LQG Design

L7:1
Lecture 7
LQG Design
Linear Quadratic Gaussian regulator
Control-estimation duality
SRL for optimal estimator
Example of LQG design for MIMO plant
L7:2
LQG regulator
If the process and measurement noises are
assumed Gaussian, then the estimate is
statistically most likely, as well as producing the
minimum squared error
This assumption is not necessary, but has led to
the name Linear Quadratic Gaussian (LQG) for a
regulator designed with an optimal control law
(u = Kx) and a Kalman estimator
Matlab provides the function lqgreg to form the
feedback regulator:
w
v
Hlqg = lqgreg(Kest, K) r = 0 u
+
G +
Form the closed-loop system
with feedback
Hlqg
Kx
L7:3
Selection of covariance matrices Q and R

Measurement noise R
noise signals from the sensors are usually
uncorrelated R is diagonal
expected value of rms noise (i) from each
sensor may be available from manufacturers
specifications R = diag(2i)
Process noise Q
more difficult to specify rational values: often
need trial-and-error with simulation studies
unlikely to have information on crosscorrelation Q is diagonal
if there is a random process disturbance which
can be approximated as white noise with a
mean-square spectral density Si, then Qi Si/T
Duality between optimal control and estimation

The time-varying optimal control equations are:
M (k ) = P (k ) P( k )[T P(k ) + R ] T P (k )
1
K ( k 1) = [T P(k ) + R ] T P (k )
1
P(k 1) = T M (k ) + Q
This is a backward recursion starting from
P(N) = Q, K(N) = 0
The time-varying optimal estimation equations are:
P(k ) = M (k ) M (k )C
[CM(k )C
+ R]
CM (k )
L(k ) = P(k )CT R 1

M (k + 1) = P( k ) T + wQwT
This is a forward recursion starting from
M (0) = E {~
x ( 0) ~
x T (0)}
L7:4
Duality between optimal control and estimation

M (k ) = P(k ) P(k )[ P (k ) + R ] T P(k )
1
Control:
K ( k 1) = [ P (k ) + R ] T P(k )
1
P(k 1) = T M (k ) + Q
P(k ) = M (k ) M (k )C [CM (k )C + R ] CM (k )
T
Estimation:
L(k ) = P(k )CT R 1

M (k + 1) = P( k ) T + wQwT
Apart from the different direction, the recursion

relationships are identical, with these dualities:
Control
Estimation T
M P
Q
P M wQwT
CT
R
R
L7:5
L7:6
Symmetric root locus for SISO system estimator

Consider a system with:
a single process noise input w with rms value w
a single sensed output y corrupted with sensor
noise v with rms value v
The uncertainty of the plant model relative to the
sensor data could then be represented by
= Q/R = (w /v )2
The controlestimation duality has w2Q / R Q / R =
The transfer function from w to y is

Y ( z)
N ( z)
1
= C( zI ) w =
, say
W ( z)
D( z )
The estimator poles will hence
1
(
)
(
)
N
z
N
z
be the stable solutions of the 1 +
=0
1
D( z ) D( z )
characteristic equation
L7:7
Symmetric root locus for SISO system estimator

If the process noise and control signal are
additive, so that w = u , the control and
estimation optimal root loci will be identical
In a pole placement design, the control and
estimator poles could then be selected from the
same symmetric root locus
In the estimator case, increasing the gain = Q/R
corresponds to putting more reliance on the
sensor data larger estimator gains L
w(k)
u(k)
+
x (k )
v(k)
Plant
Estimator
y(k)
L7:8
Example: SRL pole-placement design of controller and

estimator for flexible structure (script flex_srlc)
For controller, = 1.67 dclz = 0.79j0.41, 0.73j0.14
K = [-0.59 0.48 1.45 1.68]
For estimator, = 1349 depz = 0.43j0.57, 0.39j0.16
L = [2.18 3.97 6.49 4.08]T
Symmetric root locus
1.5
Initial Condition Results
To: disp d
estimator poles 2-3 times

faster than controller poles
0.5
0
-1
To: disp y
-2
Amplitude
2
0
-2
-0.5
To: force u
si
x
A
g
a
m
I
-1
10
0
-1.5
-1.5
-1
-0.5
0
Real Axis
0.5
1.5
control force prev design
-10
2
8
Time (sec.)
10
12
14
16
L7:9
Example: SRL pole-placement design of controller and

estimator for flexible structure (script flex_srlc)
For controller, = 1.67 dclz = 0.79j0.41, 0.73j0.14
K = [-0.59 0.48 1.45 1.68]
For estimator, = 2.5e5 depz = -0.07j0.52, 0.13j0.13
L = [3.71 9.61 20.1 18.2]T
Symmetric root locus
1.5
Initial Condition Results

1
To: disp d
estimator poles 4-5 times

faster than controller poles
0
-1
To: disp y
5
0
-5
150
-0.5
To: force u
si
x
A
g
a
m
I
Amplitude
0.5
-1
100
50
control force even bigger!
0
-50
0
-1.5
-1.5
-1
-0.5
0
Real Axis
0.5
1.5
8
Time (sec.)
10
12
14
16
L7:10
Comparison of state estimates
low L = [2.18 3.97 6.49 4.08]T high L = [3.71 9.61 20.1 18.2]T
Response of states and predictive estimates
to x (0) = 1
1
4
Response of states and predictive estimates

1 to x (0) = 1
x 1=d
x 1hat
x 2=ddot
x 2hat
x 1=d
x 1hat
x 2=ddot
x 2hat
5
0
0
-5
-2
-4
10
-10
0
5
Time (s)
10
10
-5
5
Time (s)
10
10
x 3=y
x 3hat
-2
x 4=ydot
x 4hat
-5
-5
-20
10
-10
0
5
Time (s)
10
-15
0
5
Time (s)
10
50
x 3=y
x 3hat
20
5
Time (s)
10
40
-10
0
5
Time (s)
x 4=ydot
x 4hat
-50
-40
0
5
Time (s)
10
-100
0
5
Time (s)
State estimates converge more

rapidly, but initial errors larger
larger initial control force
10
LQG design of MIMO regulator for flexible

structure (script flex_lqg)
w(k)
v(k)
u(k)
+
x (k )
Plant
Estimator
Controller design:
weight displacements, not velocities
Qc = diag(Q11, 0, Q22, 0)
scalar input Rc = 1
K = dlqr(Phi, Gam, Qc, Rc)
Kalman estimator design:
assume 1% rms sensor noise on d and y
Re = diag(0.012, 0.012)
trial-and-error values for process noise Qe
L7:11
y(k) = [d y]T
LQG design of MIMO regulator for flexible

structure (script flex_lqg)
Kalman estimator design:
Plant model must have process noise input:
P = ss(Phi, [Gam Gam], C, [D D], T)
[Kest, L] = kalman(P, Qe, Re)
Form regulator:
Hreg = lqgreg(Kest, K)
Close loop:
Gcl = feedback(Gpd, Hreg, +1)
w(k)
v(k)
u(k)
+
x (k )
Plant
Estimator
y(k) = [d y]T
L7:12
L7:13
Simulate (model sim_flex_lqg)

4
up
Sensor noise
v
Process noise
2
Gp(1:2,1)
Sum
Continuous
plant model
Sum1
1
yp
Hreg
Discrete regulator
3
u
ys
L7:14
Q11 = 1.67, Q22 = 0

Qe = 0.15, Re = diag(0.012, 0.012)
As per actual disturbances
L7:15
Actual noise corresponds

with estimator design
Sensor rms noise ten times

estimator design value
Some further considerations
L7:16
Pole placement control design

For arbitrary placement of controller poles, the plant (A, B) or
(
, ) must be controllable
For arbitrary placement of estimator poles, the plant (A, C) or
(
, C) must be observable
Optimal control design

The state-weighting matrix Q must be positive semidefinite
(xTQx 0)
The control-weighting matrix R must be positive definite
(uTRu > 0)
The plant (A, B) or (
, ) must at least be stabilisable (i.e., any
unstable modes are controllable)
The plant (A, C) or (
, C) must at least be detectable (i.e., any
unstable modes are observable)
Robustness
For an LQ regulator with a diagonal R, the closedloop system will have a
gain margin of
gain-reduction margin of 0.5
phase margin 60
in each plant input control channel
Given the duality between optimal control and
estimation, the Kalman filter has similar robust
properties
However, the LQG combination can have
arbitrarily poor stability margins!
Increasing the speed of the estimator dynamics
may reduce stability margins
L7:17
Loop transfer recovery

It is possible to recover the robustness of the LQ
regulator, for minimum phase systems, by a procedure
called loop transfer recovery (LTR)
This involves cancellation of the plant zeros by some of
the filter poles; the remaining poles may become
arbitrarily fast
LTR concept:
suppose an LQG regulator is designed for realistic process and
sensor noise values
now add increasing amounts of fictitious process noise
adjustment to estimator design
in coping with added noise, LQG controller becomes more
robust to gain and phase changes at plant input
however, it is no longer optimal for actual noise levels
LTR can be targeted (say to vicinity of gain crossover
frequency) by frequency shaping of fictitious noise
L7:18
LTR procedure
w
r=0 +
For the LQR, u = Kx,

so the loop gain is HG = K (sI A )1 B
L7:19
H
We have seen that this gives desirable
robustness properties (phase and gain margins)
For the LQG regulator, u = Kx , so that the loop
1
1
gain is HG = K [sI A + BK + LC] L C(sI A ) B
It can be shown that this approaches the ideal LQR
loop gain if L = B, as . (Glad & Ljung, ch. 9)
One way of achieving this is to design the Kalman filter
assuming w = (process noise is additive with control
input), and setting Q = R.
As increases (large process noise), the optimal
estimation solution approaches L = B . The advantage
of this approach is that the estimator is guaranteed
stable for any .
Case studies
Hot steel rolling mill
Matlab Control Systems Toolbox Users Guide
(On-line pdf documentation)
milldemo.m
Magnetic tape drive
Franklin, Powell & Workman (3rd edn), sec. 9.5.4
L7:20

LQG Design

Uploaded by

Copyright:

Available Formats

LQG Design

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

LQG Design

Uploaded by

Copyright:

Available Formats

L7:1

Selection of covariance matrices Q and R

Duality between optimal control and estimation

L(k ) = P(k )CT R 1

Duality between optimal control and estimation

L(k ) = P(k )CT R 1

Apart from the different direction, the recursion

Symmetric root locus for SISO system estimator

The transfer function from w to y is

Symmetric root locus for SISO system estimator

Example: SRL pole-placement design of controller and

Initial Condition Results

estimator poles 2-3 times

control force prev design

Example: SRL pole-placement design of controller and

Initial Condition Results

estimator poles 4-5 times

control force even bigger!

Comparison of state estimates

Response of states and predictive estimates

State estimates converge more

LQG design of MIMO regulator for flexible

LQG design of MIMO regulator for flexible

Simulate (model sim_flex_lqg)

Q11 = 1.67, Q22 = 0

Actual noise corresponds

Sensor rms noise ten times

Some further considerations

Pole placement control design

Optimal control design

Loop transfer recovery

For the LQR, u = Kx,

You might also like

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.