Anda di halaman 1dari 6

17th IEEE International Conference on Control Applications ThB04.

Part of 2008 IEEE Multi-conference on Systems and Control
San Antonio, Texas, USA, September 3-5, 2008

Robust stability analysis with Integral Quadratic Constraints (IQCs):

A design example.
Matthias Löhning and Johannes Ruder and Johann Bals

Abstract— This paper gives an overview of the applicability fˆ v̂

of the stability analysis based on integral quadratic constraints ∆
(IQCs) using the IQCβ toolbox in MATLAB. One main ad-
vantage of the IQC theory is the possibility to deal with
different kinds of perturbations - like linear time invariant
(LTI) and linear time variant (LTV) parametric uncertainties, ŵ ê
nonlinearities and unmodelled dynamics - in one stability
analysis. Before using the IQCβ toolbox the perturbated system
has to be split in a nominal LTI part and the remaining parts. Fig. 1. Basic feedback configuration.
This step can be performed with the LFR toolbox in MATLAB,
especially for large systems. In this paper a new procedure is
presented how to import objects of the LFR toolbox into the
IQCβ toolbox in a general sense. As design example a controlled
two-mass-spring system with friction is taken, which serves as Parametric uncertainties are considered, since model param-
the model for many mechanical components in real life. The eters are often known only approximately. By presuming the
main result of this paper is the stability analysis of the presented controller parameters as uncertain a continuous parameter
model with respect to uncertain LTI and slowly LTV parameters space can be given, in which the closed loop system remains
and nonlinearities. The stability region of the IQC analysis is
compared to the one expected by extensive simulations and
stable. This will be of interest, if the performance of a real
worst case optimizations. life application shall be improved by changing the controller
parameters without getting unstable. Also the case of LTV
I. I NTRODUCTION uncertain parameters and nonlinearities are examined. This
combination of perturbations appears for example in robotic
IQCs provide a unifying framework for stability analysis
applications, where the power train is modelled elastically
in modern robust control. They represent a tool for studying
with highly nonlinear friction. The time-variance is due to
the stability of feedback connections of stable LTI systems
different mass inertia in different tool center point positions
and perturbation blocks as shown in Fig. 1. For many
(see Fig. 2). Since the IQC theory only offers sufficient
different perturbation blocks, IQCs are already available in
stability conditions, the conservatism of the IQC analysis
the literature (for example see [1]) and there exists a MAT-
is studied, that is how much smaller the IQC stability region
LAB toolbox for the IQC based stability and performance
is compared to the stability region obtained by extensive
analysis called IQCβ (see [2]). For the different kinds of
perturbations different multipliers, e.g. constant or frequency
dependent, are necessary to get non-conservative results [3]. The paper starts with a brief introduction of IQCs. Section
Within the IQCβ toolbox the different multipliers are chosen III presents how the LFR toolbox objects can be imported
automatically, such that the user has not to worry about. in the IQCβ toolbox. In Section IV the two-mass-spring
Instead the user has to choose the type of IQC, which is model and the used controller are described. In section V,
more physically related, e.g. LTI uncertain parameter. the stability of the controlled two-mass-spring system with
Since a lot of mechanical systems like gearboxes or robots various types of uncertainties is analyzed and evaluated. The
can be modelled as two-mass-spring systems, this model is paper is concluded in section VI.
chosen as the basic object of study in this paper. The nominal
two-mass-spring system is stabilized with a PID controller.
Other controllers like a state feedback with observer or
an H∞ output feedback controller were also examined but
not presented, since the results only differed quantitative.

This work was supported by the German Aerospace Center (DLR)

M. Löhning is with Institute of Robotics and Mechatronics,
German Aerospace Center (DLR), 82234 Wessling, Germany
J. Ruder is with Institute of Robotics and Mechatronics,
German Aerospace Center (DLR), 82234 Wessling, Germany
J. Bals is with Institute of Robotics and Mechatronics, German Aerospace
Center (DLR), 82234 Wessling, Germany Fig. 2. Time variant mass inertia due to different roboter positions.

978-1-4244-2223-4/08/$25.00 ©2008 IEEE. 816

A. Notation v and w at the frequency ω. In principle, Π(jω) : jR → C
In this paper â denotes the Fourier transform of the signal can be any measurable Hermitian-valued function. In most
a; A∗ is the (complex conjugate) transpose of A; In denotes applications, however, it is sufficient to use rational functions
an identity matrix of size n × n; h ∗ w means convolution that are bounded on the imaginary axis. D : L2 [0, ∞) →
of the two signals h and w, dim(v) means the dimension of D(L2 [0, ∞))1 will be the set-valued functions ŵ ∈ D(v̂),
the vector v. where D(v̂) = {ŵ ∈ L2 [0, ∞)|(2)}.
RH∞ denotes the set of stable real-rational transfer func- The main theorem from [1] is stated as follows
tions; Ln2 [0, ∞) denotes the Hilbert space of measurable Theorem 2.1: Let G(jω) ∈ RH∞ and let ∆ be a bounded
functions f : [0, ∞) → Rn satisfying causal operator. Assume that:
Z ∞ 1) for every τ ∈ [0, 1], the interconnection of G(jω) and
kf k = |f (t)|2 dt < ∞. τ ∆ is well posed;
0 2) for every τ ∈ [0, 1], the IQC defined by Π(jω) is
This is a subspace of Ln2e [0, ∞),whose members only need satisfied by τ ∆;
to be square integrable on finite intervals [0, T ]. 3) there exists ǫ > 0 such that
The transfer function of a linear time invariant system is  ∗  
denoted as G(jω) G(jω)
Π(jω) ≤ −ǫI, ∀ω ∈ R. (3)
  I I
G(jω) = C(jωI − A) B + D =
. Then the feedback interconnection of G and ∆ is stable.
Note that if the upper left corner, Π11 (jω), of Π(jω) is
The lower linear fractional transformation (LFT) of a
positive semi definite for all ω ∈ R then ∆ = 0 satisfies (2).
If further the lower right corner, Π22 (jω), is negative semi
M11 M12
M= definite for all ω ∈ R, then any convex combination of ∆’s
M21 M22
will satisfy (2) for τ ∈ [0, 1] if and only if ∆ does so. This
and a matrix K with compatible dimension is defined as simplifies the second assumption in theorem 2.1.
Fl (M, K) = M11 + M12 K(I − M22 K)−1 M21 . The search for multipliers Π(jω), can be carried out as a
convex optimization problem by parameterizing
In [1] several IQCs for robustness analysis were proposed. Π(jω) = xi Πi (jω)
The IQCs form a generalizing approach of the three classical
research fields input-output theory, absolute stability theory where xi are positive real parameters and Πi (jω) is a set
and finally the robust control theory. Using IQCs one may of basis multipliers. Usually Πi (jω) and G(jω) are proper
form a stability criterion for the interconnection of a stable rational functions and G(jω) has no poles on the imaginary
system G(jω) ∈ RH∞ and a bounded causal operator ∆, axis, so that the left part of equation (3) can be rewritten as
see Fig. 1.  ∗  
G(jω) G(jω)
Πi (jω)
v̂ = G(jω)ŵ + fˆ I I
ŵ = ∆(v̂) + ê. (1)  ∗  
D + C(jωI − A)−1 B D + C(jωI − A)−1 B
The interconnection of G(jω) and ∆ is called well-posed = Mi .
if the map (v, w) → (e, f ) defined by (1) has a causal
inverse on L2e . The interconnection is stable if, in addition, By applying the Kalman-Yakubovich-Popov lemma (see
the inverse is bounded, that is, if there exists a constant C [4]), the search for xi , can be implemented using linear
such that matrix inequalities (LMIs). Then (3) is equivalent to the
Z T Z T existence of P = P ∗ such that
|v(t)|2 + |w(t)|2 dt ≤ C |f (t)|2 + |e(t)|2 dt
P A + A∗ P P B C D C D
0 0 + M > 0 (4)
B∗P 0 0 I 0 I
for any T ≥ 0 and for any solution of (1).
Depending on the particular application, various types of holds, where
IQCs are available. Two signals w ∈ L2 [0, ∞) and v ∈ 
M11 M12
 2 [0, ∞) are said tosatisfy the IQC defined by Π(jω) = M= = xi Mi .

Π11 (jω) Π12 (jω) i
, if
Π21 (jω) Π22 (jω) In a more general setting, M can also be defined as a
Z ∞
 convex set specified by an LMI. For instance, the constraints
Π(jω) ≥0 (2) Π11 (jω) > 0 and Π22 (jω) < 0 for all ω ∈ R may be added.
−∞ ŵ(jω) ŵ(jω)
See [1] or [5] for examples.
holds, where absolute integrability is assumed. Here v̂(jω)
and ŵ(jω) represent the harmonic spectrum of the signals 1 P(L [0, ∞))
2 denotes the set of all subsets of L2 [0, ∞)

A(δ)x x

For the IQC analysis every uncertainty has to be iso- ⇓

lated in one block, which is described by the associated
IQC. Commonly parameter dependent models are obtained AA,LF R
by physical motivated differential equations. For numerical
reasons the models are normaly implemented in state space, A(δ)x
CA,LF R ∆(δ)
such that the model parameters canappear more than once. +
Ai (δ) Bi (δ)
The general systems Gi (jω, δ) = with DA,LF R
Ci (δ) Di (δ)
uncertain parameters δ = [δ1 . . . δk ] can be represented
in a linear fractional representation (see Fig. 3)2 , which is
commonly in the robust control theory. With this represen-
tation the uncertain parameters are extracted in one block Fig. 4. Conversion of the uncertain matrix A(δ) in a MATLAB LFR-object.
as necessary for the IQC analysis. The transformation can
be done automatically with the LFR toolbox. Since the LFR
toolbox does not provide the in- and outputs of the ∆(δ)- where 0 denotes zero matrices of appropriate dimensions. If
block with the corresponding system matrices, this step has the uncertain parameters occur in all matrices of the original
to be done by hand to use the objects of the LFR toolbox in system the new matrices become
the IQCβ toolbox. The procedure is done in the following
way. Ã = DA,LF R
In MATLAB, an uncertain matrix A(δ) is represented by B̃ = CA,LF R CB,LF R 0 0 DB,LF R
the four matrices AA,LF R , BA,LF R , CA,LF R , DA,LF R and
 
a uncertain ∆(δ)-Block (see Fig. 4). The first index of these  0 
matrices denotes the name of the corresponding uncertain
 
 BC,LF R 
C̃ =  
matrix.  0 
To do the IQC analysis the ∆(δ) block has to be replaced DC,LF R
by the appropriate IQC block. Before that, the remaining  
AA,LF R 0 0 0 0
certain matrices of the original system (for example B, C  0 AB,LF R 0 0 BB,LF R 
and D) are rearranged and combined with the LFR-matrices  
 0
D̃ =  0 AC,LF R 0 0 
in that way, that the input signals of the removed ∆(δ) block  0

0 0 AD,LF R BD,LF R 
become new system outputs and the output signals of ∆(δ)
become new system inputs.
For example, if the uncertain parameters exclusively occur After that, one can use IQC blocks to relate the new in- and
in the system matrix A, the new matrices of the IQC  output signals to each other (see Fig. 5).
w To handle both parametric and non parametric uncertain-
convenient system Ã, B̃, C̃ and D̃ with input ũ =
  u ties the original model has to be adapted, more precisely
v appropriate in- and output signals for the non parametric
and output ỹ = become
y uncertainties have to be included in u and y respectively.
à = DA,LF R , B̃ = CA,LF R B ,
C̃ = , D̃ = , The mechanical two-mass-spring system studied in this
C 0 D
paper is shown in Fig. 6. The exact definition of the used
2 Nonlinear systems can be included in this representation via quasi LPV
parameters is given in table I. The engine- and output-torque
models [6]

w v w v
∆(δ) Ddim(v) (v)

Mi (jω) Mi (jω)
u y u y

Gi (jω, δ) Gi (jω, δ)

Fig. 3. The linear fractional representation. Fig. 5. The setup for IQC analysis.

GP ID (jω, δ) by choosing the uncertainty δ as one or more
of the above described parameters. The system has one single
Abbr. Nominal value Unit Parameter
input u(t) = ϕe,ref erence and one single output y = ϕe .
kg m2
Je 43 [ rad
] engine sided mass inertia
Jo 52 [ rad
] output sided mass inertia
cg 20000 [ rad
] stiffness of the gearbox In the following, different types of uncertainties are added
dg 320 [ ] damping of the gearbox
Nms in the above nominal system. The stability of the resulting
ce 350 [ rad
] engine sided viscous friction coef.
co 20 [ Nms
] output sided viscous friction coef.
uncertain closed loop systems are analyzed using the IQC
framework and the results are compared with system analysis
and simulations in MATLAB/Simulink and Dymola (Dynasim
AB). In the following by conservatism we mean how much
are denoted as τe and τo respectively where as ϕe and ϕo smaller the stability region of the IQC analysis is compared
are representing the angles of engine and output. to the expected region of the system. Since the conditions in
Using the law of conservation of angular momentum the theorem 2.1 are only sufficient, the stability region is defined
following system equations can be obtained by the resolvability of (4). Outside this region no statement
can be made.
Je · ϕ̈e = cg · (ϕo − ϕe ) + dg · (ϕ̇o − ϕ̇e ) − ce · ϕ̇e + τe
Jo · ϕ̈o = −cg · (ϕo − ϕe ) − dg · (ϕ̇o − ϕ̇e ) − co · ϕ̇o + τo . A. Single constant uncertain parameter

Choosing the four states x1 = (ϕo − ϕe ), x2 = (ϕ̇o − ϕ̇e ), In the first analysis, stability of the system GP ID (jω, δ)
x3 = ϕo , x4 = ϕ̇o , the input u = τe , the output y = ϕe due to one uncertain but constant parameter is proved. For
and τo = 0 yields the nominal state space model of the two- this purpose Jo is chosen as the uncertain parameter, which
mass-spring system can take its values from the interval [1, 365]. To illustrate
this, the bode plot of GP ID (jω, Jo ) is shown in Fig. 7.
0 1 0 0
 
To do the analysis, the IQC for constant uncertain param-
− Jcg − Jcg − dgJ+ce − Jdg 0 Jce − Jco  eters is chosen as
A= e o e o e o
 0 0 0 1   
c d X(jω) Y (jω)
− Jgo − Jgo 0 − Jcoo Π(jω) =
 Y (jω)∗ −X(jω)
− 1    where X(jω) = X(jω)∗ ≥ 0 and Y (jω) = −Y (jω)∗ are
B= Je  C = −1 0 1 0 ,
 0 , D=0 bounded and measurable matrix functions (see [1]).
0 After that, the interval of Jo for which stability can be
proved is maximized (before normalization). The minimum
A standard PID controller 2
  value Jo,l = 1 kg m / rad is chosen as the lower bound while
1 the upper bound Jo,u of the parameter interval is maximized,
KP P I (jω) = kpi 1 + (kp + s)
Ti s denoted with Jo,maxIQC . This was done using the bisection
with the nominal parameters Ti = 0.1, kp = 75 and method in combination with the criterion whether stability
kpi = 11100 is used. The analysis below are also done with of the interval could be proven or not.
an observer based state feedback and H∞ controller and
differ only the resulting L2 gain. Due to the limited space
these results cannot be presented. With negative feedback Bode Diagram

the closed loop nominal system is represented by the trans- 40

fer function GP ID (jω). One obtains the uncertain system 20

Magnitude (dB)

Je Jo
dg , cg 0
τe τo
Phase (deg)




0 1 2 3 4
10 10 10 10 10

ϕ̈e ϕ̈o Frequency (rad/sec)

Fig. 7. Bode plot of uncertain system GP ID (jω, δ) with normalized

Fig. 6. The two-mass-spring system Jo ∈ [−1, 1].

To verify the results of the IQC analysis the maximum was introduced to consider nonlinear friction as described
parameter value of Jo is increased further while their eigen- below. As the sector IQC of the IQCβ toolbox is used, the
values are checked. The last value of Jo , where the associated corresponding multiplier is described in [2]. Ti is varied
system possesses exclusively eigenvalues with negative real inside the interval [0.001, 0.011] while the appropriate kp
part, is denoted as Jo,maxEV . With this value, one obtains a interval is maximized. By this, a kind of stability map for
relative error of the IQC analysis of the two controller parameters is created (see Fig. 9). This
Jo,maxIQC − Jo,maxEV
stability map shows controller parameter pairs (Ti , kp ) for
= 0.003. which stability is guaranteed. This parameter space can be
taken to improve performance on real applications.
The same analysis is done with other single uncertain pa- The solid line obtained by the IQC analysis represents
rameters cg and dg . With the IQC analysis it is possible to the stability border, e.g. all pairs on or below this line (in a
prove stability due to these uncertain parameters for huge continuous manner) will result in a robustly stable system.
uncertainty intervals with an upper bound far beyond the For pairs above the line no statement can be made by the IQC
nominal value. Checking the eigenvalues of these systems, theory. This stability map provides assistance in choosing the
it turned out that even above the received bounds the system controller parameters to get a stable system.
remained stable. The IQC analysis with the above model is To verify this result the linear engine sided friction is
conservative due to numerical problems by solving the LMI replaced by a nonlinear friction ϕ̇e → τo
(4) with the default LMI solver of the MATLAB Robust 104
Control Toolbox, but only when the parameters are getting τo (t) = 350 · ϕ̇e (t) + ϕ̇e (t)
− 52,
1+e −280·
very huge.
within the sector constraints 350 [ Nradms
] and 160000
B. Multiple constant uncertain parameters [ rad ]. After that, the closed loop system is simulated for
several parameter pairs (Ti , kp ) with MATLAB and Dymola.
Motivated by the last results stability of the above systems
The results of these simulations are represented in Fig. 9 by
due to more than one uncertainty is studied. First, two
crosses and circles. The crosses denote that the appropriate
uncertain parameters Jo and cg were considered before
simulation discovered a stable system whereas the circles
finally all four parameters Jo , cg , dg and co were stated
stand for an unstable closed loop system.
as uncertain. The appropriate setup of the IQC analysis is
It can be seen that all simulations with parameter pairs
shown in Fig. 8. The potentially vector signals v1 till vn
below the solid line show stability of the related systems.
result from the procedure of III and are no more physically
This verifies the stability map. With Ti ≤ 0.003 s the
interpretable as the parameter itself.
simulations with kp below the solid line discovered stable
Using this setup and IQCβ stability for the nominal system
systems, where as all simulations above showed unstable
GP ID (jω, δ) and each combination of uncertain parameters
systems. That is, for Ti ≤ 0.003 s the IQC analysis meets the
is proved. With this kind of uncertainty the stability can be
real stability border with good accuracy. For Ti > 0.003 s
proven with the mu analysis [7], too. A comparison of the
one can see a big difference between the two stability
results with the IQC and the mu theory is not possible due
to different stability quantities (L2 gain and µ value).
D. Slowly time varying parameter
C. Controller robustness with nonlinearity
In the next analysis, Jo is stated as slowly time variant,
As a special case of the analysis with uncertain parameters, that is
the two PID-controller parameters Ti and kp were chosen as w(t) = ∆(t)v(t) (5)
constant uncertain parameters and a sector constraint on τo

w2 v2 IQC stab. bound

Ddim(vn ) (vn ) 6000 stable simulation
.. instable simulation
w1 v1

Ddim(v1 ) (v1 )

Mi (jω)
u y
Gi (jω, δ) 1 2 3 4 5 6 7 8 9 10 11
Ti [s] ×10−3

Fig. 8. Setup for a structured uncertainty. Fig. 9. Stability map for the PID controller.

where ∆(t) is a diagonal matrix of size nd with δ(t on the which has to be minimized. Each population consists of 30
main diagonal, where δ(t) = Jo is a scalar function which individuals. Tournament selection is used as selection scheme
satisfies |δ(t)| ≤ D and |δ̇(t)| ≤ d ∀ t and nd results from with a shuffling technique for choosing random pairs for
the procedure of III. mating. After 200 generations, the optimal design parameter
The following IQC formulation for slowly time varying 
aopt = 4.5 40 510
parameters is taken from [2]
Z ∞
vext (t)
T  2
D K1 M1

vext (t)
 obtained the minor improvement of the upper bound
wext (t) M1T −K1 wext (t)
dt ≥ 0 (6) |J˙o |max = 17.8 N m s. Choosing different optimization al-
Z ∞ T  2   gorithms, optimization parameters or an increased order of
y(t) d K2 M 2 y(t) the design parameter a hardly improved this upper bound of
dt ≥ 0 (7)
0 u(t) M2T −K1 u(t) |J˙o |. In [1] it is proven that under certain conditions quadratic
with positive symmetric K1 , K2 and skew symmetric M1 , stability and IQC stability is equivalent. Since the multipliers
M2 matrices. vext , wext , y and u are defined by are constant for pure slowly time variant parameters and
    the uncertainty block has diagonal structure these conditions
y(t) z(t) + x(t)
vext (t) = wext (t) = are satisfied. As well-known the quadratic stability analysis
v(t) w(t)
of slowly linear parameter varying (LPV) systems will be
ẏ(t) = Ay(t) + Bv(t) y(0) = 0 (8) conservative, if a single quadratic Lyapunov functions is
ż(t) = Az(t) + Bw(t) z(0) = 0 (9) taken4 (see [9] and [3]).
ẋ(t) = Ax(t) + u(t) x(0) = 0 (10) VI. CONCLUSIONS
u(t) = δ̇(t)y(t) A comprehensive case study of the IQC stability analysis
   
−a1 1 ··· 1 of a two-mass-spring system model with a PID controller and
A= .. B= .. diverse types of uncertainties and combinations of them was
   
.  . presented. To easily get IQC convenient systems a LFR based
−an 1 ··· 1 procedure was introduced. In conjunction with the chosen
(11) model the IQC stability analysis with constant parametric
where the size and values if the parameter a = [a1 . . . an ]T uncertainties obtained very low conservative results, where
has to be chosen. This IQC results from differentiation of (5) the analysis with time varying parameters seems to possess
in combination that 1s 3 is estimated by a transfer function more conservatism, in spite of the appropriate IQC design
with the poles a1 till an . parameter a was optimized and an extensive worst case op-
An adequate Jo interval is chosen while the upper bound timization was performed. As there is a relation to quadratic
of |J˙o | is maximized to |J˙o |max with bisection optimization. stability a plausible reason for the conservatism is given.
Unfortunately, it was not possible to prove stability for any
upper bound |J˙o | > 0 with the default parameter values of
the appropriate IQC from IQCβ. [1] A. Megretski and A. Rantzer, “Sytem analysis via integral quadratic
constraints,” IEEE Transactions on Automatic Control, vol. 42, no. 6,
With an manual adapted IQC design parameter a, which pp. 819–830, 1997.
sets up the IQC multiplier poles, finally stability of the [2] U. Jönsson, C. Kao, A. Megretski, and A. Rantzer, A Guide to IQCβ:
feedback loop could be proved with an upper bound of A MATLAB Toolbox for Robust Stability and Performance Analysis,
August 2004.
|J˙o |max = 8.6 N m s. By simulation with different Jo (t), [3] A. Helmersson, “Methods for robust gain scheduling,” Ph.D. disser-
the system only gets unstable for |J˙o | > 7003 N m s. This tation, Department of Electrical Engineering, Linköping University,
bound can be decreased to about 1200 N m s by worst case Schweden, 1995.
[4] J. Willems, “The least squares stationary optimal control and the
optimization with MOPS (see [8]) by taking a sinus and algebraic riccati equation,” IEEE Transactions on Automatic control,
triangular signal with the optimization parameters frequency, vol. 16(6), pp. 621–634, 1971.
amplitude and phase. The stability is analysed by an increas- [5] A. Helmersson, “An iqc-based stability criterion for systems with slowly
varying parameters,” Linköping University, Departement of Electrical
ing of the local maxima of two periods of the time signals Engineering, Tech. Rep., 1999.
after a long enough simulation time. Nearly the same bound [6] J.-F. Magni, User Manual of the Linear Fractional Representation
is achieved by taking twice of the resonance frequency of the Toolbox, System Control and Flight Dynamics Department, October
nominal closed loop system. Thus the real stability boundary [7] G. Balas, Robust Control Toolbox 3 (Usert’s Guide). The MathWorks,
is expected about this value. 2007.
Due to that, the IQC design parameter a is optimized. a [8] H. Joos, J. Bals, G. Looye, K. Schnepper, and A. Varga,
“A multi-objective optimisation-based software environment
is stated to be a vector of dimension three and a genetic for control systems design,” 2002. [Online]. Available:
algorithm of MOPS is used to find optimal values ai . The
quality function is chosen to [9] F. Wu, “Control of linear parameter varying systems,” Ph.D.
dissertation, University of California at Berkeley, 1995. [Online].
1 Available:
Ψ(a) = ,
|Jo |max (a)
3s is the Laplace variable. 4P in (4) is constant, i.e. independent of the LTV parameters