Adaptative high-gain extended Kalman filter and applications
Adaptative high-gain extended Kalman filter and applications
Adaptative high-gain extended Kalman filter and applications
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
tel-00559107, version 1 - 24 Jan 2011<br />
− the machine is fed 54V for 30 seconds<br />
− at t = 30 the input voltage is switched to 42V for another 30 seconds<br />
4.3 real-time Implementation<br />
− the voltage is then raised to 66V <strong>and</strong> finally shut down after 30 seconds.<br />
At each step, for a period of about 10 seconds, an undetermined frictional force is applied<br />
to the shaft of the motor (those perturbations are applied by h<strong>and</strong>, which explains why we<br />
operate at such low speeds).<br />
Figure 4.27 shows the estimations provided by the Luenberger observer in an open loop 26 .<br />
The <strong>high</strong>-<strong>gain</strong> parameter was set to θ =2.5 <strong>and</strong> the application was run with a 0.001 seconds<br />
time sampling.<br />
Because of the additional computations needed to dynamically solve the Ricatti equation,<br />
the <strong>high</strong>-<strong>gain</strong> <strong>Kalman</strong> <strong>filter</strong> is often seen as a very slow observer. Compared to a Luenberger<br />
<strong>filter</strong> this is indeed the case. Thus, we ran the real-time task at time samplings of 0.01<br />
seconds. On the positive side, this observer is more efficient than a Luenberger when dealing<br />
with systems with a matrix A that depends on the input variable u(t).<br />
The results presented in Figure 4.28 show the estimation of the rotational speed computed<br />
by both an <strong>extended</strong> <strong>Kalman</strong> <strong>filter</strong> (θ = 1) <strong>and</strong> its <strong>high</strong>-<strong>gain</strong> counterpart (θ =2.5). Since<br />
our perturbations are done by h<strong>and</strong> (<strong>and</strong> are thus not reproducible), the only way we can<br />
display such information is to run the two observers in parallel. We therefore feel that it<br />
is possible to tune the code of the <strong>extended</strong> <strong>Kalman</strong> <strong>filter</strong> to make it run efficiently with a<br />
sample time of 0.001 seconds. As expected, the non <strong>high</strong>-<strong>gain</strong> <strong>filter</strong> reacts more slowly to the<br />
applied perturbations. As the measurement noise is not large, it is difficult to distinguish the<br />
(bad) influence of a large value of θ on the estimation. This influence can still be noticed in<br />
the time windows [5; 10] <strong>and</strong> [25; 30].<br />
The adaptive-<strong>gain</strong> <strong>extended</strong> <strong>Kalman</strong> <strong>filter</strong> is more dem<strong>and</strong>ing in terms of computational<br />
time even though this observer runs at the sample time 0.01 seconds. Table 4.2 shows the<br />
values selected for this experiment.<br />
The results of this last experiment are displayed in Figure 4.29. Because this run was<br />
done with no other observer in parallel, a comparison is not easy to make. Still, we can see<br />
that, when dealing with perturbations, the observer has a speed of convergence comparable<br />
to that of the <strong>high</strong>-<strong>gain</strong> <strong>extended</strong> <strong>Kalman</strong> <strong>filter</strong> <strong>and</strong> responds in a nature that is as smooth<br />
as the one of the <strong>extended</strong> <strong>Kalman</strong> <strong>filter</strong>.<br />
If we take a look at Figure 4.30 we see that the parameter θ reacts 9 times during<br />
the experiment: once for every modification plus one extra time corresponding to a bad<br />
initialisation. Those reactions correspond to:<br />
− measured changes of the input voltage,<br />
− non-measured changes in the load torque (i.e. perturbations).<br />
It seems regular that changes occur in the second situation but not in first. In fact, this is<br />
due to modeling errors which means that the innovation may not vanish. This problem is<br />
solved by <strong>filter</strong>ing the innovation 27 in the following way:<br />
26 During perturbations estimation is less precise but stays in a range of 10 to 15% of the real value.<br />
Remember that we had no sensor to measure the torque precisely when the model was calibrated.<br />
27 The convergence result can be proven with such a <strong>filter</strong>ing process. Once the time d ∗ of Theorem 36 has<br />
been set, we have to design the <strong>filter</strong>ing procedure in such a way that I doesn’t vanish in a time less than d ∗ .<br />
88