Linear Prediction Model
Given the knowledge of the measured variable
yn-1, ..., yn-m,
the prediction of yn is denoted
Yn , and the prediction error is
the difference between the newly measured value and its prediction,
an = fm(n) is called the innovation. We assume that the prediction is a linear function of the measured values (linear prediction), with coefficients cn,k. The innovation can therefore be written
The coefficients of the linear prediction are obtained by minimizing the correlation of the innovation with the measures,
If the process is stationary the prediction coefficicents
cn,k do not depend on the index n, but only on
the delay index k. In this case they are computed by solving
the linear equation
The estimation of the parameter variable x is assumed a linear function of the measures, i.e., of the innovations,
bk is found by minimizing the distance between xn and Xn. This is a mean square value problem. In order to find the minimum of
the derivatives of F with respect to the bk are computed and equated to zero. This gives the equations
Since E[akaj]=0 if k is different from j, this system is immediately solved
and
which is suitable for a recursive implementation since Xn(y1 ... yn) is expressed as Xn-1(y1 ... yn-1) and a coefficient, bn , that depends on xn and an, i.e., on xn and yn, and is multiplied by an.
Suppose to have a model described by the (linear) evolution equation
The measured variable is
The covariance of the system state is Pn+1,n = E[ xn+1 xn ].
The a-priori estimate of the system state (ie, the estimate at time step
n+1 given the knowledge up to time n) is
The estimate errors are
To derive these results we begin by writing the a-posteriori estimate
as a linear combination of the a-priori estimate and the difference
between the actual measurement and the predicted measurement,
To find it we substitute this equation in the definition of the covariance, the derivative with respect to K is taken and equated to zero. THe result is the expression of the Kalman gain matrix written above.
When the process is not linear the Kalnam filter theory can be applied to the linearized system (about te current mean and covariance). This is called Extended Kalman Filter.
Marco Corvi - Page hosted by geocities.com.