The file Demo_014.m is found in IQClab’s folder demos; it performs a robust estimator synthesis with the function fRobest for a mass-spring-damper problem derived from the example presented in [24]. The setup is shown in the following figure.

The numerical values are ,
,
,
and
with
. The first mass is disturbed by an external force
. The objective is to estimate
using a measument of
corrupted by measurement noise
. The linear fractional representation can then be written as
with ,
.
For comparison reasons, we design a nominal estimator,
, using the function ‘h2syn’, which is part of MATLAB’s robust control toolbox, as well as a robust
estimator,
, using the function ‘fRobest’, which is part of IQClab.
The code for designing the nominal estimator is found in the file Demo_014.m. Designing the robust estimator proceeds as follows:
% Define plant (see script for system matrices) H = ss(A,[Bp,Bw],[Cq;Cz;Cy],[Dqp,Dqw;Dzp,Dzw;Dyp,Dyw]); % Define uncertainties de1 = iqcdelta('de1','InputChannel',1,'OutputChannel',1, 'Bounds',[-1,1]); de1 = iqcassign(de1,'ultis','Length',3); de2 = iqcdelta('de2','InputChannel',2,'OutputChannel',2, 'Bounds',[-1,1]); de2 = iqcassign(de2,'ultis','Length',3); % set options options.perf = 'H2'; options.StrProp = 'yes'; options.subopt = 1.03; options.constants = 1e-8*ones(1,4); options.Pi11pos = 1e-8; options.FeasbRad = 1e9; [Erob,gamrob] = fRobest(H,{de1,de2},[2,1,1],[2,2],options);
We obtain a nominal and robust estimator, and
respectively, which guarantee a nominal and robust
performance level of
and
respectively. The corresponding Bode-magnitude plots are depicted in the following figure.

Given and
we can also perform a robustness analysis using the function ‘iqcanalysis’. This yields the following computed worst-case
perfomormance levels for increasing levels of
with
,
.

As can be seen, the nominal estimator, , shows better performance levels for smaller values of
, while for
the robust estimator,
, outperforms
.
To continue, let us also show some time-domain simulations, where we perturb the system with a random force disturbance and for different values of . The results are shown in the following two figures for
and
respectively.


As can be seen, the robust estimator does a better job, if compared to
. This is confirmed by the computed RMS values which are given in the following table.
![]() | ![]() | ![]() | ![]() | ![]() | ![]() | Mean |
RMS value ![]() | 0.059 | 0.069 | 0.074 | 0.111 | 0.068 | 0.076 |
RMS value ![]() | 0.022 | 0.026 | 0.028 | 0.043 | 0.026 | 0.029 |