LQG Lecture

Download as pdf or txt
Download as pdf or txt
You are on page 1of 41

LQG Control

 Some critical Issues in LQR Control Implementation:


1. Complete State x(t) availability is necessary for full state feedback. This
requires a number of independent sensors equal to the system’s dimensions
(complexity). If the number of sensors is lower, then the remaining states must
be reconstructed (estimated) by an additional dynamic system, requiring
observability and/or detectability.
2. The loop may have signals/inputs that are not deterministic but stochastic
representing a uncertainty in the process or inherent noise due to sensor
dynamics.
• Question: Can we design a dynamic compensator, which is optimal (in a quadratic
sense)?

• Answer 1: From a controller standpoint the answer is YES, since we could design a
LQR (constant feedback gain matrix, guaranteed stability, good stability margins,...)

• Answer 2: From the state estimation standpoint the answer is NO, since we only
have a Luenberger observer, which is guaranteed to be stable but not optimal
LQG Control

 The answer to the second question is the development of an optimal state


estimator (under some conditions) called Kalman Filter, and the use of the
separation principle to connect it with an optimal LQR controller leading to the so-
called Linear Quadratic Gaussian (LQG) compensator.
http://www.stengel.mycpanel.princeton.edu/MAE546Seminars.html

• Although derived for a specific estimation problem, the Kalman Filter and its derivatives, is
probably the most commonly used estimator in industry

• The Kalman filter, the linear-quadratic regulator, and the linear–quadratic–Gaussian


controller are solutions to what arguably are among the most fundamental problems in
control theory.
Kalman Filter and Linear Quadratic Gaussian Optimal Controller – LQG
LQG Control

 Optimal State Estimation in a Quadratic


 Note: Properties of E [.]
Sense.

ìïx = Ax + Bu + w ìïw(t )  é 0,W ù


ï ëê ûú ïï { }
ïïìE x T (t )Q(t )x (t ) = tr éQ(t )C (t, t )ù > 0
êë x úû
ï ï
ïv(t )  é 0,V ù Markov Process
í í ê ú íx (t ) stochastic process
ïïy = Cx + v ï ë
ïx (t )  é h ,Q ù
û ïï
î ï êë 0 0 úû ïïQ(t ) > 0
ï
î 0 î

 w(t), v(t) are wide sense stationary, white { }


C x (t, t ) = E xT (t )x (t )
noise, uncorrelated, ergodic processes
{ }
E x TQx = tr éêëQRx (0)ùúû
W ³ 0,V > 0 é +¥ ù
 are symmetric constant matrices, called the { T
}
E x Ax = tr êA ò Sx (w)df úú
ê
intensity matrices of the two white noise êë -¥ úû
processes (indicating the variance). The initial
state is assumed to be a random vector.

 We wish to find:
xˆ = Axˆ + Bu + K F (y - Cxˆ)
xˆ(t0 ) = xˆ0
Find: K F , xˆ0
min J = min E eT Ze
K F ,x 0 K F ,x 0
{ } ì
ï
ïe = x - xˆ
í
ï
ïZ >0
î
LQG Control (Extra)

 Markov Process

 NOTE: The above result requires that A be Hurwitz. [A < 0]


LQG Control

• Insert the Error Dynamics to modify the  Minimization of heT Z he


Performance Index to be minimized.

ìïe = (A - K C )e + w - K v
( )
min heT Z he  he (t ) = 0
ï F F
í h e = (A - K FC )he
ïïe(t0 ) = x 0 - xˆ0 = e0
î
he (t ) = 0  he (t0 ) = 0
{ }
e(t ) = (he (t ),Q(t )) E e(t ) = he (t )
{ }
E {xˆ0 - x 0 } = 0 E xˆ0 = E x 0 { }
{
Q(t ) = E éêëe - he ùúû ée - h ù
êë e úû
T
}
 Minimization of tr éêëQZ
 ù
úû
Q(t ) ³ Q(t )
E ee { T
} = Q(t ) + he heT
ìïQ. = AQ + QAT + W - QC TV -1CQ
ïï
í
ïïQ(t ) = Q
ïî 0
{ }
0
J = E eT Ze = heT Z he + tr éêëQ(t )Z ùúû
Q(t ) º Q(t )  K F = QC TV -1

• Minimize each term separately


LQG Control

 Lemma 3.1. ‘Linear Optimal Control’, Kwakernaak& Sivan, p. 218.


Given the matrix differential equation:
ì
ï 
ï-Q = -R1 + K R2K + Q(A - BK ) + (A - BK ) Q
T T

í
ï
ïQ(t ) = Q1 ³ 0; R1 ³ 0; R2 > 0
ï
î 1
Given an arbitrary continuous matrix K(t). Then for t0 ≤t ≤ t1,

Q(t ) ³ Q(t )

With Q(t) solution of the matrix differential equation:

ì
ï  T T -1
ïQ = AQ + QA + W - QC V CQ
í
ï
ïQ(t ) = Q0
ï
î 0
The equality holds if : K = QC TV -1

• For a controllable system, and infinite horizon optimization


Q(t )  Q > 0 (const )

AQ + QAT + W - QC TV -1CQ = 0 • KF Riccati equation (ARE)


LQG Control

• (t ) Q(t ) º Q(t )  K = QC TV -1


Proof of Q (t ) £ Q F
LQG Control

 Summary
Given the system:
c. If Q exists, it is ≥ 0 and solution of
ïìïx = Ax + Bu + w
í AQ + QA T + W − QC T V −1CQ = 0
ïïy = Cx + v
î
( unique solution in case b.)

w = [0,W ] uncorrelated, wide sense stationary White Noise


v = [0,V ] uncorrelated, wide sense stationary White Noise d. If and only if the system is controllable. Then:
x (t0 ) = [h(t0 ),Q(t0 )] uncorrelated, wide sense stationary
Q>0

for Q≥0 the following holds: e. If Q exists, the optimal estimator is:
 . .
Q = AQ + QA T + W − QC TV −1CQ
 xˆ = Axˆ + Bu + K F (y - Cxˆ)
Q(t 0 ) = Q0
K F = QC TV -1

a. Let: Q0 = 0 , for t → −∞ If the system is stabilizable and detectable,


the estimator is asymptotically stable.
the above solution tends to a constant matrix
Q>0
f. The asymptotic estimator minimizes:
If and only if there are no poles that are simultaneously
unstable, unobservable, and controllable; { }
lim E eT Ze = tr {QZ }
t -¥

b. If the system is stabilizable and detectable, .


the solution tends to: The Point f. Is true iff E {xˆ0 } = E {x 0 }
Q for t → −∞ ∀Q0 ≥ 0
LQG Control (Extra)
LQG Control (Extra)

 Summary
 Kalman Filter Implementation: the discrete case

Initial Previous New Predicted


State State State

x0 x k −1 x kP = Ax k −1 + Buk + wk
P0 Pk −1 PkP = APk −1AT + Qk

Update with new


Output of Current State Meas. And Kalman Gain
Updated State Becomes Previous
PkP H
xk Pk = [I − KH ]P P K =
k
HPkP H T + R Zk
Pk xk x k = x kP + K [Z k − Hx kP ]
LQG Control

 LQG Control Synthesis • Find the optimal controller u(t) that minimizes:
• Given the LTI stochastic system
{
J = E xT (t f )S (t f )x (t f ) +
ìïx = Ax + Bu + Dw ìïw(t )  é 0,W ù ü
ï
ï ëê ûú
tf
ï ï
ïv(t )  é 0,V ù +ò ( )
ï
x (t )Q(t )x (t ) + u (t )R(t )u(t ) d t ï
í
T T
í ý
ïïy = Cx + v ï êë úû ï
ï
î ï
ïx (t )  êë h0 ,Q0 ùúû
é
t0
ï
þ
ï
î 0 DWDT
Q ³ 0, R > 0, S (t f ) = S f ³ 0
 Main Result
LQG Control

 Outline of Proof: From KF

{ } { } {
E xTQx = E (x - xˆ + xˆ) Q (x - xˆ + xˆ) = E (x - xˆ) Q (x - xˆ) + 2E (x - xˆ) Qxˆ + E xˆTQxˆ
T T
} { T
} { }
 Since noises are uncorrelated: E eTQxˆ = 0 { }
• We know that:
{ } { }
E (x - xˆ) Q (x - xˆ) = E eTQe = tr éêëQ Sùúû
T

• We obtain:
{ }
E x TQx = tr [Q S] + E xˆTQxˆ { }

• Similarly: {
E x Tf S f x f } = tr [S S ] + E {xˆ
f
T
f
S f xˆ f }
ì
ï tf ü
ï é tf ù
ï ï ê ú
J =Eïíò (xˆ Qxˆ + u Ru )dt + xˆf S f xˆf ï
T T T
ý + tr ê ò Q Sdt + S f S f ú
ï ïþïï êt ú
ïï t0
î ëê 0 ûú
ì tf
ï ü
ï
ï ï
J (u ) = E íò (xˆ Qxˆ + u Ru )dt + xˆf S f xˆf ï
 ï T T T
ý u = -Kcxˆ K c = R −1 B T S
ï
ï ï
ï
ï t0
î ï
þ
ì
ï   .
ïxˆ = Axˆ + Bu + K F h S = A T S + SA + Q − SBR −1 B T S
í 
ï
ïh = y - Cxˆ  S (t f ) = S f
î
LQG Control

 The performance of the closed loop system is computed in a stochastic sense, since the state
vector and the output are now stochastic processes

ée ù ée ù éx - h ù
ê ú
x = ê ú x 0 = êê 0 úú = êê 0 0ú
• The closed loop dynamics are now a
ˆ
x xˆ0 ú ê h0 ú
êë úû ê
ë û ë úû Linear System driven by white noise

éA - K C ù é ù éw ù x = AC x + BC x
ê
x = ê F
0 ú x + êD -K F ú ê ú
êë K FC A - BKc úú ê0 K ú
êë F úû
êv ú
êë úû • The statistics are given by the closed loop
û
covariance matrix
Note: x is the augmented state vector of dimension 2n éQ x Q x ù
x é
{ ù
T
é
ë
é ù
}
Q (t ) = E êëx - E (x )úû ê êëx - E (x )úû ú = êê 11x
ù
û
12 ú
x ú
êëQ21 Q22 úû

ìï éW 0 ù
ï
ï x x x T ê ú T
ïQ = ACQ + Q AC + BC ê 0 V ú BC
í ëê úû
ï
ï
ï
ïQ x (t0 ) = Q0x
î

ìïQ x = [A - K C ]Q x + Q x (A - K C )T + DWDT + K VK T Q11x ( t 0 ) = Q0


ïï 11 F 11 11 F F F
ïQ x = Q x C T K T + Q x (A - BK )T + (A - K C )Q x - K VK T Q12 (t 0 ) = Q22 (t 0 ) = 0
í 12
ïï x 11 F 12 C F 12 F F

ïïQ 22 = Q12 C K F + Q22 (A - BKC ) + K FCQ12 + (A - BKC )Q22x + K FVK FT


x T T T x T x
î
LQG Control

• Analysis of the Covariance Matrix differential equation solution:

K F = QC T V −1 Q11x C T K FT − K F VK FT = 0

• Q11x (t ) can be solved independently:


ì
ï x x x T T T
ïQ11 = [A - K FC ]Q11 + Q11(A - K FC ) + DWD + K FVK F
í x
ï
ïQ (t ) = Q0
ï
î 11 0
• Error and state estimate are uncorrelated since:
x
12
x
Q (t 0 ) = 0  Q (t ) = 0
12
E { ée - E (e)ù éxˆ - E (xˆ)ù
êë úû êë
T
úû }=0
ìïQ x = Q x (A - BK )T + (A - BK )Q x + K VK T
ï 22 22 C C 22 F F
í x
ïïQ22 (t0 ) = 0
ïî :

éQ x ù é ù
ê 11 0 ú = êFilter Riccati Solution 0 ú
ê 0 Qx ú ê 0 Lyapunov Equation Solutionúú
êë 22 úû êë û

 Once Q x (t ) is computed all the statistics are known since:


x = e + xˆ
NOTE: In the case of infinite horizon the equations are algebraic and not differential
LQG Control

 LQG Statistics (see Kwakernaak and Sivan, ‘Linear Optimal Control’ , page 95, Theorem 1.50)

• State Vector Mean Square Value


• Relationship between Mean Square
{ T
ë } û { ë }
E x Qx = tr éêQE xxT ùú = tr éêQ(hhT + Q11x + Q22x )ùú
û
Value and Spectral Density
Given a vector stochastic process v(t) and a positive
h = E {xˆ}
Semidefinite symmetric Matrix W(t), then:
• Output Vector Mean Square Value
{ }
E vT (t )W (t )v(t ) = tr {W (t )C v (t, t )}
{ T
ë}
E y y = tr éêC C (hh + Q + Q )ùú
T
û
T x
11
x
22 { }
C v (t1, t2 ) = E v (t1 )v T (t2 )
• Input Mean Square Value If v(t) is wide-sense stationary with zero Mean,
and Covariance Rv (t1 - t2 )
{ } {
E uT Ru = E xˆT KCT RKC xˆ = }
= tr éêKCT RKC (hhT + Q22x )ùú And W(t)= W = constant, then:
ë û

• State Vector Expected Value


{ }
E vT (t )Wv(t ) = tr {WRv (0)}

If v(t) has zero Mean, and PSD å v (w )


E {x(t )} = e E {x (t0 )} ì
ï ü
ï
AC t +¥

é E {e }ù é 0 ù { }
E vT (t )Wv(t ) = tr ïíò
ï
W å v
(w )df ï
ý,
ï
ê 0 ú ê ú ï
ï-¥
î ï
ï
þ
E {x(t0 )} = ê ú =ê ú
êëE {xˆ0 }ûú êë h0 = E {x 0 }úû

w
f = , R (0) = ò åv (w)df
2p v -¥
LQG Control

-1
K LQG (s ) = -KC (sI - A + BKC + K FC ) K F

éA - K C ù é ù éw ù ée ù ée ù éx - h ù
ê
x = ê F
0 ú x + êD -K F ú ê ú x = êê úú x 0 = êê 0 úú = êê 0 0ú
ú
êë K FC A - BKc úú ê0 K ú
êë F úû
êv ú
êë úû êëxˆúû êëxˆ0 úû êë h0 úû
û
8. LQG Control Examples
LQG Control Commands
LQG Control Commands
LQG Control Examples

 Example No. 1: Single Channel Rigid Antenna with noisy controller


y = x1
ì
ï é0 1 ù é ù
ï
ïx = êê úx + ê 0 úu 0.1 x2
ï
í ê 0 -1úú ê0.1ú
êë úû - s(s + 1)
ï ë û
ï é ù
ïïîïy = êë1 0úû x
w1 w2
+ +
u(t ) = - éê20 10ùú {x (t ) + w(t )}
ë û + 20 10

é.0025rad 2 / Hz 0 ù
w(t ) := [0, S w ]  S w = êê ú
êë 0 0.05rad / sec / Hz úú
2 2
û

• Compute the solution to the associated Lyapunov equation, to get the state vector statistics

AQ + QAT + DWDT = 0
é0 1 ùú é 0 0ù é.0025 0 ù
ê
A= ê ê
,D = ê ú ,W = êê ú
-2 -2ú ú ú
2 1ú .05úú
ëê û ëê û êë 0 û
LQG Control Examples

• The solution is found to be: {


Q¥ = diag .0075 .015 }
• The Standard Deviation of the output is 0.087 rad = 5 degrees. This means that the controller is
keeping the antenna position within 5 degrees from desired position for 67 % of the time.

 Example No. 2: Consider now the same controller, but we add a disturbance torque τd, (white
noise with intensity Vd). The output is the angular position with measurement noise vm.

é0 1 ù é 0ù é0 ù
ê
x = ê ú
ú x + ê ú m + êê úú td
ê ú y = éê1 0ùú x + vm ; vm = [0,Vm ]
ë û
êë 0 -a úû êëk úû êë g úû

• Kalman Filter finite time Riccati equation:


ì
ï 1 2
ï
ïq11(t ) = 2q12 (t ) - q (t )
é0 1 ù é0 0 ù é0 0 ùú é1 ù 1 ï
ï Vm 11

Q(t ) = êê ú ê ú ê ê ú é ù ï
0 -a ú Q(t ) + Q(t ) ê1 -a ú + ê 0 g 2V ú - Q(t ) ê0ú V ëê1 0úû Q(t ) ï
ï
.
1
ëê úû ëê úû êë d úû ëê ûú m í 12 (t ) = q22 (t ) - aq12 (t ) -
q q (t )q (t )
ï
ï Vm 11 12
ï
ï
ï 1 2
ï
ïq22 (t ) = -2a q22 (t ) + g 2Vd - q12 (t )
• Asymptotic Solution: Q (t ) = 0  lim Q(t ) = Q¥ ï
ï
î Vm
t ¥

é 2 ù
ê -a + a + 2b a 2 + b - a a 2 + 2b ú Vd
Q¥ = Vm ê ú b=g
ê a 2 + b - a a 2 + 2b
êë (
-a 3 - 2ab + a 2 + b ) a2 + 2b úú
û Vm
LQG Control Examples

é 2 ù
• ê -a + a + 2b ú
Kalman Filter Gain Matrix: KF = ê ú
ê a 2 + b - a a 2 + 2b ú
ëê ûú

• Kalman Filter Poles: det (sI - A + KC ) = s 2 + s ( )


a2 + 2b + b
1
2
(
- a 2 + 2b  a 2 - 2b )
• Numerical Example:
ìïk = 0.787 rad / (Vs 2 )
• Disturbance torque RMS value of ïï
ïïa = 4.6 s -1 é40.36ù
31.6 Nm with constant spectral
ïï K F = êê ú
ú -22.48  j 22.24
density between ± 50 Hz ïíg = 0.1 Kg -1m -2 êë814.3úû
• Measurement noise RMS value of ïï
0.01 rad. With constant spectral ïïVd = 10 N 2m 2s
ïï -7 2
density between ± 500 Hz ïïîVm = 10 rad s

• From the square root of the diagonal terms of Q we can determine the statistics of the reconstruction
error, which are 0.002 rad and 0.06 rad/sec in position and velocity respectively.

 Use Matlab to simulate and plot position and rate


LQG Control Examples
Antenna Comparative Control Design

NASA JPL DSN


Motivation / Considerations

 The performance of a controller improves with the gain increase. However, high gains excite structural vibrations.
For a PI controller, structural vibrations cannot easily be controlled since the structural deformations are not
directly measured by encoders.

 An LQG controller uses a Kalman filter to estimate telescope vibrations, overcoming the difficulty of direct
measurement. The Kalman filter consists of a telescope analytical model that needs to be accurate to produce an
accurate estimate of the telescope structural dynamics.

 LQG controllers guarantee wide bandwidth and good wind-disturbance rejection properties; thus, they can
be used for antennas with stringent pointing requirements in the presence of wind disturbances (modeled
as random processes).

 For antennas and radio telescopes, the LQG control systems can be implemented in two different ways: at
the telescope rate loop or at the position loop (as any other controller).

 The analysis is augmented with the performance of the proportional-and-integral (PI) controller. The latter
analysis is necessary not only for comparison purposes (the PI controller is a standard antenna industry
feature), but also because the implementation of the LQG controller requires preliminary installation of PI
controllers in both rate and position loops.

 Remainder: The implementation of the LQG algorithm needs an accurate telescope model, and it can be
obtained from the field test of the telescope.
General Controller Structure

The PP control system is a typical telescope


control system configuration.
The PL case is the configuration of the NASA
Deep Space Network (DSN) antenna control
system, and it was implemented at the
34-meter antennas in Goldstone, California.
The LQG controller has been considered
by MAN Technologies. The LP and LL
control systems were implemented in 2010, and
2012
 A Command Preprocessor (CPP) deals with the rate and acceleration limits imposed at the drives. During tracking, the telescope
motion is within the rate and acceleration limits. However, during slewing, the large position-offset commands exceed the
acceleration limit or both the acceleration and rate limits. When limits are exceeded, the telescope dynamics are no longer linear, and
the telescope becomes unstable, which is observed in the form of limit cycling. The CPP modifies the telescope commands such that
they remain unaltered if they do not exceed the rate and acceleration limits, and it processes the command to the maximum
acceleration and rate limits if the limits are exceeded by the command. The CPP algorithm represents an integrator, rate and
acceleration limits, a variable-gain controller, and a feed-forward gain.
PP Controller

 The PP control system consists of a PI controller in the position loop and a PI controller in the rate loop.

The rate-loop model consists of the finite-element


model (FEM) of the telescope structure, which includes
the drives and the azimuth (AZ) and elevation (EL) rate-
loop controllers. It is a discrete-time control system with
a 0.001-s sampling time. The proportional-and-integral
gains of the azimuth controller are 300; for the elevation
controller, the proportional gain is also 300, and the
integral gain is 400. The bandwidth of the rate-loop
transfer function is 1.0 Hz, both in azimuth and in
elevation.

 A Command Preprocessor (CPP) deals with the rate and acceleration limits imposed at the drives. During tracking, the telescope
motion is within the rate and acceleration limits. However, during slewing, the large position-offset commands exceed the
acceleration limit or both the acceleration and rate limits. When limits are exceeded, the telescope dynamics are no longer linear, and
the telescope becomes unstable, which is observed in the form of limit cycling. The CPP modifies the telescope commands such that
they remain unaltered if they do not exceed the rate and acceleration limits, and it processes the command to the maximum
acceleration and rate limits if the limits are exceeded by the command. The CPP algorithm represents an integrator, rate and
acceleration limits, a variable-gain controller, and a feed-forward gain.
PP Controller

A feed-forward loop is added to improve tracking accuracy,


especially at high rates. The feed-forward loop differentiates
the command and forwards it to the rate-loop input. The
derivative is the inversion of the rate-loop transfer function.
In this way, we obtain the open-loop transfer function from the
command to the encoder, approximately equal to 1. Indeed, the
magnitude of the rate-loop transfer function Gr is shown in Fig.
It can be approximated (up to 1 Hz) with an integrator
(Grapprox = 1/s), which is shown in the figure (short-dashed line).
The feed forward transfer function is a derivative (Gff(s) = s) shown
as (long-dashed line), so that the overall open-loop transfer
function is a series connection of the feed-forward and the rate
loop Go(s) = Gr(s)Gff(s), which is approximately equal to 1 up to
a frequency of 1 Hz. In this way, the transfer function of the
system is equal to 1 (up to 1 Hz) without applying the position
feedback. The position feedback is added to compensate disturbances
and system imperfections.
PP Controller

The position-loop model consists of the rate-loop model, PI and feed-forward controllers in azimuth and elevation,
command preprocessors in azimuth and elevation, and rate and acceleration limiters in azimuth and elevation.
The telescope rate limit is 1.0 deg/s, and the acceleration limit is 0.5 deg/s2, both in azimuth and elevation. The PI
controller gains were selected to minimize settling time and servo error in wind gusts. They also guarantee zero
steady-state error for constant rate tracking.

Note that the position-loop bandwidth is higher than the rate-loop


bandwidth, both in azimuth and elevation. Note also that the
elevation-axis bandwidth is higher than the azimuth-axis bandwidth,
mainly because the elevation transfer function has only a few
resonances and the resonances are well damped. Azimuth bandwidth
is lower since the controller gains will be low in order to prevent
excitation of multiple resonances.
PP Controller
PL Controller

 The PL control system consists of the LQG controller in the position loop and the PI Controller in the rate
loop.

In this case, the rate-loop model of the telescope is as shown before in the PP control.

The position loop is similar to the previous one, but the PI controllers are replaced with the LQG controllers. The
LQG controller structure is shown below. The controller includes the estimator, which is an analytical model of the
telescope.

The estimator is driven by the same input (uc) as the


telescope, but also by the estimation error ε (the difference
between the actual rate y and the estimated rate yest). The
error is amplified with the estimator gain ke to correct for
transient dynamics. The estimator outputs are the estimated
telescope states that consist of the telescope rate and the
telescope flexible deformations xf. The latter are the missing
vibration measurements, which allow for suppression of the
telescope vibrations
PL Controller

PL

PP
PL Controller

 The results obtained on the LQG position controller


produce a settling time of 1.6 s, bandwidth of 1.3 Hz, and
wind servo error of 0.76 mdeg or 2 times smaller than that
of the PP control system.

 The wind-gust simulations show 0.15-mdeg rms servo error in azimuth and 0.74-mdeg rms servo error
in elevation. These numbers are compared with the PP control system (0.35 mdeg in azimuth and 1.4
mdeg in elevation). This means that the LQG controller improves the servo error in wind over the PI
controller by a factor of 2.3 in azimuth and a factor of 1.9 in elevation.
LP Controller

 The LP control system consists of the PID (proportional-integral-derivative) controller in the position loop
and the LQG controller in the rate loop.

 The position-loop transfer functions for azimuth and elevation show a wide bandwidth of 200 Hz in azimuth and 20
Hz in elevation. The steady-state error due to rate offsets is zero. The wind-gust simulations for a 12-m/s wind
show 0.012-mdeg rms servo error in azimuth and 0.150-mdeg rms servo error in elevation. These small numbers
show that, as compared with the PP control system, the LQG controller in the rate loop improves the servo error
in wind by a factor of 30 in azimuth and a factor of 10 in elevation.
LL Controller (implemented in 2010)

 Finally, consider the complete telescope control system with a LQG controller in the rate and
position loops. This is also a novel configuration in the antenna industry.

 The rate loop is the same as for the LP control system. For the given rate loop, the position-loop
controller was designed to minimize the servo error in the wind gusts.
 The position-loop characteristics are plotted below, it follows that the system settling time is 0.5 s, and
there is no overshoot in either azimuth or elevation. The bandwidth is 20 Hz in azimuth and 40 Hz in
elevation.
LL Controller

 Finally, the wind-gust simulations for a 12-m/s wind


are plotted in the zoomed insert. The figure shows
0.0012-mdeg rms servo error in azimuth and 0.0057-
mdeg rms servo error in elevation, which give a total
rms error of 0.0058 mdeg. It is 250 times smaller
than the error of the PP control system.

 Thus, the LL control system performance is the best


of all the systems presented, although the system is
the most complicated and will require careful tuning
of both rate- and position-loop LQG controllers in
order to obtain the predicted performance.
LL Controller
References
Extra

Measurement
input
X0 X k −1 X kP = AX k −1 + BU k + wk
P0 Pk −1 PkP = APk −1AT + Qk
Yk = CXkM + z k
Initial Prev.
State State

Xk PkP H
Xk KG = = AX k −1 + BU k + wk
Pk = [I − KG H ]P P HPkP H T + R
Pk k
X k = X kP + KG [Yk − HX kP ]
Update covariance Update state

X, U, Y= Process state, input, measurement

w, z = predicted process noise and measurement uncertainty

P = predicted process covariance or error in the process estimate


Q = error in the prediction of covariance matrix due to noise
R = sensor noise covariance matrix

You might also like