linear quadratic gaussian - university of colorado
TRANSCRIPT
ECE5530: Multivariable Control Systems II. 6–1
LINEAR QUADRATIC GAUSSIAN
6.1: Deriving LQG via separation principle
■ We will now start to look at the design of controllers for systems
Px.t/ D A.t/x.t/ C Bu.t/u.t/ C Bw.t/w.t/
y.t/ D Cy.t/x.t/ C v.t/:
where w.t/, v.t/ are white, Gaussian, zero-mean, independent.
v.t/ ! N .0; Sv/; w.t/ ! N .0; Sw/; x.0/ ! N .x0; „x;0/:
■ Time-varying plant dynamics can be handled quite easily, but we will
concentrate on time-invariant case.
OBJECTIVE: Design a compensator U D Gc.s/Y so that the closed-loop
system is stable, and system achieves “good performance.”
■ Our goal will be to minimize the cost function
J D E
!
1
2xT .tf /Hx.tf / C
1
2
Z tf
0
xT .t/Qx.t/ C uT .t/Ru.t/ dt
"
;
where H; Q; R > 0 and we measure only the system output.
■ We have seen two versions of this problem already:
1. Noise-free: u.t/ D "K.t/x.t/.
2. Noisy, but with complete state observations: u.t/ D "K.t/x.t/.
Same K.t/ but cost is increased.
3. Now, noisy output observations. We will show that
u.t/ D "K.t/ Ox.t/ where Ox.t/ is the output of a Kalman filter.
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett
ECE5530, LINEAR QUADRATIC GAUSSIAN 6–2
■ This is called the separation principle—using estimate Ox.t/ as if it
were the true state.
KEY POINT: This is exactly the same separation operation that we
performed before when we designed state-space controllers via
pole-placement techniques.
➠ Validation of that procedure was based only on the stability of the
closed-loop system.
■ We are now saying that the separation into two independent optimal
sub-problems is actually what optimizes the original objective.
➠ On one hand, the separation is intuitively appealing: Ox.t/ is our
best guess of what is going on inside the system. So, in some sense,
it is clear we should use it in the regulator.
➠ But to extend this intuition further and postulate that it is actually
the optimal procedure to follow is not obvious. (e.g., Noise )
Uncertainty ) Reduced control gains??)
Separation Principle for Optimal Performance
■ Our approach to solving the optimal control problem with finite and
noisy measurements will be
1. Rewrite cost and system in
terms of the estimator states and
dynamics ➠ Can access these.
2. Design a stochastic LQR
controller for this revised system
➠ Full state feedback of Ox.t/.
3. Connection u.t/ D "K.t/ Ox.t/.
Ox.t/
w.t/
u.t/y.t/
v.t/
"K.t/ Kalmanfilter
Plant
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett
ECE5530, LINEAR QUADRATIC GAUSSIAN 6–3
■ Start with the cost
EŒxT Qx! D EŒfx " Ox C OxgT Qfx " Ox C Oxg!
D EŒfx " OxgT Qfx " Oxg! C 2EŒfx " OxgT Q Ox! C EŒ OxT Q Ox!
D EŒ QxT Q Qx! C 2EŒ QxT Q Ox! C EŒ OxT Q Ox!:
■ Each term is a scalar, and unchanged by the trace operator,
EŒxT Qx! D trace.EŒ QxT Q Qx!/ C trace.EŒ QxT Q Ox!/ C EŒ OxT Q Ox!
D trace.EŒ Qx QxT Q!/ C trace.EŒ Ox QxT Q!/ C EŒ OxT Q Ox!
D trace.„xQ/ C EŒ OxT Q Ox!
because Ox and Qx are orthogonal if a Kalman filter is used to estimate
x (Proved in text).
■ We can now write the LQG cost as
J D E
!
1
2OxT .tf /H Ox.tf / C
1
2
Z tf
0
OxT .t/Q Ox.t/ C uT .t/Ru.t/ dt
"
C1
2trace
#
„x.tf /H C
Z tf
0
„x.t/Q dt
$
:
■ The terms within the tracefg operator are independent of the control
input u.t/ so J is minimized whenever
J D E
!
1
2OxT .tf /H Ox.tf / C
1
2
Z tf
0
OxT .t/Q Ox.t/ C uT .t/Ru.t/ dt
"
is minimized.
■ This looks like a stochastic LQR problem in the (measurable) state Ox.
But, note that the dynamics of Ox are different from the dynamics of x.
Need to show that the assumptions of LQR are met.
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett
ECE5530, LINEAR QUADRATIC GAUSSIAN 6–4
■ Estimator dynamics:
POx.t/ D A Ox.t/ C Buu.t/ C L.t/%
y.t/ " Cy Ox.t/&
:
KEY FACT: The innovations process r.t/ D y.t/ " Cy Ox.t/ is white,
Gaussian ! N .0; Sv/. (Shown in section A12 of book appendix).
■ So, LQG problem is to minimize
J D E
!
1
2OxT .tf /H Ox.tf / C
1
2
Z tf
0
OxT .t/Q Ox.t/ C uT .t/Ru.t/ dt
"
C const
subject to the dynamics
POx.t/ D A Ox.t/ C Buu.t/ C L.t/r.t/
L.t/ known;
r.t/ ! N .0; Sv/:
■ This is a (strange looking) stochastic LQR problem.
➠ Solution independent of the white driving noise.
➠ Optimal solution a linear feedback law u.t/ D "K.t/ Ox.t/.
➠ K.t/ is from solution of LQR with data (A, B, Q, R) . . . same.
➠ Therefore, LQG D combination of LQR/LQE is
PERFORMANCE OPTIMAL.
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett
ECE5530, LINEAR QUADRATIC GAUSSIAN 6–5
6.2: Satellite-tracking example
EXAMPLE: A satellite tracking antenna, subject to random wind torques,
can be modeled as"
P".t/R".t/
#
D
"
0 1
0 "0:1
# "
".t/P".t/
#
C
"
0
0:001
#
u.t/ C
"
0
0:001
#
w.t/
y.t/ Dh
1 0i
"
".t/P".t/
#
C v.t/;
where
■ ".t/ is the pointing error of the antenna in degrees,
■ u.t/ is the control torque in N m, and
■ w.t/ is the wind torque (disturbance input) in N m.
■ The wind torque is assumed to be white noise with a spectral density
of Sw D 5000 N2 m2 Hz"1.
■ The measurement noise is white with a spectral density
Sv D 1 deg2 Hz"1.
■ The cost function to be minimized is
J D E
!Z 100
0
q"2.t/ C ru2.t/ dt
"
:
■ The simulation is initialized with
Ox.0/ D
"
0
0
#
; „x.0/ D
"
0 0
0 0
#
:
■ The MATLAB code is from Burl. Results follow.
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett
ECE5530, LINEAR QUADRATIC GAUSSIAN 6–6
■ First, the baseline example with: q D 180; r D 1; Sw D 5000; Sv D 1.
0 20 40 60 80 1000
20
40
60
80
100
Time (s)
Feed
back
gai
ns
0 20 40 60 80 1000
0.1
0.2
0.3
0.4
Time (s)
Kalm
an g
ains
0 20 40 60 80 100−2
−1
0
1
2
3
4
Time (sec)
Angl
e er
ror (
deg)
ActualEstimated
0 20 40 60 80 100−100
−50
0
50
Time (s)
Con
trol (
N-m
)
■ Increased regulation requirement: q D 18000; r D 1; Sw D 5000;
Sv D 1. State feedback gains are larger and approach steady state
more rapidly, but Kalman gains remain the same. Larger control
inputs and better control performance.
0 20 40 60 80 1000
200
400
600
800
1000
Time (s)
Feed
back
gai
ns
0 20 40 60 80 1000
0.1
0.2
0.3
0.4
Time (s)
Kalm
an g
ains
0 20 40 60 80 100−1
−0.5
0
0.5
1
1.5
Time (sec)
Angl
e er
ror (
deg)
ActualEstimated
0 20 40 60 80 100−400
−200
0
200
400
Time (s)
Con
trol (
N-m
)
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett
ECE5530, LINEAR QUADRATIC GAUSSIAN 6–7
■ Increased control-effort penalty: q D 180; r D 100; Sw D 5000; Sv D 1.
State feedback gains smaller, converge more slowly. Kalman gains
remain the same. Smaller control inputs, increased angle errors.
0 20 40 60 80 1000
2
4
6
8
10
12
Time (s)
Feed
back
gai
ns
0 20 40 60 80 1000
0.1
0.2
0.3
0.4
Time (s)
Kalm
an g
ains
0 20 40 60 80 100−15
−10
−5
0
5
Time (sec)
Angl
e er
ror (
deg)
ActualEstimated
0 20 40 60 80 100−2
0
2
4
6
8
10
Time (s)
Con
trol (
N-m
)
■ High disturbance: q D 180; r D 1; Sw D 500000; Sv D 1. Kalman gains
larger, converge much more rapidly. State feedback gains remain the
same. Large gains make estimator faster but larger plant noise
means that estimate is less accurate. Control input also increased
due to larger gains and increased plant noise. Angle error increased
by roughly 10, which is square-root of increase in spectral density.
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett
ECE5530, LINEAR QUADRATIC GAUSSIAN 6–8
0 20 40 60 80 1000
20
40
60
80
100
Time (s)
Feed
back
gai
ns
0 20 40 60 80 1000
0.5
1
1.5
Time (s)
Kalm
an g
ains
0 20 40 60 80 100−20
−10
0
10
20
Time (sec)
Angl
e er
ror (
deg)
ActualEstimated
0 20 40 60 80 100−400
−200
0
200
400
Time (s)C
ontro
l (N
-m)
■ Large sensor noise: q D 180; r D 1; Sw D 5000; Sv D 100. Kalman
gains smaller, converge more slowly. State feedback gains remain
same. Smaller Kalman gains make estimator slower and larger
measurement noise makes estimates less accurate. Control is poorer
due to increased measurement noise.
0 20 40 60 80 1000
20
40
60
80
100
Time (s)
Feed
back
gai
ns
0 20 40 60 80 1000
0.01
0.02
0.03
0.04
0.05
0.06
Time (s)
Kalm
an g
ains
0 20 40 60 80 100−4
−2
0
2
4
6
Time (sec)
Angl
e er
ror (
deg)
ActualEstimated
0 20 40 60 80 100−40
−20
0
20
40
Time (s)
Con
trol (
N-m
)
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett
ECE5530, LINEAR QUADRATIC GAUSSIAN 6–9
6.3: Steady-state LQG control and the compensator
■ Assume time-invariant plant dynamics.
■ Use steady-state cost. Analysis (Stengel) shows (where
´.t/ D C´x.t/)
JLQG D limt!1
EŒ´T .t/Q´´.t/ C uT .t/Ru.t/!
D trace
#
†x;ss LssSvLTss C „x;ss C T
´ Q´C´
$
D tracen
†x;ss BwSwBTw C „x;ss KT
ssRKss
o
where terms are labeled regulator or estimator and
AT †x;ss C †x;ssA C C T´ Q´C´ " †x;ssBuR"1BT
u †x;ss D 0
A„x;ss C „x;ssAT C BwSwBT
w " „x;ssCTy S"1
v Cy„x;ss D 0
and Lss D „x;ssCTy S"1
v ; Kss D R"1BTu †x;ss.
■ Can evaluate steady-state cost from the solution of two Riccati
equations. More complicated than the stochastic LQR result
JLQR D trace˚
.Q C KTssRKss/†x;ss!
'
:
■ Reason: JLQG must also account for cost increase because of
estimation error.
■ In LQG, Ox ¤ x in general. Concepts of (1) Regulation error (x ¤ 0)
and (2) Estimation error ( Ox ¤ x). Both contribute to the cost.
Further interpretations
FAST REGULATOR: To see that both errors contribute, consider the result
of a very fast regulator ➠ Control effort cheap, R ! 0.
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett
ECE5530, LINEAR QUADRATIC GAUSSIAN 6–10
■ Precise analysis requires we find impact of R ! 0 on †x;ss but
JLQG D tracen
†x;ssLssSvLTss C „x;ssC
T´ Q´C´
„ ƒ‚ …
not fn of R
o
# trace˚
„x;ssCT´ Q´C´
'
for all R
In particular limR!0
JLQG # trace˚
„x;ssCT´ Q´C´
'
.
■ Even in the limit of no control penalty, the cost is lower-bounded by a
term associated with the estimation error in the state „x;ss.
FAST ESTIMATOR: Low sensor noise and fast estimation, Sv ! 0.
■ Use second identity to show limSv!0
JLQG # trace˚
†x;ssBwSwBTw
'
.
■ Regulation error remains: lower bound on performance with fast
estimator.
■ Both cases illustrate that it is futile to make either the estimator or the
regulator much faster than the other.
■ Ultimate performance limited and quickly reach point of diminishing
returns.
RULE OF THUMB: Select R, Sv so that estimator and regulator poles are
roughly the same (order of magnitude) distance to origin ➠ Similar
levels of authority. Since Sv is generally physical, this places a design
constraint on R.
The compensator
■ Classical control system design techniques create a compensator
D.s/ to control a plant G.s/.
■ State-space methods use state feedback and estimation.
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett
ECE5530, LINEAR QUADRATIC GAUSSIAN 6–11
■ Can combine the state-feedback law and estimation law to derive a
compensator. Usually called K.s/.
REGULATOR: u.t/ D "K.t/ Ox.t/ where
K.t/ D R"1BTu P.t/
" PP .t/ D AT P.t/ C P.t/A C C T´ Q´C´ " P.t/BuR"1BT
u P.t/I P.tf / D H:
ESTIMATOR: Dynamics
POx.t/ D A Ox.t/ C Buu.t/ C L.t/%
y.t/ " Cy Ox.t/&
; Ox.0/ D EŒx0!
L.t/ D „x.t/C Ty S"1
v
P„x.t/ D A„x.t/ C „x.t/AT C BwSwBTw " „x.t/C T
y S"1v Cy„x.t/I „x.0/ D „x;0:
COMPENSATOR: Combine the above to get (where xc D Ox)
Pxc.t/ D .A " BuK.t/ " L.t/Cy/xc.t/ C L.t/y.t/
u.t/ D "K.t/xc.t/:
■ In steady-state, L.t/ D Lss and K.t/ D Kss and we get
Pxc.t/ D .A " BuKss " LssCy/xc.t/ C Lssy.t/
u.t/ D "Kssxc.t/:
■ We can then generate a transfer function for the compensator
K.s/ D NC .sI " NA/"1 NB:
(what are NA, NB, and NC ?)
r.t/y.t/ y.t/u.t/
: : : or : : :K.s/ "K.s/
G.s/ G.s/
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett
ECE5530, LINEAR QUADRATIC GAUSSIAN 6–12
Closed-loop poles (steady-state)
■ The regulator has poles at det.sI " A C BuKss/ D 0;
■ The estimator has poles at det.sI " A C LssCy/ D 0;
■ The closed-loop poles are the combination of regulator and estimator
poles (via the separation principle).
■ The compensator poles are at det.sI " A C BuKss C LssCy/ D 0 which
are neither the estimator nor the regulator poles. In particular, the
compensator might be unstable!
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett
ECE5530, LINEAR QUADRATIC GAUSSIAN 6–13
6.4: Reference tracking
■ So far, we have discussed only regulator design—forcing the system
state to zero as quickly as possible and keeping it there.
■ This is the same as “tracking” the reference input 0.
■ What if we want to track a more complicated reference signal?
Tracking constant reference inputs
TRACKING VIA COORDINATE TRANSLATION: A constant plant output can
be generated if the plant state is constant.
■ Generate the desired constant plant output with a specific state xd
and regulate around that state.
u.t/ D "KŒx.t/ " xd .t/!:
■ The desired state is chosen so that Cyxd D r .
■ Often, many different xd will work in this equation.
■ There is usually steady-state error when using this reference-tracking
method. Form Nx.t/ D x.t/ " xd.t/.
■ Then,
PNx.t/ D Px.t/ D AŒ Nx.t/ C xd ! C BuŒ"K Nx.t/! D .A " BuK/ Nx.t/ C Axd:
■ Notice that this differential equation is not homogeneous—it has input
Axd . So, in general, Nx.t/ 6! 0.
■ We have an additional constraint Axd D 0. This can be met only if xd
is in the null-space of A.
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett
ECE5530, LINEAR QUADRATIC GAUSSIAN 6–14
■ In order for A to have a null-space, it must be singular—have an
eigenvalue of 0. An integrator! (More on this later if A does not have
an integrator).
■ In general, A must have as many 0 eigenvalues as there are
reference inputs.
■ If Cyxd D r and Axd D 0 then the plant model becomes
PNx.t/ D A Nx.t/ C Buu.t/
and standard LQR theory can be applied to generate u.t/.
FEEDFORWARD CONTROL: An alternate approach, which works even
when A does not contain integrators, is
u.t/ D "Kx.t/ C Krr
where Krr is feedforward control (like the NN approach in ECE5520).
The closed-loop system is
Px.t/ D .A " BuK/x.t/ C BuKrr
y.t/ D Cyx.t/:
■ Assuming stability, using the final value theorem,
y.1/ D "Cy.A " BuK/"1BuKrr:
The inverse exists because the closed-loop Acl matrix is stable, so
has no eigenvalues at 0.
■ Set r D y.1/ and solve for Kr . For a square system,
Kr D "ŒCy.A " BuK/"1Bu!"1:
■ Kr is the inverse of the dc-gain of the system.
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett
ECE5530, LINEAR QUADRATIC GAUSSIAN 6–15
■ Not a robust method. Feed-forward control is more sensitive to plant
modeling errors than feedback control.
EXAMPLE: Consider the state equation
Px.t/ D
"
0 1
"10 "1
#
x.t/ C
"
0
1
#
u.t/
y.t/ Dh
1 0i
x.t/
and the LQR cost function
J D
Z 1
0
y2.t/ C 0:001u2.t/ dt:
■ Using feedforward control,
u.t/ D "h
23 5:9i
x.t/ C 33r:
0 0.5 1 1.5 20
0.2
0.4
0.6
0.8
1
1.2
1.4
Time (sec)
y(t)
0 0.5 1 1.5 20
5
10
15
20
25
30
35
Time (sec)
Con
trol t
orqu
e
■ Zero steady-state error has been achieved since our model is
“perfect”.
INTEGRAL CONTROL: We saw in the first method that constant reference
tracking could be achieved if the plant contained integrator(s).
■ If the plant does not contain integrators, we can add them:
u.t/ xI .t/
r.t/
y.t/ e.t/Plant
1
s
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett
ECE5530, LINEAR QUADRATIC GAUSSIAN 6–16
■ The integrator error equation is
PxI .t/ D e.t/ D r.t/ " y.t/
D r.t/ " Cyx.t/:
Note that there are as many integrators as there are reference inputs.
■ We can include the integral state into our normal state-space form by
augmenting the system dynamics
"
Px.t/
PxI .t/
#
D
"
A 0
"Cy 0
# "
x.t/
xI .t/
#
C
"
Bu
0
#
u.t/ C
"
0
I
#
r.t/:
■ Note that the new “A” matrix has an open-loop eigenvalue at the
origin. This corresponds to increasing the system type, and integrates
out steady-state error.
■ LQR may be used to generate a state feedback for the augmented
plant. Use a cost function that penalizes the integral of the error
J D
Z tf
0
xTI .t/xI .t/ C uT .t/Ru.t/ dt
D
Z tf
0
h
xT .t/ xTI .t/
i"
0 0
0 I
# "
x.t/
xI .t/
#
C uT .t/Ru.t/ dt:
■ The control weighting matrix R is generated by trial-and-error since
the physical meaning of the cost function is nebulous.
■ The control law is,
u.t/ D "K.t/
"
x.t/
xI .t/
#
D "h
Kx KI
i"
x.t/
xI .t/
#
:
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett
ECE5530, LINEAR QUADRATIC GAUSSIAN 6–17
■ Generally, integral control is more robust and precise than the other
two methods.
EXAMPLE: For the same system considered in the example for
feedforward control, use integral control instead.
■ The augmented system becomes
"
Px.t/
PxI .t/
#
D
2
64
0 1 0
"10 "1 0
"1 0 0
3
75
"
x.t/
xI .t/
#
C
2
64
0
1
0
3
75 u.t/ C
2
64
0
0
1
3
75 r.t/:
■ Ignoring the reference input and applying LQR gives
u.t/ D "h
31 7i
x.t/ C 100
Z t
0
fr " y.t/g dt:
The control weighting R D 0:0001 was chosen by trial and error to
yield desirable pole locations.
0 0.5 1 1.5 2 2.5 30
0.2
0.4
0.6
0.8
1
1.2
1.4
Time (sec)
y(t)
0 0.5 1 1.5 2 2.5 30
2
4
6
8
10
12
Time (sec)
Con
trol t
orqu
e
Tracking time-varying reference inputs
■ The integration method of eliminating steady-state error may be
generalized in order to track more complicated reference-inputs.
■ Integration works with dc signals since it puts an open-loop pole at dc.
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett
ECE5530, LINEAR QUADRATIC GAUSSIAN 6–18
■ Generally, increasing the open-loop gain in a certain band of
frequencies will improve closed-loop tracking of those frequencies.
■ The system dynamics are augmented by a filter:
u.t/ xf .t/
r.t/
y.t/ e.t/Plant Filter
■ Zero steady-state tracking error is obtained when the poles in the
Laplace transform of the reference input are also poles of the loop
transfer function.
■ This is called the internal (signal) model principle.
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett
ECE5530, LINEAR QUADRATIC GAUSSIAN 6–19
6.5: Designing for disturbance rejection
■ LQG is designed to reject white-noise disturbances, but may be
suboptimal when applied to nonwhite disturbances.
■ Can modify LQG to reject nonwhite disturbance.
Feedforward disturbance cancellation with full information
■ Assume that the state and disturbance are measured exactly.
Px.t/ D Ax.t/ C Buu.t/ C Bwowo
y.t/ D Cyx.t/
where wo is a constant disturbance input.
■ Use LQR to generate a state-feedback K matrix, ignoring disturbance
input.
■ Use
u.t/ D "Kx.t/ C Kww0:
■ The closed-loop state equation is
Px.t/ D .A " BuK/x.t/ C BuKww0 C Bw0w0:
■ The steady-state output due to disturbance input is
yw.1/ D "Cy.A " BuK/"1Bw0w0
and the steady-state output due to the feedforward control is
yu.1/ D "Cy.A " BuK/"1BuKww0:
■ These two terms should cancel to eliminate steady-state error due to
disturbance
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett
ECE5530, LINEAR QUADRATIC GAUSSIAN 6–20
"Cy.A " BuK/"1Bw0w0 " Cy.A " BuK/"1BuKww0 D 0
which gives
Kw D "fCy.A " BuK/"1Bug"1Cy.A " BuK/"1Bw0:
■ As with reference tracking, feedforward disturbance canceling is not
very robust.
EXAMPLE: The angular velocity of a field-controlled dc motor is required
to equal a set point regardless of the load torque. The state equation
of the motor is"
P!.t/
P#M .t/
#
D
"
"1 1
0 "0:1
# "
!.t/
#M .t/
#
C
"
0
0:1
#
u.t/ C
"
1
0
#
#L.t/
y.t/ Dh
1 0i
x.t/;
where ! is the angular velocity; #M is the motor torque and #L is the
load torque.
■ The LQR cost function is
J D
Z 1
0
!2.t/ C 0:01u2.t/ dt:
■ We get
u.t/ D "h
14:6 16:1i
x.t/ " 17:1#L.t/:
■ Simulation of step load torque with set-point of zero:
0 5 10 15 20−0.05
0
0.05
0.1
0.15
0.2
0.25
0.3
Time (s)
Angu
lar v
eloc
ity (r
ad/s
)
0 5 10 15 20−20
−15
−10
−5
0
Time (s)
Fiel
d vo
ltage
(V)
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett
ECE5530, LINEAR QUADRATIC GAUSSIAN 6–21
■ Simulation output “perfect” due to exact model used.
Feedforward disturbance cancellation with partial information
■ Usually we have only noisy partial state measurements.
■ Estimate both the plant state and disturbance using a Kalman filter.
■ To estimate disturbance, the plant model must be augmented with a
filter that generates a nearly-constant disturbance.
wo.t/
u.t/w.t/ y.t/
v.t/
Plant
1
s
■ The integral has an artificial noise input that allows the disturbance
input to change slowly.
■ The plant state model for designing the Kalman filter is then
"
Px.t/
Pw0.t/
#
D
"
A Bw0
0 0
# "
x.t/
w0.t/
#
C
"
Bu
0
#
u.t/ C
"
0
I
#
w.t/
y.t/ Dh
Cy 0i
"
x.t/
w0.t/
#
C v.t/
■ The disturbance-canceling controller that uses the noisy partial
measurements is then
u.t/ D "K Ox.t/ " Kw Ow0.t/;
where K is calculated via LQR and Kw is calculated as before.
Integral control for disturbance canceling
■ Integral control may also be used to eliminate steady-state errors due
to step-like disturbances.
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett
ECE5530, LINEAR QUADRATIC GAUSSIAN 6–22
■ The Kalman filter must
estimate both the plant
state and the disturbance
input; if it estimates only the
plant state it will be biased
by the disturbance (the
assumptions made when
deriving the Kalman filter
are violated by non
white-noise disturbance).
Ox.t/
w.t/
O#L.t/
u.t/y.t/
v.t/#L.t/
Oe
r.t/
"Kx
"Kw Cy
Kalmanfilter
Plant
1=s
EXAMPLE: The disturbance canceling example from before."
P!.t/
P#M .t/
#
D
"
"1 1
0 "0:1
# "
!.t/
#M .t/
#
C
"
0
0:1
#
u.t/
C
"
1
0
#
#L.t/ C
"
0
1
#
w.t/
y.t/ Dh
1 0i
x.t/ C v.t/;
■ The augmented state for the purpose of designing a Kalman filter is2
64
P!.t/
P#M .t/
P#L.t/
3
75 D
2
64
"1 1 1
0 "0:1 0
0 0 0
3
75
2
64
!.t/
#M .t/
#L.t/
3
75 C
2
64
0
0:1
0
3
75 u.t/
C
2
64
0 0
1 0
0 1
3
75
"
w.t/
wT .t/
#
y.t/ Dh
1 0 0i
x.t/ C v.t/;
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett
ECE5530, LINEAR QUADRATIC GAUSSIAN 6–23
where wT .t/ is the white noise that drives changes in the load torque.
The noise spectral densities are
S2
64
w
wT
3
75
D
"
0:001 0
0 0:01
#
I Sv D 0:001;
where the spectral density for wT was selected by trial and error.
■ The resulting steady-state Kalman gain matrix is
L Dh
1:8 0:1 3:2iT
:
■ Incorporating this information, we have so far the dynamics of the
estimator:
POx.t/ D A Ox.t/ C Bu.t/ C L.y.t/ " Oy.t//
D .A " LC / Ox.t/ C Bu.t/ C Ly.t/2
64
PO!.t/PO#M .t/PO#L.t/
3
75 D
2
64
"2:8 1 1
"0:1 "0:1 0
"3:2 0 0
3
75
2
64
O!.t/
O#M .t/
O#L.t/
3
75 C
2
64
0
0:1
0
3
75 u.t/ C
2
64
1:8
0:1
3:2
3
75 y.t/:
■ For LQR integral state feedback, the augmented state equation is2
64
P!.t/
P#M .t/
PxI .t/
3
75 D
2
64
"1 1 0
0 "0:1 0
"1 0 0
3
75
2
64
!.t/
#M .t/
xI .t/
3
75 C
2
64
0
0:1
0
3
75 u.t/ C
2
64
0
0
1
3
75 r.t/;
and the cost function is
J D
Z 1
0
x2I .t/ C 0:0001u2.t/ dt:
■ For this cost function, we find that
K Dh
60 34 "100i
:
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett
ECE5530, LINEAR QUADRATIC GAUSSIAN 6–24
■ Note that the integrator state implements POxI .t/ D r.t/ " O!.t/ so we
can augment the estimator dynamics:
2
66664
PO!.t/PO#M .t/PO#L.t/POxI .t/
3
77775
D
2
66664
"2:8 1 1 0
"0:1 "0:1 0 0
"3:2 0 0 0
"1 0 0 0
3
77775
2
66664
O!.t/
O#M .t/
O#L.t/
OxI .t/
3
77775
C
2
66664
1:8
0:1
3:2
0
3
77775
y.t/ C
2
66664
0
0
0
1
3
77775
r.t/
C
2
66664
0
0:1
0
0
3
77775
h
"60 "34 0 100i
2
66664
O!.t/
O#M .t/
O#L.t/
OxI .t/
3
77775
2
66664
PO!.t/PO#M .t/PO#L.t/POxI .t/
3
77775
D
2
66664
"2:8 1 1 0
"6:1 "3:5 0 10
"3:2 0 0 0
"1 0 0 0
3
77775
2
66664
O!.t/
O#M .t/
O#L.t/
OxI .t/
3
77775
C
2
66664
1:8
0:1
3:2
0
3
77775
y.t/ C
2
66664
0
0
0
1
3
77775
r.t/
u.t/ Dh
"60 "34 0 100i
2
66664
O!.t/
O#M .t/
O#L.t/
OxI .t/
3
77775
:
■ Response to step load torque.
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett
ECE5530, LINEAR QUADRATIC GAUSSIAN 6–25
0 5 10 15 20−0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
Time (s)
Angu
lar v
eloc
ity (r
ad/s
)
0 5 10 15 20−12
−10
−8
−6
−4
−2
0
2
Time (s)
Fiel
d vo
ltage
(V)
Frequency-shaped LQG
■ One final “trick” we look at is used to shape the frequency response
of the LQG system.
■ For example, a system may have a resonance frequency we do not
wish to excite. Standard control design would place a notch filter at
this frequency.
■ To perform frequency-shaped LQG design, we place a bandpass filter
in parallel with the plant dynamics.
u.t/ y.t/
xf .t/
Plant
Bandpassfilter
Augmentedplant
■ This adds states that we wish to penalize.
■ When designing a Kalman filter, we need only to estimate the true
plant states as we have access to the filter states.
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett
ECE5530, LINEAR QUADRATIC GAUSSIAN 6–26
Appendix: Steady-state LQG cost
■ We earlier stated (where ´.t/ D C´x.t/)
J D limt!1
EŒ´T .t/Q´´.t/ C uT .t/Ru.t/!
D trace
#
†x;ss LssSvLTss C „x;ss C T
´ Q´C´
$
D tracen
†x;ss BwSwBTw C „x;ss KT
ssRKss
o
:
■ Here, we show that the second and third lines are the same (not yet
sure how to show that the first and second are the same).
■ Start with the second line:
trace˚
†x;ssLssSvLTss C „x;ssC
T´ Q´C´
'
D trace
8
ˆ<
ˆ:
†x
0
B@ „xC T
y S"1v Cy„x
„ ƒ‚ …
A„xC„xAT CBwSwBTw
1
CA C „xC T
´ Q´C´
9
>=
>;
■ But,
C T´ Q´C´ D †xBuR"1BT
u †x„ ƒ‚ …
KT RK
"AT †x " †xA
so
trace˚
†x;ssLssSvLTss C „x;ssC
T´ Q´C´
'
D trace˚
†xBwSwBTw C †x.A„x C „xAT / C „x
(
KT RK " AT †x " †xA)'
D trace˚
†xBwSwBTw C „xKT RK
'
C trace˚
†x.A„x C „xAT /'
" trace˚
„x.AT †x C †xA/'
„ ƒ‚ …
equal
D trace˚
†xBwSwBTw C „xKT RK
'
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett
ECE5530, LINEAR QUADRATIC GAUSSIAN 6–27
where we used the knowledge that
AT †x;ss C †x;ssA C C T´ Q´C´ " †x;ssBuR"1BT
u †x;ss D 0
A„x;ss C „x;ssAT C BwSwBT
w " „x;ssCTy S"1
v Cy„x;ss D 0
and Lss D „x;ssCTy S"1
v ; Kss D R"1BTu †x;ss.
Lecture notes prepared by Dr. Gregory L. Plett. Copyright © 2016, 2010, 2008, 2006, 2004, 2002, 2001, 1999, Gregory L. Plett