Matlab Robust Control Toolbox

April 2006

Nathan Sorensen Kedrick Black

Purpose
Increase Enthusiasm for Robust Controls Learn how to simulate control algorithms with uncertainty Increase your knowledge in Robust Controls

April 2006

Nathan Sorensen Kedrick Black

Outline
What is the Robust Control Toolbox Uncertainty
Uncertain Elements Uncertain Matricies and Systems Manipulation of Uncertain Models Interconnection of Uncertain Models Model Order Reduction Robustness and Worst Case Analysis Parameter-Dependent Systems Controller Synthesis -synthesis

Robustness

MIMO Control Sampled Data Systems Gain Scheduling Supporting Utilities LMI
Specification of Systems of LMIs LMI Characteristics LMI Slovers Validation of Results Modification of Systems of LMIs Nathan Sorensen Kedrick Black 3

April 2006

What is the Robust Control Toolbox

A collection of functions and tools that help you to analyze and design MIMO control systems with uncertain elements Ability to build uncertain LTI system models containing uncertain parameters and uncertain dynamics You get tools to analyze MIMO system stability margins and worst case performance Ability to simplify and reduce the order of complex models with model reduction tools that minimize additive and multiplicative error bounds It provides tools for implementing advanced robust control methods like H , H2, linear matrix inequalities (LMI), and -synthesis robust control You can shape MIMO system frequency responses and design uncertainty tolerant controllers

April 2006

Nathan Sorensen Kedrick Black

Modeling Uncertainty
Arises when system gains or other parameters are not precisely known, or can vary over a given range Examples of real parameter uncertainties include uncertain pole and zero locations and uncertain gains With the Robust Control Toolbox you can create uncertain LTI models as objects specifically designed for robust control applications

April 2006

Nathan Sorensen Kedrick Black

Sources of Uncertainty
Uncertainty in an system can occur in various forms and from various sources.

Additive Uncertainty Multiplicative Uncertainty

Feedback Uncertainty

April 2006

Nathan Sorensen Kedrick Black

Uncertain Elements
1) UComplex() is a function to define complex uncertain parameters

2)
3)

Ucomplexm() is a function for the creation of complex valued uncertain matricies.

Udyn() creates an unstructured uncertain dynamic system class, with input/output dimension specified by iosize.

4)

Ultidyn() is a function to create an uncertain linear time inveriant object where only bounds on the frequency response are known.
Ureal() is a function to define real uncertain parameters used in various analysis and design functions in the robust control toolbox.

5)

April 2006

Nathan Sorensen Kedrick Black

UComplex(Name,nominal value)
UComplex() is a function to define complex uncertain parameters
Sampled Complex Uncertain Parameter A 4.5 4

Name Variable name for the uncertain parameter Nominal Value center nominal value for the parameter radius paramter name to specify a radius of uncertainty around the nominal. Radius value should follow the parameter specification.
2.5 3 3.5 4 4.5 5 5.5

3.5

2.5

1.5

EX:
A = ucomplex('A',4+3*j) Uncertain Complex Parameter: Name A, Nominal Value 4+3i, Radius 1 April 2006

percentage parameter specifying an uncertainty percentage of the nominal for the radius. Value should follow the parameter specificaiton
8

Nathan Sorensen Kedrick Black

Ucomplexm(Name, nominal value)

Ucomplexm() is a function for the creation of complex valued uncertain matricies.
EX: F = ucomplexm('F',[1 2 3;4 5 6],'WL',diag([.1 .3]),... 'WR',diag([.4 .8 1.2])) Uncertain Complex Matrix: Name F, 2x3
Sampled F(1,1) - F(1,1).NominalValue 8 6 4 2 0

0.05

0.1

0.15

0.2

0.25

Sampled F(2,3) - F(2,3).NominalValue 10

Name Variable name for the uncertain parameter Nominal Value center nominal value for the parameter WL,WR WL and WR are square, invertible, and weighting matrices that quantify the size and shape of the ball of matrices represented by this object. Each parameter name is to be followed by the respective matrix. (Assumed to be the identity if left out.) AutoSimplify parameter to set the type of simplification to be done on outputs from the uncertain matrix. Setting values are off, basic, and full.
9

0.05

0.1

0.15

0.2

0.25

April 2006

Nathan Sorensen Kedrick Black

n = udyn('name',iosize)
Udyn() creates an unstructured uncertain dynamic system class, with input/output dimension specified by iosize.
Input: Name Variable name for output Iosize - input/output dimension specification. [Input,Output] EX: N = udyn('N',[2 3]) Uncertain Dynamic System: Name N, size 2x3 size(N) ans = 2 3 get(N) Name: 'N' NominalValue: [2x3 double] AutoSimplify: 'basic'

Output: N - an unstructured uncertain dynamic system class

April 2006

Nathan Sorensen Kedrick Black

10

ultidyn('Name',iosize)
Ultidyn() is a function to create an uncertain linear time inveriant object where only bounds on the frequency response are known.
Parameters Include:
Nyquist Diagram 40 30

Name variable name Iosize the size of [input, output] GainBounded upper bound on the magnitude of the response.

20

10

Imaginary Axis

-10

-20

-30

PositiveReal lower bound on the real part

0 10 20 30 40 50 60 70 80 Real Axis

-40 -10

EX:
B = ultidyn('B',[1 1],'Type','PositiveReal','Bound',2.5) B.SampleStateDim = 3; nyquist(usample(B,30)) Uncertain PositiveReal LTI Dynamics: Name B, 1x1, M+M' >= 2*(2.5)

Bound Specifies the actual bound for the frequency response. SampleStateDim Specifies the state dimension to be used by usample(). AutoSimplify specifies the amount of result simplification.
11

April 2006

Nathan Sorensen Kedrick Black

Ureal(Name,Nominal Value)
Ureal() is a function to define real uncertain parameters used in various analysis and design functions in the robust control toolbox.
Parameters Include:
EX: c = ureal('c',4,'Mode','Range','Percentage',25); Name: 'c' NominalValue: 4 Mode: 'Range' Range: [3 5] PlusMinus: [-1 1] Percentage: [-25 25] AutoSimplify: 'basic' Name variable name Nominal Value Mode Choose uncertainty type amongst: Percentage percentage area around the nominal. Range a given interval with the nominal inside.

PlusMinus add and subtract from nominal

Autosimplify choose amount of simplification for result.
April 2006 Nathan Sorensen Kedrick Black 12

Uncertain Matricies and Systems

1. 2. 3. Umat() Uncertain matrices Uss() creates uncertain state space objects given the uncertain state space system matricies a,b,c,d. Ufrd() is a function to create an uncertain frequency response model which often arises when converting uncertain state space objects to frequency response objects. Randatom() generates a 1-by-1 type uncertain object. Randumat() Generate random uncertain umat objects Randuss() Generate stable, random uss objects

4. 5. 6.

April 2006

Nathan Sorensen Kedrick Black

13

h = umat(m)
Uncertain matrices are usually created by manipulation of uncertain atoms (ureal, ucomplex, ultidyn, etc.), double matrices, and other uncertain matrices The command umat is rarely used There are two situations where it may be useful If M is a double, then H = umat(M) recasts M as an uncertain matrix (umat object) without any uncertainties In both cases, simplify(H,'class') is the same as M
April 2006 Nathan Sorensen Kedrick Black 14

uss(a,b,c,d)
Uss() creates uncertain state space objects given the uncertain state space system matricies a,b,c,d.
Parameters Include: A,B,C,D umat or other uncertain state space parameters to describe the system. Ts sample time for a discrete time state space object. RefSys if a refrence uss is given, the parameters of that object are included. EX: p1 = ureal('p1',5,'Range',[2 6]); p2 = ureal('p2',3,'Plusminus',0.4); A = [-p1 0;p2 -p1]; B = [0;p2]; C = [1 1]; usys = uss(A,B,C,0);

Note: if only one matrix is include, the function assumes that it represents D and a static gain matrix.
April 2006 Nathan Sorensen Kedrick Black 15

ufrd(usys,frequency)
Ufrd() is a function to create an uncertain frequency response model which often arises when converting uncertain state space objects to frequency response objects.
Bode Diagram 0

Magnitude (dB)

-10

-20

Parameters Include: Name uncertain system

-2

-30

-40 45 0

Phase (deg)

-45 -90 -135 -180 10 10

-1

10

10

10

Frequency (rad/sec)

Frequency Range of frequencies for ufrd model. Units specify units: rad/s default

EX: p1 = ureal('p1',5,'Range',[2 6]); p2 = ureal('p2',3,'Plusminus',0.4); p3 = ultidyn('p3',[1 1]); Wt = makeweight(.15,30,10); A = [-p1 0;p2 -p1]; B = [0;p2]; C = [1 1]; usys = uss(A,B,C,0)*(1+Wt*p3); usysfrd = ufrd(usys,logspace(-2,2,60)); bode(usysfrd,'r',usysfrd.NominalValue,'b+')

hz

April 2006

Nathan Sorensen Kedrick Black

16

A = randatom(Type)
Generate random uncertain atom objects generates a 1-by-1 type uncertain object Valid values for Type include 'ureal', 'ultidyn', 'ucomplex', and 'ucomplexm'
Example
xr = randatom('ureal') Uncertain Real Parameter: Name BMSJA, NominalValue -6.75, Range [-7.70893 -1.89278]

April 2006

Nathan Sorensen Kedrick Black

17

um = randumat(ny,nu)
Generate random uncertain umat objects generates an uncertain matrix of size ny-by-nu andumat randomly selects from uncertain atoms of type 'ureal', 'ultidyn', and 'ucomplex
Example
The following statement creates the umat uncertain object x1 of size 2by-3. Note that your result can differ because a random seed is used x1 = randumat(2,3) UMAT: 2 Rows, 3 Columns ROQAW: complex, nominal = 9.92+4.84i, radius = 0.568, 1 occurrence UEPDY: real, nominal = -5.81, variability = [-1.98681 0.133993], 3 occurrences VVNHL: complex, nominal = 5.64-6.13i, radius = 1.99, 2 occurrences

April 2006

Nathan Sorensen Kedrick Black

18

usys = randuss(n,p,m)
Generate stable, random uss objects generates an nth order uncertain continuous-time system with p outputs and m inputs
Example
The statement creates a fifth order, continuous-time uncertain system s1 of size 2-by-3. Note your display can differ because a random seed is used. s1 = randuss(5,2,3) USS: 5 States, 2 Outputs, 3 Inputs, Continuous System CTPQV: 1x1 LTI, max. gain = 2.2, 1 occurrence IGDHN: real, nominal = -4.03, variability = [-3.74667 22.7816]%, 1 occurrence MLGCD: complex, nominal = 8.36+3.09i, +/- 7.07%, 1 occurrence OEDJK: complex, nominal = -0.346-0.296i, radius = 0.895, 1 occurrence

April 2006

Nathan Sorensen Kedrick Black

19

Manipulation of Uncertain Models

1.

2.
3.

4.
5. 6.

7.

isuncertain() Checks whether argument is an uncertain class type simplify() performs model-reduction-like techniques to detect and eliminate redundant copies of uncertain elements usample() substitutes N random samples of the uncertain objects in A, returning a certain array of size [size(A) N] usubs() used to substitute a specific value for an uncertain element of an uncertain object. gridureal() Grid ureal parameters uniformly over their range lftdata() Decompose uncertain objects into fixed normalized and fixed uncertain parts Ssbal() yields a system whose input/output and uncertain properties are the same as usys, a uss object.

April 2006

Nathan Sorensen Kedrick Black

20

B = isuncertain(A)
Checks whether argument is an uncertain class type (checks the class) does not actually verify that the input argument is truly uncertain Returns true if input argument is uncertain, false otherwise Uncertain classes are umat, ufrd, uss, ureal, ultidyn, ucomplex, ucomplexm, and udyn
Example
isuncertain(rand(3,4)) ans = 0 isuncertain(ureal('p',4)) ans = 1 isuncertain(rss(4,3,2)) ans = 0 isuncertain(rss(4,3,2)*[ureal('p1',4) 6;0 1]) ans = 1

April 2006

Nathan Sorensen Kedrick Black

21

B = simplify(A)
Simplify representations of uncertain objects performs model-reduction-like techniques to detect and eliminate redundant copies of uncertain elements After reduction, any uncertain element which does not actually affect the result is deleted from the representation
Example
Create a simple umat with a single uncertain real parameter. Select specific elements, note that result remains in class umat. Simplify those same elements, and note that class changes. p1 = ureal('p1',3,'Range',[2 5]); L = [2 p1]; L(1) UMAT: 1 Rows, 1 Columns L(2) UMAT: 1 Rows, 1 Columns p1: real, nominal = 3, range = [2 5], 1 occurrence simplify(L(1)) ans = 2 simplify(L(2)) Uncertain Real Parameter: Name p1, NominalValue 3, Range [2 5]

April 2006

Nathan Sorensen Kedrick Black

22

B = usample(A,N)
substitutes N random samples of the uncertain objects in A, returning a certain array of size [size(A) N]
Example
A = ureal('A',5); Asample = usample(A,500); size(A) ans = 1 1 size(Asample) ans = 1 1 500 class(Asample) ans = double hist(Asample(:))

April 2006

Nathan Sorensen Kedrick Black

23

B = usubs(M,StrucArray)
Substitute given values for uncertain elements of uncertain objects
Example
p = ureal('p',5); m = [1 p;p^2 4]; size(m) ans = 2 2 m1 = usubs(m,'p',5) m1 = 1 5 25 4 NamesValues.p = 5; m2 = usubs(m,NamesValues) m2 = 1 5 25 4 m1 - m2 ans = 0 0 0 0

April 2006

Nathan Sorensen Kedrick Black

24

B = gridreal(A,N)
Grid ureal parameters uniformly over their range substitutes N uniformly spaced samples of the uncertain real parameters in A The N samples are generated by uniformly gridding each ureal in A across its range

Example
gamma = ureal('gamma',4); tau = ureal('tau',.5,'Percentage',30); p = tf(gamma,[tau 1]); KI = 1/(2*tau.Nominal*gamma.Nominal); c = tf(KI,[1 0]); clp = feedback(p*c,1); subplot(2,1,1); step(gridureal(p,20),6) title('Open-loop plant step responses') subplot(2,1,2); step(gridureal(clp,20),6)

April 2006

Nathan Sorensen Kedrick Black

25

[M,Delta] = lftdata(A)
Decompose uncertain objects into fixed normalized and fixed uncertain parts partially decompose an uncertain object into an uncertain part and a normalized uncertain part Uncertain objects (umat, ufrd, uss) are represented as certain objects in feedback with block-diagonal concatenations of uncertain elements separates the uncertain object A into a certain object M and a normalized uncertain matrix Delta such that A is equal to lft(Delta,M), as shown below Example
p1 = ureal('p1',-3,'perc',40); p2 = ucomplex('p2',2); A = [p1 p1+p2;1 p2]; [M,Delta] = lftdata(A); simplify(A-lft(Delta,M)) ans = 0 0 0 0 M M= 0 0 1.0954 1.0954 0 0 0 1.0000 1.0954 1.0000 -3.0000 -1.0000 0 1.0000 1.0000 2.0000

April 2006

Nathan Sorensen Kedrick Black

26

usysout = ssbal(usys,Wc)
Ssbal() yields a system whose input/output and uncertain properties are the same as usys, a uss object.
Input: Usys - a continuous-time uncertain system Wc - the critical frequency for the bilinear prewarp transformation from continuous time to discrete time. (Default is 1)
EX: p2=ureal('p2',-17,'Range',[-19 -11]); p1=ureal('p1',3.2,'Percentage',0.43); A = [-12 p1;.001 p2]; B = [120 -809;503 24]; C = [.034 .0076; .00019 2]; usys = ss(A,B,C,zeros(2,2)) USS: 2 States, 2 Outputs, 2 Inputs, Continuous System p1: real, nominal = 3.2, variability = [-0.43 0.43]%, 1 occurrence p2: real, nominal = -17, range = [-19 -11], 1 occurrence usysout = ssbal(usys) USS: 2 States, 2 Outputs, 2 Inputs, Continuous System p1: real, nominal = 3.2, variability = [-0.43 0.43]%, 1 occurrence p2: real, nominal = -17, range = [-19 -11], 1 occurrence

Output:
Usysout - a system whose input/output and uncertain properties are the same as usys, a uss object. (The numerical conditioning of usysout is usually better than that of usys.)

April 2006

Nathan Sorensen Kedrick Black

27

ndist = actual2normalized(A,V)
the normalized distance between the nominal value of the uncertain atom A and the given value V If V is an array of values, then ndist is an array of normalized distances

April 2006

Nathan Sorensen Kedrick Black

28

avalue = normalizedactual2(A,NV)
Convert the value for an atom in normalized coordinates to the corresponding actual value converts the value for atom in normalized coordinates, NV, to the corresponding actual value
Example
a = ureal('a',3,'range',[1 5]); actual2normalized(a,[1 3 5]) ans = -1.0000 -0.0000 1.0000 normalized2actual(a,[-1 1]) ans = 1.0000 5.0000 normalized2actual(a,[-1.5 1.5]) ans = 0.0000 6.0000

April 2006

Nathan Sorensen Kedrick Black

29

umatout = stack(arraydim,umat1,umat2,...)
Construct an array by stacking uncertain matrices, models, or arrays along array dimensions of an uncertain array All models must have the same number of columns and rows
Example
Consider usys1 and usys2, two single-input/single-output uss models: zeta = ureal('zeta',1,'Range',[0.4 4]); wn = ureal('wn',0.5,'Range',[0.3 0.7]); P1 = tf(1,[1 2*zeta*wn wn^2]); P2 = tf(zeta,[1 10]); You can stack along the first dimension to produce a 2-by-1 uss array.

stack(1,P1,P1) USS: 2 States, 1 Output, 1 Input, Continuous System [array, 2 x 1] wn: real, nominal = 0.5, range = [0.3 0.7], 3 occurrences zeta: real, nominal = 1, range = [0.4 4], 1 occurrence

April 2006

Nathan Sorensen Kedrick Black

30

B = squeeze(A)
Remove singleton dimensions for umat objects returns an array B with the same elements as A but with all the singleton dimensions removed A singleton is a dimension such that size(A,dim)==1 2-D arrays are unaffected by squeeze so that row vectors remain rows
April 2006 Nathan Sorensen Kedrick Black 31

Interconnection of Uncertain Models

1. 2. 3. 4. Imp2exp() transforms a linear constraint between variables into an explicit input/output relationship Sysic Builds interconnections of certain and uncertain matrices and systems Iconnect Builds complex interconnections of uncertain matrices and systems Icsignal Specifies signal constraints described by the interconnection of components

April 2006

Nathan Sorensen Kedrick Black

32

B = imp2exp(A,yidx,uidx)
transforms a linear constraint between variables Y and U of the form A(:,[yidx;uidx])*[Y;U] = 0 into an explicit input/output relationship Y = B*U The vectors yidx and uidx refer to the columns (inputs) of A as referenced by the explicit relationship for B The constraint matrix A can be a double, ss, tf, zpk and frd object as well as an uncertain object, including umat, uss and ufrd The result B will be of the same class

April 2006

Nathan Sorensen Kedrick Black

33

sysout = sysic
Build interconnections of certain and uncertain matrices and systems requires that 3 variables with fixed names be present in the calling workspace: systemnames, inputvar and outputvar systemnames is a char containing the names of the subsystems that make up the interconnection nputvar is a char, defining the names of the external inputs to the interconnection outputvar is a char, describing the outputs of the interconnection Also requires that for every subsystem name listed in systemnames, a corresponding variable, input_to_ListedSubSystemName must exist in the calling workspace This variable is similar to outputvar - it defines the input signals to this particular subsystem as linear combinations of individual subsystem's outputs and external inputs

April 2006

Nathan Sorensen Kedrick Black

34

H = iconnect
Create empty iconnect (interconnection) objects Interconnection objects (class iconnect) are an alternative to sysic, and are used to build complex interconnections of uncertain matrices and systems An iconnect object has 3 fields to be set by the user, Input, Output and Equation Input and Output are icsignal objects, while Equation is a cell-array of equality constraints (using equate) on icsignal objects the System property is the input/output model, implied by the constraints in Equation relating the variables defined in Input and Output

April 2006

Nathan Sorensen Kedrick Black

35

v = icsignal(n,'name')
Create an icsignal object creates an icsignal object of length n, which is a symbolic column vector used with iconnect objects to specify signal constraints described by the interconnection of components internal name identifiers given by the character string argument name

April 2006

Nathan Sorensen Kedrick Black

36

Model Order Reduction and Application

Complex models are not always required for good control Optimization methods generally tend to produce controllers with at least as many states as the plant model The Robust Control Toolbox offers you an assortment of model-order reduction commands Less complex low-order approximations to plant and controller models
April 2006 Nathan Sorensen Kedrick Black 37

Model Order Reduction Tools

The Robust Control Toolbox has various controller synthesis tools outlined below: 1. reduce() returns a reduced order model 2. balancmr() Balanced model truncation via square root method 3. bstmr() Balanced stochastic model truncation (BST) via Schur method

4. hankelmr() Hankel Minimum Degree Approximation (MDA) without

balancing

5. hankelsv() Computes Hankel singular values for stable/unstable or

continuous/discrete system

April 2006

Nathan Sorensen Kedrick Black

38

Model Order Reduction Tools cont:

6. modreal() returns a set of state-space LTI objects in modal form 7. ncfmr() Balanced model truncation for normalized coprime factors. 8. shurmr() Balanced model truncation via Schur method 9. slowfast() computes the slow and fast modes decompositions of a
system

10. stabproj() computes the stable and antistable projections of a minimal

realization

11. imp2ss() produces an approximate state-space realization of a given

impulse response

April 2006

Nathan Sorensen Kedrick Black

39

GRED = reduce(G,order)
Returns a reduced order model of G Groups all the Hankel SV based model reduction routines Hankel singular values of a stable system indicate the respective state energy of the system. In many cases, the additive error method GRED=reduce(G,ORDER) is adequate to provide a good reduced order model. for systems with lightly damped poles and/or zeros, a multiplicative error method (namely, GRED=reduce(G,ORDER,'ErrorType','mult')) that minimizes the relative error between G and GRED tends to produce a better fit.
April 2006 Nathan Sorensen Kedrick Black 40

GRED = balancmr(G,order)
balanced truncation model reduction for continuous/discrete & stable/unstable plant. With only one input argument G, the function will show a Hankel singular value plot of the original model and prompt for model order number to reduce.

April 2006

Nathan Sorensen Kedrick Black

41

GRED = bstmr(G,order)
Balanced stochastic truncation (BST) model reduction for continuous/discrete & stable/unstable plant. bstmr returns a reduced order model GRED of G With only one input argument G, the function will show a Hankel singular value plot of the phase matrix of G and prompt for model order number to reduce This method guarantees an error bound on the infinity norm of the multiplicative || GRED-1(G-GRED) || or relative error || G-1(G-GRED) || for well-conditioned model reduction problems

April 2006

Nathan Sorensen Kedrick Black

42

GRED = hankelmr(G,order)
Optimal Hankel norm approximation for continuous/discrete & stable/unstable plant hankelmr returns a reduced order model GRED of G With only one input argument G, the function will show a Hankel singular value plot of the original model and prompt for model order number to reduce This method guarantees an error bound on the infinity norm of the additive error || G-GRED || for wellconditioned model reduced problems

April 2006

Nathan Sorensen Kedrick Black

43

hankelsv(G)
Compute Hankel singular values for stable/unstable or continuous/discrete system draws a bar graph of the Hankel singular values such as the following:

April 2006

Nathan Sorensen Kedrick Black

44

[G1,G2] = modreal(G,cut)
State-space modal truncation/realization returns a set of state-space LTI objects G1 and G2 in modal form given a state-space G and the model size of G1, cut The modal form realization has its A matrix in block diagonal form with either 1x1 or 2x2 blocks. The real eigenvalues will be put in 1x1 blocks and complex eigenvalues will be put in 2x2 blocks. These diagonal blocks are ordered in ascending order based on eigenvalue magnitudes G can be stable or unstable

April 2006

Nathan Sorensen Kedrick Black

45

GRED = ncfmr(G,order)
balanced truncation model reduction for normalized coprime factors of continuous/discrete & stable/unstable G ncfmr returns a reduced order model GRED formed by a set of balanced normalized coprime factors With only one input argument G, the function will show a Hankel singular value plot of the original model and prompt for model order number to reduce The left and right normalized coprime factors are defined as:

April 2006

Nathan Sorensen Kedrick Black

46

GRED = schurmr(G,order)
Schur balanced truncation model reduction for continuous/discrete & stable/unstable plant schurmr returns a reduced order model GRED of G With only one input argument G, the function will show a Hankel singular value plot of the original model and prompt for model order number to reduce This method guarantees an error bound on the infinity norm of the additive error || G-GRED || for well-conditioned model reduced problems
April 2006 Nathan Sorensen Kedrick Black 47

G1,G2] = slowfast(G,cut)
State-space slow-fast decomposition computes the slow and fast modes decompositions of a system G(s) such that: Gs Gs Gs
s f

,B ,D ,C ) Gs f : ( A 22 2 2 2

,B ,D ,C ) Gs s : ( A 11 1 1 1

April 2006

Nathan Sorensen Kedrick Black

48

[G1,G2,m] = stabproj(G)
State-space stable/anti-stable decomposition stabproj computes the stable and antistable projections of a minimal realization G(s) such that:

April 2006

Nathan Sorensen Kedrick Black

49

[a,b,c,d,totbnd,hsv] = imp2ss(y,ts,nu,ny,tol)
System identification via impulse response The function imp2ss produces an approximate state-space realization of a given impulse response imp=mksys(y,t,nu,ny,'imp'); A continuous-time realization is computed via the inverse Tustin transform if t is positive; otherwise a discrete-time realization is returned In the SISO case the variable y is the impulse response vector; in the MIMO case y is an N+1-column matrix containing N + 1 time samples of the matrix-valued impulse response H0, ..., HN of an nuinput, ny-output system stored row-wise. The variable tol bounds the H norm of the error between the approximate realization (a, b, c, d) and an exact realization of y The inputs ts, nu, ny, tol are optional; if not present they default to the values ts = 0, nu = 1, ny = (number of rows of y)/nu, tol = 1.01sigmabar1
April 2006 Nathan Sorensen Kedrick Black 50

Robustness and Worst Case Analysis

To be robust, your control system should meet your stability and performance requirements for all possible values of uncertain parameters Monte Carlo parameter sampling via usample can be used for this purpose, but Monte Carlo methods are inherently hit or miss The Robust Control Toolbox gives you a powerful assortment of robustness analysis commands that let you directly calculate upper and lower bounds on worst-case performance without random sampling:

April 2006

Nathan Sorensen Kedrick Black

51

Performance Analysis

The hyperbola is used to define the performance margin. Systems whose performance degradation curve intersects high on the hyperbola curve represent "non-robustly-performing systems" in that very small deviations of the uncertain elements from their nominal values can result in very large system gains. Conversely, an intersection low on the hyperbola represent "robustly-performing systems."

April 2006

Nathan Sorensen Kedrick Black

52

Small Gain Theorem

When the plant modeling uncertainty is not too big, you can design high-gain, high-performance feedback controllers. High loop gains significantly larger than 1 in magnitude can attenuate the effects of plant model uncertainty and reduce the overall sensitivity of the system to plant noise. But if your plant model uncertainty is so large that you do not even know the sign of your plant gain, then you cannot use large feedback gains without the risk that the system will become unstable. Thus, plant model uncertainty can be a fundamental limiting factor in determining what can be achieved with feedback. The Small Gain Theorem asserts that the overall system gain must therefore be kept below 1 to assure stability.
April 2006 Nathan Sorensen Kedrick Black 53

Robustness and Worst Case Performance Tools

The Robust Control Toolbox has various controller synthesis tools outlined below:
1. cpmargin() calculates the normalized coprime factor/gap metric robust stability of the multivariable feedback loop consisting of C in negative feedback with P. 2. Gapmetric() calculates upper bounds on the gap and nugap metric between systems 3. loopmargin() analyzes the multivariable feedback loop consisting of the loop transfer matrix. 4. loopsens() Creates a Struct, loops, whose fields contain the multivariable sensitivity, complementary and open-loop transfer functions. 5. mussv() calculates upper and lower bounds on the structured singular value, or , for a given block structure.

6. Mussvextract() is used to extract the compressed information within muinfo into a readable form.

April 2006

Nathan Sorensen Kedrick Black

54

Robustness and Worst Case Performance Tools cont:

7. ncfmargin() Calculates the normalized coprime stability margin of the plant-controller feedback loop 8. popov() uses the Popov criterion to test the robust stability of dynamical systems with possibly nonlinear and/or time-varying uncertainty. 9. robustperf() Calculates the robust performance margin of an uncertain multivariable system. 10. robuststab() Calculate robust stability margins of an uncertain multivariable system. 11. robopt() Creates an options object for use with robuststab and robustperf 12. wcnorm() Calculate the worst-case norm of an uncertain matrix

April 2006

Nathan Sorensen Kedrick Black

55

Robustness and Worst Case Performance Tools cont:

13. wcgain() Calculates bounds on the worst-case gain of an uncertain system

14. wcgopt() creates a wcgain, wcsens and wcmargin options object called options in which specified properties have specific values.
15. wcmargin() calculates the combined worst-case input and output loop-at-a-time gain/phase margins of the feedback loop 16. wcsens() Calculate the worst-case sensitivity and complementary sensitivity functions of a plant-controller feedback loop

April 2006

Nathan Sorensen Kedrick Black

56

[MARG,FREQ] = cpmargin(P,C,TOL)
Cpmargin() calculates the normalized coprime factor/gap metric robust stability of the multivariable feedback loop consisting of C in negative feedback with P.

Input Parameters: P LTI plant C only feedback controller Tol - specifies a relative accuracy for calculating the normalized coprime factor/gap metric robust stability margin

Output Parameters:

Marg - contains upperand lower bound for the normalized coprime factor/gap metric robust stability margin
Freq - the frequency associated with the upper bound

April 2006

Nathan Sorensen Kedrick Black

57

[gap,nugap] = gapmetric(p0,p1,tol)
Gapmetric() calculates upper bounds on the gap and nugap metric between systems p0 and p1.
Bode Diagram Bode Diagram

Phase (deg) Magnitude (dB)

Phase (deg) Magnitude (dB)

50 0 -50 0 -45 -90 -2 10

0 2

Phase (deg) Magnitude (dB)

EX: p1 = tf(1,[1 -0.001]); p2 = tf(1,[1 0.001]); [g,ng] = gapmetric(p1,p2) K = 1; H1 = loopsens(p1,K); H2 = loopsens(p2,K); subplot(2,2,1); bode(H1.Si,'-',H2.Si,'--'); subplot(2,2,2); bode(H1.Ti,'-',H2.Ti,'--'); subplot(2,2,3); bode(H1.PSi,'-',H2.PSi,'--'); subplot(2,2,4); bode(H1.CSo,'-',H2.CSo,'--');
g = 0.0029 ng = 0.0020

0 -50 -100 180 90 0 -5 10

0

Phase (deg) Magnitude (dB)

50 0 -50 0 -45 -90 -2 10

0 2

10

10

10

Frequency (rad/sec) Bode Diagram 0 -50 -100 180 90 0 -5 10

Frequency (rad/sec) Bode Diagram

10

10

10

Frequency (rad/sec)

Frequency (rad/sec)

Input: P0,P1 plants of the same size Tol tolerance for the gap metric (default is .001)
April 2006

Output: Gap gap metric value Nugap nugap metric value (For gap or nugap, a value close to 0 means the plants are far apart and a value close to 1 means they are close together.)
58

Nathan Sorensen Kedrick Black

loopmargin(L)

[SM, DM, MM] = Loopmargin() is a tool to determine the closed loop margins for an uncertain system. Outputs include: SM A structure containing the nominal system margins. DM also called disk margin. Describes the disk of deviations from the nominal system. MM establishes guaranteed bounds on the system with variations from all input channels.
April 2006 Nathan Sorensen Kedrick Black 59

loopsens(P,C)
Loopsens() creates a structure which contains the multivariable sensitivity, complementary and open-loop transfer functions.
Structure Output Fields:
Poles - Closed-loop poles. NaN for frd/ufrd objects. Stable - 1 if nominal closed loop is stable, 0 otherwise. NaN for frd/ufrd objects. Si - Input-to-plant sensitivity function Ti - Input-to-plant complementary sensitivity function Li - Input-to-plant loop transfer function So - Output-to-plant sensitivity function To - Output-to-plant complementary sensitivity function Lo - Output-to-plant loop transfer function PSi - Plant times input-to-plant sensitivity function CSo - Compensator times output-toplant sensitivity function 60

EX: int = tf(1,[1 0]); addunc = ultidyn('addunc',[1 1],'Bound',0.2); uncint = int + addunc; int3 = blkdiag(int,uncint,int); p = [1 .2 .5;-.5 0 1;1 0 0;0 1 0]*int3; gain = ureal('gain',0.9, 'Range',[-3 1.5]); c = [.02 .3 0 0;1 -.3 .32 gain; 0 1 -.2 .1]; loops = loopsens(p,c) April 2006

Nathan Sorensen Kedrick Black

[bounds,muinfo] = mussv(M,BlockStructure)
Mussv() calculates upper and lower bounds on the structured singular value, or , for a given block structure.
Input: M - is a double or a frd object BlockStructure - a matrix specifying the perturbation block structure consisting of 2 rows and as many columns as uncertainty blocks. Output: Bounds upper and lower bounds on the mu value of the system.

EX: M = randn(5,5) + sqrt(-1)*randn(5,5); F = randn(2,5) + sqrt(-1)*randn(2,5); BlockStructure = [-1 0;-1 0;1 1;2 0]; [ubound,Q] = mussv(M,F,BlockStructure); points completed (of 1) ... 1 bounds = mussv(M,BlockStructure); optbounds = mussv(M+Q*F,BlockStructure,'C5');
[optbounds(1) ubound] 1.5917 1.5925 [bounds(1) bounds(2)] 3.8184 3.7135

Muinfo a structure containing additional information

April 2006

Nathan Sorensen Kedrick Black

61

[VDelta,VSigma,VLmi] = mussvextract(muinfo)
Mussvextract() is used to extract the compressed information within muinfo into a readable form.
Input: Muinfo a variable returned by mussv() with information Ouput:

VDelta used to verify the lower bound

VSigma a structure of information used to verify the Newlin/Young upper bound

EX: M = randn(4,4) + sqrt(-1)*randn(4,4); BlockStructure = [1 1;1 1;2 2]; [bounds,muinfo] = mussv(M,BlockStructure); [VDelta,VSigma,VLmi] = mussvextract(muinfo);

VLmi a structure of information used to verify the LMI upper bound

April 2006 Nathan Sorensen Kedrick Black 62

[marg,freq] = ncfmargin(P,C,tol)
Ncfmargin() calculates the normalized coprime factor/gap metric robust stability margin b(P, C), marg, of the multivariable feedback loop consisting of C in negative feedback with P.
Input:

P LTI plant model

C Compensator in the feedback path Tol tolerance for the ncfmargin must be between .00001 and .01

Output:
Marg - the normalized coprime factor/gap metric robust stability margin b(P, C) Freq destabilizing frequency

EX: x = tf(4,[1 0.001]); clp1 = feedback(x,1); clp2 = feedback(x,10); [marg1,freq1] = ncfmargin(x,1); [marg2,freq2] = ncfmargin(x,10);
marg1 = 0.7071 freq1 = Inf marg2 = 0.0995 freq2 = Inf
63

April 2006

Nathan Sorensen Kedrick Black

[t,P,S,N] = popov(sys,delta,flag)
Perform the Popov robust stability test uses the Popov criterion to test the robust stability of dynamical systems with possibly nonlinear and/or time-varying uncertainty The uncertain system must be described as the interconnection of a nominal LTI system sys and some uncertainty delta
April 2006 Nathan Sorensen Kedrick Black 64

[perfmarg,perfmargunc,Report] = robustperf(sys)
Robustperf() is a tool to determine the robust performance margin which sets bounds on the robustness of a nominally stable system to given uncertainty.
EX:

Structure Output Fields: Perfmarg performance margin: lower bound, upper bound, and critical frequency. Perfmargunc structure of values of critical uncertain elements of the system. Report A text description of the robustness analysis.

P = tf(1,[1 0]) + ultidyn('delta',[1 1],'bound',0.4); BW = 0.8; K = tf(BW,[1/(25*BW) 1]); S = feedback(1,P*K); [perfmargin,punc] = robustperf(S); perfmargin = UpperBound: 7.4305e-001 LowerBound: 7.4305e-001 CriticalFrequency: 5.3096e+000

Note: Output may not contain true critical frequency because of the assigned frequency range over which the function operates. For limitations, see the function definition in the matlab help.

April 2006

Nathan Sorensen Kedrick Black

65

[stabmarg,destabunc,Report] = robuststab(sys)
Robuststab() is used to determine the robust stability margin for a nominally stable uncertain system up to the closest instability from the stable nominal.
Structure Output Fields: Stabmarg robust stability margin: lower bound, upper bound, destabilizing frequency. Destabunc structure of uncertain values closest to the nominal that cause instability. Report text description of the stability analysis
Note: Output may not contain all unstable frequencies because of the assigned frequency range over which the function operates. For limitations, see the function definition in the matlab help. EX: P = tf(4,[1 .8 4]); delta = ultidyn('delta',[1 1],'SampleStateDim',5); Pu = P + 0.25*tf([1],[.15 1])*delta; C = tf([1 1],[.1 1]) + tf(2,[1 0]); S = feedback(1,Pu*C); [stabmarg,destabunc,report,info] = robuststab(S);

report = Uncertain System is NOT robustly stable to modeled uncertainty. -- It can tolerate up to 81.8% of modeled uncertainty. -- A destabilizing combination of 81.8% the modeled uncertainty exists, causing an instability at 9.13 rad/s. -- Sensitivity with respect to uncertain element ... 'delta' is 100%. Increasing 'delta' by 25% leads to a 25% decrease in the margin.

April 2006

Nathan Sorensen Kedrick Black

66

opts = robopt('name1',value1,'name2',value2,...)
Robopt() creates an options object for use with robuststab and robustperf.

EX: opt = robopt Property Object Values: Display: 'off' Sensitivity: 'on' VaryUncertainty: 25 Mussv: 'sm9' Default: [1x1 struct] Meaning: [1x1 struct]

April 2006

Nathan Sorensen Kedrick Black

67

[maxnorm,wcu] = wcnorm(m)
Wcnorm() determines the maximum norm over all allowable values of the uncertain elements and is referred to as a worst-case norm analysis.
Input: M - a umat or a uss Output:
EX: a=ureal('a',5,'Range',[4 6]); b=ureal('b',2,'Range',[1 3]); b=ureal('b',3,'Range',[2 10]); c=ureal('c',9,'Range',[8 11]); d=ureal('d',1,'Range',[0 2]); M = [a b;c d]; Mi = inv(M); [maxnormM] = wcnorm(M) maxnormM = LowerBound: 14.7199 UpperBound: 14.7327 [maxnormMi] = wcnorm(Mi) maxnormMi = LowerBound: 2.5963 UpperBound: 2.5979

Maxnorm a structure with fields:

Wcu - a structure that includes values of uncertain elements and maximizes the matrix norm
April 2006

Nathan Sorensen Kedrick Black

68

[maxgain,maxgainunc] = wcgain(sys)
Wcgain() calculates the worst case gain frequency response of a given system due to uncertain elements. Pointwise-in-frequency analysis creates a response curve for the system.
Structure Output Fields: Maxgain maximum gain: lower bound, upper bound, critical frequency. Maxgainunc structure containing uncertain elements that maximize the gain.
EX: P = tf(1,[1 0]) + ultidyn('delta',[1 1],'bound',0.4); BW1 = 0.8; K1 = tf(BW1,[1/(25*BW1) 1]); S1 = feedback(1,P*K1); BW2 = 2.0; K2 = tf(BW2,[1/(25*BW2) 1]); S2 = feedback(1,P*K2); [maxgain1,wcunc1] = wcgain(S1); maxgain1 = LowerBound: 1.5070e+000 UpperBound: 1.5080e+000 CriticalFrequency: 5.3096e+000 Note: Output may not contain all unstable frequencies because of the assigned frequency range over which the function operates. For limitations, see the function definition in the matlab help .

Options set in wcopt():

Pointwise-in-frequency- creates a response curve for the system. Peak-over-frequency computes the largest gain over all freqencies.

April 2006

Nathan Sorensen Kedrick Black

69

options = wcgopt('name1',value1,'name2',value2,...)
Wcgopt() creates a wcgain, wcsens and wcmargin options object called options in which specified properties have specific values.
EX: opt = wcgopt('MaxTime',10000,'Sensitivity','off') Property Object Values: Sensitivity: 'off' LowerBoundOnly: 'off' FreqPtWise: 0 ArrayDimPtWise: [] VaryUncertainty: 25 Default: [1x1 struct] Meaning: [1x1 struct] AbsTol: 0.0200 RelTol: 0.0500 MGoodThreshold: 1.0400 AGoodThreshold: 0.0500 MBadThreshold: 20 ABadThreshold: 5 NTimes: 2 MaxCnt: 3 MaxTime: 10000

April 2006

Nathan Sorensen Kedrick Black

70

[wcmargi,wcmargo] = wcmargin(L)
Wcmargin() describes a disk within with the nominal gain margin and phase margin are stable for an uncertain system.
EX: a = [0 10;-10 0]; b = eye(2); c = [1 8;-10 1]; d = zeros(2,2); G = ss(a,b,c,d); K = [1 -2;0 1]; ingain1 = ureal('ingain1',1,'Range',[0.97 1.06]); b = [ingain1 0;0 1]; Gunc = ss(a,b,c,d); unmod = ultidyn('unmod',[2 2],'Bound',0.08); Gmod = (eye(2)+unmod)*Gunc; Gmodg = ufrd(Gmod,logspace(-1,3,60)); [wcmi,wcmo] = wcmargin(Gmodg,K); wcmi(1) = GainMargin: [0.3613 2.7681] PhaseMargin: [-50.2745 50.2745] Frequency: 0.1000 Sensitivity: [1x1 struct]

Output Structure Fields: Wcmargi contains gain margin bound, phase margin in degrees, frequency associate with worst case, and a structure containing the sensitivity to all uncertain elements.

Wcmargo same structure as wcmargi but gives single loop values for each.
April 2006

Nathan Sorensen Kedrick Black

71

wcst = wcsens(L)
Wcsens() determines the worst case sensitivity of a system to various input and disturbance parameters in an uncertain system L.
Bode Diagram

Output Structure Fields:

Magnitude (dB)

Si - Worst-case input-to-plant sensitivity function

Ti - Worst-case input-to-plant complementary sensitivity function So - Worst-case output-to-plant sensitivity function To - Worst-case output-to-plant complementary sensitivity function

-5

-10

-15

-20

-25

-30

Nominal Sensitivity Worst-Case Sensitivity

-35 -1 10

10

10

Frequency (rad/sec)

EX: delta = ultidyn('delta',[1 1]); tau = ureal('tau',5,'range',[4 6]); P = tf(1,[tau 1])*(1+0.25*delta); PSi - Worst-case plant times input-to-plant C=tf([4 4],[1 0]); sensitivity function looptransfer = loopsens(P,C); CSo - Worst-case compensator times output- Snom = looptransfer.Si.NominalValue; wcst = wcsens(P,C) ; to-plant sensitivity function Swc = wcst.Si.BadSystem; Stable - 1 if nominal closed loop is stable, 0 omega = logspace(-1,1,50); bodemag(Snom,'-',Swc,'-.',omega); otherwise. NaN for frd/ufrd objects. legend('Nominal Sensitivity','Worst-Case Sensitivity',... 'Location','SouthEast')

April 2006

Nathan Sorensen Kedrick Black

72

Robust Analysis for Parameter Dependent Systems

1. 2. 3. psys() specifies state-space models where the state-space matrices can be uncertain, time-varying, or parameter-dependent. psinfo() a multi-usage function for queries about a polytopic or parameter-dependent system ispsys() returns 1 if sys is a polytopic or parameter-dependent system pvec() used in conjunction with psys to specify parameterdependent systems. pvinfo() retrieves information about a vector of real parameters declared with pvec and stored in pv polydec() takes an uncertain parameter vector PV taking values ranging in a box, and returns the corners or vertices of the box as columns of the matrix vertx

4.
5. 6.

April 2006

Nathan Sorensen Kedrick Black

73

Robust Analysis for Parameter Dependent Systems cont.

6. 7. 8. aff2pol() Derives a polytopic representation polsys of the affine parameter- dependent system. quadstab() seeks a fixed Lyapunov function that establishes quadratic stability quadperf() Compute the quadratic H-inf performance of a polytopic or parameter-dependent system pdlstab() uses parameter-dependent Lyapunov functions to establish the stability of uncertain state-space models over some parameter range or polytope of systems. decay() returns the quadratic decay rate drate pdsimul() simulates the time response of an affine parameterdependent system

9.
10. 11.

April 2006

Nathan Sorensen Kedrick Black

74

pols = psys(syslist) affs = psys(pv,syslist)

Specify a parameter-dependent system psys specifies state-space models where the state-space matrices can be uncertain, time-varying, or parameter-dependent.

a polytopic model with vertex systems S1, . . ., S4 is created by pols = psys([s1,s2,s3,s4]) while an affine parameter-dependent model with 4 real parameters is defined by affs = psys(pv,[s0,s1,s2,s3,s4]) The output is a structured matrix storing all the relevant information
April 2006 Nathan Sorensen Kedrick Black 75

[type,k,ns,ni,no] = psinfo(ps)
Query characteristics of a P-system displays the type of system (affine or polytopic) the number k of SYSTEM matrices involved in its definition the numbers of ns, ni, no of states, inputs, and outputs of the system
April 2006 Nathan Sorensen Kedrick Black 76

bool = ispsys(sys)
True for parameter-dependent systems returns 1 if sys is a polytopic or parameterdependent system

April 2006

Nathan Sorensen Kedrick Black

77

pv = pvec('box',range,rates) pv = pvec('pol',vertices)
Specify a vector of uncertain or time-varying parameters used in conjunction with psys to specify parameter-dependent systems The type 'box' corresponds to independent parameters ranging in intervals (extremal values) The parameter vector p then takes values in a hyperrectangle of Rn called the parameter box range is an n-by-2 matrix that stacks up the extremal values type 'pol' corresponds to parameter vectors p ranging in a polytope of the parameter space Rn efined by a set of vertices V1, . . ., Vn corresponding to "extremal" values of the vector p second argument is the concatenation of the vectors v1,...,vn

April 2006

Nathan Sorensen Kedrick Black

78

[typ,k,nv] = pvinfo(pv)
pvinfo(pv) displays the type of parameter vector ('box' or 'pol'), the number n of scalar parameters, and for the type 'pol', the number of vertices used to specify the parameter range

April 2006

Nathan Sorensen Kedrick Black

79

vertx = polydec(PV)
Compute polytopic coordinates wrt. box corners takes an uncertain parameter vector PV taking values ranging in a box, and returns the corners or vertices of the box as columns of the matrix vertx

April 2006

Nathan Sorensen Kedrick Black

80

polsys = aff2pol(affsys)
Convert affine P-systems to polytopic representation derives a polytopic representation polsys of the affine parameter- dependent system

Where p = (p1, . . ., pn) is a vector of uncertain or time- varying real parameters taking values in a box or a polytope. The description affsys of this system should be specified with psys

April 2006

Nathan Sorensen Kedrick Black

81

[tau,P] = quadstab(ps,options)
Assess quadratic stability of parameter-dependent system Quadratic stability of polytopic or affine parameter-dependent systems

April 2006

Nathan Sorensen Kedrick Black

82

[perf,P] = quadperf(ps,g,options)
Assess quadratic Hinf performance of P-systems

April 2006

Nathan Sorensen Kedrick Black

83

[tau,Q0,Q1,...] = pdlstab(pds,options)
Assess the robust stability of a polytopic or parameter-dependent system uses parameter-dependent Lyapunov functions to establish the stability of uncertain state-space models over some parameter range or polytope of systems

April 2006

Nathan Sorensen Kedrick Black

84

[drate,P] = decay(ps,options)
Compute quadratic decay rate

April 2006

Nathan Sorensen Kedrick Black

85

pdsimul(pds,'traj',tf,'ut',xi,options)
Time response of a parameter-dependent system along a given parameter trajectory along a parameter trajectory p(t) and for an input signal u(t) The parameter trajectory and input signals are specified by two time functions p=traj(t) and u=ut(t) The affine system pds is specified with psys but can also use the polytopic representation of such systems as returned by aff2pol(pds) or hinfgs The final time and initial state vector can be reset through tf and xi options gives access to the parameters controlling the ODE integration

April 2006

Nathan Sorensen Kedrick Black

86

Robust Controller Synthesis Tools

The Robust Control Toolbox has various controller synthesis tools outlined below:
1. Augw() computes a state-space model of an augmented LTI plant with weighting functions
2. Hinfmix() creates a suboptimal controller K(s) based on the optimization of a mixed H2/ H criterion. 3. H2syn() computes a stabilizing LTI/SS controller K for a partitioned plant P(s). 4. Hinfsyn() creates an optimal H-inifinity controller for a partitioned plant matrix P. 5. Sdhinfsyn() synthesizes a sampled data H-infinity controller. 6. Loopsyn() is an H infinity optimal way for loopshaping controller synthisis.

April 2006

Nathan Sorensen Kedrick Black

87

Robust Controller Synthesis Tools cont:

7. Ltrsyn() computes a reconstructed-state output-feedback controller K for LTI plant G using Kalman filtering to asymptotically recover the lost states.
8. Mixsyn() is H-infinity mixed-sensitivity synthesis method for robust control loopshaping design. 9. Ncfsyn() is a normalized coprime factorization based parameterized controller synthesis using the Glover-McFarlane method for loopshaping. 10.Mkfilter() returns a single-input, single-output analog low pass filter sys as an ss object.

April 2006

Nathan Sorensen Kedrick Black

88

P = AUGW(G,W1,W2,W3)
Augw() computes a state-space model of an augmented LTI plant P(s) with weighting functions W1(s), W2(s), and W3(s) penalizing the error signal, control signal and output signal respectively.
Singular Values 60 S = 1/(1+L) GAM/W1 40 T=L/(1+L) GAM*G/W2 20

Singular Values (dB)

-20

-40

-60

Input:
10
-2

-80 -4 10

10

10

10

10

Frequency (rad/sec)

G Nominal Plant Model W1 weight for the error signal W2 weight for the control signal W3 weight for the output signal

EX: s=zpk('s'); G=(s-1)/(s+1); W1=0.1*(s+100)/(100*s+1); W2=0.1; W3=[]; P=augw(G,W1,W2,W3); [K,CL,GAM]=hinfsyn(P); [K2,CL2,GAM2]=h2syn(P); L=G*K; S=inv(1+L); T=1-S; sigma(S,'k',GAM/W1,'k-.',T,'r',GAM*G/W2,'r-.') April 2006

Output:
P augmented plant model 89

Nathan Sorensen Kedrick Black

[gopt,h2opt,K,R,S] = hinfmix(P,r,obj,region,dkbnd,tol)
Hinfmix() creates a suboptimal controller K(s) based on the optimization of a mixed H2/ H criterion that optimizes the following function: subject to ||T || < gamma,
||T2||2 < v. The closed-loop poles lie in some prescribed LMI region D. Input Parameters:
P lmi object such as ss or tf.

R [z2, y, u]
Obj [gamma, v, alpha, beta] Region identifies the pole placement region. Default is the open LHP.

Dkbnd set bound on norm of controller feed through matrix DK.

Tol tolerance for the trade-off parameter
April 2006

Output Parameters:
Gopt - guaranteed Hinf performance H2opt guaranteed H2 performance K Controller K(s) coefficients R,S optimal values of the LMI objects.
90

Nathan Sorensen Kedrick Black

[K,CL,GAM,INFO]= H2SYN(P,NMEAS,NCON)
H2syn() computes a stabilizing LTI/SS controller K for a partitioned plant P(s).
Input Parameters:
P LTI/SS plant model Nmeas B2 column size Ncon C2 row size Output Parameters: K vector of Feedback Controller Coefficients CL Closed loop system matrix Plant Partitioning (if done by mktito(), nmeas and ncon can be omitted): B1 Inputs are disturbances

Gam the H2 norm of the closed loop system

Info structure containing additional information.
April 2006

B2 Inputs are control signals

C1 Outputs are errors to be minimized C2 Outputs are inputs to control K
91

Nathan Sorensen Kedrick Black

[K,CL,GAM,INFO] = hinfsyn(P,NMEAS,NCON)
Hinfsyn() creates an optimal h-inifinity controller for a partitioned plant matrix P.
Input Parameters: P Partitioned plant matrix Nmeas row size of the C2 matrix Ncon column size of the B2 matrix Output Parameters: K controller coefficient vector CL closed loop system matrix Gam H-inifinity norm for the closed loop system Info - structure containing additional information
April 2006

Plant Partitioning (if done by mktito(), nmeas and ncon can be omitted): B1 Inputs are disturbances B2 Inputs are control signals C1 Outputs are errors to be minimized C2 Outputs are inputs to control K
92

Nathan Sorensen Kedrick Black

[K,GAM]=sdhinfsyn(P,NMEAS,NCON)
Sdhinfsyn() synthesizes a sampled data H-infinity controller.

Input Parameters: P partitioned system matrix Nmeas row size of C2 NCon column size of B2 Output parameters: K H infinity controller coefficients Gam H infinity norm cost for the system
April 2006

Where:
Plant Partitioning: B1 Inputs are disturbances B2 Inputs are control signals C1 Outputs are errors to be minimized C2 Outputs are inputs to control K
93

Nathan Sorensen Kedrick Black

[K,CL,GAM,INFO]= loopsyn(G,Gd,RANGE)
Loopsyn() is an H infinity optimal way for loopshaping controller synthisis.
Singular Values 50 40 30 20

Input Parameters:
G LTI plant

Gd desired loop shape Range [Wmin,Wmax] desired loop shaping frequency range. (10*Wmin <= Wmax) Output Parameters:
10
0

Singular Values (dB)

10 0 -10 -20 -30 -40 -1 10

10 Frequency (rad/sec)

10

K Controller Coefficient vector CL Closed Loop System Matrix

EX:
rand('seed',0);randn('seed',0); s=tf('s'); w0=5; Gd=5/s; G=((s-10)/(s+100))*rss(3,4,5); [K,CL,GAM,INFO]=loopsyn(G,Gd); sigma(G*K,'r',Gd*GAM,'k-.',Gd/GAM,'k-.',{.1,100}) April 2006

Gam H infinity norm for the closed loop system

Info structure containing additional information.
94

Nathan Sorensen Kedrick Black

[K,SVL,W1] = ltrsyn(G,F,XI,THETA,RHO)
Ltrsyn() computes a reconstructed-state output-feedback controller K for LTI plant G using Kalman filtering to asymptotically recover the lost states.
Input Parameters: G LTI System matrix F - LQ full-state-feedback gain matrix XI plant noise intensity Theta sensor noise intensity Rho vector containing a set of recovery gains EX:

s=tf('s');G=ss(1e4/((s+1)*(s+10)*(s+100)));[A,B,C, D]=ssdata(G); Output Parameters: F=lqr(A,B,C'*C,eye(size(B,2))); K controller coefficient vector L=ss(A,B,F,0*F*B); XI=100*C'*C; THETA=eye(size(C,1)); Svl sigma plot data for recovered system RHO=[1e3,1e6,1e9,1e12];W=logspace(-2,2); nyquist(L,'k-.');hold; W1 - frequencies for SVL plots [K,SVL,W1]=ltrsyn(G,F,XI,THETA,RHO,W); April 2006 Nathan Sorensen Kedrick Black 95

[K,CL,GAM,INFO]= mixsyn(G,W1,W2,W3)
Mixsyn() is H-infinity mixed-sensitivity synthesis method for robust control loopshaping design.

Input Parameters:
G LTI system W1,W2,W3 Weighting matricies associated with S(s), R(s) and T(s)

Mixsyn() creates the h-infinity controller based on the sensitivity function S(s) and the complimentatry sensitivitiy T(s) function according to the weighting matricies W1, W2, W3.;
April 2006

Output Parameters:

K Controller Coefficient vector CL Closed Loop System Matrix Gam H infinity norm for the closed loop system Info structure containing additional information. Nathan Sorensen Kedrick Black 96

[K,CL,GAM,INFO]= ncfsyn(G,W1,W2)
Ncfsyn() is a normalized coprime factorization based parameterized controller synthisis using the Glover-McFarlane method for loopshaping.
Singular Values 60

Input Parameters:
G LTI plant model W1,W2 Pre and Post compensator to give Gd = W1GW2 a good high frequency and low frequency response Output Parameters: K controller coefficient vector

40

20

Singular Values (dB)

-20

-40

-60

-80

-100 -2 10

10

-1

10

10

10

EX:

Frequency (rad/sec)

CL closed loop system matrix

s=zpk('s'); G=(s-1)/(s+1)^2; W1=0.5/s; [K,CL,GAM]=ncfsyn(G,W1); sigma(G*K,'r',G*W1,'r-.',G*W1*GAM,'k.',G*W1/GAM,'k-.') April 2006

Gam H infinity optimal cost

Info structure containing additional information
97

Nathan Sorensen Kedrick Black

sys = mkfilter(fc,ord,type,psbndr)
Mkfilter() returns a single-input, single-output analog low pass filter sys as an ss object.
Bode Diagram Bode Diagram

Phase (deg) Magnitude (dB)

0 -50 -100 180 90 0 -5 10

0

Phase (deg) Magnitude (dB)

50 0 -50 0 -45
0 2

Input:
Fc cutoff frequency of the filter Ord order of the filter Type type of the filter. 'butterw- Butterworth filter cheby - Chebyshev filter 'bessel - Bessel filter 'rc - Series of resistor/ capacitor filters Psbndr - the magnitude of the passband

10

-90 -2 10 Butterw orth Chebyshev RC filter

10

10

Frequency (rad/sec) Bode Diagram

Frequency (rad/sec) Bode Diagram

Phase (deg) Magnitude (dB)

50 0 -50 0 -45 -90 -2 10

0 2

Phase (deg) Magnitude (dB)

0 -100 -200 0 -180 -360

0 2

10

10

10

10

Frequency (rad/sec)

Frequency (rad/sec)

EX:
butw = mkfilter(2,4,'butterw'); cheb = mkfilter(4,4,'cheby',0.5); rc = mkfilter(1,4,'rc'); bode(butw,'-',cheb,'--',rc,'-.') 98

Output:
Sys Designed filter April 2006

Nathan Sorensen Kedrick Black

Mu-synthesis
Mu-synthesis functions enable the synthesis of robust optimal controllers. Approximation tools for this type of design are shown below:
1. Cmsclsyn() approximately solves the constant-matrix, upper bound synthesis problem. 2. Dksyn() synthesizes a robust controller via D-K iteration which is an approximation to -synthesis control design.

3. Dkitopt() creates a dkitopt object called options with specific values assigned to certain properties.
4. Drawmag() interactively uses the mouse in the plot window to create pts and sysout, which approximately fits the frequency response (magnitude) in pts. 5. Fitfrd() fits D-scaling frequency response data with state-space model. 6. Fitmagfrd() fits frequency response magnitude data with a stable, minimum-phase state-space model.
April 2006 Nathan Sorensen Kedrick Black 99

[qopt,bnd] = cmsclsyn(R,U,V,BlockStructure);
Cmsclsyn() approximately solves the constant-matrix, upper bound synthesis problem by minimization, for given matrices applies to constant matrix data in R, U, and V. Input: R,U,V Constant matrices of appropriate structure BlockStructure a matrix specifying the perturbation blockstructure as defined for mussv. and . This

Output: Qopt - the optimum value of Q Bnd - the upper bound of mussv(R+U*Q*V,BLK)

April 2006

Nathan Sorensen Kedrick Black

100

[k,clp,bnd] = dksyn(p,nmeas,ncont)
Dksyn() synthesizes a robust controller via D-K iteration which is an approximation to -synthesis control design.

Input: P Partitioned plant matrix Nmeas row size of the C2 matrix Ncon column size of the B2 matrix Output: K synthesized controller coefficients Clp Closed loop system Bnd robust performance bounds on the closed loop system
April 2006

EX: G = tf(1,[1 -1]); Wu = 0.25*tf([1/2 1],[1/32 1]); InputUnc = ultidyn('InputUnc',[1 1]); Gpert = G*(1+InputUnc*Wu); Wp = tf([1/4 0.6],[1 0.006]); P = [Wp; 1 ]*[1 Gpert]; [K,clp,bnd1] = dksyn(P,1,1); P = [Wp; 1 ]*[1 Gpert]; [K,clp,bnd2] = dksyn(P,1,1); bnd1 =0.6768 Nathan Sorensen Kedrick Black bnd2 =0.6768 101

options = dkitopt('name1',value1,'name2',value2,...)
Dkitopt() creates a dkitopt object called options with specific values assigned to certain properties.

April 2006

Nathan Sorensen Kedrick Black

102

[sysout,pts] = drawmag(data)
interactively uses the mouse in the plot window to create pts and sysout, which approximately fits the frequency response (magnitude) in pts data: Either a frequency response object that is plotted as a reference, or a constant matrix of the form [xmin xmax ymin ymax] specifying the plot window on the data. sysoutStable: minimum-phase ss object that approximately fits, in magnitude, the pts data ptsFrequency: response of points While drawmag is running, all interaction with the program is through the mouse and/or the keyboard Examples of commands Clicking the mouse
Typing r Typing integers 0 9
See help file for more info

Typing w
Typing p Typing k April 2006 Nathan Sorensen Kedrick Black 103

B = fitfrd(A,N)
Fit D-scaling frequency response data with state-space model state-space object with state dimension N, where A is an frd object and N is a nonnegative integer The frequency response of B closely matches the D-scale frequency response data in A A must have either 1 row or 1 column, although it need not be 1-by-1

Example
sys = tf([1 2 2],[1 2.5 1.5])*tf(1,[1 0.1]); sys = sys*tf([1 3.75 3.5],[1 2.5 13]); omeg = logspace(-1,1); sysg = frd(sys,omeg); bode(sysg,'r-'); b1 = fitfrd(sysg,1); b3 = fitfrd(sysg,3); b1g = frd(b1,omeg); b3g = frd(b3,omeg); bode(sysg,'r-',b1g,'k:',b3g,'b-.')

April 2006

Nathan Sorensen Kedrick Black

104

B = fitmagfrd(A,N)
Fit frequency response magnitude data with a stable, minimum-phase state-space model a stable, minimum-phase ss object, with state-dimension N, whose frequency response magnitude closely matches the magnitude data in A A is a 1-by-1 frd object, and N is a nonnegative integer
Example
sys = tf([1 2 2],[1 2.5 1.5])*tf(1,[1 0.1]); sys = sys*tf([1 3.75 3.5],[1 2.5 13]); omeg = logspace(-1,1); sysg = abs(frd(sys,omeg)); bodemag(sysg,'r'); ord = 3; b1 = fitmagfrd(sysg,ord); b1g = frd(b1,omeg); bodemag(sysg,'r',b1g,'k:');

April 2006

Nathan Sorensen Kedrick Black

105

Sampled Data Systems

The synthesis of robust optimal controllers with sampled data systems can be analyzed using the following functions.
1. Sdhinfnorm() computes the L2 induced norm of a continuous-time LTI plant, sdsys, in feedback with a discrete-time controller, k, connected through an ideal sampler and a zero-order hold. 2. Sdlsim() plots the time response of the hybrid feedback system. lft(p,k), is forced by the continuous input signal described by w and t.

April 2006

Nathan Sorensen Kedrick Black

106

[gaml,gamu] = sdhinfnorm(sdsys,k,delay)
Sdhinfnorm() computes the L2 induced norm of a continuous-time LTI plant, sdsys, in feedback with a discrete-time controller, k, connected through an ideal sampler and a zero-order hold.

Input: Sdsys a strictly proper partitioned plant model, such that the constant feedback gain is be zero K feedback controller gain matrix
2006 Nathan delay Sorensen April Delay controller computational time

Output: Gamu - upper bound on the induced L2 norm of the sampleddata closed-loop system Gaml - lower bound on the induced L2 norm of the sampledKedrick Black data closed-loop system 107

Sdhinfnorm() Example
p = ss(tf(30,[1 30])*tf([1],[1 0])); k = ss(tf(4,[1 4])); cl = feedback(p,k); norm(cl,'inf') ans = 1 kd = c2d(k,0.75,'zoh'); [gu,gl] = sdhinfnorm([1; 1]*p*[1 1],-kd); [gu gl] ans = 3.7908 3.7929 kd = c2d(k,0.2,'zoh'); [gu,gl] = sdhinfnorm([1; 1]*p*[1 1],-kd); [gu gl] ans = 1.0044 1.0049
April 2006 Nathan Sorensen Kedrick Black 108

sdlsim(p,k,w,t,tf,x0,z0,int)
Sdlsim() plots the time response of the hybrid feedback system. lft(p,k), is forced by the continuous input signal described by w and t.
Input: P - a continuous-time LTI system K a discrete-time LTI system with a specified sampling time W,T input forcing functions Tf simulation end time X0 - initial state vector of p, at time t(1)

Z0 - initial state vector of k, at time t(1)

Int - the continuous-time integration step size
April 2006 Nathan Sorensen Kedrick Black 109

Sdlsim() example cont:

Step response: disturbance (dashed) & output (solid) 1.4 1.2 1 0.8

EX: P = tf(1,[1, 1e-5,0]); T = 1.0/20; C = ss([-1.5 T/4; -2/T -.5],[ .5 2;1/T 1/T],... [-1/T^2 -1.5/T], [1/T^2 0],T); Pd = c2d(P,T,'zoh'); systemnames = 'Pd C'; inputvar = '[ref]'; outputvar = '[Pd]'; input_to_Pd = '[C]'; input_to_C = '[ref ; Pd]'; sysoutname = 'dclp'; cleanupsysic = 'yes'; sysic;

0.6 0.4 0.2 0 -0.2

0.1

0.2

0.3

0.4 0.5 0.6 Time: seconds

0.7

0.8

0.9

EX cont: M2 = [0,1,1;1,0,0;0,1,1]*blkdiag(1,1,P); t = [0:.001:1]'; dist = 0.1*sin(41*t); u = ones(size(t)); [y3,meas,act] = sdlsim(M2,C,[u dist],t,1); plot(y3{:},'-',t,dist,'b--',t,u,'g-.') xlabel('Time: seconds') title('Step response: disturbance (dashed) & output (solid)') Nathan Sorensen Kedrick Black 110

April 2006

Gain Scheduling
Gain scheduling is a useful tool for controller design. The Robust Control Toolbox offers one function to optimally achieve this.
1. Hinfigs() synthesizes gain-scheduled H controllers.

April 2006

Nathan Sorensen Kedrick Black

111

[gopt,pdK,R,S] = hinfgs(pdP,r,gmin,tol,tolred)
Synthesis of gain-scheduled H controllers returns the optimal closed-loop quadratic performance gopt and a polytopic description of the gain-scheduled controller pdK To test if a closed-loop quadratic performance is achievable, set the third input gmin to The arguments tol and tolred control the required relative accuracy on gopt and the threshold for order reduction also returns solutions R, S of the characteristic LMI system
April 2006 Nathan Sorensen Kedrick Black 112

Synthesis Supporting Utilities

The following functions are useful in setting up systems for controller synthesis:

1. Bilin() gives the multivariable bilinear transform of frequency (s or z).

2. Mktito() adds TITO (two-input-two-output) partitioning to LTI system SYS, assigning OutputGroup and InputGroup properties.
3. Sectf() gives the state-space sector bilinear transformation. 4. Skewdec() forms an m-by-m skew-symmetric matrix.

5. Symdec() forms an m-by-m symmetric matrix.

April 2006

Nathan Sorensen Kedrick Black

113

GT = bilin(G,VERS,METHOD,AUG)
Multivariable bilinear transform of frequency (s or z) bilin computes the effect on a system of the frequency-variable substitution The variable VERS denotes the transformation direction: VERS= 1, forward transform or VERS=-1, reverse transform computes several state-space bilinear transformations such as backward rectangular, etc., based on the METHOD you select

April 2006

Nathan Sorensen Kedrick Black

114

SYS=mktito(SYS,NMEAS,NCONT)
Mktito() adds TITO (two-input-two-output) partitioning to LTI system SYS, assigning OutputGroup and InputGroup properties.

EX:
P=rss(2,4,5); P=mktito(P,2,2); disp(P.OutputGroup); disp(P.InputGroup); Y1: [1 2] Y2: [3 4] U1: [1 2 3] U2: [4 5] April 2006

Plant Partitioning: B1 Inputs are disturbances B2 Inputs are control signals C1 Outputs are errors to be minimized C2 Outputs are inputs to control K
Nathan Sorensen Kedrick Black 115

[G,T] = sectf(F,SECF,SECG)
State-space sector bilinear transformation computes a linear fractional transform T such that the system lft(F,K) is in sector SECF if and only if the system lft(G,K) is in sector SECG where G=lft(T,F,NU,NY) where NU and NY are the dimensions of uT2 and yT2, respectively

April 2006

Nathan Sorensen Kedrick Black

116

x = skewdec(m,n)
Skewdec() forms the m-by-m skew-symmetric matrix:

NOTE: This function is useful to define skew-symmetric matrix variables. In this case, set n to the number of decision variables already used.

April 2006

Nathan Sorensen Kedrick Black

117

x = symdec(m,n)
Symdec() forms an m-by-m symmetric matrix of the form:

NOTE: This function is useful to define symmetric matrix variables. n is the number of decision variables.

April 2006

Nathan Sorensen Kedrick Black

118

LMIs in the Robust Control Toolbox

The Robust Control Toolbox LMI functionality serves two purposes: Provide state-of-the-art tools for the LMI-based analysis and design of robust control systems Offer a flexible and user-friendly environment to specify and solve general LMI problems (the LMI Lab) Examples of LMI-based analysis and design tools include: Functions to analyze the robust stability and performance of uncertain systems with varying parameters (popov, quadstab, quadperf ...) Functions to design robust control with a mix of H2, H, and pole placement objectives (h2hinfsyn) Functions for synthesizing robust gain-scheduled H controllers (hinfgs)
April 2006 Nathan Sorensen Kedrick Black 119

Linear Matrix Inequalities

A linear matrix inequality (LMI) is any constraint of the form

where x = (x1, . . . , xN) is a vector of unknown scalars (the decision or optimization variables) A0, . . . , AN are given symmetric matrices < 0 stands for "negative definite," i.e., the largest eigenvalue of A(x) is negative Note that the constraints A(x) > 0 and A(x) < B(x) are special cases of (1-1) since they can be rewritten as -A(x) < 0 and A(x) - B(x) < 0, respectively.

April 2006

Nathan Sorensen Kedrick Black

120

Specifying/Modifying LMI Systems

LMI Systems can be specified and modified in various ways briefly outlined below
The specification of an LMI system involves two steps: 1. Declare the dimensions and structure of each matrix variable X1, . . . , XK. 2. Describe the term content of each LMI. This process creates the so-called internal representation of the LMI system. This computer description of the problem is used by the LMI solvers and in all subsequent manipulations of the LMI system. It is stored as a single vector called LMISYS. There are two ways of generating the internal description of a given LMI system: 1. by a sequence of lmivar/lmiterm commands that build it incrementally 2. via the LMI Editor lmiedit where LMIs can be specified directly as symbolic matrix expressions. Though somewhat less flexible and powerful than the commandbased description, the LMI Editor is more straightforward to use, hence particularly well-suited for beginners.
April 2006 Nathan Sorensen Kedrick Black 121

Specification Functions
The following functions are used for the LMI Specificiation:
1. Lmiedit is a graphical user interface for the symbolic specification of LMI problems. 2. Setlmis() is a command to be used to initialize or add to an internal representation for an lmi. 3. Lmivar() defines a new matrix variable X in the LMI system currently described. 4. Lmiterm() specifies the term content of an LMI one term at a time. 5. Newlmi adds a new LMI to the LMI system currently described and returns an identifier tag for this LMI. 6. Getlmis returns the internal representation of a lmi system declared using lmivars() and lmiterm().

April 2006

Nathan Sorensen Kedrick Black

122

lmiedit
Lmiedit is a graphical user interface for the symbolic specification of LMI problems.
Instructions:
1. 2. Give it a name (top of the window). Declare each matrix variable (name and structure) in the upper half of the window. The structure is characterized by its type (S for symmetric block diagonal, R for unstructured, and G for other structures) and by an additional structure matrix similar to the second input argument of lmivar. Please use one line per matrix variable in the text editing areas. Specify the LMIs as MATLAB expressions in the lower half of the window. An LMI can stretch over several lines. However, do not specify more than one LMI per line. 123

3.

April 2006

Nathan Sorensen Kedrick Black

setlmis(lmi)
Setlmis() is a command to be used to initialize or add to an internal representation for an lmi.

Two Uses: 1. setlmis([ ]) - used to initialize an internal representation for an lmi to be followed by lmivar and lmiterm. 2. setlmis(lmi0) adds to an existing lmi where lmi0 is the internal representation of a lmi. (Subsequent lmivar and lmiterm commands will then add new variables and terms to the initial LMI system lmi0.)
April 2006 Nathan Sorensen Kedrick Black 124

X = lmivar(type,struct)
Lmivar() defines a new matrix variable X in the LMI system currently described.
Input:

EX: setlmis([]) X1 = lmivar(1,[3 1]) % Type 1 X2 = lmivar(2,[2 4]) % Type 2 of dim. 2x4 X3 = lmivar(1,[5 1;1 0;2 0]) % Type 1

Type - selects among available types of variables and the second argument Options: 1- Symmetric matrices with a blockdiagonal structure 2 - Full m-by-n rectangular matrix 3 - Other structures Struct - gives further information on the structure of X depending on its type Ouput: X Matrix variable in LMI system

April 2006

Nathan Sorensen Kedrick Black

125

lmiterm(termID,A,B,flag)
Lmiterm() specifies the term content of an LMI one term at a time.

EX:
lmiterm([1 1 1 X2],2*A,A') % 2*A*X2*A' lmiterm([1 1 1 x3],-1,E) % -x3*E lmiterm([1 1 1 0],D*D') % D*D' lmiterm([1 2 1 -X1],1,B) % X1'*B lmiterm([1 2 2 0],-1) % -I lmiterm([-1 0 0 0],M) % outer factor M lmiterm([-1 1 1 X1],C,C','s') % C*X1*C'+C*X1'*C' lmiterm([-1 2 2 X2],-f,1) % -f*X2 April 2006 Nathan Sorensen Kedrick Black 126

tag = newlmi
Newlmi adds a new LMI to the LMI system currently described and returns an identifier tag for this LMI. (Tagging LMIs is optional and
only meant to facilitate code development and readability.)
Purpose: To give individual names to lmis so that they can be easily distinguished in code. Use: When parts of LMIs are deleted from a LMI, identifiers help to identify the remaining parts of the system.

April 2006

Nathan Sorensen Kedrick Black

127

lmisys = getlmis
Getlmis returns the internal representation of a lmi system declared using lmivars() and lmiterm(). This can then be used in any LMI solver or processing function.

April 2006

Nathan Sorensen Kedrick Black

128

LMI Characteristics
LMI System characteristics can be determined using the following functions:
1. Lmiinfo() provides qualitative information about the system of LMIs lmisys. 2. Lminbr returns the number k of linear matrix inequalities in the LMI problem described in lmisys. 3. Matnbr() returns the number K of matrix variables in the LMI problem described by lmisys. 4. Decnbr() returns the number ndec of decision variables (free scalar variables) in the LMI problem described in lmisys. 5. Dec2mat() is a function that, given values of the decision variables, derives the corresponding values of the matrix variables.

April 2006

Nathan Sorensen Kedrick Black

129

LMI Characteristics cont:

6. Mat2dec() returns the corresponding value decvec of the vector of decision variables if given an LMI system with variables X1, . . ., XK given values of x1,...,xk. 7. DecX() expresses the entries of a matrix variable X in terms of the decision variables x1, . . ., xN.

April 2006

Nathan Sorensen Kedrick Black

130

lmiinfo(lmisys)
Lmiinfo() provides qualitative information about the system of LMIs lmisys. This includes the type and structure of the matrix variables, the number of diagonal blocks in the inner factors, and the term content of each block.

EX:

lmiinfo(lmis) LMI ORACLE ------This is a system of 1 LMI with 3 variable matrices Do you want information on (v) matrix variables (l) LMIs (q) quit?> v Which variable matrix (enter its index k between 1 and 3) ? 1 X1 is a 2x2 symmetric block diagonal matrix its (1,1)-block is a full block of size 2
------April 2006 Nathan Sorensen Kedrick Black 131

LMI Info Example Cont:

This is a system of 1 LMI with 3 variable matrices Do you want information on (v) matrix variables (l) LMIs (q) quit?> l Which LMI (enter its number k between 1 and 1) ? 1 This LMI is of the form 0 < R(x) where the inner factor(s) has 2 diagonal block(s) Do you want info on the right inner factor ? (w) whole factor (b) only one block (o) other LMI (t) back to top level?> w Info about the right inner factor block (1,1) : I + a1*X1 + A2*X2*B2 + B2'*X2'*A2' block (2,1) : A3*X1 block (2,2) : x3*A4 (w) whole factor (b) only one block (o) other LMI (t) back to top level

April 2006

Nathan Sorensen Kedrick Black

132

k = lminbr(lmisys)
Lminbr returns the number k of linear matrix inequalities in the LMI problem described in lmisys.

April 2006

Nathan Sorensen Kedrick Black

133

K = matnbr(lmisys)
Matnbr() returns the number K of matrix variables in the LMI problem described by lmisys.

April 2006

Nathan Sorensen Kedrick Black

134

ndec = decnbr(lmisys)
Decnbr() returns the number ndec of decision variables (free scalar variables) in the LMI problem described in lmisys.
Input: Lmisys System of LMIs EX: For an LMI system LMIs with two matrix variables X and Y such that X is symmetric block diagonal with one 2by-2 full block, and one 2-by2 scalar block Y is 2-by-3 rectangular, the number of decision variables is: ndec = decnbr(LMIs) ndec = 10
April 2006 Nathan Sorensen Kedrick Black 135

Output: Ndec - the length of the vector of decision variables.

valX = dec2mat(lmisys,decvars,X)
Dec2mat() is a function that, given values of the decision variables, derives the corresponding values of the matrix variables.
Input: Lmisys System of LMIs Decvars - the vector of decision variables X - the matrix variable identifier with which the function computes valX Output:

ValX Values of the matrix variable with the identifier X.

April 2006

Nathan Sorensen Kedrick Black

136

decvec = mat2dec(lmisys,X1,X2,X3,...)
Mat2dec() returns the corresponding value decvec of the vector of decision variables if given an LMI system with variables X1, . . ., XK given values of x1,...,xk.
Input: Lmisys System of LMIs X1,..,Xn system matrix variables evaluated at x1,,xn EX: Given

Output:
Decvec - vector of decision variables

decv = mat2dec(lmisys,X0,Y0) decv = [ 1 3 -1 5 1 2 4 5 6]

3
137

April 2006

Nathan Sorensen Kedrick Black

decX = decinfo(lmisys,X)
DecX() expresses the entries of a matrix variable X in terms of the decision variables x1, . . ., xN.
EX:

setlmis([]) X = lmivar(1,[3 0]) Y = lmivar(2,[2 1]) : : lmis = getlmis dX = decinfo(lmis,X) dX = 1 0 0 0 1 0 0 0 1 dY = decinfo(lmis,Y) dY = 2 3

April 2006

Input: Lmisys System of LMIs X - each entry of X is either a hard zero, some decision variable xn, or its opposite xn

Output: DecX - clarifies the structure of X as well as its entry-wise dependence on x1, . . ., xN

Nathan Sorensen Kedrick Black

138

LMI Solvers
LMI solvers are provided for the following three generic optimization problems
Feasibility problem Find x RN (or equivalently matrices X1, . . . , XK with prescribed structure) that satisfies the LMI system A(x) < B(x) The corresponding solver is called feasp. Minimization of a linear objective under LMI constraints Minimize cTx over x RN subject to A(x) < B(x) The corresponding solver is called mincx. Generalized eigenvalue minimization problem

Minimize over x RN subject to C(x) < D(x) 0 < B(x) A(x) < B(x). The corresponding solver is called gevp.
April 2006 Nathan Sorensen Kedrick Black 139

[tmin,xfeas] = feasp(lmisys,options,target)
Feasp() computes a solution xfeas (if any) of the system of LMIs described by lmisys.
Input: Lmisys LMI system to solve Options zeros set to default parameters

1 not used
2 - sets the maximum number of iterations (100 by default) 3 - resets the feasibility radius 4 - helps speed up termination 5 a 1 turns off the trace of execution of the optimization procedure. 0 is on. Target - sets a target value for tmin
April 2006

Output: Tmin Constraints are feasible if tmin<=0 and strictly feasible if tmin<0 Xfeas - a solution in terms of the decision variables
140

Nathan Sorensen Kedrick Black

Feasp() Example
EX: setlmis([]) p = lmivar(1,[2 1]) lmiterm([1 1 1 p],1,a1,'s') % LMI #1 lmiterm([2 1 1 p],1,a2,'s') % LMI #2 lmiterm([3 1 1 p],1,a3,'s') % LMI #3 lmiterm([-4 1 1 p],1,1) % LMI #4: P lmiterm([4 1 1 0],1) % LMI #4: I lmis = getlmis [tmin,xfeas] = feasp(lmis) tmin = -3.1363 P = dec2mat(lmis,xfeas,p)

April 2006

Nathan Sorensen Kedrick Black

141

[copt,xopt] = mincx(lmisys,c,options,xinit,target)
Mincx() solves the convex program

where x represents a vector of scalar

decision variables.
Output: Copt the global minimum for the objective cTx Xopt - the minimizing value of the vector of decision variables Input:
Lmisys System of LMIs C vector the same length as x and defined by defcnbr() or defcx() Options zeros set to default parameters 1 not used 2 - sets the maximum number of iterations (100 by default) 3 - resets the feasibility radius 4 - helps speed up termination 5 a 1 turns off the trace of execution of the optimization procedure. 0 is on. Xinit intial guess for x Target - sets a target value for tmin
142

April 2006

Nathan Sorensen Kedrick Black

[lopt,xopt] = gevp(lmisys,nlfc)
Gevp() solves a generalized eigenvalue minimization problem described by: Minimize LMIs. EX:
setlmis([]); p = lmivar(1,[2 1]) lemiterm([1 1 1 0],1) % P > I : I lemiterm([-1 1 1 p],1,1) % P > I : P lemiterm([2 1 1 p],1,a1,'s') % LFC # 1 (lhs) lemiterm([-2 1 1 p],1,1) % LFC # 1 (rhs) lemiterm([3 1 1 p],1,a2,'s') % LFC # 2 (lhs) lemiterm([-3 1 1 p],1,1) % LFC # 2 (rhs) lemiterm([4 1 1 p],1,a3,'s') % LFC # 3 (lhs) lemiterm([-4 1 1 p],1,1) % LFC # 3 (rhs) lmis = getlmis [alpha,popt]=gevp(lmis,3) alpha = -0.122 April 2006

subject to C(x) < D(x) and A(x) <

B(x) > 0 denote systems of Output: Lopt - the global minimum Xopt - the minimizing value of the vector of decision variables x.

Input:

Lmisys LMI System

Nlfc number of linear fractional constraints

Nathan Sorensen Kedrick Black

143

[V1,...,Vk] = defcx(lmisys,n,X1,...,Xk)
Defcx() is useful to derive the c vector needed by mincx when the objective
is expressed in terms of the matrix variables. Input: Lmisys system of LMIs Output: V1,,Vk value of these variables when the n-th decision variable is set to one and all others to zero.

N the number of the nth decision variable

X1,,Xk identifiers of the matrix variables involved in this objective

April 2006

Nathan Sorensen Kedrick Black

144

LMI Solution Validation

Once a solution to an LMI system is obtained, the following functions assist in the validation of the given solution.

1. Evallmi() evaluates all LMI constraints for a particular instance determined by decvars of the vector of decision variables. 2. Showlmi() evaluates all variable terms in a system of LMIs for given values of the decision variables.

April 2006

Nathan Sorensen Kedrick Black

145

evalsys = evallmi(lmisys,decvars)
Evallmi() evaluates all LMI constraints for a particular instance determined by decvars of the vector of decision variables.
EX: Find X s.t. ATXA - X + I < 0
setlmis([]) X = lmivar(1,[2 1]) % full symmetric X lmiterm([1 1 1 X],A',A) % LMI #1: A'*X*A lmiterm([1 1 1 X],-1,1) % LMI #1: -X lmiterm([1 1 1 0],1) % LMI #1: I lmiterm([-2 1 1 X],1,1) % LMI #2: X lmis = getlmis [tmin,xfeas] = feasp(lmis) tmin = -4.7117e+00 xfeas' = [1.1029e+02 -1.1519e+01 1.1942e+02] evals = evallmi(lmis,xfeas) [lhs1,rhs1] = showlmi(evals,1) [lhs2,rhs2] = showlmi(evals,2) eig(lhs1-rhs1) ans = -8.2229e+01, -5.8163e+01 April 2006

Input:

Lmisys System of LMIs

Decvars particular instance of the decision variables

Output:
Evalsys - LMI system containing only constant terms
Nathan Sorensen Kedrick Black 146

[lhs,rhs] = showlmi(evalsys,n)
Showlmi() evaluates all variable terms in a system of LMIs for given values of the decision variables.

Input: Evalsys - the output of evallmi N the number of lmis

Output: Lhs, Rhs - The left- and right-hand sides of the n-th LMI constant matrices

April 2006

Nathan Sorensen Kedrick Black

147

LMI Modification
Once an LMI is formally declared, the following functions allow for the modification of LMI objects.

1. Dellmi() deletes the nth LMI from the system of LMIs described in lmisys. The updated system is returned in newsys. 2. Delmvar() removes the matrix variable X with identifier X from the list of variables defined in lmisys. 3. Setmvar() sets the matrix variable X with identifier X to the value Xval.

April 2006

Nathan Sorensen Kedrick Black

148

newsys = dellmi(lmisys,n)
Dellmi() deletes the nth LMI from the system of LMIs described in lmisys. The updated system is returned in newsys. EX: Input: Suppose that the following is an LMI Lmisys system of LMIs N the number of the deletion LMI System:

Output: Newsys new LMI system after the nth LMI is deleted.

Then give the command: lmis = dellmi(lmisys,LMI2)

LMIS is then reduced to:

April 2006 Nathan Sorensen Kedrick Black 149

newsys = delmvar(lmisys,X)
Delmvar() removes the matrix variable X with identifier X from the list of variables defined in lmisys.
EX: Consider the LMI: Input: Lmisys system of LMIs X name of LMI to be deleted

Ouput: lmisys = delmvar(lmisys,X) Gives the system: Newsys system of LMIs after the LMI X has been deleted

April 2006

Nathan Sorensen Kedrick Black

150

newsys = setmvar(lmisys,X,Xval)
Setmvar() sets the matrix variable X with identifier X to the value Xval. EX:
Consider the system: and the problem of finding a stabilizing statefeedback law u = Kx where K is an unknown matrix. This LMI is entered by the commands: n = size(A,1) % number of states ncon = size(B,2) % number of inputs setlmis([]) P = lmivar(1,[n 1]) % P full symmetric Y = lmivar(2,[ncon n]) % Y rectangular lmiterm([1 1 1 P],A,1,'s') % AP+PA' lmiterm([1 1 1 Y],B,1,'s') % BY+Y'B' lmiterm([1 1 1 0],1) % I lmis = getlmis news = setmvar(lmis,P,1) [tmin,xfeas] = feasp(news) Y = dec2mat(news,xfeas,Y)

Input: Lmisys system of LMIs X name of an LMI in Lmisys

Xval value to evaluate X at

Output Newsys LMI system after X is evaluated at Xval.

April 2006

The computed Y is feasible whenever tmin < 0. Nathan Sorensen Kedrick Black 151

LMI Supporting Functions

Within the LMI functionality, various functions are employed to assist in the creation and analysis of the LMI system.

1. Matnbr() matnbr returns the number K of matrix variables in the LMI problem described by lmisys.

2. Lmireg() is an interactive facility to specify the LMI regions involved in multi-objective H infinity synthesis with pole placement constraints.

April 2006

Nathan Sorensen Kedrick Black

152

K = matnbr(lmisys)
Matnbr() matnbr returns the number K of matrix variables in the LMI problem described by lmisys

April 2006

Nathan Sorensen Kedrick Black

153

region = lmireg
Lmireg() is an interactive facility to specify the LMI regions involved in multi-objective H infinity synthesis with pole placement constraints.

Recall that an LMI region is any convex subset D of the complex plane that can be characterized by an LMI in z and , i.e.,

for some fixed real matrices M and L = LT. This class of regions encompasses half planes, strips, conic sectors, disks, ellipses, and any intersection of the above.

April 2006

Nathan Sorensen Kedrick Black

154

Plotting Functions
Various Plotting functions are useful with the Robust Control Toolbox. Some of particular interest are shown here.

1. Loglog() is same as plot(...), except a logarithmic (base 10) scale is used for the x-axis and y-axis. 2. Semilog() is same as plot(...), except a logarithmic (base 10) scale is used for the x-axis. 3. Uplot() allows one to plot multiple frequency response objects and doubles on the same graph.

4. Dmplot() plots the maximum allowable phase variation as a function of the actual gain variation for a given disk gain margin diskgm.

April 2006

Nathan Sorensen Kedrick Black

155

loglog(sys)
log-log scale plot of frd objects same as plot, except a logarithmic (base 10) scale is used for both the X- and Yaxis Example
omega = logspace(-2,2,40); sys1 = tf(0.9,[.1 1]) sys1g = frd(sys1,omega); sys2 = tf(0.9,[1 1]) sys2g = frd(sys2,omega); sys3 = tf(0.9,[10 1]) sys3g = frd(sys3,omega); loglog(abs(sys1g),abs(sys2g),'r+',sys3g.Freq,... abs(sys3g.Resp(:)),'--'); Sys1: 0.9 --------0.1 s + 1 Sys2: 0.9 ----s+1 Sys3: 0.9 -------10 s + 1

April 2006

Nathan Sorensen Kedrick Black

156

semilogx(...) same as plot

Semilog scale plot of a frd object same as plot(...), except a logarithmic (base 10) scale is used for the x-axis

April 2006

Nathan Sorensen Kedrick Black

157

uplot(type,G1,linetype,Xdata,Ydata, linetype)
Plot multiple frequency response objects and doubles on the same graph

April 2006

Nathan Sorensen Kedrick Black

158

Uplot example
Example
a1 = [-1,1;-1,-0.5]; b1 = [0;2]; c1 = [1,0]; d1 = 0; sys1 = ss(a1,b1,c1,d1); a2 = [-.1,1;-1,-0.05]; b2 = [1;1]; c2 = [-0.5,0]; d2 = 0.1; sys2 = ss(a2,b2,c2,d2); omega = logspace(-2,2,100); sys1g = frd(sys1,omega); omega2 = [ [0.05:0.1:1.5] [1.6:.5:20] [0.9:0.01:1.1] ]; omega2 = sort(omega2); sys2g = frd(sys2,omega2); sys3 = rss(1,1,1); rspot = frd(sys3,2); uplot('liv,lm',sys1g,'b-.',rspot,'r-*',sys2g); xlabel('log independent variable') ylabel('log magnitude') title('axis specification: liv,lm')

April 2006

Nathan Sorensen Kedrick Black

159

dmplot(diskgm)
Dmplot() plots the maximum allowable phase variation as a function of the actual gain variation for a given disk gain margin diskgm.
EX:
dmplot(1.414)

April 2006

Nathan Sorensen Kedrick Black

160

Miscellaneous
This is a group of functions employed throughout the Robust Control Toolbox that do not fit into any general category.
1. Rcond() creates an estimate for the reciprocal of the condition of the frd object x in the 1-norm . 2. Schur() applies the schur command to frd objects. 3. Diag(x) puts x on the main diagonal of a matrix.

4. Genphase() fit SISO magnitude data with a real, rational, minimum-phase transfer function.
5. Ssv() computes the singular value decomposition of an frd object. 6. Repmat() creates a large matrix B consisting of an M-by-N tiling of copies of some matrix A.

April 2006

Nathan Sorensen Kedrick Black

161

r = rcond(x)
is an estimate for the reciprocal of the condition of the frd object x in the 1-norm obtained by the LAPACK condition estimator operates on x.ReponseData of the x frd at each frequency to construct r If x is well conditioned, rcond(x) is near 1.0 If x is badly conditioned, rcond(x) is near EPS
April 2006 Nathan Sorensen Kedrick Black 162

[u,t] = schur(x)
applies the schur command to frd objects operates on the x.ReponseData of the frd object at each frequency point to construct u and t u and t are frd objects. x must be square

April 2006

Nathan Sorensen Kedrick Black

163

v = diag(x)
If x is a vector of uncertain system models or matrices, diag(x) puts x on the main diagonal If x is a matrix of uncertain system models or matrices, diag(x) is the main diagonal of v diag(diag(x)) is a diagonal matrix of uncertain system models or matrices Example
x = rss(3,4,1); xg = frd(x,logspace(-2,2,80)); size(xg) FRD model with 4 output(s) and 1 input(s), at 80 frequency point(s). mxg = diag(xg); size(mxg) FRD model with 4 output(s) and 4 input(s), at 80 frequency point(s). xxg = [xg(1:2,1) xg(3:4,1)]; m = diag(xxg); size(m) FRD model with 2 output(s) and 1 input(s), at 80 frequency point(s).

April 2006

Nathan Sorensen Kedrick Black

164

resp = genphase(d)
Fit single-input/single-output magnitude data with a real, rational, minimum-phase transfer function uses the complex-cepstrum algorithm to generate a complex frequency response resp whose magnitude is equal to the real, positive response d, but whose phase corresponds to a stable, minimum-phase function
April 2006 Nathan Sorensen Kedrick Black 165

[U,S,V] = svd(X)
Singular value decomposition of an frd object produces a diagonal frd S that has the same dimensions as X and includes positive diagonal elements in decreasing order U and V are unitary matrices and frd objects, such that X = U*S*V'
April 2006 Nathan Sorensen Kedrick Black 166

B = repmat(A,M,N)
Replicate and tile an array creates a large matrix B consisting of an M-by-N tiling of copies of A
Example
repmat(randumat(2,2),2,3) repmat(ureal(`A',6),[4 2])

UMAT: 4 Rows, 6 Columns BJVAD: real, nominal = 5.95, variability = [-3.97443 15.0948]%, 3 occurrences FRHOD: complex, nominal = 3.34+5.95i, radius = 0.757, 3 occurrences LAVLQ: complex, nominal = -2.16-8.33i, radius = 1.84, 2 occurrences WNFRV: complex, nominal = 0, radius = 1.36, 2 occurrences UMAT: 4 Rows, 2 Columns A: real, nominal = 6, variability = [-1 1], 1 occurrence

April 2006

Nathan Sorensen Kedrick Black

167

References
www.mathworks.com/ Robust Control Toolbox Tutorial

April 2006

Nathan Sorensen Kedrick Black

168