Wise 2012

AIAA Guidance, Navigation, and Control Conference AIAA 2012-4889
13 - 16 August 2012, Minneapolis, Minnesota
Asymptotic Properties of LQG/LTR Controllers in Flight

Control Problems
Kevin A. Wise *
Eugene Lavretsky †
The Boeing Company P.O. Box 516, St. Louis, MO 63166
This paper presents asymptotic properties of Linear Quadratic Gaussian with Loop
Downloaded by KUNGLIGA TEKNISKA HOGSKOLEN KTH on July 30, 2015 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4889
Transfer Recovery controllers as applied to the design of flight systems. In many flight
control problems output feedback design methods are needed when the states are not
available for feedback. Optimal control based designs have proven to be useful in practice
due to their robustness and performance. Their implementation using output feedback
through a full order observer or Kalman filter can degrade stability robustness properties of
these controllers. Recently, a new method of Loop Transfer Recovery was published that
offers different asymptotic properties as compared to conventional Loop Transfer Recovery.
This method asymptotically achieves positive real system behavior at certain loop break
points for minimum phase systems, as well as for non-minimum phase plants where the
number of output measurements exceeds the number of controls. This paper compares these
approaches and highlights the design features that engineers can use to tune and achieve
excellent robustness and performance in output feedback flight control systems. A design
example using the X-45A Joint Unmanned Combat Air System pitch-plane dynamics is
utilized to illustrate the control designs.
I. Introduction
O PTIMAL control theory has been shown to provide excellent stability, performance, and robustness properties
in the design of modern flight control systems. Implementing these designs using an output feedback control is
still a challenge. When only the output is available for feedback, a full-order observer can be designed to estimate
the state. For linear time invariant (LTI) systems with Gaussian models for disturbances and measurement noise, the
Kalman filter is the optimal state estimator. When optimal control (LQR) is combined with optimal state estimation
(Kalman filter) the control design is called the Linear Quadratic Gaussian (LQG) problem.
The Kalman filter algorithm is an excellent state estimator. It is widely used in estimation problems, such as
GPS navigation, where accurate state estimates are desired. However, when used to estimate the state in output
feedback control design problems, the optimal state estimator (optimal in the sense of minimizing the error
covariance) may not exhibit the best overall control and frequency domain properties needed for a control design. It
is well known that the LQG controller captures the excellent time domain characteristics of the state feedback
design, but the Kalman filter degrades the frequency domain properties (stability margins) of the design. In Doyle1
and Safanov and Athans2 stability margins for LQG regulators were analyzed. It was shown, that in some cases,
these controllers exhibited poor stability margins. Analysis of observer-based control system designs and of these
LQG control systems led to the development of a control system design process called Loop Transfer Recovery3-12
(LTR). This design process published in Doyle and Stein3 and Stein and Athans4 is a tuning process that
asymptotically recovers the state feedback frequency domain robustness properties (such as stability margins) in the
output feedback design. These design methods, fundamentally, provide a systematic process to place the poles and
zeros of the dynamic compensator such that robust loop shapes are obtained.
Prior methods (e.g. eigenstructure assignment) provided advice regarding the design of the observer to place the
eigenvalues of the observer 4-5 times faster than those of the desired dynamics for the closed–loop state feedback
design. The (stable) state estimation error dynamics then would decay quickly, decoupling the observer dynamics
*
Senior Technical Fellow, Phantom Works, St. Louis MO, 63166. Assoc. Fellow AIAA
†
Senior Technical Fellow, Boeing Research and Technology, Huntington Beach CA, Assoc. Fellow AIAA
Copyright © 2012 by The Boeing Company. Published by the American Institute of Aeronautics and Astronautics, Inc., with permission.
and preventing undesirable transients. This design heuristic was shown to be false in Doyle and Stein13 in which
fast observer dynamics were shown, in some cases, to degrade stability margins.
There are several methods available towards applying LTR to the LQG problem (called LQG/LTR), depending
upon if the system is minimum phase or non-minimum phase (see Zhang and Freudenberg8), square or non-square,
with several text books available11,12. These methods introduce some form of tuning mechanism either in the
controller or the observer for recovering the frequency domain properties (to provide loop shaping). Unfortunately,
high gains are often obtained somewhere in the control architecture. Care must be taken to avoid large feedback
gains. Also, a thorough analysis must be performed to make sure the system is truly robust and implementable.
When using LQG/LTR, the Kalman filter is no longer thought of as an optimal state estimator, but a dynamic
compensator/observer tuned for performance and robustness. The fact that algebraic Riccati equations are used in
the control and filter design makes the method attractive, and provides the mechanism for proving asymptotic
properties of the system analytically. Although H∞ design methods have in a large part replaced LQG/LTR
methods for dynamic compensator design, loop shaping in multivariable control is very important and LTR methods
still play an important role.

This paper reviews and compares a popular LTR tuning mechanism attributed to Doyle and Stein3 and Stein and
Athans4, and compares it with a new method recently developed by Lavretsky14,15. These methods are demonstrated
in a command tracking flight control example using the X-45A Joint Unmanned Combat Air System (J-UCAS)
pitch-plane dynamics. This aircraft is shown in Figure 1.
Fig. 1 X-45A Joint Unmanned Combat Air System.
This paper is organized as follows. Section 2 presents and reviews asymptotic properties in the design of LQ
controllers, and certain definitions for positive real systems. This summary of well known properties is provided to
highlight the mechanisms used in the asymptotic analysis of the LQ-based controllers. Section 3 presents a summary
of the Doyle and Stein3 LTR method. Section 4 presents a summary of the LTR method of Lavretsky14,15. Section 5
demonstrates and compares these LTR design methods by applying them to the X-45A J-UCAS longitudinal
dynamics. This is followed by a summary.
II. Asymptotic Properties of Linear Optional Control

The Linear Quadratic Regulator (LQR) is one of the most widely used control design methods in aerospace.
Trade studies have been performed comparing properties of controllers (performance, robustness, control usage) in
many different applications. In flight control system design problems, LQRs have demonstrated excellent
performance, robustness, and minimize the control usage.
The numerical values in the LQR penalty matrices Q and R determines the eigenstructure of the closed loop
system. The eigenstructure specifies the system’s performance and robustness properties. It is very important to
properly choose the numerical choices for Q and R , and more importantly, to learn how to exploit these matrices
to tune the gains to achieve the desired performance and robustness. This section will investigate how the
eigenstructure evolves as the weighting matrices are varied numerically. Detailed asymptotic analysis of LQRs may
be found in Kwakernak and Sivan16.
Consider the following LTI system
x =
Ax + Bu x ∈ R nx u ∈ R nu (1)
along with the infinite-time LQR problem
∞
=J ∫ (x
T
)
Qx + uT Ru dτ (2)
0
with Q =QT ≥ 0, R =RT > 0 , ( A, B ) stabilizable, and ( A, Q1 2 ) detectable having no transmission zeros on the
jω axis. Then Algebraic Riccati Equation (ARE) for this optimal control problem is
PA + AT P − PBR −1BT P + Q =
0 (3)
Associated with this ARE is the 2nx × 2nx Hamiltonian matrix H given by
 A − BR −1BT 
H =  (4)
 −Q − AT 
which can be used to determine the solution to the ARE. The optimal state feedback control is given by
− R −1BT Px =
u= − Kx (5)
which when substituted into (1) yields the closed loop system
( A BK ) x =
x =− Acl x (6)
The nx eigenvalues of the closed loop system, λ ( Acl ) , are the stable eigenvalues of the Hamiltonian matrix H . In
fact, the Hamiltonian matrix H has 2nx eigenvalues in which nx have negative real parts (stable) and nx have
positive real parts (unstable, but stable backwards in time). Let
s ) det [ sI − A + BK ]
φcl (= (7)
then
det [ sI=
− H ] φcl ( s ) φcl ( − s ) (8)
The asymptotic properties of interest are those associated with the migration of these eigenvalues as the
numerical values in the LQR penalty matrices Q and R are varied. These eigenvalues can be examined (roots of
φcl ( s ) ) through the polynomial formed by expanding the det [ sI − H ] , and begins with some elementary row and
column operations on H . Multiply the first row of H by −Q ( sI − A ) H −1 and add it to the second row. This
yields
 sI − A BR −1BT 
det [ sI − H ] =
det  
 Q sI + AT 
(9)
 sI − A BR −1BT 
= det  
 0
 ( ) −1 −
sI + A − Q ( sI − A ) BR B 
T 1 T

Then,
( )
] det [ sI − A] det  sI + AT − Q ( sI − A)−1 BR −1BT 
det [ sI − H=

( ) ( ) −1 T  
 −1 −1
= det [ sI − A] det  sI + AT  I − sI + AT Q ( sI − A ) BR B  (10)
  
( )
 −1 −1 − 
= det [ sI − A] det  sI + AT  det  I − sI + AT Q ( sI − A ) BR 1BT 
   
Next, factor the LQR penalty matrix Q into square roots, Q = Q1T Q1 , and BR −1BT into square roots,
BR −1BT = R1R1T . Then, using the identity det [ I − AB ] = det [ I − BA] , yields
 
( )
 −1 T −1 
det  I − sI + AT Q1 Q1 ( sI − A ) R1R1  =T
   
 B A  (11)
( )
 −1 −1 T 
det  I − Q1 ( sI − A ) R1R1T sI + AT Q1 
 
Then
 
( )
 − −1 T 
det [ sI − H=
] det [ sI − A] det  sI + A  det  I − Q1 ( sI −
A ) R1 R1 sI + A
T 1 T T
Q1 
 
 H1( s )  (12)
 
=φ ( s )( −1) φ ( − s ) det  I + H1 ( s ) H1 ( − s ) 
nx T
 
φ ( s ) det [ sI − A] . Thus,
where =
− s ) φ ( s ) φ ( − s ) det  I + H1 ( s ) H1T ( − s ) 
φcl ( s ) φcl ( = (13)
 
and let
ψ (s)
det  H1 ( s )  = (14)
φ (s)
Next, consider the performance index where a positive scalar weighting ρ > 0 is introduced:
∞
=J ∫ (x
T
Qx + ρ 2uT Ru dτ ) (15)
0
We are interested in the asymptotic behavior as ρ → 0 and as ρ → ∞ . The roots of (13) are also the roots of
φ ( s ) φ ( − s ) det  ρ I + H1 ( s ) H1 ( − s )  (16)
As ρ → 0 , some of the roots will go to infinity; those that stay finite approach the transmission zeros of the transfer
function matrix H1 ( s ) and their negative values. These finite zeros control the dynamic response of the optimal
regulator. As ρ → ∞ , the roots of φcl ( s ) are the nx stable roots of φ ( s ) φ ( − s ) . That is if roots of φ ( s ) have
positive real part, then the mirror image of them in φ ( − s ) will become the stable roots in φcl ( s ) .
We see that shaping the zeros of H1 ( s ) plays a crucial role in the design of the optimal control. This is done
through selection of the LQR penalty matrix weightings Q and R . Kwakernak and Sivan16, section 3.8, provides an
excellent review of these properties and describe the Butterworth patterns for the roots that move to infinity (where
no finite transmission zeros exist).
III. Doyle and Stein LQG/LTR

In the previous section an optimal control, (5), was presented which uses state feedback. When only the output
is available for feedback, a full-order observer can be designed to estimate the state. For LTI systems with Gaussian
models for disturbances and measurement noise, the Kalman filter is the optimal state estimator. When optimal
control (LQR) is combined with optimal state estimation (Kalman filter) the design is called the Linear Quadratic
Gaussian (LQG) problem. LQG controllers capture the excellent time domain characteristics of the state feedback
design, but the Kalman filter degrades the frequency domain properties. Loop Transfer Recovery (LTR)
asymptotically recovers these state feedback frequency domain properties. This section presents the tuning
mechanism attributed to Doyle and Stein3 and Stein and Athans4. This LTR method asymptotically inverts the
Kalman filter dynamics to recover the LQR frequency domain properties.
Consider the following linear time invariant Gaussian design model
x = Ax + Bu + w
(17)
=y Cx + v
where w and v are zero mean, white, uncorrelated Gaussian random processes with covariances given by
{
E w ( t ) wT= }
(τ ) Q0δ ( t − τ )
(18)
E {v ( t ) vT=
(τ )} R0δ ( t − τ )
The state estimate, x̂ , is formed using the following Kalman filter state estimator
xˆ = Axˆ + Bu + K f ( y − yˆ )
K f = Pf C T R0−1 (19)
0 = APf + Pf AT + Q0 − Pf C T R0−1CPf
where ŷ is the estimate of the output, { } is the steady state error covariance, which results from
Pf = E xxT
solving the algebraic filter Riccati equation (covariance equation), and Q0 and R0 are the process and measurement
noise covariances from (18), respectively. The optimal control is formed using the LQR state feedback control gain
matrix K c and the estimated state feedback x̂ , given as
u = − K c xˆ (20)
Figure 2 combines the LQR controller with the Kalman filter state estimator (19) into a block diagram. This is the LQG
control architecture.
The frequency-domain properties of the LQG system do not equal that of the LQR system primarily due to the observer
dynamics introduced by the Kalman filter state estimator. For the state feedback controlled system, the LQR loop transfer
function matrix (LTFM) at the plant input is
−1
=
LLQR ( s ) Kc ( sI − A) B (21)
For the output feedback controlled system, the LQG LTFM at the plant input is
−1
LLQG = (
( s ) Kc sI − A + BKc + K f C ) K f C ( sI − A )
−1
B (22)
Clearly the dynamics introduced by the dynamic compensator alters the frequency domain characteristics for the LQG
system.
For this control architecture there are two approaches for applying LTR to the LQG control problem. One modifies the
Kalman filter (state observer) to recover the state feedback loop properties, and the other modifies the LQR controller.
Here the method of modifying the Kalman filter is discussed. The tuning procedure consists of designing Kalman filters
with the plant process disturbance covariance matrix Q f parameterized with a scalar ρ as
1
Q=
f Q0 + BBT (23)
ρ
where Q0 is the nominal plant process disturbance covariance from (18), B is the control input distribution matrix, and
ρ is the LTR filter compensation parameter. This parameter is adjusted to recover the LQR frequency-domain
characteristics over the frequency range of interest. The modified matrix Q f is used to compute the steady-state
covariance Pf and filter gain matrix K f to be used in the LQG controller.
Considering the loop broken at the plant input, LTR modifies K f to create a system that has stability properties that
asymptotically approach those of the LQR. The method uses a trial and error procedure in which the filter design is
parameterized by a scalar ρ > 0 such that when ρ → 0 , LLQG → LLQR asymptotically, but not necessarily uniformly.
It is evident that the location of the Kalman filter eigenvalues, (22), alters the closed-loop frequency characteristics of the
system.
The LQG/LTR approach requires that the controlled system (plant) be minimum phase (i.e., no RHP transmission
zeros). The requirement for minimum phaseness occurs because the LTR procedure asymptotically inverts the plant
dynamics of the Kalman filter and substitutes the linear regulator dynamics. If there was an RHP transmission zero, an
RHP pole would be created, causing an unstable system. The procedure may still be applied to nonminimum phase
systems, but care must be taken to prevent instability in the LQG compensator. This limits the amount of recovery.
The LQG/LTR loop transfer function matrix at the plant input, LLQG , will asymptotically recover the LQR frequency
domain characteristics as ρ → 0 . This can be shown as follows. As ρ → 0 , the process covariance Q f in (23) gets
large dominated by the second term 1 BBT . As these elements of Q f get large, the covariance matrix Pf has
ρ
elements that get large, resulting in the Kalman gain matrix K f getting large, with the following result:
( ) ( )
−1 −1
LLQG (=
s ) K c sI − A + BK
 +K C
c f K f C sI − A B
( sI − A )
−1
LLQG ( s ) ≈ K c B
(24)
It is this process that inverts the plant (within the Kalman filter) resulting in recovering the LQR LLQR . It is important to
( ) ( )
note that as ρ → 0, σ Pf → ∞ and σ Pf → 0 , creating a singular covariance matrix. In the next section a new
method of LTR design is presented which prevents this condition from occurring during the recovery process.
The LQG controller transfer function matrix that relates the measurement y to the control u is
u= (  + K C −1 K y
− K c sI − A + BK ) (25)
c f f
y Cx + v and letting ρ → 0 as in (24) yields

Substituting for the measurement =
( ) K ( Cx + v )
−1
u=− K c sI − A + BK
 +K C
c f f
− K ( sI − A + BK
 + K C ) K Cx − K ( sI − A + BK )
−1 −1
=  + KfC Kfv
c c f f c c
− K x − K ( sI − A + BK
 + K C) K v
−1
= c c c f f
(26)
−1
in which the first term is inverted and cancelled K f C ( ) K f C = I resulting in − K c x . However, the second term is not
−1
(
exactly cancelled; K f C ) K f ≠ I , and the sensor noise v can be amplified. This feature limits the amount of recovery
possible in real applications. In the use of this design method for making the LQG system robust, the sensor noise
amplification in (26) must be examined.
The LQG/LTR controller design, examining the loop properties at the plant input, may be realized through the following
synthesis technique:
Step 1: LQR controller design: K c

Follow the robust servomechanism design approach outlined in chapter 3. Design LQR weighting matrices Q and R
−1 
=
such that the resulting LTFM LLQR ( s ) Kc sI − A ( ) B meets performance and stability robustness requirements,
and exhibits the desired bandwidth. The frequency domain properties of the LQG system will be no better than those
of the LQR system.
Step 2: Kalman filter design: K f
Design the Kalman filter state estimator using (19), with (23) defining the plant disturbance covariance. The LTR
filter recovery parameter ρ is used to recover the LQR frequency-domain characteristics over the frequency range of
interest. Examine plant input and output frequency domain criteria and the sensor noise amplification in (26) and limit
the LTR recovery so that the sensor noise is not amplified.
IV. LQG/LTR Method of Lavretsky

In the previous section an optimal control (LQR) was combined with an optimal state estimator (Kalman filter)
to form the LQG controller. A LTR tuning process was then used to recover the LQR frequency domain properties
in the LQG controlled system by inverting the filter dynamics. In this section an alternate method is explored for this
observer-based LTR which is referred to as the LTRLM method [14, 15].
The design goal is to achieve the best performance and stability robustness properties for a process or a system
via control design. Among linear time invariant systems, there is a special class of dynamics, called Positive Real
(PR) and Strictly Positive Real (SPR)17,18. These systems have very interesting properties that enable robust output
feedback control design. Here PR and SPR definitions are presented as they are stated in Khalil18.
Definition 6.1 A ( p × p) proper rational transfer function matrix G ( s ) of the complex variable s= σ + jω is
called Positive Real if
i) Poles of all elements of G ( s ) are in the left half complex plane.
ii) For all real ω for which j ω is not a pole of any element of G ( s ) , the matrix G ( j ω ) + GT ( − j ω ) is
positive semi-definite.
iii) Any pure imaginary pole j ω of any element of G ( s ) is a simple pole, and the residue matrix
lim (s − j ω )G (s) is positive semi-definite Hermitian.
s→ j ω
Definition 6.2 The transfer function G ( s ) is called Strictly Positive Real if G ( s − ε ) is Positive Real, for some
ε > 0.
For scalar systems ( p = 1) , PR and SPR dynamics have their Nyquist frequency response locus located entirely in
the right half complex plane. This condition for G ( s ) can be satisfied only if the system relative degree is zero or
one. Thus, encirclements of (-1, j0) cannot occur. In other words, such a system will remain stable under a large set
of uncertainties, which is a highly desirable property for any system to possess.
The relationship between PR, SPR transfer functions, and Lyapunov stability theory of the corresponding
dynamical system has lead to the development of several stability criteria for feedback systems with LTI and
nonlinear components. These criteria include the Popov’s criterion and its variations18. The link between PR, SPR
transfer function matrices and the existence of a Lyapunov function for studying stability can be established by the
following two lemmas18.
−1
Positive Real Lemma Let G ( s ) = C T ( sI − A ) B + D be a ( p × p) transfer function matrix, where ( A, B ) is
controllable and ( A, C ) is observable. Then, G ( s ) is Positive Real if and only if there exist matrices =
P PT > 0 ,
L , and W such that
P A + AT P =
− LT L
B CT − LT W
P=
W T W= D + DT
(27)
Kalman-Yakubovich-Popov (KYP) Lemma18. Let,
−1
G ( s ) = C T ( sI − A ) B+D
be a ( p × p) transfer function matrix, where ( A, B ) is controllable and ( A, C ) is observable. Then, G ( s ) is
P PT > 0 , L , W , and a positive constant ε such that

Strictly Positive Real if and only if there exist matrices =
P A + AT P =
− LT L − ε P
B CT − LT W
P=
W T W= D + DT
(28)
Clearly, if D is the zero matrix then the SPR conditions (28) reduce to
P A + AT P =
− LT L − ε P
P B = CT (29)
and in this case, setting ε = 0 , gives the PR conditions in the form:
P A + AT P =
− LT L
P B = CT (30)
The first relation in (30) is the algebraic Lyapunov equation, and V ( x ) = xT P x is the Lyapunov function [18]. The
second relation in (30) enables output feedback control design, whereby the system output y = C x can be fed back
into the input to control the system, while preserving closed-loop stability. Also note that the matrices B and C
−1
( s ) C ( sI − A) B .
define the transmission zeros of the system transfer function matrix G=
The LTRLM modifies the control design such that, for a class of restricted systems, the PR property is obtained
asymptotically, Pv Bv → C T , with the positive tuning parameter ν → 0 . In addition, the method ensures that Pv
remains symmetric and strictly positive definite, uniformly in v . These are the distinguishing features of LTRLM
design. Similar to the previous section, in this design the Kalman filter is no longer treated as a filter. It will continue
to estimate the system state and serve as a dynamic compensator, tuned to improve the frequency domain properties
of the system. The Gaussian covariance matrices for w and v are altered significantly to improve the controller
robustness and to limit sensor noise amplification. So, these matrices no longer “model” the stochastic processes of
the system.
The LTRLM design approach uses the linear time invariant Gaussian design model,
x = Ax + Bu + w
=y Cx + v (31)
where w and v are zero mean, white, uncorrelated Gaussian random processes with covariances given by
{
E w ( t ) wT= }
(τ ) Q0δ ( t − τ )
E {v ( t ) vT=
(τ )} R0δ ( t − τ )
(32)
The state estimate x̂ is formed as before, using the state estimator,
xˆ = Axˆ + Bu + K f ( ymeas − yˆ )

(33)
and the control input is calculated using the LQR state feedback gain matrix K c , with the estimated state feedback
x̂ .
u = − K c xˆ (34)
In LTRLM, parameterize the process and measurement noise covariance matrices using a positive scalar ν as,
 v +1 T v
Qv =
Q0 +   B B , Rv = R0
 v  v +1 (35)
where B is a matrix formed by adding “fictitious” columns to B , to make B = [ B X ] have its column rank equal
to the row rank of C , such that CB becomes invertible and the corresponding extended system C ( s I − A )−1 B is
minimum phase, that is all its transmission zeros are located in the left half complex plane. This is the “squaring-up”
step of the method. Substituting the weights from (35) into the filter Riccati equation yields
 1  1
Pv AT + A Pv − 1 +  Pv CT R0−1 C Pv + Q0 + 1 +  B BT =
0
 v   v (36)
or, equivalently
Pv AT + A Pv − Pv CT R0−1 C Pv + Q0 + B BT +  B BT − Pv CT R0−1 C Pv  =
1
0
v  (37)
The gains in (33) are computed as
K f = Pν C T Rν −1
(38)
Now as ν → 0 , one can show that the filter covariance matrix Pv asymptotically approaches a constant symmetric
positive definite matrix P0 , that is
=
P0 lim =
Pv T PT > 0
lim P=
v 0
v →0 v →0 (39)
( )
This behavior is in contrast to the previous section, where as the LTR parameter ρ → 0, σ Pf → ∞ , σ Pf → 0 ( )
, and the Pf matrix became singular.
The important properties of P0 in (39) are listed below without proof, (see Lavretsky and Wise15, Chapter 13,
Theorem 13.1 for formal derivations):
• P0 is the unique symmetric strictly positive definite solution of the following algebraic Lyapunov
equation:
( ) + ( A − CT R0−1 C P1 ) P0 + Q0 =
T
P0 A − CT R0−1 C P1 0
(40)
• There exists a unitary matrix W ∈ R m×m such that:
P0 CT = B W T R0
(41)
• The unitary matrix W in (41) can be chosen as:

W = (U V )
T
(42)
where U and V are two unitary matrices, defined by the singular value decomposition,
1
−
T T
B C R0 2= U Σ V
(43)
and Σ represents the diagonal matrix of the corresponding singular values.
For minimum phase systems, the SPR property is implied by (41). What the LTRLM design is trying to do is to
shape the transmission zeros of the state estimator, such that the original system with the extended input becomes
SPR asymptotically, as v → 0 . To do this, the design process must “square-up the system” by adding extra columns
to B (to form B ) and then apply the LTR tuning process, whereby the tuning parameter v is decreased in (35) until
the system becomes almost SPR.
It was discussed earlier that in the LQR design problem, with the penalty matrix Q factored as Q = N T N , the
poles of the closed-loop system, λ ( A − BK c ) , would approach the transmission zeros defined by N ( sI − A )−1 B
asymptotically as the gains grew large. If no finite transmission zeros existed, the roots would form a Butterworth
pattern (or combinations of Butterworth patterns) in the left half complex plane. Thus, by the proper selection of Q ,
the designer places these zeros to achieve the desired response of the system. So, the selection of the LQR penalty
matrix is a key tuning mechanism in the LQR controller design.
This same basic idea is in work under LTRLM. For the state estimator (aka Kalman Filter), the process
covariance Q f is the equivalent to the LQR penalty matrix. Factoring the process covariance Q f as Q f = LT L ,
( )
the eigenvalues of the Kalman filter, λ A − K f C , will approach the finite transmission zeros defined by
−1
C ( sI − A ) L . Thus, the selection of the process covariance Q f is an ideal tuning mechanism in the design of the
LTRLM controller. Placing the zeros of the system in a desirable location is the key to achieving a robust design.
This is achieved through the modified process covariance and measurement noise matrices in (35).
The LQG/LTR controller design using the LTR Method of Lavretsky, examining the loop properties at the plant input,
may be realized through the following synthesis technique:
Step 1: LQR controller design: K c

Follow the robust servomechanism design approach outlined in chapter 3. Design LQR weighting matrices Q and R
−1 
=
such that the resulting LTFM LLQR ( s ) Kc sI − A ( ) B meets performance and stability robustness requirements,
and exhibits the desired bandwidth.
Step 2: Kalman filter design. K f
Select columns X to make B = [ B X ] have column rank equal to the row rank of C . Design the Kalman filter
state estimator using (36), with (35) defining the plant process and measurement noise covariance matrices. The LTR
filter recovery parameter ν is used to recover the LQR frequency-domain characteristics over the frequency range of
interest. Ad hoc adjustment of the sensor noise covariance magnitude may be needed to scale the Kalman gains to
prevent large gains from occurring. Examine plant input and output frequency domain criteria and the sensor noise
amplification in and limit the LTR recovery so that the sensor noise is not amplified.
V. X-45A Design Example

In this section the two LQG/LTR controllers presented in the previous two sections are demonstrated using the
X-45A J-UCAS aircraft.
Doyle and Stein LQG/LTR

Consider the model:
x = Ax + Bu + w
(44)
=y Cx + v
T
q δ e δe  , u = δ ec , y = [ Az q ] , and where w and v are white Gaussian process and
T
where x =  Az
measurement noise, respectively.
The RSLQR (which uses integral control) design model is
0 Cc    0 
=A =  B B
0 A   
0 1 0 0 0   0 
0 −1.053 −346.5 0 −11.29   0  (45)
  
= A =0 0.007 −1.033 −1.093 0  B  0 
   
0 0 0 0 1   0 
0 0 0 −6672. −98.02  6672.
Solving an LQR using the above plant model, the state feedback gain matrix is
K c = [ 0.49482 0.17904 -14.061 2.2089 1.8036e-003] (46)
The control law is implemented using
( )
T
− K c  ∫ Azm − r
u= xˆ  (47)
 
where the first gain in K c multiplies the integral error, and the remaining gains multiply estimates of Az , q , δ e ,
and δ , respectively.
e
To design the Kalman filter state estimator models of the process and measurement noise covariance matrices
from (44) are needed. At this flight condition the process noise modeled in the state equations is:
1.94 × 10−4 0 0 0  ( fps )2 / s 

  
 0 2.5 × 10−7 0 0   ( rps )2 / s 
Q0 =    (48)
 0 0 1.0 × 10−8 0  ( rad )2 / s 
  
 0 0 0 1.0 × 10   ( rps ) / s 
−6   2
The numerical values in Q0 are often adjusted in the design process to tune the Kalman filter. For a typical inertial
measurement unit, the measurement noise in Az and qm are modeled as:
m
6.25 × 10−2 0  ( fps )2 

R0 =    (49)
 −6   2
0 1.0 × 10  ( rps ) 
Figure 3 shows Az and q simulation time histories of the state feedback controlled system without process and
measurement noise, along with simulation time histories of the measured values that contain process and
measurement noise. The Kalman filter state estimator is
xˆ = Axˆ + Bu + K f ( y − yˆ ) (50)
where u is formed using (47), and was implemented using steady state matrices obtained from the filter covariance
equation
0 = APf + Pf AT + Q0 − Pf C T R0−1CPf
K f = Pf C T R0−1
 2.0442e-003 -6.0760e-006 2.4929e-010 -5.6592e-008 

 -6.0760e-006 7.8188e-008 -1.2264e-012 1.3375e-010 
Pf =  
 2.4929e-010 -1.2264e-012 1.2523e-010 -5.0000e-009  (51)
 
 -5.6592e-008 1.3375e-010 -5.0000e-009 3.4544e-007 
 3.2707e-002 -6.0760e+000 
 -9.7217e-005 7.8188e-002 
Kf = 
 3.9887e-009 -1.2264e-006 
 
 -9.0547e-007 1.3375e-004 
Figure 4 shows state estimates using the estimator with the nominal process noise matrix Q0 (LQG design). The
Kalman filter does an excellent job estimating the states from the noisy measurements. However, the full order
observer (Kalman filter) has degraded the excellent frequency domain properties of the LQR state feedback design.
To recover the frequency domain properties (at the plant input), Loop Transfer Recovery (LTR) is used. The LTR
procedure consists of designing Kalman filters with the plant process covariance matrix Q f parameterized with a scalar
ρ as
1
Q=
f Q0 + BBT (52)
ρ
where Q0 is the nominal covariance, B is the control input distribution matrix, and ρ is the LTR filter compensation
parameter. This parameter is adjusted, ρ → 0 , to recover the LQR frequency-domain characteristics over the frequency
range of interest. The modified matrix Q f is used to compute steady-state covariance matrices Pf and filter gain
matrices K f to be used in the LQG controller. In this example, values of ρ were chosen to be
ρ = ∞ 105 104 103 102  (53)

 
The following controller combines the robust servo controller and Kalman filter estimator
 x1   0 0   x1  1 0   −1
 − B K 1: n + +
 ˆ Ap − K f C p − B p K c (nr + 1: nx )   xˆ   K f  meas  0 
z r
 x   p c( r) (54)
x 
u = − Kc  c 
 xˆ 
Note that (54) is valid for plant models with no D matrix, i.e. D p = 0 . The first state x1 is the robust servo
integrator, the vector x̂ is the estimated state, zmeas contains the acceleration and pitch rate measurements, and r
is the acceleration command. Writing the controller in a generic form gives
xc =
Ac xc + Bc1 zmeas + Bc2 r
(55)
u=
Cc xc + Dc1 zmeas + Dc2 r
For the LQG design ( ρ = ∞ )
 0 0 0 0 0 
 0 -1.0854e+000 -3.4041e+002 0 -1.1289e+001 

 
Ac =  0 6.8202e-003 -1.1116e+000 -1.0925e+000 0 
 
 0 -3.9887e-009 1.2264e-006 0 1.0

 -3.3010e+003 -1.1944e+003 9.3804e+004 -2.1408e+004 -1.1005e+002 
 1.0 0   -1 
 3.2707e-002 -6.0760e+000   0
   
Bc =  -9.7217e-005 7.8188e-002  ; Bc =  0  (56)
1 2
   
 3.9887e-009 -1.2264e-006   0
 -9.0547e-007 1.3375e-004   0 
Dc = [ 0 0 ] ; Dc = [ 0 ]
1 2
Note that in the above controller the robust servo error, e = yc − r = Az − Az , is formed using the measured
meas cmd
acceleration. This error is formed from the top row in Bc1 and Bc2 . An alternate controller would be to use the
estimate of Az from the Kalman filter, =
e Aˆ z − Az . This would change the control architecture significantly.
cmd
Next, analyze the LQG/LTR design in the frequency domain, and determine the desired amount of LTR to be
applied at this flight condition. Figure 5 shows a Nyquist plot of the LQR, LQG, and LQG/LTR designs using
values of ρ from (53). The red circle is a unit circle centered at (-1,j0) for reference. The LQR locus (blue)
demonstrates infinite gain margin (at the plant input) and excellent phase margin. The LQG design (blue) shows the
decrease in gain margin and phase margin from inserting the Kalman filter state estimator into the controller. The
locus for the LTR designs show initially, ρ = 105 , that the margins are worse than those of the LQG. As the LTR
parameter is reduced further, the margins improve and approach those of the LQR design. This demonstrates that the
LTR recovery process is not uniform in its recovery. Figures 6 and 7 show the analysis results examining the return
difference dynamics I + L and stability robustness matrix I + L−1 at the pant input, respectively. Both figures show
the recovery of the LQR characteristics at the plant input.
To further examine the effects of LTR, examine the sensitivity and complementary sensitivity at the plant
output, and the noise transmission through the controller. The sensitivity and complementary sensitivity are given by
e
= S (s)
r
(57)
y
= T (s)
r
The noise transmission through the controller is given by

u
v ( −1
=σ Cc ( sI − Ac ) Bc + Dc
1 1 ) (58)
Figures 8, 9, and 10 show the analysis results at the plant output. The LTR process only guarantees recovery of
the LQR properties at the plant input. The sensitivity function in Figure 8 shows undesirable peaking in S ( s ) as the
recovery is made. From this figure the value of ρ would need to be limited to 103. The complementary sensitivity
function in Figure 9 shows undesirable peaking in T ( s ) as the recovery is made. This peak is similar to a peak
resonance in under- damped second order systems. Even though the stability margins at the plant input are getting
better with LTR, the margins at the plant output are getting worse. Figure 9 also shows the value of ρ would need
to be limited to 103 to keep the peak small. Finally, Figure 10 shows the noise transmission through the controller,
and shows that as ρ → 0 the noise amplification increases. This would be quite undesirable. This figure indicates
that the value of ρ would need to be limited to 104 or larger.
( )
To finalize a choice of ρ , the decision should be made on maximizing σ ( I + L ) and σ I + L−1 at the plant
input, minimizing σ ( S ) and σ (T ) at the plant output, and preventing noise amplification over a frequency range
of interest. The following table summarizes these peak values:
Design σ ( I + L) σ ( I + L−1 ) σ (S ) σ (T )
LQR 1.0000 0.7963 1.4936 1.0480
LQG 0.5506 0.9808 1.0791 1.0000
ρ = 105 0.5233 0.7136 1.0923 1.0000
ρ = 104 0.5853 0.6567 1.0599 1.0000
ρ = 103 0.7920 0.7301 1.4581 1.0000
ρ = 102 0.9160 0.7715 2.9361 2.1570
From the σ ( I + L ) values, ρ ≤ 104 is needed to meet plant input stability margin requirements. It is desired to have
( )
σ I + L−1 to be as large as possible, which is also satisfied by ρ ≤ 104 . It is desired to minimize σ ( S ) , which
points to ρ = 104 as the desired recovery level. If ρ = 103 the peak in σ ( S ) would be too large. Thus, ρ = 104 is
selected as the design. For comparison, the following table lists the Kalman filter gains:
Kalman Filter Gains
LQG LQG/LTR ρ=104
3.2707e-002 -6.0760e+000 6.9018e+000 1.7004e+002

-9.7217e-005 7.8188e-002 2.7206e-003 9.6745e+000
3.9887e-009 -1.2264e-006 -5.5183e-001 -5.1037e+001
-9.0547e-007 1.3375e-004 -2.8572e+000 3.0166e+003
It is evident that this method increases the gains to large values. Further analysis would be needed to determine of
gains of this magnitude could actually be used in a real flight control system.
LQG/LTR Lavretsky Method (LTRLM)

The above flight control design example does not exactly satisfy the requirements for using the LTRLM design
method. The requirements are: 1) that the system be minimum phase; and 2) that columns X can be added to B to
make C [ B X ] be invertible. To satisfy the requirement for the system to be nonminimum phase the aircraft model
must be approximated with one that is minimum phase. This is easily done by neglecting the tail vertical force Zδ
in the dynamics. Normally the acceleration transfer function has a RHP zero. When Zδ is zeroed, the acceleration
transfer function no longer has any finite zeros. To satisfy the requirement for C [ B X ] be invertible requires
removing the actuator model. This reduces the number of states, creating a second order design model.
Consider the longitudinal dynamics given as
Zα
α + δ δ +q
Z
α =
V V (59)
q = Mα α + M δ δ + M q q
where angle-of-attack α and pitch rate q are the states, and elevator δ the control. The outputs are acceleration
Az and pitch rate q given by
 Az   Zα 0  α   Zδ 
=q  0 +
1   q   0 δ (60)
   
( )
Neglect the tail vertical force ( Zδ = 0 ) , and form the minimum phase plant model Ap , B p , C p , D p as
-1.0527 1 0 
 Ap B p  -2.3294 -1.0334 -1.1684 
 = . (61)
C p D p  -346.48 0 0 
 
 0 1 0 
We begin first with the LQR state feedback design. To track an acceleration command, the RSLQR state feedback
design model is
0 C p (1,:)   0 
=A =  B  
 0 Ap  Bp 
0 -346.48 0   0  (62)
= 
A = 0 -1.0527 1  B  0 
 
 0 -2.3294 -1.0334   -1.1684 
Using a LQR penalty matrix Q = diag [ 0.2448 0 0] and R = 1 , the state feedback gain matrix is
K c = [ 0.31623 -33.261 -6.7127 ] (63)
The above control law is implemented using
( )
T
− K c  ∫ Azm − r
u= α q (64)
 
To analyze and compare the controllers, each one will be implemented in the following model: :
xc = Ac xc + Bc1 y + Bc 2 r
(65)
u = Cc xc + Dc1 y + Dc 2 r
Using the gains from (63), the RSLQR state feedback controller is:
xc
= [0] xc + [1 0] y + [ −1] r
(66)
[-0.31623] xc + [33.261
u= 6.7127 ] y + [ 0] r
To form the closed loop system the above controller is connected to the plant model.
Next is the LQG design. The nominal process and measurement noise covariance matrices are
0.000196 0  ( rad ) / s 

2
Q0 = 
0.0025  ( rps )2 / s 
(67)
 0
 
and
6.25 × 10−2  ( fps )2 

R0 =    (68)
 1.0 × 10−6  ( rps )2 
The Kalman filter state estimator is
xˆ = Axˆ + Bu + K f ( y − yˆ ) (69)
where u is formed using (64), and was implemented using steady state matrices obtained from the filter covariance
equation
0 = APf + Pf AT + Qν − Pf CT Rν −1CPf
(70)
K f = Pf C T Rν −1
The steady state covariance and Kalman filter gains design (using Q0 and R0 ) are:
 9.5843e-006 3.8344e-007 
Pf =  
 3.8344e-007 4.8957e-005 
(71)
-0.053132 0.38344 
Kf =
-0.0021257 48.957 
For the LQG controller, the RSLQR control law is given by
− K1 ∫ e − K x xˆ
u=
(72)
e = yc − r = Az − Azcmd
where x̂ is the estimated state, the RSLQR gain matrix is partitioned as K c = [ K1 K x ] . To form the estimated
state, substitute the control (72) into the state estimator (69). Doing so gives
xˆ = Ap xˆ + B p u + K f ( y − yˆ )
( ) ( ( (
xˆ = Ap xˆ + B p − K1 ∫ e − K x xˆ + K f y − C p xˆ + D p − K1 ∫ e − K x xˆ ))) (73)
( (
 
) ) ( 
)
xˆ = Ap − B p − K f D p K x − K f C p xˆ − B p − K f D p K1 ∫ e + K f y
A22 A21
T
The LQG controller states are xc =  ∫ e xˆ  . The controller state space model using (72) and (73) is
e   0 0   ∫ e  1 0   −1
= ˆ 
 +  ymeas +   r

 x   A21 A22   xˆ   K f  0
(74)
 e
− [ K1 K x ]  ∫  + [ 0] ymeas + [ 0] r
u=
 xˆ 
where A21 and A22 are defined as in (73). Substituting the gains into (74) yields
 0    1 0 0 0   -1
e  
 ˆ=   0
 ∫ e  
-19.462 0.61656   + -0.053132 0.38344 ymeas +  0  r
  xˆ     
 x   0.36948
-41.928 -57.833     -0.0021257 48.957   0 
 (75)
 e
=u [ -0.31623 33.261 6.7127 ]  ∫  + [ 0 0 ] ymeas + [ 0 ] r
 xˆ 
Next is the Doyle and Stein LQG/LTR design. This method (from the previous section) adds a term to the process
noise covariance matrix as
1 T
Q=
f Q0 + B B (76)
ρ p p
In the previous section the LTR parameter ρ was selected to be ρ = 25 . This produces
0.000196 0 
Qf =  (77)
 0 0.057106 
which results in
 9.5933e-006 8.3416e-007 
Pf =  
 8.3416e-007 0.00023793 
(78)
-0.053183 0.83416 
Kf = 
-0.0046243 237.93 
Note the magnitude increase in the gain K f ( 2, 2 ) .

The last controller in this example uses the LTR Method of Lavretsky (LTRLM). The first step in the design
process is to design the LQR control law. This is the same for both controllers. The second step is to select columns
X to make B =  B p X  have column rank equal to the row rank of C p . To do this the numbers within these
matrices must be examined:
-346.48 0   0   b21 
=Cp  =  B p =  B Bp  (79)
 0 1 -1.1684   b22 
For the second column in B any values are possible, except b21 = 0 . If b21 = 0 , then
 0 0 
B= ' (80)
-1.1684 b22 
which is rank 1. To evaluate the effect of this selection, we will examine two designs, described
 0 1.   0 0.1 
=B1 =  and B2 -1.1684 10.  (81)
 -1.1684 0   
To improve the numerical scaling between Q f and R f , R0 is scaled by 250. The process noise and measurement
covariance matrices are given by
 v +1 T v
Qv =
Q0 +   B B , Rv = 250 R0 (82)
 v  v +1
For ν = 2.5 and using B = B1 yields
0.000196 0  1 0  1.4002 0 
Qv =   + 1.4   =
 0 0.0025 0 1.3652   0 1.9137 
(83)
0.0625 0  11.161 0 
=Rv (= 0.71429 )( 250 )    
 0 1e-006   0 0.00017857 
Solving for the steady state covariance and gain matrix from (70) yields
0.011312 -3.5561e-005
Pf =  
-3.5561e-005 0.018303 
(84)
-0.35116 -0.19914 
Kf = 
 0.001104 102.5 
The controller is formed by substituting the gains K f into (74) and results in
 0 0 0   1 0   -1
e  
=
 ˆ  0 -19.48 0.16584  
∫ e  + -0.053183 0.83416 ymeas +  0  r
 x   0.36948   xˆ     
 
-42.794 -246.8    -0.0046243 
237.93   0  (85)
∫ e
=u [ -0.31623 33.261 6.7127 ]   + [ 0 0 ] ymeas + [ 0 ] r
 xˆ 
For ν = 2.5 and using B = B2 yields
0.000196 0  0.01 1  0.014196 1.4 

Qv =  + 1.4   =
 0 0.0025  1 101.37   1.4 141.91
(86)
0.0625 0  11.161 0 
=Rv (=0.71429 )( 250 ) 
 0 1e-006   0 0.00017857 
 
Solving for the steady state covariance and gain matrix from (70) yields
0.00017018 0.0017429 
Pf = 
0.15898 
0.0017429
(87)
-0.0052832 9.7605
Kf =
-0.054109 890.31
The controller is formed by substituting the gains K f into (74) and results in
 0 0 0    1 0   -1
e 
=   ∫ e  +  -0.35116 -0.19914  y + 0  r
 ˆ  0 -122.72 1.1991
  xˆ    meas  
 x   0.36948 -40.809 -111.37       0 
   0.001104 102.5 (88)
 ∫ e
=u [ -0.31623 33.261 6.7127 ]   + [ 0 0 ] ymeas + [ 0 ] r
 xˆ 
We see from varying the columns in B that significantly different controllers result. The second choice has a larger
Qν which results in larger gains. One can expect that these larger gains will recover the LQR properties more than
the designs with smaller gains.
Figure 11 shows a step response for the closed loop system using all 5 controllers (LQR, LQG, LQG/LTR, and
the two LTRLM controllers). The plot shows that the time domain simulation results are identical for all the designs.
Figures 12 through 16 show the frequency domain analysis of these controllers. Figures 12 shows the Nyquist plot
which shows that the LQR design (black line) does not enter the red unit circle centered at (-1,j0). The LQG design
(also black line) is the locus to the left which has the degraded gain margin and phase margins properties. Note that
the LQG design here is not as bad as in the previous section due to the actuator being neglected within this model.
The LQG/LTR design with LTR parameter ρ = 25 is the red curve. The two designs using the LTRLM approach
(blue and green curves) bracket the LQG/LTR locus. The LTRLM method with B = B2 (green curve) has the most
recovery (closest to the LQR black curve).
We see from the figures that the LQG/LTR and LGQ/LTRLM methods can all be tuned to recover the LQR
properties. These methods all recover the properties by increasing the Kalman filter gains. Care must be taken to
prevent the gains from getting too large. It seems from our use of the LTRLM method that it can achieve the
recovery with smaller overall gains as compared to the conventional LQG/LTR method. The LQG/LTR method
adds B p BTp to Q0 while the LTRLM method adds B BT . For this example these are:
0 0  0.01 1 
=B p BTp =  ; B BT   (89)
0 1.3652   1 101.37 
The additional parameters seem to offer improvements by distributing the recovery into additional loops within the
architecture.
VI. Summary and Conclusions
This paper demonstrates the design and analysis of observer-based LQG/LTR controllers for flight control
applications. Two methods were examined. The first was the Doyle and Stein method which modifies the Kalman
filter design matrices to asymptotically recover the state feedback loop gain at the plant input. In the Doyle and
Stein LTR method, the resulting Kalman filter covariance matrix becomes singular in the limit of the recovery
parameter. The plant dynamics of the observer are inverted using large gains, and sensor noise can be amplified.
We have found that the sensor noise amplification limits the amount of recovery possible.
In the second method developed by Lavretsky the Kalman filter design matrices are modified in a different
manner. The resulting Kalman filter covariance matrix does not become singular in the limit of the recovery
parameter, but approaches a fixed matrix. In the limit the observer gains also get large, as in the Doyle and Stein
method, so sensor noise amplification must be examined. It was found that the LTRLM method offers
improvements in the amount of recovery as compared to the Doyle and Stein method, and the method offers
improvements in distributing the recovery into additional loops within the architecture.
References
1
Doyle, J.C., “Guaranteed Margins for LQG Regulators,” IEEE Trans. Automat. Contr., 23(4) , 1978, pp. 756-757.
2
Safanov, M.G., Athans, M., “Gain and Phase Margins of Multiloop LQG Regulators,” IEEE Trans. Automat. Contr., 22(1) ,
1977, pp. 173-179.
3
Doyle, J.C., Stein, G., “Multivariable feedback design: Concepts for a classical/modern synthesis,” IEEE Trans. Automat.
Contr., 26(1) , 1981, pp. 4 – 16.
4
Stein, G., M. Athans, “The LQG/LTR procedure for multivariable feedback control design,” IEEE Trans. Automat. Contr., 32(2)
, 1987, pp. 105 – 114.
5
Madiwale, A.N., Williams, D.E., “Some Extensions of Loop Transfer Recovery,” Proc, of the American Control Conference,
June, 1985, pp 790-795.
6
Athans, M., “A Tutorial on the LQG/LTR Method,” Proc, of the American Control Conference, June, 1986, pp 1289-1296.
7
Shafai, B., Keel, L. H., Beale, S., “Zero Assignment and Loop Transfer Recovery in LQG Design,” Proceedings of the 29th
IEEE Conference on Decision and Control, Vol. 3. Dec. 1990, pp. 1217-1221.
8
Zhang, Z., Freudenberg, J., Loop Transfer Recovery for Nonminimum Phase Systems,” IEEE Trans. Automat. Contr., 35(5),
1990, pp. 547-553.
9
O'Leary, D.P., Monahemi, M.M.; Barlow, J.B., “On The Precise Loop Transfer Recovery and Transmission Zeros”, Proc of the
First IEEE Conference on Control Applications, Vol. 1, Sept. 1992, pp. 264-270.
10
Monahemi, M.M.; Barlow, J.B., O'Leary, D.P., “Design of Reduced Order Observers with Precise Loop Transfer Recovery,”
Journal of Guid,. Control, and Dynamics, 15(6), 1992, pp. 1320-1326.
11
Maciejowski, J.M., (1989) Multivariable Feedback Design, Addison Wesley, New York.
12
Saberi, A., Chen, B., Sannuti, P., (1993) Loop Transfer Recovery: Analysis and Design, Springer Verlag, Berlin.
13
Doyle, J.C., Stein, G., “Robustness with Observers,” IEEE Trans. Automat. Contr., 24(1) , 1979, pp. 607-611.
14
Lavretsky, E “Adaptive Output Feedback Design Using Asymptotic Properties of LQG / LTR Controllers,” In Proceedings of
AIAA Guidance, Navigation and Contr. Conf., Toronto, Ontario, Canada, 2010.
15
Lavretsky, E., and K. A. Wise, Robust and Adaptive Control, Springer-Verlag, 2012 (to be published).
16
Kwakernak, H., and R. Sivan (1972), Linear Optimal Control Systems, Wiley, New York.
17
Zhou, K., Doyle, J. C., Glover, K. (1996), Robust and Optimal Control, Prentice Hall, New Jersey.
18
Khalil, H (2002) Nonlinear systems, Third Edition, Prentice Hall, Upper Saddle River, NJ 07458
Plant w x0
u + x x
B
+
∫ C
+ v
+
A y
Control +
KF Gain
Kf
-
x̂0
u + x̂ x̂ ŷ
B
+
∫ C
A
Controller
− Kc
LQR Gain
Fig. 2 Robust servo LQG using integral control-estimated state feedback.
State Feedback simulation State Feedback simulation

1.4 0.3
1.2
0.2
1
0.1
0.8
Accel (ft/sec**2)
Pitch Rate (dps)
0.6 0
0.4 -0.1
0.2
-0.2
0
-0.3
-0.2
-0.4 -0.4
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time (sec) Time (sec)
Stochastic Simulation Stochastic Simulation
2 0.6
0.4
1.5
0.2
Measured Accel (ft/sec**2)
Measured Pitch Rate (dps)
0
0.5
-0.2
0
-0.4
-0.5
-0.6
-1 -0.8
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Fig. 3 State feedback and measured acceleration and pitch rate time histories.
1.4 0.3
1.2
0.2
1
0.1
0.8
Az (ft/sec**2)
0.6 0
q (dps)
0.4 -0.1
0.2
-0.2
0
-0.3
-0.2
-0.4 -0.4
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

2 25
20
1.5
15
1 10
Deltadot (dps)
5
Delta (deg)
0.5
0
0
-5
-0.5 -10
-15
-1
-20
-1.5 -25
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Fig. 4 State estimates using the nominal Kalman Filter process noise Q0 .
Nyquist
3
1
Im(L)
0
ρ = 104
ρ = 103
-1 ρ = 102
LQG
-2
LQR ρ = 105
-3
-3 -2 -1 0 1 2 3
Re(L)
Fig. 5 Nyquist plot comparing LQR and LQG designs.

|I+L| at input
60
01 50
40
Magnitude (dB)
30 LQR
20
ρ = 102
ρ = 103
10
ρ = 104
ρ = 105
0
LQG
-10
-1 0 1 2 3
10 10 10 10 10
Frequency (rps)
Fig. 6 Return difference dynamics I + L at the plant input for LQR, LQG/LTR designs.
|I+inv(L)| at Plant Input

90
80
70
60 ρ = 102
ρ = 103
Magnitude (dB)
50
ρ = 104
40
ρ = 105
30
20
LQG
10 LQR
0
-10
-1 0 1 2 3
10 10 10 10 10
Frequency (rps)
Fig. 7 Stability robustness I + L

−1 at the plant input for LQR and LQG/LTR designs.
|S| at Plant Output
10
0 LQG
-10
Magnitude (dB)
-20 LQR
ρ= 102
-30 ρ = 103
ρ = 104
-40
ρ = 105
-50
-60
-1 0 1 2 3
10 10 10 10 10
Frequency (rps)
Fig. 8 Sensitivity S at the plant output for LQR and LQG/LTR designs.
|T| at Plant Output

10
0
LQR
-10
LQG
-20
Magnitude (dB)
-30 ρ = 102
-40 ρ = 103
ρ = 104
-50
ρ = 105
-60
-70
-80
-90
-1 0 1 2 3
10 10 10 10 10
Frequency (rps)
Fig. 9 Complementary sensitivity T at the plant output for LQR and LQG/LTR designs.
Noise-to-Control Gain Matrix
60
40 LQR
20
Magnitude (dB)
0
-20 LQG
ρ = 102
-40
ρ = 103
ρ = 104
-60
ρ = 105
-80
-4 -2 0 2 4
10 10 10 10 10
Frequency (rps)
Fig. 10 Noise transmission through the controller for LQR and LQG/LTR designs.
Closed Loop Sim Test

1.4
1.2
1
Az (ft/sec**2)
0.8
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time (sec)
Fig. 11. Step response for the LQR, LQG, LQG/LTR, and LQG.LTRLM controller designs
Nyquist
0
Nyquist -0.1
LQG
2
-0.2
LQG/LTRLM - 2
-0.3
1.5
-0.4
LQG/LTRLM - 1
Im(L)
-0.5
1
-0.6
-0.7
LQG/LTR
0.5
-0.8
LQR
Im(L)
-0.9
0
-1
-1 -0.9 -0.8 -0.7 -0.6 -0.5 -0.4 -0.3 -0.2 -0.1 0
Re(L)
-0.5
-1
-1.5
-2
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
Re(L)
Fig. 12. Nyquist plot for the LQR, LQG, LQG/LTR, and LQG.LTRLM controller designs
Bode Bode
60 -90
-100 LQG/LTRLM - 2
40
-110
20
LQG/LTR -120
LQR
Magnitude dB
Phase deg
0 -130
LQR
-140
-20 LQG
LQG/LTRLM - 2 -150
-40
LQG/LTRLM - 1
LQG/LTRLM - 1 -160
-60
LQG -170 LQG/LTR
-80 -180
-1 0 1 2 3 -1 0 1 2 3
10 10 10 10 10 10 10 10 10 10
Frequency (rps) Frequency (rps)
Fig. 13. Bode plot for the LQR, LQG, LQG/LTR, and LQG.LTRLM controller designs
|I+L| at input |I+inv(L)| at Plant Input

60 70
50 60 LQG
40
50 LQG/LTRLM - 1
40
LQG/LTR
Magnitude dB
Magnitude dB
30
30
20
LQG/LTRLM - 2
20
10
LQR
10
0 0
-10 -10
-1 0 1 2 3 -1 0 1 2 3
10 10 10 10 10 10 10 10 10 10
Fig. 14. I + L and I + L−1 at the plant input for the LQR, LQG, LQG/LTR, and LQG.LTRLM controller designs
|S| at Plant Output |T| at Plant Output
0 0
-20
-10
-40
-20
LQG
Magnitude dB
Magnitude dB
-60
LQG/LTRLM - 1
-30
-80
LQG/LTR
-40
-100
LQG/LTRLM - 2
-50
-120
LQR
-60 -140
-1 0 1 2 3 -1 0 1 2 3
10 10 10 10 10 10 10 10 10 10
Fig. 15. S and T at the acceleration output for the LQR, LQG, LQG/LTR, and LQG.LTRLM controller designs
Noise-to-Control Transfer Function Matrix

35
30
LQG/LTRLM - 2 LQR
25
20
Magnitude dB
15
LQG/LTR
10
LQG/LTRLM - 1
5
-5 LQG
-10
-1 0 1 2 3
10 10 10 10 10
Frequency (rps)
Fig. 16. Noise transmission through the controller for the LQR, LQG, LQG/LTR, and LQG.LTRLM controller designs

Wise 2012

Uploaded by

Document Information

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

Wise 2012

Uploaded by

Copyright:

Available Formats

AIAA Guidance, Navigation, and Control Conference AIAA 2012-4889

13 - 16 August 2012, Minneapolis, Minnesota

Asymptotic Properties of LQG/LTR Controllers in Flight

still play an important role.

Fig. 1 X-45A Joint Unmanned Combat Air System.

II. Asymptotic Properties of Linear Optional Control

along with the infinite-time LQR problem

III. Doyle and Stein LQG/LTR

function matrix (LTFM) at the plant input is

y Cx + v and letting ρ → 0 as in (24) yields

Step 1: LQR controller design: K c

IV. LQG/LTR Method of Lavretsky

i) Poles of all elements of G ( s ) are in the left half complex plane.

Kalman-Yakubovich-Popov (KYP) Lemma18. Let,

be a ( p × p) transfer function matrix, where ( A, B ) is controllable and ( A, C ) is observable. Then, G ( s ) is

P PT > 0 , L , W , and a positive constant ε such that

and in this case, setting ε = 0 , gives the PR conditions in the form:

The state estimate x̂ is formed as before, using the state estimator,

xˆ = Axˆ + Bu + K f ( ymeas − yˆ )

The gains in (33) are computed as

• There exists a unitary matrix W ∈ R m×m such that:

• The unitary matrix W in (41) can be chosen as:

and Σ represents the diagonal matrix of the corresponding singular values.

Step 1: LQR controller design: K c

V. X-45A Design Example

Doyle and Stein LQG/LTR

K c = [ 0.49482 0.17904 -14.061 2.2089 1.8036e-003] (46)

The control law is implemented using

1.94 × 10−4 0 0 0  ( fps )2 / s 

6.25 × 10−2 0  ( fps )2 

xˆ = Axˆ + Bu + K f ( y − yˆ ) (50)

 2.0442e-003 -6.0760e-006 2.4929e-010 -5.6592e-008 

ρ = ∞ 105 104 103 102  (53)

For the LQG design ( ρ = ∞ )

 0 -1.0854e+000 -3.4041e+002 0 -1.1289e+001 

The noise transmission through the controller is given by

that the value of ρ would need to be limited to 104 or larger.

Kalman Filter Gains

LQG LQG/LTR ρ=104

3.2707e-002 -6.0760e+000 6.9018e+000 1.7004e+002

LQG/LTR Lavretsky Method (LTRLM)

Az and pitch rate q given by

K c = [ 0.31623 -33.261 -6.7127 ] (63)

The above control law is implemented using

0.000196 0  ( rad ) / s 

6.25 × 10−2  ( fps )2 

The Kalman filter state estimator is

xˆ = Axˆ + Bu + K f ( y − yˆ ) (69)

For the LQG controller, the RSLQR control law is given by

Note the magnitude increase in the gain K f ( 2, 2 ) .

For ν = 2.5 and using B = B1 yields

0.000196 0  0.01 1  0.014196 1.4 

Fig. 2 Robust servo LQG using integral control-estimated state feedback.

State Feedback simulation State Feedback simulation

Pitch Rate (dps)

Measured Pitch Rate (dps)

Stochastic Simulation Stochastic Simulation

Fig. 5 Nyquist plot comparing LQR and LQG designs.

|I+inv(L)| at Plant Input

Fig. 7 Stability robustness I + L

|T| at Plant Output

Closed Loop Sim Test

|I+L| at input |I+inv(L)| at Plant Input

Noise-to-Control Transfer Function Matrix