You are on page 1of 3

2004).

In these equations there are two groups: a time update group (sometimes called the pre-
diction) and a measurement update group (sometimes called the correction), as detailed in the
following example.

Example 12.13
In its simplest form, without all the state matrices the standard KF approach uses the following
equation,

xk = Kk zk + 1 − Kk xk − 1 12 22

where x is the current model estimate of the parameter of interest, Kk is the Kalman gain, zk is the
measured value of the parameter of interest and xk − 1 is the previous estimate of the parameter of
interest. Equation 12.22 shows that the value x is a weighted sum of the measurement value, zk,
and the previous estimated value, and the Kalman gain is used to determine the weights in this
linear combination. Hence, the Kalman gain is an important quantity. If it is large, more weight
is placed on the measurement, zk, and if it is small the most weight is on the previous estimate,
xk − 1 . Equation 12.22 is recursive, meaning that the values of the process at time step k are
dependent on the values from the previous time step k−1.
In a standard KF approach the following equation shows that the signal value of interest at a
time step k, xk, is the linear combination of it previous value at time step k-1, xk-1, a control
signal uk, and a process noise, wk−1.

xk = Axk −1 + Buk + wk −1 12 23

The second equation, below, indicates that any measurement value, zk, is a linear combin-
ation of the signal value, xk, and the measurement noise, vk.

zk = Hxk + vk 12 24

The solution process proceeds as follows. There are two groups of equations: a Time Update
(prediction) and a Measurement Update (correction). The Time Update equations are:

xk = Axk − 1 + Buk 12 25a


Pk = APk − 1 A T + Q 12 25b

where Pk is the previous error covariance (the error is the difference between the predicted
model value at two subsequent time steps). Pk is a necessary parameter because it is used in
determining the Kalman gain, Kk, as will be shown. Q is a “waste basket” for unknown mod-
eling errors, and will not be needed in the example here.
The Measurement Update equations are:

−1
Kk = Pk H T HPk H T + R 12 26a
xk = xk + Kk zk − Hxk 12 26b

Pk = 1 − Kk H Pk 12 26c
where R is a parameter to model the standard deviation in the experimental noise (the noise in
the measurement, z). Equation (12.26b) is found by replacing the measurement noise, vk, by the
equality in Equation (12.24).
The process noise and the measurement noise are presumed to be statistically independent.
The matrices A, B, and H are used to model multidimensional problems. Because this example
is a simple one-dimensional problem being used for prediction (not for control) the following
simplifications are realized.

A = B = H = 1, Q = 0 and uk = 0, wk − 1 = 0 12 27

The Time Update (prediction) equations are simplified based on the assumptions of identities
in Equation (12.27) to the following. Equations (12.25a) and (12.25b) become

xk = xk − 1 12 28a

Pk = Pk −1 12 28b

The Measurement Update (correction) Equations (12.26) are simplified, as shown as fol-
lows, on the assumptions of identities in Equation (12.27); the expression inside the paren-
theses of Equation (12.26b) is the value of vk in Equation (12.24).

Pk
Kk = 12 29a
Pk + R
xk = xk + Kk zk − xk 12 29b

Pk = 1 − Kk Pk 12 29c

For the voltage example here, the value of R will be given a constant value of 0.1 volts. The
uncertainty in R will be assumed to be Gaussian. The calculations will start with the initial con-
ditions of P0 = 1 and x0 = 0 at time step k = 1. The measurement data, from a voltmeter for the
time steps (k = 1, 2, …,10) in milliseconds (ms) and the values of the voltmeter readings, in
volts (V), are shown in Table 12.2.
Equations (12.28) then become, as shown in the fifth column of Table 12.3 (Time Update),

xk = xk − 1 = 0

Pk = Pk − 1 = 1

and the Measurement Update values are shown in the sixth column of Table 12.3

Table 12.2 Experimental readings for example problem. Adapted from Esme, 2009.

TIME (ms) 1 2 3 4 5 6 7 8 9 10
VALUE (V) 0.39 0.50 0.48 0.29 0.25 0.32 0.34 0.48 0.41 0.45
Table 12.3 One-parameter KF simulation for a voltage reading. Adapted from Esme, 2009.

k Zk xk − 1 Pk Time Update Measurement Update xk Pk


1 0.39 0 1 xk = xk −1 = 0 Kk = 1 1 + 0 1 = 0 909 0.355 0.091
Pk = Pk − 1 = 1 xk = 0 + 0 909 0 390 − 0 = 0 355
Pk = 1− 909 1 = 0 091
2 0.50 0.355 0.091 xk = 0 355 Kk = 091 091 + 0 1 = 0 476 0.424 0.048
Pk = 091 xk = 0 355 + 0 476 0 500 − 0 355
= 0 424
Pk = 1 − 476 0 091 = 0 048
3 0.48 0.424 0.048 0.442 0.032
4 0.29 0.442 0.032 0.405 0.024
5 0.25 0.405 0.024 0.375 0.02
6 0.32 0.375 0.020 0.366 0.016
7 0.34 0.366 0.016 0.362 0.014
8 0.48 0.362 0.014 0.377 0.012
9 0.41 0.377 0.012 0.38 0.011
10 0.45 0.38 0.011 0.387 0.010

Kk = 1 1 + 0 1 = 0 909
xk = 0 + 0 909 0 390− 0 = 0 355
Pk = 1 − 909 1 = 0 091

These values are also placed in the third and fourth columns for time step k = 2. The meas-
urement update values for k = 2 are given in the calculations that follow,

Kk = 091 091 + 0 1 = 0 476


xk = 0 355 + 0 476 0 500− 0 355 = 0 424
Pk = 1 − 476 0 091 = 0 048

For the remaining time steps (k = 3 to 10) the values are shown in Table 12.3.
The predicted voltages of this simple 10-step simulation are shown in the seventh column of
Table 12.3 and also in Figure 12.21. In Figure 12.21 the convergence of the voltage to about
0.38 volts seems obvious. Now, we will see how this same fusion method would work using a
fuzzy KF approach.
The expression zk − Hxk in Equation 12.26b is sometimes referred to as the residual (Raol,
2010), and it is multiplied by the Kalman gain, Kk. The expression HPk H T + R in
Equation 12.26a is referred to as the covariance of the residuals. Both of these quantities
are expressed in Example 12.13 in their simpler forms of zk − xk in Equation (12.29b) and
Pk + R in Equation 12.29a, because H = 1 for the example. The residuals are an important quan-
tity in the KF process because when the residuals gradually decrease the solution of the KF
approaches the converged, or true, value of the state variable. The expression zk − xk is also
equivalent to the measurement noise (see Equation 12.24) and when this noise reaches a

You might also like