You are on page 1of 11

3

Robust Multivariable Control

O s c a r R. Gonzfilez 3.1 Introduction ....................................................................................... 1037


Department of Electrical and 3.2 Modeling ............................................................................................ 1037
Computer Engineering, 3.3 Performance Analysis ........................................................................... 1039
Old Dominion University,
Norfolk, Virginia, USA 3.3.1 MIMO Frequency Response and System Gains • 3.3.2 Performance Measures
3.4 Stability Theorems ............................................................................... 1042
3.4.1 Nyquist Criteria • 3.4.2 Small Gain Criteria
3.5 Robust Stability ................................................................................... 1042
Atul G. Kelkar
3.6 Linear Quadratic Regulator and Gaussian Control Problems ....................... 1043
Department of Mechanical
Engineering, Iowa State 3.6.1 Linear Quadratic Regulator Formulation • 3.6.2 Linear Quadratic Gaussian Formulation
University, Ames, Iowa, USA 3.7 H ~ Control ........................................................................................ 1044
3.8 Passivity-Based Control ........................................................................ 1045
3.8.1 Passivity of Linear Systems • 3.8.2 State-Space Characterization of PR Systems •
3.8.3 Stability of PR Systems • 3.8.4 Passification Methods
3.9 Conclusion ......................................................................................... 1047
References .......................................................................................... 1047

3.1 Introduction tions relating the inputs to a standard set of variables, such as
loop currents and the configuration variables (e.g., displace-
A mathematical model of the physical process that needs to be ments and velocities). In the time domain, the differential
controlled is needed for control system design and analysis. This equations can be written as follows:
model could be derived from first principles, obtained via system
identification, or created by one engineer and then given to P(79)~(t) = Q(79)u(t), (3.1)
another. Regardless of the origin of the model, the control
engineer needs to completely understand it and its limitations. where 79 a__d / d t is a differential operator, u(t) E T4m is a vector
In some processes, it is possible to consider models with a single of inputs, ~(t) E 7-4l is a vector of the standard variables called
input and output (SISO) for which a vast amount of literature is the partial state variables, and P(79) and Q(79) are differ-
available. In many other cases, it is necessary to consider models ential operator matrices of compatible dimensions. The vector
with multiple inputs and/or outputs (MIMO). This chapter of output variables is, in general, represented by a linear,
presents a selection of modeling, analysis, and design topics for differential combination of the partial state variables and the
multivariable, finite-dimensional, causal, linear, time-invariant inputs:
systems that generalize the results for SISO systems. The reader is
encouraged to consult the references at the end of this chapter for y(t) = R(79)~(t) + W ( D ) u ( t ) , (3.2)
additional topics in multivariable control systems and for the
proofs to the lemmas and theorems presented. where y(t) E RP is the vector of outputs and where R(79) and
W ( D ) are differential operator matrices of compatible dimen-
sions. The system representation introduced in equations 3.1
3.2 Modeling and 3.2 is c o m m o n l y referred to as the polynomial matrix
description (PMD) and is the most natural representation for
The derivation of a mathematical model for a linear, time- many engineering processes. For analysis and design, a state-
invariant system typically starts by writing differential equa- space representation that is an equivalent representation of

Copyright© 2005 by AcademicPress. 1037


All rights of reproductionin any form reserved.
1038 Oscar R. Gonzdlez and Atul G. Kelkar

proper I PMDs is more appropriate. The state-space represen- The transmission zero definition states that s = Sz is a zero if
tation of a system with input u(t) and output y(t) is as follows: the rank of NL(sz) is less than the rank of NL(s). An important
difference between SISO and transmission zeros is that a trans-
ic(t) = Ax(t) + Bu(t), y(t) = Cx(t) + Du(t), (3.3) mission zero does not have to be a zero of any SISO transfer
function entry of G(s). Another difference is that it is possible
where x(t) C ~ is a vector of state variables, any minimal set for G(s) to have a pole equal to a zero. To understand other
of variables at time to that together with the input u(t), t > to properties of poles and transmission zeros, such as multiplici-
completely characterizes the response of the system for t > to. ties, the Smith-McMillan form of G(s) can be used.
The state-space representation as defined in equation 3.3 will To understand the effect of the poles of G(s) on a particular
be represented by the four-tuple (A, B, C, D). The study of the output, consider y/(s), the ith response given by:
fundamental properties of the state-space representation is
covered, for example, in a first graduate course in linear
systems (Antsaklis and Michel, 1997.) Because any state-
yi(s) = ~ Gq(s)uj(s), (3.6)
)-1
space and PMD representation of a given system are equiva-
lent, they have similar properties. For example, the eigenvalues
where GO(s) is the ijth entry of G(s) and where uj(s) is the
of A are the roots of [P(k) I, where P(M is the polynomial
Laplace transform of the jth input. Equation 3.6 shows that in
matrix in ~ found by replacing D with k. Furthermore, the
MIMO systems, the outputs are linear combinations of the
state-space representation is state controllable (observable) if
inputs where the weights are rational transfer functions. Only
and only if the equivalent PMD is controllable (observable).
the poles of the entries in the ith row of G(s) can affect yi(s),
The generalization of an SISO transfer function is a transfer
depending on which inputs are nonzero. More insight can be
function matrix, which is found by taking the Laplace trans-
gained by solving for yi(s) in a state-space representation of the
form of equation 3.3 with zero initial conditions. This oper-
system. Because the poles correspond only to the controllable
ation yields:
and observable eigenvalues of A, let equation 3.3 be a minimal,
that is, a controllable and observable realization of G(s) with n
G(s) = C(sI - A)-lB + D, (3.4) states. The Laplace transform of the ith output is given by:

where G(s) E ~(s) p×m is a proper rational transfer function yi(s) = ( Ci(sI - A)-I B + Di)u(s), (3.7)
matrix consisting o f p x rn SISO transfer function entries. The
transfer function matrix can also be written in terms of an where Ci and Di are the ith rows of C and D, respectively. If A
equivalent PMD. Because a transfer function matrix as in the has distinct eigenvalues, the partial fraction expansion of the
SISO case represents only the controllable and observable response is given by:
dynamics of a system, consider the following controllable
and observable PMD:
~
yi(s) = (~-~. ~
1 )
Cit)twHB + Di u(s), (3.8)
D~(s)t;(s) = NL(s)u(s), v(s) = t2(s), (3.5)
where DL(s) C T~[s]/'xp and NL(s) E 7"~[S]pxm are left coprime where re, we E C ~ are right and left eigenvectors of A, respect-
polynomial matrices in s. The PMD in equation 3.5 is con- ively, and m denotes complex-conjugate transpose. The effect of
trollable since DL(s) and NL(s) and NL(s) are left coprime; the gth pole on the ith output is determined by the 1 x n residue
otherwise, it would only be observable. The PMD is a realiza- matrix Ciotw[tB and the product of this residue times the vector
tion of the transfer function matrix in equation 3.4 since of inputs u(s). If each entry in the residue matrix is small, then
G(s) = DL(s) 1NL(s). The poles of a transfer function matrix this pole will have little effect on Yi (t). It is also possible for some
G(s) are the roots of the pole polynomial of G(s) where the entries of the residue matrix not to be small but the product of a
pole polynomial is given by the least common denominator of residue with the vector of inputs to still be small. This indicates
all nonzero minors of G(s). So, every pole appears as the pole that for some directions of the input vector, a pole may have a
of at least one SISO transfer function entry of G(s). The more significant effect than in other directions. The concept of
MIMO poles can also be found as the roots of [DL(s)[. As in input directions is unique to MIMO systems, and it will be
SISO systems, the set of poles is a subset of the set of eigen- discussed again in the next subsection.
values of A irA and G(s) satisfy equation 3.4. There are several Stability is an important system property. This section is
definitions of multivariable zeros (Schrader and Sain, 1989). mostly interested in two types: asymptotic and bounded-
input and bounded-output (BIBO) stability. Asymptotic sta-
1
A PMD is proper if lim (R(k)P 1 (MQ(k) + W(M) exists and is finite. In
bility is a property of an internal representation, such as a
fact, if the PMD in eq~ua't~ons 3.1 and 3.2 is equivalent to the state-space state-space. The representation in 3.3 is said to be asymptoti-
in equation 3.3, then klim
~:x?
(R(k)P l(k)Q(k) + W(k)) = D. cally stable if the solutions of 2(t) = Ax(t) approach the origin
3 Robust Multivariable Control 1039

trivial example, the classical unity feedback is represented in


I d' I do
the general block diagram with the following transfer function
matrices:

uK(s) J
FIGURE 3.1 ClassicalUnity Feedback Configuration where:

for all initial conditions x(0). The test for asymptotic stability [I -I -@(s) :-@(s)
is that the eigenvalues of A have negative real parts. Asymptotic P(s) = =/0 .... 0 ...... 0 .... i...: . . . .

stability implies BIBO stability, which is a property of an Le2'(s) i P22(s)] LI -I -Gp(s) :-Gp(s)
external representation. For BIBO stability, only the poles of
a transfer function matrix need to have negative real parts. In this example, K(s) = Gc(s), the exogenous inputs vector
Consider now the classical unity feedback configuration in is w(t) = [r(t), do(t), di(t)] r, the vector of performance vari-
Figure 3.1, where r(t) E ~/' denotes a vector of reference ables has been taken to be z(t)= J r ( t ) - y(t), y1((t)] T, and
inputs and where di(t) E T¢m and do(t) E T~p denote vectors the vector of controller's inputs and outputs is simply given by
of disturbance inputs at the input and output of the plant, uK(t) = uc(t) and yK(t) = yc(t), respectively. The first per-
respectively. Assume that the state-space representations . , A
formance variable xs the tracking error, e(s)=r(s)- y(s).
(transfer function matrices) of the plant and controller are A state-space representation of the two-input and two-output
given by (A;, B;, Cp, Dp)(Gp(s)) and (Ao Bo Co Dc)(Gc(s)), plant is the following:
respectively. The representation of the closed system will
be well-formed if the dimensions (@(s)E ~;(s) ;×m and •
Gc(s) ~ 7Ep(s)re×p) are compatible and if fl + DpD~[ ~ O. The ke(t) = Apxv(t) + [0 0 Bp:BP] LyK(t) j (3.10)
state representation of the closed-loop system is of the form:

2d(t) = Aclxd(t) + Bdw(t), y(t) = Cclxd(t) + Ddw(t), (3.9) Lu :(t)J =


[z('t)l . .... 0
Ly/dt)]
. . . . . .

where xd(t) = [xp(t), &(t)] T and w(t) = [r(t), do(t), di(t)] T.


L-Cpj L1 0 -Dp! -D~J
The properties of the dosed-loop system are determined (3.11)
by analyzing (Acl(t), Bcl(t), Cd(t), Dd(t)). For example, the
The closed-loop transfer function matrix from the exogen-
closed-loop system is asymptotically stable if the eigenvalues
ous inputs to the performance variables is as follows:
ofAd (t) have negative real parts. The dosed-loop system is then
said to be internally stable.
z(s) = Tzw(S)W(S).
The analysis and design of complex feedback configurations
is simplified by using the general dosed-loop block diagram in The closed-loop transfer function matrix is in fact a lower
Figure 3.2, where w(t) is the vector of all exogenous inputs, linear fractional transformation (LFT) of P(s) with respect to
y~c(t) is the vector of controller outputs, z(t) is the vector of K(s) as given by:
performance variables, and u/((t) is the vector of inputs to the
controller. The top block is called the two-input and two- Tzw(S) =Srg(P(s), K(s)) =Pll(S)
output plant, P, and the bottom one corresponds to the matrix (3.12)
of controllers denoted by K that is formed after all the con-
+P12(s)K(s)(I-P22(s)K(s))-1p21(s).
trollers have been pulled out of the closed-loop system. As a
Based on this LFT, the general closed-loop system is well-
formed if ]I - P22(oc)K(oc)] ~ O. In the trivial example of a
classical unity feedback system, this reduces to [ I - P22(0c)
W b Z
P(s) K ( ~ ) I = ]i + D~DcL ~ o.
YK - - UK

3.3 Performance Analysis

The purpose of the controller is to make the closed-loop system


K(s)
1 meet the desired specifications regardless of the uncertainty
present. Examples of specifications are that the closed-loop
FIGURE 3.2 GeneralClosed-Loop Block Diagram system should be asymptotically stable, the steady-state error
1040 Oscar R. Gonz~lez and Atul G. Kelkar

for each output channel should be small, and these stability and Zero steady-state error is possible for step inputs by appro-
performance specifications should be maintained not only for priately including an exogenous model of the exosystem in the
the nominal plant model but also for all models in a specified feedback loop (Gonzalez and Antsaklis, 1991). If zero steady-
uncertainty set. In this case, the closed-loop system will be said to state error is not needed, then it will be necessary to make
have the properties of robust stability and robust performance. the mappings from the four exogenous inputs to the tracking
Consider the classical unity feedback configuration in Figure error in equation 3.14 small. Three ways to determine the size
3.1 with an additional sensor noise vector rl(s) so that Uc(S) = of a transfer function matrix are presented in the following
r(s) - (rl(S) + y(s)). The performance analysis is dependent on subsection.
the following three types of transfer function matrices: To simplify the presentation, assume from now on that the
plant is square with p = m. In addition, since the physical units
• Return ratio: Lo(s) = Gp(s)Gc(s).
used for input and output signals may lead to errors of different
• Sensitivity: So(s) = (I + Lo(s)) -1.
orders of magnitude, it is useful to normalize or scale the
• Complementary sensitivity: To(s) = Lo(s)(I + Lo(s)) -1.
magnitudes of the plant's inputs and outputs. Procedures to
These transfer function matrices are defined when the loop perform scaling of MIMO systems are presented in, for example,
is broken at the output to the plant. Similar definitions follow Skogestad and Postlethwaite (1996). One approach is to nor-
when the loop is broken at the plant's input. The input/output malize the plant's inputs and outputs so that the magnitude of
and input/error relations can be written as shown next: each error is less than one. An alternative and common choice
is to include the normalization in the frequency-dependent
y(s) = So(s)do(s) + So(s)Gp(s)di(s) + To(s)r(s) - To(s)rl(S). (3.13) weights to be introduced for control system design.

e(s) = -So(s)do(s) - So(s)Gp(s)di(s) + (I - To(s))r(s) + To(s)~l(S).


(3.14) 3.3.1 MIMO Frequency Response
and System Gains
Notice that the mapping from the reference inputs, r(s), to
To determine the frequency response ofa BIBO stable transfer
the errors contributed by them, er(S), is given by I - To(s). In
function matrix, G(s), let its input be the vector of complex
this case, if the closed-loop system is internally stable, the
exponentials u(t)=u_eJ°~t,u_E CP; then, the steady-state
steady-state tracking error contributed by a vector of step refer-
response is a complex exponential vector of the same fre-
ence inputs r( s) = Ro ~ , Ro E 7-4p is found using Laplace's final
quency with amplitudes and phases changed by G(S)ls~j~o.
value theorem to be er(cxD) = ( 1 - To(O))Ro. Thus, the error
Let the steady-state response be given by y s / t ) : y ' e J~°t,
contributions to each output channel will be zero if in addition
y E CP, then the complex vectors _u and y are related by:
to internal stability, To(O) = I. Because the sensitivity and com-
plementary sensitivity transfer function matrices satisfy:
y = G(jto)u. (3.16)
So(s) + To(s) = I, (3.15)
The complex matrix G(jto) E C p×p is called the frequency
then So(O) = Op×p results in e~(oo) = 0. A sufficient condition response matrix, and it can be used to determine the size of
for the dc gain of the sensitivity matrix to vanish is that every G(s) at a particular frequency to.
entry of Lo(s) must have at least one pole at the origin. This is In general, the size of G(s) is defined as the gain from an
the generalization of system type to MIMO systems. Further- input to its corresponding output. If the input is u(t) = _ueTM,
more, if So(O) = 0 p x p , the steady-state error contributed by a the gain of G(s) at to can be defined at steady-state to be the
vector of step functions at the output disturbances, do(s), will ratio of the Euclidean vector norms [[y[[/][_u[[.This concept of
also be zero. An additional condition is needed for zero steady- gain is bounded as follows:
state error contributed by a vector of step functions at the
input disturbances, di(s). A sufficient condition is that the if(to) = rain IIG(jto)ull ~ IIG(jto)_ull ~ max IIG(jto)_ull _ ~(to),
plant has no entries with poles at the origin. Jr_~ll~0 I1_~11 tl~,It II_~ll~0 I1~11
Making T ( 0 ) = I results in the desired zero steady-state (3.17)
errors to vectors of step functions at the reference inputs and
at the input and output disturbances. This choice, however,
where 6-(to) and ¢(to) are the largest and smallest singular
has the undesirable effect of making the error contributed by
values of G(jto). These bounds are used to define the size of
the dc component of the sensor noise not to be attenuated.
a TFM as follows:
This is a common trade-off in control systems, which does not
affect the desired performance as long as the signal-to-noise • Large G(s) is said to be large at co ifff(G(jto)) is large.
ratio for low frequencies is made sufficiently high. • Small G(s) is said to be small at to if 6"(G(jto)) is small.
3 Robust Multivariable Control 1041

The application and the scaling of the model determine where Ijull2 is the norm of the vector signal and Ilu(~)ll is the
what is meant by large or small singular values. In general, Euclidean norm of a vector in T4;. To pose and solve opti-
(r(G(jto)) >> 1 indicates that G(s) is large at co. Similarly, mal control problems, it is useful to consider the most general
#(G(jto)) << 1 indicates that G(s) is small at co. set of energy signals. The desired set is the Lebesgue space
A graphical representation of the frequency response con- of all square integrable functions. This set is denoted by
A . . . .
sists of plotting the maximum and minimum singular values £2+ =£2[0, 00). For convenience, the spatial dxmenslon of
of G(jto) versus log (co). This is the generalization of Bode's vectors with entries in £2+ will not be included. A useful
magnitude plot to MIMO systems. measure of the size of the transfer function matrix is the
To get a better understanding of the role of singular values, induced system gain. If the system is represented by the map-
substitute the singular value decomposition (SVD) of G(jto) in ping G: £2 -"+ £2, then the induced system gain is as follows:
equation 3.16:

y = Y(e0)E(to) U(to)Hu_, (3.18) sup IlCull2 _ sup jIGull2 = IIGIt , (3.21)


Ilull2 ,<<
where E(to) E Cp×p is diag {ff(to)=o-l(to), ¢r2(to). . . . .
Crp(to) = ~(to)} and where Y(to) and U(to) C C/'×p are unitary where II C;lloo is the induced system norm. This is the 7-/oo norm
matrices (see Section 9, Chapter 1). The diagonal entries of of the system that gives the maximum gain of the system to
E(to) are called the singular values of G(s), and they depend on a vector of sinusoids in the worst possible direction and
frequency. They are ordered from the largest to the smallest. worst possible frequency. If G(s) is proper and stable, then
Now, rewrite (equation 3.18 in terms of the columns of []G]]o~ = sup 6-{G(jo~)} is finite. Another popular measure of
(o
Y(to) and U(to), and let Y(to)= [Yl(to)'"Yp(to)] and
a system is the ~-/2 n o r m . One interpretation of the ~-{2 n o r m is
U(to) = [ u l ( t o ) - ' . up(to)]. Since the columns of U(to) form
as the gain from a white noise input with unit variance to the
an orthonormal basis of C p, let ~ be the representation of _u
power of the output. The power is defined as:
with respect to {Ul (to) . . . . . up(to) }. Substituting _u = U(to)a_ in
equation 3.18 gives the following:
p
___= Y(to)E(to)ot = Z¢i(to)oLi(to)yi(to), (3.19) Ilyllpow = (lim
\ r-+oo --'
2T Jri Ily('r)n2 dT
i=1

where ~ = [Otl(to)-.. e~p(to)]T. This equation shows that the where UyUpo,,is only a seminorm.
representation of y in terms of the columns of Y(to) (also an
orthonormal basis) is as written here:
3.3.2 Performance Measures
~(to)Ot = [ff(to)Otl(to), O'2(to)Ot2(to)"". o-(to)(to)ccp(to)] r.
Consider again the classical unity feedback configuration
The columns of U(to) can be called the principal input in Figure 3.1 with an additional sensor noise vector r/(s) so
directions, and the columns of Y(to) are the principal out- that Uc(S) = r(s) - (rl(s) + y(s)). If r(t), di(t), and do(t) are
put directions. If the input is parallel to a principal input vectors of sinusoids up to a frequency tolow, then adequate
direction, then the steady-state response will be along the steady-state performance to these inputs requires that
corresponding principal output direction scaled by the corres- ~(So(jto)) << 1 and ~(So(jto)Gp(jto)) << 1 for co < tolow. The
ponding singular value. In this sense, the singular values quan- former requirement is met if and only if ~(Lo(jto)) >> 1 for
tify the size of the effect of a transfer function matrix on co < totow. If Gc(s) is invertible, the second requirement is met
particular input directions. if ~(Gc(jto)) >~ 1 for co < tolow. If the signal-to-noise ratio
This characterization of size of a TFM will be used to develop becomes poor for co > tohigh, then acceptable performance at
quantitative measures of performance when the inputs are steady-state requires attenuating the high-frequency compon-
sinusoids in a specified frequency band. Two other measures ents of the noise by making ~(To(jto)) << 1 for co > tohigh. This
will also be useful, and they can be defined as induced gains is accomplished by making ff(Lo(jto)) << 1 for to > tohigh.
when the inputs and outputs belong to specified signal spaces. These are just some of the design guidelines that will lead to
Consider first the set of energy signals that are vectors of acceptable designs. There are additional trade-offs and per-
functions u that map 7-4 into 7-4p. The signal u is said to have formance limitations that need to be taken into account,
finite energy if: including guidelines for roll-offs during the mid-frequencies.
A convenient way to combine the low-, mid-, and high-
frequency requirements is to introduce frequency-dependent
[lull2 = Ilu(~')ll2d-r < oo, (3.20) weights. These weights are included in P(s) when the gen-
eral block diagram in Figure 3.2 is formed. The weights are
1042 Oscar R. Gonzfilez and Atul G. Kelkar

typically added on the exogenous input channels and the performance of the closed-loop system with the real plant as
performance output channels. The former weights serve to discussed in the following section.
include in the model spectral information about the inputs.
The latter weights are important in design to emphasize the
frequency bands where the performance outputs need to be
3.5 Robust Stability
minimized. A possible design problem that results is to find a
Controller design uses a nominal plant model. The error
proper compensator Gc(s) so that IITz,o[l~ < 1. This and other
between the nominal model and the real plant arises primarily
control problems will be discussed in the following sections.
from two sources: unmodeled dynamics and parametric un-
certainties. If the controller design does not take these errors
into account, it cannot guarantee the performance of the
3.4 Stability Theorems
closed-loop system with the real plant nor guarantee that
the closed-loop system will be stable. Therefore, it is important
In this section, a brief introduction will be given to various
to design controllers that will maintain closed-loop stability in
stability criteria that can be used to determine the stability of a
spite of erroneous design models and uncertainties in param-
closed-loop control system.
eter values. Controllers so designed are said to impart stability
robustness to the closed-loop system.
3.4.1 Nyquist Criteria To analyze stability robustness, consider the unity feedback
system in Figure 3.1. A basic plant uncertainty representation
Consider a negative feedback interconnection of a plant Gp(s)
is G;(s) = Gpo(S) + A~(s), where G;o(S) is the nominal plant
and controller Go(s) in Figure 3.1. Let nLo+ denote the number
model and where Aa(s) is the additive uncertainty. Other
of unstable pole s of the return ratio Lo(s). Then the Nyquist
uncertainty representations include the output multiplicative
stability criteria is given by the following theorem:
Stability Theorem: The closed-loop system consisting of the uncertainty, Gp(s) = (I + Ao(s))Gpo(S), and the input multi-
plicative one, Gp(s) = Gpo(S)( I + Ai(s) ). For design purposes,
negative feedback interconnection of Gt,(s) and G,-(s) is inter-
it helps to normalize the uncertainty representations. For
nally stable if and only if nLo+ = ncp+ + nGc+ and the Nyquist
example, consider that the real plant is represented with an
plot of 1I + Lo(s) I encircles the origin nLo+ times in the anti-
output multiplicative uncertainty, and let Ao(s) = Wo(s)Ao(s),
clockwise direction and does not pass through the origin.
where Ao(s) and Wo(s) are proper and stable with
The first condition guarantees that the closed-loop system
[I/~o(s)[[~ < 1 and where Wo(s) is a frequency-dependent
has no unstable hidden modes. The second condition gives a
scaling matrix. In this case, the unity feedback system in Figure
MIMO generalization of the Nyquist plot in terms of the
3.1 will be robustly stable if and only if IIToWo[I~ <_ 1.
determinant of I + Lo(s). Because of the determinant, the
A more general result is possible that is independent of the
Nyquist plot of a scalar times I + Lo(s) is not simply a scaled
particular types of uncertainties needed to represent the real
version of the Nyquist plot of [I + Lo(s)[.
plant. Consider the general block diagram in Figure 3.2: if all
the normalized uncertainty blocks are pulled out, it results
3.4.2 Small Gain Criteria in a new general block diagram shown in Figure 3.3 that is
Another criteria that is often used to determine internal stabil-
ity of the feedback interconnection is the small gain theorem,
which is based on limiting the loop gain of the system. This
theorem is central to the analysis of robust stability. The small
;~(s)
gain theorem states that if the feedback interconnection of
two proper and stable systems has a loop-gain product less
than unity, then the closed-loop system is internally stable.
There exist several versions of this theorem. One version is
given next.
Theorem: Consider the feedback system in Figure 3.1, where
systems Gp and Gc are proper and stable. Then, the feedback
YK
! Po(S) I,,

UK
Z

system is internally stable if:

IIG/l llGcll < 1


K(s)
The stability theorems already given can be used in the
analysis and synthesis of control systems. These theorems are
also useful to determine conditions for robust stability and FIGURE3.3 Closed-Loop System with Uncertainties Pulled Out
3 Robust Multivariable Control 1043

X(s) I~
[p(t)
~(t)]
J -Q - Lp(t)]
with initial conditions: x(O) = Xo; p(tf) = Sx(tf). (3.24)

The optimal controller uses full-state feedback and is given


by:

M(s) u(t) = - R 1BTp(t)x(t)=-K(t)x(t),


W [I b Z

where P(t) is the solution of the matrix Riccati equation:


FIGURE 3.4 Simplified Closed-Loop System with Uncertainties
Pulled Out P(t) = - P ( t ) A - ATP(t) + P(t)BR-IBTp(t) - Q,

which is solved backward in time starting at P(tf) = S. The


useful for robustness analysis. In Figure 3.3, /~(s) is the block optimal feedback gain matrix K(t) is given by:
diagonal matrix of all the normalized uncertainty blocks
([IA(s)]]oo _< 1), and Po(s) is the augmented nominal plant, K = R-1BrP(t),
including the uncertainty frequency weights. Po(s) is assumed
to be stabilized by K(s). Using a lower LFT, the bottom loop in and the optimal cost is as follows:
Figure 3.3 can be closed resulting in Figure 3.4, where M(s) is
proper and stable. If M(s) is partitioned corresponding to
Jopt ~- xT(o)P(O)x(O).
the two vector inputs and outputs and if ][MH]Ix < 1, then
the closed-loop system is robustly stable. These results that
make use of the unstructured uncertainties lead to very con-
servative results. One way to reduce the conservativeness is to 3.6.2 Linear Quadratic Gaussian Formulation
take advantage of the structure in the block diagonal matrix The linear quadratic Gaussian (LQG) control problem is an
/~(s) as done with the structured singular values. optimal control problem where a quadratic cost function is
minimized when the plant has random initial conditions,
white noise disturbance input, and white measurement noise.
3.6 Linear Quadratic Regulator and The typical implementation of the LQR solution requires that
Gaussian Control Problems the plant states be estimated, which can be posed as an LQG
problem. The plant is described by the following state and
3.6.1 Linear Quadratic Regulator Formulation output equations:
The linear quadratic regulator (LQR) is a classical optimal
2(t) = Ax(t) + B.u(t) + Bww(t).
control problem used by many control engineers. Solutions
to LQR are easy to compute and can typically be used to ym(t) = Crux(t) + v(t) (measurement output). (3.25)
compute a baseline design useful for comparison. A formula- yp( t ) = Cpx( t ) (performance output).
tion of the LQR problem considers the state equation of the
plant: The v(t) and w(t) are uncorrelated zero-mean Gaussian
noise processes; that is, w(t) and v(t) are white noise processes
2(t) = Ax(t) + Bu(t). (3.22) with covariances satisfying:

The formulation also considers the following quadratic cost


function of the states and control input:
[ v(t) ] [wr(t + "r), [0
ts The quadratic cost function that is to be minimized is given
] = xT(tf)Sx(tf) + ~ xr(t)Qx(t) + uT(t)Ru(t)dt, (3.23) by:
0

where S = Sr _> 0, Q = Qr > 0 and R R T > O. =

To minimize the cost function, consider the Hamiltonian


system with state and costate (p(t)) dynamics given by:
] = E xT(tf)Sx(tf) + ~
i
0
xT(t)Qx(t) + ur(t)Ru(t)dt ,
]
(3.26)
1044 Oscar R. Gonzdlez and Atul G. Kelkar

where S = ST ~ 0, Q = Qr _> 0 and R = R T > 0. The optimal The integral LQG can be used both for disturbance rejection
controller is a full-state feedback controller and is given by: and tracking. Similar modifications are possible to handle
tracking of time-varying reference inputs.
u(t) = -K(t)2(t),

where 2(0 is the Kalman state estimate. The closed-loop state 3.7 H~ Control
equation can then be given by:
The LQR and LQG can also be posed as two-norm optimiza-
Ix(t)] +Bc'(t)[wl:l]'
=Ad(t)[e(t)J tion problems (referred to as H2 control problems). If the
optimization problem is posed using the Hoo norm as the
cost function, the H~ formulation results.
where: The H~ control problem can be defined in terms of the
general closed-loop block diagram in Figure 3.2 where the
Ac, 0
Bu ,t,] ' B d =
A-G(t)Cm Bw -G(t)
0] "
exogenous signals are included in the vector co(s) and an
appropriate choice of performance variables is given by the
vector z(s). The TFM P(s) includes frequency-dependent
The closed-loop state covariance matrix is as follows: weights and appropriate normalization as described previ-
ously. Consider the following realization of P(s):
~-~(t) =Acl(t) Z ~ (t)+ Z ~ (t)A~(t)+Bcl(t)[ W O]BT(t),
2p(t)=Axp(t)+[Bl!B2] [.w.! t.).] (3.27)
IrK(t)]
where ~(t) = [x(t), e(t)] r, and:
.Z.(.t).] {.C12.]Xp(t)_}_[ 0 i a,2 1
[x(O)xr(o) x(O)er(o) 1 UK(t)j= (/:52; • i • .6..j [~Ki:~], (3.28)
y]~(O) = ELe(O)xr(O) e(O)eT(O)j.
where:
The performance output covariance is given by:
D21BT = 0. (3.29)
D21DT = I. (3.30)
DTc1 = 0. (3.31)
and the input covariance is given by:
T T = I.
D12D12 (3.32)
,, [--KT(t)]
Zu(t) = [ - K(t) K(t)] Z ~ ( U [ KT(t) J. The control objective is to design a feedback controller that
internally stabilizes the closed-loop system such that the
The cost is as follows: oc-norm of the mapping Tz~ is bounded:

,=~Tr [~ 00]K-" 't'+ f[Q+KT(t)RK(t) --KT(t)RK(t)I (ITzwil~ = sup I)z(t)ll~ < Y. (3.33)
Z-~ tf' J[ KT"(t)RK(') KT(t)RK(,)J Zs, (')dt• I]w(t)[12~0 IIw(t) 112

A suboptimal controller satisfying the above mentioned


Remarks objective exists if positive semidefinite solutions to the
For convenience of implementation, the steady-state solution following two Riccati equations are possible:
obtained in the limit as tf---* oo is used. The LQG control
design is optimized to reject white noise disturbances; how- P(t) = P(t)A + ATp(t) -- P(t)(B2B T - y-2B1SW)p(t) + Crl C1.
ever, it can be modified to handle constant disturbances via
(3.34)
feed-forward and integral control. The selection of gains in
the feed-forward case can be done in a similar way as in the 0(t) = AQ(t) + Q(t)A T - Q(t)( cT c2 - ,I-2CT C1)Q(t) + B1BT.
case of tracking system design. Prior to implementation, the (3.35)
robustness of LQG designs needs to be evaluated since there is
no guarantee that any useful robustness will be obtained. For Moreover, the solutions P(t) and Q(t) satisfy the following:
constant disturbance rejection via integral control, one needs
to use integral Kalman filter and also integral state feedback. p(P(t)Q(t)) < y2. (3.36)
3 Robust Multivariable Control 1045

The Hamiltonian systems corresponding to the two Riccati function (Newcomb, 1966; Desoer and Vidyasagar, 1975).
equations can be obtained similarly to how they were obtained The concept of strict positive realness has also been defined
for the LQG case. The Hoo controller, which satisfies the in the literature and is closely related to strict passivity.
mentioned bound, is given by: Let G(s) denote a p x p matrix whose elements are proper
rational functions of the complex variable s. The G(s) is said
kc = Ac(t)xc(t) + Bc(t)uK(t). to be stable if all its elements are analytic in Re(s) >_ O. Let
(3.37)
yK(t) = Cc(t)&(t). the conjugate-transpose of a complex matrix H be denoted
by H H.
The matrices At(t), Bc(t), and C~(t) are given as follows:
Definition 1
Ac(t) = A + ~/-2BIBTp(t) - B2BY2P(t) A p × p rational matrix G(s) is said to be positive real (PR) if."
- [ I - y-2Q(t)P(t)] 'Q(t)cTc2. • All elements of G(s) are analytic in Re(s) > 0
(3.38)
Sc(t) = [I - y - 2 Q ( t ) n ( t ) ] - l Q ( t ) C T. • G(s) + GH(s) > 0 in Re(s) > 0, or equivalently:
• Poles on the imaginary axis are simple and have non-
Co(t) = --BTp(t).
negative-definite residues
• G(jto) + GH(jto) > 0 for to E ( -- oo, c~)
In practice, the steady-state solution of the Hoo control
problem is often desired. The steady-state solution not only Various definitions of strictly positive real (SPR) systems
simplifies the controller implementation but also renders a are found in the literature (Kelkar and Joshi, 1996). Given
closed-loop system time-invariant, simplifying the robustness below is the definition of a class of SPR systems: marginally,
and performance analysis. For the steady-state Ho~ control strictly, and positive-real (MSPR) systems.
problem, a suboptimal solution exists if and only if the
following conditions are satisfied. The first condition is that Definition 2
the algebraic Riccati equation: A p × p rational matrix G(s) is said to be marginally strictly
positive real (MSPR) if it is positive real and the following is
0 = PA + A T p -- P(B2BT2 -- "y-2B1B~)P + cTc1, true:

has a positive semidefinite solution P such that [A (B2BT- -


G(jto) + GH(jto) > 0 for to E ( - o c , oc).
y-iB~BT)p] is stable. The second condition is that the alge-
braic Riccati equation: Definition 2 (Joshi and Gupta, 1996) gives the least restrict-
ive class of SPR systems. If G(s) is MSPR, it can be expressed
0 = A Q + QA r - Q(c2Tc2 - "y-2cTC1)Q + B1BT, as G ( s ) = G l ( s ) + G2(s), where G2(s) is weak SPR (Kelkar
and Joshi, 1996) and where all the poles of G1(s) are purely
has a positive semidefinite solution Q such that [A - Q(CeTC2 imaginary (Joshi and Gupta, 1996).
- y - 2 c T c 1 ) ] is stable. The third condition is p(PQ) < ,/2.

3.8.2 State-Space Characterization of PR Systems


3.8 P a s s i v i t y - B a s e d C o n t r o l For LTI systems, the state-space characterization of positive
real (PR) conditions results in the Kalman-Yakubovich-Popov
Passivity is an important property of dynamic systems. A large (KYP) lemma. In Lozano-Leal and Joshi (1990), the KYP
class of physical systems, such as flexible space structures with lemma was extended to WSPR systems, in Joshi and Gupta
collocated and compatible actuators and sensors, can be clas- (1996), it was extended to MSPR systems. These extensions are
sified as being naturally passive. A passive system can be given next.
robustly stabilized by any strictly passive controller despite Let (A, B, C, D) denote an nth-order minimal realization of
unmodeled dynamics and parametric uncertainties. This im- the p × p transfer function matrix G(s). The following lemma
portant stability characteristic has attracted much attention of then gives the state-space characterization of WSPR system.
researchers in the control of passive systems. This section The Lozano-Leal and Joshi (1990) Lemma: The G(s) is
presents selected definitions and stability theorems for passive WSPR if and only if there exist real matrices: P = p r > 0,
linear systems. P E R '~×n, L ~ R px'~, and W E R p×i', such that:

3.8.1 Passivity of Linear Systems A T p + PA = - L T L .


C = BTp + WTL. (3.39)
For finite-dimensional linear, time-invariant (LTI) systems,
passivity is equivalent to positive realness of the transfer wTw = D + D T.
1046 Oscar R. Gonzdlez and Atul G. Kelkar

In these equations, (A, B, L, W) is controllable and observable P = P r > 0, (3.42)


or minimal, and F ( s ) = W + L ( s I - A ) - I B is minimum
phase. If G(s) is MSPR, it can be expressed as where U = 0, V = 0, and W = - I . This LMI condition is
G(s) = Gl(s) + G2(s), where Gi(s) is WSPR and all the poles convenient to use in the case of checking PR-ness of MIMO
of Gl(s) are purely imaginary (Josh and Gupta, 1996). Let systems. This LMI is a special case of dissipativity LMI (Kelkar
(A2, B2, C> D) denote an n2th-order minimal realization of and Joshi, 1996).
G2(s), the stable part of G(s). The following lemma is an
extension of the KYP lemma to the MSPR case. Stability Theorem
The Josh and Gupta (1996) Lemma: If G(s) is MSPR, there The closed-loop system consisting of negative feedback inter-
exist real matrices: P = P r > 0 , P E R ~ × n , g C R p×< and connection of Gp(s) and Go(s) (Figure 3.1) is globally asymp-
W E R p×p, such that equation 3.39 holds with: totically stable if Gp(s) is PR, Gc(s) is MSPR, and none of the
purely imaginary poles of Gc(s) is a transmission zero of Gp(s)
L = [0p×nl, "~pxn2], (3.40) (Joshi and Gupta, 1996).
Note that in the theorem systems Gp(s) and Gc(s) can be
where (A2, B2, 9, W) is minimal and F(s) = W + interchanged. Some nonlinear extensions of these results are
L(sI - A)-IB = W + 58(si - A2)-lB2 is minimum phase. also obtained in Isidori et al. (1999). Passivity-based control-
lers based on these fundamental stability results have proven to
be highly effective in robustly controlling inherently passive
3.8.3 Stability of PR Systems linear and nonlinear systems.
Most physical systems, however, are not inherently passive,
The stability theorem for a feedback interconnection of a PR
and passivity-based control methods cannot extend directly to
and a MSPR system is given next.
such systems. For example, unstable systems and acoustic
LMI form of PR Lemma: An alternate form of KYP Lemma
systems are not passive. One possible method of making
can be given in terms of the following Linear Matrix Inequality
these nonpassive systems amenable to passivity-based control
(LMI). A system (A, B, C, D) is said to be PR if it satisfies:
is to passify them using suitable compensation. If the compen-
sated system is ensured to be robustly passive despite plant
Arp+PA PB]+
BrP [0C 7]r[W r
U
W](0C 7] <0
(3.41) uncertainties, it can be robustly stabilized by any MSPR con-
troller. In Kelkar and Joshi (1997), various passification tech-

. . . . . . . . . . . . . . . . . . . . . . . . .
G(s). . . . . . . . . . . . . . . . . . . . . . . . . . . . .

......................................................
G(s)
U
G(S) ~-ym=y
G,(s) l J~ Gp(S) Ym = Y

Gfb(S)

(A)Series (B)Feedback
.................................................... a

, . . . . . . . . . . . . . . . . . . . . . . . . . .
G(s)
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . , . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G(s)
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

GAs) +f_~ Gp(S) 1 ~y

i
4 Gn(s) ~ Ym Gfb(S) Gs(S) ,~Ym

(C) Feed-Forward (D)Hybrid


FIGURE 3.5 Methods of Passification
3 Robust Multivariable Control 1047

niques are presented, and some numerical examples are given, time, robust, multivariable control systems. The chapter starts
demonstrating the use of such techniques. A brief review of with an introduction to modeling. The analysis tools include
these methods is given next. basic measures of performance, frequency response, and sta-
bility theorems. Linear quadratic, H~ and passivity-based
3.8.4 Passification Methods control synthesis techniques were also introduced.

The four passification methods in Figure 3.5 series, feedback,


feed-forward, and hybrid passification are given in Kelkar References
and Joshi (1997) for fnite-dimensional linear time-invariant Antsaklis, P.J., and Michel, A.N. (1997). Linear systems. New York:
nonpassive systems as shown. Once passified, the system can McGraw-Hill.
be controlled by any MSPR or weakly SPR (WSPR) Controller Desoer, C.A., and Vidyasagar, M. (1975). Feedback systems: Input
(Isidori et al., 1999). In Figure 3.5, the system with input u(t) output properties. New York: Academic Press.
and output ym(t) (G(s)) represents the passified system. The Gonzalez, O.R., and Antsaldis, p.J. (1991). Internal models in regula-
type of passification to be used depends on the dynamic tion, stabilization, and tracking. International Journal of Control
53(2), 411-430.
characteristics of the unpassified plant. For example, the
Green, M., and Limebeer, D.J.N. (1995). Linear robust control. Engle-
system having unstable poles will require feedback passifica- wood Cliffs, NJ: Prentice Hall.
tion, whereas the system having nonminimum phase zeros Isidori, A., loshi, S.M., and Kelkar, A.G. (1999). Asymptotic stability
(i.e., having unstable zero dynamics) will require feed-forward of interconnected passive nonlinear systems. International ]ournal
passification. Some systems may require a combination of the of Robust and Nonlinear Control 9, 261-273.
basic passification methods. For SISO systems, the passifica- loshi, S.M., and Gupta, S. (1996). On a class of marginally stable
tion process is easier than for MIMO systems. The reason is positive-real systems. IEEE Transactions on Automatic Control
that for SISO systems, only the phase plot needs to be checked 41(1), 152-155.
to determine passivity, whereas in the case of MIMO systems, Kelkar, A.G., and Joshi, S.M. (1996). Control of nonlinear multibody
the KYP lemma conditions have to be checked. One numerical flexible space structures, Lecture notes in control and information
sciences, Vol. 221. New York: Springer-Verlag.
technique that can be used to check the KYP lemma is linear
Kelkar, A.G., and loshi, S.M. (1997). Robust control of nonpassive
matrix inequality (LMI)-based PR conditions. The solution of
systems via passification. Proceedings of the American Control Con-
the LMI can be done using the LMI tool box in MATLAB or ference 5, 2657-2661.
another semidefinite programming package. Kelkar, A.G., and Joshi, S.M. (1998). Robust passification and control
One important thing to be noted here is that, in the case of of nonpassive systems. Proceedings of the American Control Confer-
inherently passive systems, the use of an MSPR controller ence, 3133-3137.
guarantees stability robustness to unmodeled dynamics and Lozano-Leal, R., and Joshi, S.M. (1990). Strictly positive real func-
parametric uncertainties; however, in the case of nonpassive tions revisited. IEEE Transactions on Automatic Control 35(11),
systems that are rendered passive using passifying compen- 1243-1245.
sation, stability robustness depends on the robustness of the Lozano-Leal, R., and Joshi, S.M. (1990). On the design of dissipative
passification. That is, the problem of robust stability is trans- LQG-type controllers. In P. Dorato and R.K. Yedavalli (Eds.) Recent
advances in robust control. New York: IEEE Press.
formed into the problem of robust passification. In Kelkar
Maciejowski, J.M. (1989). Multivariable feedback design. Reading, MA:
and Joshi (1998) a number of sufficient conditions are derived
Addison-Wesley.
to check the robustness of the passification. Newcomb, R.W. (1966). Linear multiport synthesis. New York:
McGraw-Hill.
Schrader, C.B., and Sain, M.K. (1966). Research on system zeros:
3.9 C o n c l u s i o n A survey. International Journal of Control 50(4), 1407-1433.
Skogestad, S., and Postlethwaite, I. (1996). Multivariable feedback
This chapter has presented some of the fundamental tools in control. West Sussex, England: John Wiley Sons.
the analysis and design of linear, time-invariant, continuous-

You might also like