Control System Design - LQG

Bo Bernhardsson, K. J. Åström

Department of Automatic Control LTH,

Lund University

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Lecture - LQG Design

Introduction
The H2 -norm
Formula for the optimal LQG controller
Software, Examples
Properties of the LQ and LQG controller
Design tricks, how to tune the knobs
What do the “technical conditions” mean?
How to get integral action etc
Loop Transfer Recovery (LTR)
More Examples

For theory and more information, see PhD course on LQG

Reading tips: Ch 5 in Maciejowski

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Linear Quadratic Gaussian Design

Process model

ẋ = Ax + Bu + v
y = C x + Du + w

where v,w is white gaussian noise with mean zero

  T  
 v ( t)   v (τ )   R 11 R 12 
E
 
 
 =
 
 δ (t − τ )
w( t) w(τ ) R 21 R 22

Optimization criterion
Z T  T   

 x
 
 Q 11 Q 12 
  x
min E      dt
0 u Q 21 Q 22 u

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Short on Stochastics

ẋ = Ax + v
y = Cx + e

v white noise, Ev( t)vT ( t − τ ) = R 1δ (τ )

e white noise, E e( t) eT ( t − τ ) = R 2δ (τ )

State covariance

E x( t) xT ( t) = R( t), Ṙ = A R + R AT + R 1
Kalman filter, x̂˙ = A x̂ + L ( y − C x̂)

x̃ = x − x̂, E x̃( t) x̃T ( t) = P ( t)

Ṗ = A P + P AT + R 1 − PC T R− 1
2 C P, L = PC T R−
2
1

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Separation Principle

Process

Kalman Filter

Nice structure of the optimal controller: u = − K b

x
Linear feedback combined with state estimation

Certainty equivalence principle

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Linear Quadratic Gaussian Design - History

In the late 50s and early 60s computers where starting to be used to find
“optimal” controllers.

Classical Reference: Newton, Gould, Kaiser (1957)

Wiener-Kolmogorov

Kalman-Bucy

Bellman, Wonham, Willems, Andersson, Åström, Kucera and many others

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Why so popular?

Gives “optimal” controller

Automized design method. Works for MIMO.

Nice formulas for the optimal controller, reasonable computational effort

Gives absolute scale of merit - know limits of performance

Used for space program, aircraft design - Good models often available

LQ control give good robustness margins (with Q 1 2 = 0)

[1/2, ∞]-gain margin

60 degree phase margin

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Lecture - LQG Design

Introduction
The H2 -norm
Formula for the optimal LQG controller
Software, Examples
Properties of the LQ and LQG controller
Design tricks,how to tune the knobs
What do the “technical conditions” mean?
How to get integral action etc
Loop Transfer Recovery (LTR)
More Examples

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 LQG is optimizing the H2 norm

Consider the system

Y = G(s)U
y = ˆ⋆u
ẋ = Ax + Bu
y = C x + ( Du)

The L 2 -norm (LQG-norm) is defined as

X X Z∞
qGq22 = pGi j ( jω )p2 dω /2π =
i j −∞
Z∞
= trace {G∗ ( jω )G( jω )} dω /2π
−∞

H2 : Equals L 2 -norm if G asymptotically stable, equals ∞ otherwise

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 H2 -norm as noise power gain

u y

u: stationary white noise, mean zero

E (u(τ1 )u(τ2 )T ) = δ (τ1 − τ2 ) I

S u (ω ) = 1, ∀ω

then Pow( y) = E ( yT y) = qGq22 .

”Amplification of noise power”

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Proof

Z
T
E (tr yy ) = tr S y(ω ) dω /2π =
Z
= tr G∗ ( jω ) S u (ω )G( jω ) dω /2π

= qGq22

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Another interpretation of the H2 -norm

By Parseval’s formula we have

X X Z∞
qGq22 = pˆ ji ( t)p2 dt
i j −∞

Hence the H2 -norm can also be interpreted as “energy in impulse

responses”:

ui = δi(t) y=g
G : , i (t)

m Z∞
X
qGq22 = pˆ :,i p2 ( t)dt
i=1 0

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 How to compute the H2 norm
1) norm(sys) in Matlab

2) If G(s) = C (s I − A)−1 B then

qGq22 = trace (C PC T ) = trace ( BT S B)

where P is the unique solution to the Lyapunov equation

A P + P AT + BBT = 0

and S solves
S A + AT S + C T C = 0

3) Residue calculus
X 1 I
qGq22 = Gi j (−s)T Gi j (s) ds
2πi
i, j

4) Recursive formulas ala Åström-Jury-Schur

Bo Bernhardsson, K. J. Åström Control System Design - LQG
 Lecture - LQG Design

Introduction
The H2 -norm
Formula for the optimal LQG controller
Software, Examples
Properties of the LQ and LQG controller
Design tricks,how to tune the knobs
What do the “technical conditions” mean?
How to get integral action etc
Loop Transfer Recovery (LTR)
More Examples

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 A unified framework
w z

u G
y

u = Control inputs
y = Measured outputs


 Fixed commands

Unknown commands
w = Exogenous inputs =

 Disturbances

Noise . . .


 Tracking errors

Control inputs
z = Regulated outputs =

 Measured outputs

States . . .

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 The H2 Problem

Closed Loop

u = K (s) y
z = G11 + G12 K ( I − G22 K )−1 G21 w = Tzw w

The H2 problem:

Find K (s) such that the closed loop is stable and

min qTzw q2
K (s)

is obtained.

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 The Optimal Controller

ẋ = Ax + B1 w + B2 u
z = C1 x + D 12 u
y = C 2 x + D 21 w + D 22 u
Under some technical conditions the optimal controller is

u = −K b
x
xḃ = A xb + B2 u + L( y − C 2 b
x − D 22 u)
K = ( D T12 D 12 )−1 ( D T12 C1 + B2T S )
L = ( B1 D T21 + PC 2T )( D 21 D T21 )−1
where P ≥ 0 and S ≥ 0 satisfy

0 = S A + AT S + C 1T C 1 − K T D T12 D 12 K
0 = A P + P AT + B1 B1T − L D 21 D T21 L T
A − B2 K , A − LC 2 stable
Bo Bernhardsson, K. J. Åström Control System Design - LQG
 The Optimal Controller

u = − K (s I − A + B2 K + LC 2 − L D 22 K )−1 Ly

Controller has same order as process

(How to introduce reference signals later)

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 “Technical Conditions”

1) [ A, B2 ] stabilizable

2) [C 2 , A] detectable

3) “No zeros on imaginary axis” u → z

 

 jω I − A − B2 

rank  = n+m ∀ω
C1 D 12

and D 12 has full column rank (no free control)

4) “No zeros on imaginary axis” w → y

 

 jω I − A − B1 

rank  =n+p ∀ω
C2 D 21

and D 21 has full row rank (no noise-free measurements)

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Alternative formulation

Weight matrices
   T  

 Q 11 Q 12 
  C1 
 T  T
 =   C1 D 12 
Q 12 Q 22 D 12
    

 R 11 R 12 
 
 B1    T T 
 T  =   B1 D 21
R 12 R 22 D 21

If Q 12 = 0 then the notation Q 1 and Q 2 is sometimes used instead,

similar for R

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Lecture - LQG Design

Introduction
The H2 -norm
Formula for the optimal LQG controller
Software, Examples
Properties of the LQ and LQG controller
Design tricks,how to tune the knobs
What do the “technical conditions” mean?
How to get integral action etc
Loop Transfer Recovery (LTR)
More Examples

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Software

Matlab - Control system toolbox

lqr, dlqr - Linear-quadratic (LQ) state-feedback regulator

lqry - LQ regulator with output weighting
lqrd - Discrete LQ regulator for continuous plant
kalman - Kalman estimator
kalmd - Discrete Kalman estimator for continuous plant
lqgreg - LQG regulator from LQ gain & Kalman estimator

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Matlab “LQR”

LQR Linear-quadratic regulator design for state space systems.

[K,S,E] = LQR(SYS,Q,R,N) calculates the optimal

gain matrix K such that:

* For a continuous-time state-space model SYS, the

state-feedback law u = -Kx minimizes the cost function

J = Integral {x’Qx + u’Ru + 2*x’Nu} dt

subject to the system dynamics dx/dt = Ax + Bu

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Matlab “Kalman”

KALMAN Continuous- or discrete-time Kalman estimator.

[KEST,L,P] = KALMAN(SYS,QN,RN,NN) designs a Kalman

estimator KEST for the continuous- or discrete-time
plant with state-space model SYS. For a continuous-time model
.
x = Ax + Bu + Gw {State equation}
y = Cx + Du + Hw + v {Measurements}

with known inputs u, process noise w, measurement noise v, and

noise covariances

E{ww’} = QN, E{vv’} = RN, E{wv’} = NN,

By default, SYS is the state-space model SS(A,[B G],C,[D H])

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Matlab “LQGREG”

LQGREG Form linear-quadratic-Gaussian (LQG) regulator

RLQG = LQGREG(KEST,K) produces an LQG regulator by

connecting the Kalman estimator KEST designed with KALMAN
and the state-feedback gain K designed with (D)LQR or LQRY:

+----------------------------+
u | |
| +------+ |
+--->| | x_e +----+ |
| KEST |------>| -K |---+-----> u
y -------->| | +----+
+------+

The resulting regulator RLQG has input y and generates the

commands u = -K x_e where x_e is the Kalman state estimate
based on the measurements y. This regulator should be
connected to the plant using positive feedback.

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Example

Consider the following system from the pole placement lecture

1 + 0.5s
G(s) =
s2

The following controller is suggested in Åström-Murray “Feedback

Systems” p. 363

s + 11.02
C (s) = 3628
(s + 2)(s + 78.28)
To construct an LQG controller we write the system on state space form
   

 0 0
 1
ẋ =  x+  u + Gw
1 0 0
 
y = 0.5 1 x + Hw + v

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Example - slow process zero

A = [0 0 ; 1 0];
B = [1 ; 0];
C = [0.5 1];
D = 0;
sys = ss(A,B,C,D);
Q=diag([0 1]);
R=1e-5;
[k,s,e]=lqr(A,B,Q,R);
G=B;
H=0;
QN = 1;
RN = 1e-5;
syse = ss(A,[B G],C,[D H]);
[kest,l,p]=kalman(syse,QN,RN);
rlqg = lqgreg(kest,k);
PC = -rlqg*sys;

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Result

Q 11 = diaˆ([0 1]), Q 22 = 10−5 , R 11 = diaˆ([0 1]), R 22 = 10−5

20 50

0 40

-20 30

-40 20

-60 10

-80 0

-100 -10
-2 -1 0 1 2 3
10-2 10-1 100 101 102 103 10 10 10 10 10 10

Larger high-frequency gain for LQG(blue) than default controller(red)

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Result

Increased meas noise R 22 := 7 · 10−5 and Q 22 = 0.7 · 10−5

20 50

0 40

−20 30

−40 20

−60 10

−80 0

−100 −2 −1 0 1 2 3
−10 −2 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10

LQG now gives same controller as was obtained by pole-placement design

earlier

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Obsolete LQG Software - use at your own risk
Matlab - robust control toolbox and mutools

h2lqg - continuous time H_2 synthesis.

dh2lqg - discrete time H_2 synthesis.
normh2 - calculate H_2 norm.
lqg - LQG optimal control synthesis.
ltru - LQG loop transfer recovery.
ltry - LQG loop transfer recovery.

h2syn - H_2 control design

Department “LQGBOX” TFRT-7575

oldboxes (this might not work anymore)

[K,S] = lqrc(A,B,Q1,Q2,Q12)
[L,P] = lqec(A,C,R1,R2,R12)
kr = refc(A,B,C,D,K)
[Ac,By,Byr,Cc,Dy,Dyr] = lqgc(A,B,C,D,K,kr,L)
lqed, lqrd, refd, lqgd in discrete time
Bo Bernhardsson, K. J. Åström Control System Design - LQG
 Lecture - LQG Design

Introduction
The H2 -norm
Formula for the optimal LQG controller
Software, Examples
Properties of the LQ and LQG controller
Design tricks,how to tune the knobs
What do the “technical conditions” mean?
How to get integral action etc
Loop Transfer Recovery (LTR)
More Examples

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Robustness

As we will see, LQ-control u = − Lx automatically gives amazing

robustness properties: Infinite gain margin and 60 degrees phase margin

(Warning: Only if one uses block diagonal weights !)

Critique: Rosenbrock, McMorran: Good, bad or optimal, IEEE-AC 1971.

Horowitz.

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Properties of LQ control

When all states can be measured, we have nice robustness properties

Loop Gain: K (s I − A)−1 B

Return Difference: I + K (s I − A)−1 B

Compare with LQG (if D = 0)

Loop Gain: C (s I − A)−1 B K (s I − A + B K + LC )−1 L

(remark on notation: B = B2 on the LQ slides below)

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Return Difference Formula

From Riccati equation (nice matrix exercise):

M T (−s) M (s) =
( I + K (−s I − AT )−1 B)T D T12 D 12 ( I + K (s I − A)−1 B)

where M (s) = D 12 + C 1 (s I − A)−1 B

If no crossterms:

If C 1T C 1 = Q 1 , C 1T D 12 = 0 and D T
12 D 12 = Q 2

Q 2 + BT (−s I − AT )−1 Q 1 (s I − A)−1 B =

( I + K (−s I − AT )−1 B)T Q 2 ( I + K (s I − A)−1 B)

This is the return difference formula for LQ

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Consequences of RDF

( I + K (−s I − AT )−1 B)T Q 2 ( I + K (s I − A)−1 B) ≥ Q 2

For scalar system this becomes

q 2 p1 + K (s I − A)−1 Bp2 ≥ q 2
therefore
p1 + K (s I − A)−1 Bp ≥ 1

Ms ≤ 1

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 LQ Margins, Scalar case

Ms ≤ 1

Disturbance rejection performance improved for all frequencies

Gain Margin [1/2, ∞], Phase Margin ≥ 60 degrees.

Circle criterion: Stability under feedback with any nonlinear time-varying

input gain with slopes in (1/2, ∞).

Requirements: No cross-terms, Q 12 = 0. All states measurable.

TAT: Why isn’t this a violation of Bode’s integral formula?

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 LQ Gain Margin, MIMO

With
S i ( jω ) = (1 + K (s I − A)−1 B)−1
1/2 −1/2
σ̄ ( Q 2 S i ( jω ) Q 2 )≤1
If Q 2 diagonal this gives nice MIMO gain/phase margins, see LQG course.

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 High Frequency Behaviour of LQ control

If Q 12 = 0 then for large ω

K ( jω I − A)−1 B ∼ K B/ω = Q− 1 T
2 B S B/ω

LQ-controller gives loop gain with roll-off 1 (unless K = 0)

Same conclusion for

K ( jω I − A + BL)−1 B ∼ K B/ω = Q− 1 T
2 B S B/ω

Intution for the future: If the open loop system has roll-off larger than 1,
then if one forces the LQG loop gain to approach the LQ loop gain, the
LQG controller will have large high-frequency gain

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Robustness of LQG

Kalman filter producing x̂ has similar (dual) robustness properties

Since the LQG controller combines two robust parts: LQ control and
Kalman filtering, it was for a long time hoped that robustness margins for
the LQG controller would eventually be found

But, output feedback u = − L x̂ was surprisingly (?) found to have no

automatic guarantees for robustness

This was a dissappointment, especially for people hoping to automize

design

Turned attention towards robust control, e.g. H∞ in the 80s

Bo Bernhardsson, K. J. Åström Control System Design - LQG

Bo Bernhardsson, K. J. Åström Control System Design - LQG
 Example - Doyle IEEE TAC 1978

        

 ẋ1   1 1  x1   0 1
   =
     
  +  u +  
v
ẋ2 0 1 x2 1 1
   
 x1  √
y = 1 0    + σw
x2

min E [( x1 + x2 )2 + ρu2 ]

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Example Doyle IEEE TAC 1978

A = [1 1; 0 1];
B2 = [0 ; 1];
B1 = [1; 1];
C2 = [1 0];
C1 = [1 1];
G = B1;
H = 0*C2*B1;
sys=ss(A,B2,C2,0);
syse=ss(A,[B2 G],C2,[0 H]);
rho=1;sigma=1;
[K,S,E] = lqr(A,B2,C1’*C1,rho);
[Kest,L,P] = kalman(syse,1,sigma);
rlqg = -lqgreg(Kest,K);
loopgain = sys*rlqg;

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Doyle’s counter example

Loop gain with ρ = σ =1 (blue), 0.01(red), 0.0001(black)

0.8

0.6

0.4

0.2

−0.2

−0.4

−0.6

−0.8
−1.4 −1.2 −1 −0.8 −0.6 −0.4 −0.2 0

Infinitly small gain and phase margins when ρ and σ become small

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 The symmetric root locus

Assume Q 11 = C T C, Q 12 = 0, Q 22 = ρ I , then for SISO systems

B(s)
G(s) = C (s I − A)−1 B =:
A(s)
P (s)
I + H (s) := I + K (s I − A)−1 B =:
A(s)

Closed loop characteristic equation P (s) = 0 (TAT: Why?)

Riccati equation gives (return difference formula)

Q 2 + G(−s)G(s) = ( I + H (−s)) Q 2 ( I + H (s))

ρ A(−s) A(s) + B(−s) B(s) = ρ P (−s) P (s)

symlocc, symlocd in matlab (oldboxes)

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Symmetric Root Locus

Symmetric root loci for G(s) = s2 (s2s++0.1s

10
+1)
and G(s)/s.

oldboxes;robotdata
[b,a]=tfdata(sys2);b=b{1};a=a{1};
locus=symlocc(b,a,1e-6,1e10,0.003);
plot(locus(:,2:end),’b’,’Linewidth’,2)

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Cheap control ρ → 0

ρ A(−s) A(s) + B(−s) B(s) = ρ P (−s) P (s)

Eigenvalues of closed loop tend to stable zeros of B(−s) B(s) and the
rest tend to ∞ as stable roots of

s2d = const · ρ

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 An interesting formula - cheap control

Z∞ X 1
min p y( t) − 1p2 dt = 2 Re
0 zj
Rez j >0

where the sum is over all non-minimum phase zeros.

Reference: Qui-Davison, Automatica 1993 pp. 337-349

TAT: Where have you seen something similar before?

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Expensive control ρ → ∞

ρ A(−s) A(s) + B(−s) B(s) = ρ P (−s) P (s)

Eigenvalues of closed loop tend to stable zeros of A(−s) A(s)

Example
min u2 , ẋ = x + u
A(s) = s − 1 unstable.
Optimal controller u = −2x gives

ẋ = − x

P (s) = s + 1

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Energy minimizing stabilization

Consider the system

ẋ = Ax + B(u + w), u = −K x

where w is unitary white noise.

The minimal control effort needed to stabilize the system is

X
min E pup2 = 2 Re p
Re p>0

where the sum is over all unstable poles (exercise).

The optimal closed loop system A − B K has eigenvalues in the open

loop stable poles and the mirror image of the open loop unstable poles.

“The cheapest way to stabilize an unstable pole is to mirror it”

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Lecture - LQG Design

Introduction
The H2 -norm
Formula for the optimal LQG controller
Software, Examples
Properties of the LQ and LQG controller
Design tricks,how to tune the knobs
What do the “technical conditions” mean?
How to get integral action etc
Loop Transfer Recovery (LTR)
More Examples

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 How to tune the weights

Having a state-space realization where the states have a physical meaning

aids the intuition

It is helpful to choose scalings so that all interesting signal levels are

roughly the same size

Q 1 incr, or Q 2 decr. gives faster control

R 1 incr, or R 2 decr. gives faster observer

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 (Bryson’s) rule of thumb

Q1 = diag(α1 , . . . , αn )
Q2 = diag(β 1 , . . . , β m )

Let αi ∼ ( xi )−2 and β i ∼ (ui )−2 where xi and ui denote allowable

sizes on state i and input i

Similar intuition for the noise weights R 1 and R 2 .

Note that multiplying all elements of Q by the same factor does not change
the controller. Similar for the R matrices.

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Tuning Tricks

Introducing an extra punishment of

( ẋi + αxi )2

should move the system closer to ẋi = −αxi .

Gives cross-terms

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Another tuning trick

1
G(s) =
(s + 1)(s2 + 1)

Want to increase damping without moving the pole in s = −1.

This can be achieved by weights that are zero on the eigendirections to

s = −1.

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Example -continued

s =
0.7071 0.7071 0.4082
0 + 0.7071i 0 - 0.7071i -0.4082
0 0 0.8165
d =
0 + 1.0000i 0 0
0 0 - 1.0000i 0
0 0 -1.0000
   

 1
 
 2 
1
 ,   
Q 1 = q i qTi , q1 = 
  q2 = 
 0 

   
 
0 −1
2

1.5

0.5

−0.5

−1

−1.5

−2
−2 −1.8 −1.6 −1.4 −1.2 −1 −0.8 −0.6 −0.4 −0.2 0

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Example Aircraft - wind gust turbulence

Taken from Anderson-Moore Optimal Control Linear Quadratic Methods,

pp.222-224

6 state model of aircraft subject to wind gust turbulence

ẋ = Ax + Bu + Bv v, y = Cx + w
Two outputs y f and ya forward and aft accelerations, one input
Open loop resonances at 1.6 and 21 rad/s
 
 −3 · 10−8 0 

 


 −7 · 10−3   


 

 −0.12 1.57  
A=

 
 
 



 −1.57 −0.12   


 


 
 −0.71 21.3 
 

 0   
−21.3 −0.71

See home page for full model

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Open Loop

Turbulence without a controller

−3 Open loop
x 10
2

1.8

1.6
Power spectral density

1.4

1.2

0.8

0.6

0.4

0.2

0
0 5 10 15 20 25
Frequency rad/s

y f (blue) and ya (green)

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Design 1

min E [ y2f + ya2 + 0.2u2 ]

−3 Design 1 Design 1
x 10 10
2

1.8 9

1.6 8
Power Spectral Density

1.4 7

Controller gain
1.2 6

1 5

0.8 4

0.6 3

0.4 2

0.2 1

0 0
0 5 10 15 20 25 0 5 10 15 20 25
Frequency rad/s Frequency rad/s

Want to increase damping of resonance at 1.5 rad/s

Penalise x3 and x4 more

Bo Bernhardsson, K. J. Åström Control System Design - LQG
 Design 2

min E [ y2f + ya2 + 4x23 + 4x24 + u2 ]

−3 Design 2 Design 2
x 10 10
2

1.8 9

1.6 8
Power Spectral Density

1.4 7

Controller Gain
1.2 6

1 5

0.8 4

0.6 3

0.4 2

0.2 1

0 0
0 5 10 15 20 25 0 5 10 15 20 25
Frequency rad/s Frequency rad/s

Big improvement at 1.5rad/s

But is Andersson-Moore’s Design 2 any good?

Bo Bernhardsson, K. J. Åström Control System Design - LQG
 Comparison Design 1 vs 2

−3 Design 1(blue) vs 2(red)

x 10
5

4.5

Power Spectral Density u

3.5

2.5

1.5

0.5

0
0 5 10 15 20 25
Frequency rad/s

Design 2 has much more control effort around 1.5 rad/s

But that’s perhaps ok. But how about robustness?

Bo Bernhardsson, K. J. Åström Control System Design - LQG

 Comparison Design 1 vs 2
Nyquist plot − Design 1
Nyquist plot − Design 2
1
1

0
0.5

−1 0

−2 −0.5

−3 −1

−4 −1.5

−5 −2
−1 0 1 2 3 4 −2 −1 0 1 2 3

Andersson-Moore’s design 2 has very bad robustness margins, e.g.

Ms ∼ M t ∼ 20.
A change in process gain of 5 % gives an unstable loop
Conclusion: Even the masters can make a bad design with LQG. No
guarantees for robustness.
Bo Bernhardsson, K. J. Åström Control System Design - LQG
 Lecture - LQG Design

Introduction
The H2 -norm
Formula for the optimal LQG controller
Software, Examples
Properties of the LQ and LQG controller
Design tricks, how to tune the knobs

Next lecture

What do the “technical conditions” mean?

How to get integral action etc
Loop Transfer Recovery (LTR)
More Examples

Bo Bernhardsson, K. J. Åström Control System Design - LQG