1 v_2]^T$ with intensity $R$.

Theory shows it like this:

A process is given by $$\dot{x} = Ax+Bu+Nv_1$$ $$y=Cx + Du+ v_2$$ where $v = [\matrix{v_1 \\v_2}]$ is white noise with $R_0 = [\matrix{R_1& R_{12}\\ R^T_{12}& R_2}]$

I simply see the path to $R_1$, $R_2$, and $R_{12}$.

It is related to a course on Space Craft control.


Additional attempt. From further diving I have found that $$R=\frac{1}{2\pi} \int^\infty_{-\infty} |G(j\omega)|^2 d\omega$$ $$|G(j\omega)|^2 = \phi_{Q}(w)$$ $$R=\frac{1}{2\pi} \int^\infty_{-\infty} \frac{0.05}{(\frac{w}{10^5})+1} d\omega$$ $$R=\{integrating\}=2500$$ If we define $R_{12}$ as $0$, then $R=R_1 + R_2$, which is allows us to select $R_1$ and $R_2$ as desired, for example $R_1 = R_2 = \frac{1}{2}R=1250$. However I do not see how you should include $n$ in this - or how to construct the approporate matrix for dealing with $Q_x$, $Q_y$ and $n$.


1 Answer 1


I have figured out what the correct solution is for this problem. First I had missunderstood what $R_1$ was. $R_1$ should be the the $R$ the the equation below, and their purpuse.

In this scenario we have $\phi(w) = G_v(jw) R G_v^*(jw)$ where $\phi$, $R$ and $G_v$ are vectors / matricies.

If we say that $\phi(w) = \left[\matrix{\phi_x &0\\ 0& \phi_y} \right]$, $R=$$\left[\matrix{A& B\\ C& D}\right] = R_1$, $G_v = \left[\matrix{G_v(jw)& 0\\ 0& G_v(jw)}\right]$, $G_v^* = \left[\matrix{G_v^*(jw)& 0\\ 0& G_v^*(jw)}\right]$

(Re-call) $\phi_x = \phi_y = \phi$. We find that $$ \phi(w) = \left[\matrix{\phi &0\\ 0& \phi} \right] = \left[\matrix{A\cdot G_v \cdot G_v^*& B\cdot G_v \cdot G_v^*\\ C\cdot G_v \cdot G_v^*& D\cdot G_v \cdot G_v^*}\right] = \left[\matrix{A \phi& B \phi\\ C \phi& D \phi}\right],$$ this reveals that $A = D = 1$, $B = C = 0$, thous $R_1 = \left[\matrix{1& 0\\ 0& 1}\right]$. As for $R_2$ is should simply be $\phi_n = 0.625$, as it is a constant and one dimensional.

The noise can be modeld as white noise passing through a system, in this case described by $G_v$.

White noise $\cdot$ $G_v$ --> noise.

Converting the Transfer function TF to state space form we get: $$A_w = \left[\matrix{-10^5& 0\\ 0& -10^5}\right]$$ $$B_w = \left[\matrix{1& 0\\ 0& 1}\right]$$ $$C_w = \left[\matrix{\sqrt{0.05}& 0\\ 0& \sqrt{0.05}}\right].$$ Which can be used to represent the noise. The noise is thus modeled by construction the system / TF that shapes a white noise input to our noise.

$R_1$,$R_2$, and $R_{12}$ are later in this project used to check assumptions for controllability and observability of the system.


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .