2010 International Conference on Intelligent Computation Technology and Automation

LQG/LTR Flight Controller Optimal Design Based on Differential Evolution

Algorithm

Meng Zhang 1 , Peiyong Sun 2 , Ruiting Cao 1 , Jiangle Zhu 1

1. College of Automation, Northwestern Polytechnical University, Xi’an, Shaanxi, 710072, China
2. Scenic Spots and Historic Sites Management of Qianfoshan, Jinan, Shandong, 250014, China
[email protected]

Abstract—In conventional Linear Quadratic Gaussian with Storn and Kenneth Price in 1995 [3]. This algorithm is a
Loop Transfer Recovery (LQG/LTR) controller design, the population-based evolutionary algorithm for global
designer should experiment with four different weighting optimization over continuous search spaces. DE algorithm
matrices by trial-and-error method in order to get the flying has recently found extensive applications in solving global
quality requirement and the robustness. This method is a time optimization searching problem. As the same to GA, DE
consuming, inefficient and non-optimal method .To solve this algorithm is also suitable to search the best values of the
problem, a LQG/LTR flight controller optimal design method weighting matrices in the LQG and LQG/LTR optimal
based on differential evolution algorithm is proposed in this design. Unfortunately, to the best knowledge of the authors,
paper. In the optimal design, a Kalman filter is optimal
there is no research result in LQG/LTR controller optimal
designed by optimizing two weighting matrices based on a
reference model and differential evolution algorithm firstly. So
design based on DE algorithm.
the optimal target feedback loop which satisfies the In order to solve the problem above mentioned, a
performance requirement is obtained. Secondly, the principle LQGLTR flight controller optimal design method based on
of the aircraft equivalent system analog match is used for DE algorithm is proposed in this paper.
reference to design an optimal state feedback gain matrix by In the section II, the basic principle of LQG/LTR design
optimizing another two weighting matrices. To validate the method and DE algorithm are given. In order to apply the
effect of this optimal design method, a longitudinal LQG/LTR DE algorithm in the optimization of the LQG/LTR flight
flight controller is optimal designed based on differential controller two optimizing strategy based on reference model
evolution algorithm. The simulation results show the high which consider the flying quality criteria and aircraft
effectiveness of this optimal design method. equivalent system analog match are described in the section
III. In the section , a simulation on a LQG/LTR flight
Keywords-flight controller; LQG/LTR; optimal design; controller optimal design is made to validate the effect of the
differential evolution algorithm; weighting matrices optimal design method introduced in the section III. In the
last section, give the conclusion of this paper.
I. INTRODUCTION
II. THE PRINCIPLE OF LQG/LTR DESIGN AND DE
Linear Quadratic Gaussian with Loop Transfer Recovery ALGORITHM
(LQG/LTR) methodology is a method for the design of
linear multivariable control systems on the basis of Linear A. The Principle of LQG/LTR Design
Quadratic Gaussian (LQG) methodology originally proposed
by Doyle and Stein in [1]. This method shows some useful Suppose that we have a plant model in state space form:
properties of good performance and robustness in controller x (t ) = Ax(t ) + Bu (t ) + Γw(t )
(1)
or flight controller design [4] [5] [6] [7]. y (t ) = Cx(t ) + v(t )
In a conventional LQG and LQG/LTR design, the Where w and v are white noises, namely zero-mean
designer will experiment with different weighting matrices in Gaussian stochastic process which are uncorrelated in time
order to achieve the performance requirements and as well as with one another having covariance:
robustness. The parameters of the weighting matrices are set
manually by trial-and-error method. Such manual selection E[ w(t ) wT (t )] = W ≥ 0
of the matrix elements is not straightforward and, thus, E[v(t )vT (t )] = V > 0 .
Evolutionary Algorithms (EAs) such as Genetic Algorithm
(GA) and Differential Evolution (DE) Algorithm can be used E[ w(t )vT (t )] = 0
to automate the search for the best values of the weighting The structure of LQG/LTR controller is shown in figure
matrices that realize the design specifications. In one such 1.
example, Mei and Goodal [2] combined LQG design with
Genetic Algorithm which used a GA to search for the best
values for the weighting matrix of LQG design.
Differential Evolution algorithm is introduced by Rainer

978-0-7695-4077-1/10 $26.00 © 2010 IEEE 613

DOI 10.1109/ICICTA.2010.302
B. The Principle of DE algorithm
The Optimization objective function can be described as:
min f ( x1, x2 ,", xD ) s.t. xLj ≤ x j ≤ xUj ,( j = 1,2, ⋅⋅⋅, D) (8)
Let xi ( g ) = [ xi1 ( g ) , xi 2 ( g ) , " , xiD ( g )] denotes
the solutions or individuals in generation g and D denotes
the dimension of solution space, population
X ( g ) = [ x1 ( g ) , x2 ( g ) , " , xN ( g )] with population
Figure 1. The structure of LQG/LTR controller number N.
In DE algorithm, the child population is generated
The transfer function matrix of the controlled plant and through the following operations:
the control system is respectively 1) Mutation Operation
G ( s ) = C ( SI − A) −1 B (2) For the mutation operation, the mutant individual
vi ( g + 1) is generated according to (9)
Gc ( s ) = K c ( SI − A + BK c + K f C ) −1 K f (3)
vi ( g + 1) = xi ( g ) + F ⋅ ( xbest ( g ) − xi ( g ) + xr1 ( g ) − xr 2 ( g )) (9)
From equation (3), we can see that in the LQG/LTR
design we should design the matrix K f and K c . Where F is the amplification factor.
2) Crossover Operation
K f is a Kalman filter gain given by In order to increase the diversity of the perturbed
K f = PC Τ
R1−1 (4) parameter vectors, crossover operation is introduced. For the
1
crossover operation, the perturbed vector vi ( g + 1) and the
Where P1 satisfies the algebraic Riccati equation:
target vector xi , j ( g ) are both used to generate a trial vector
AP1 + P1 AΤ + Q1 − PC
1
Τ
R1−1CP1 = 0 (5)
T T T
ui , j ( g + 1) by (10).
and Q1 = Q ≥ 0 , R1 = R > 0 , P1 = P ≥ 0 .
⎧⎪vi, j (g +1), if rand ( 0,1) ≤ CR or j = rand (0, D −1)
1 1 1

K c is an optimal state-feedback matrix given by ui, j (g +1) = ⎨

⎪⎩ xi , j (g) otherwise
K c = R2−1 BT P2 (6)
(10)
Where P2 satisfies another algebraic Riccati equation:
Where rand(0,1) is a uniform random number distributed
P2 A + AT P2 − P2 BR2 −1 BT P2 + Q2 = 0 (7) between [0, 1] , rand (0, D−1) is a randomly selected index
T T
and Q2 = Q ≥ 0 , R2 = R > 0 , P2 = P ≥ 0 . T from the set {0, 1, " , D−1} , the crossover rate CR∈[0, 1]
2 2 2
controls the diversity of the population.
In order to solve the Riccati equations and then get the
3) Selection Operation
matrices K f and K c , we should select four weighting
The child individual xi ( g + 1) is selected from each pair
matrices Q1 , R1 , Q2 , R2 .
of xi ( g ) and ui ( g + 1) by using greedy selection criterion:
If the loop transfer recovery has been done at the output
of the plant, according to the separation principle, the design ⎧u (g +1) , if f (ui (g +1)) ≤ f (xi (g))
xi (g +1) = ⎨ i (11)
procedure of LQG/LTR consists of two independent steps as ⎩ xi (g) , otherwise
follows: Where f is the fitness value.
1) Design a Kalman filter gain K f by manipulating the
weighting matrices Q1 and R1 so that the target feedback III. THE PRINCIPLE OF LQG/LTR FLIGHT CONTROLLER
OPTIMAL DESIGN
loop (TFL) Cφ ( s ) K f = C ( sI − A)−1 K f meets the
performance and robustness requirements. A. The Principle of the TFL Optimal Design method
2) Design an optimal state feedback gain matrix K c by When the DE algorithm is applied in the optimization of
manipulating the weighting matrices Q2 and R2 so that the the target feedback loop design, the problem is how to
describe the requirements on the TFL which should match
return ratio G ( s )Gc ( s ) has been converged sufficiently
the military flying quality criteria such as MIL-STD-1797A,
closed to C ( sI − A) −1 K f over a sufficiently large range of GJB-185-86. Usually, when the flight controller is designed,
frequencies. This procedure is called loop transfer recovery. we can obtain the desired transfer function of the closed
From the principle and the design procedure of loop system according to the flying quality criteria. This
LQG/LTR method above mentioned, we can see that the transfer function can be regarded as the reference model.
key link of designing LQG/LTR controller is selecting the The reference model denotes the aircraft’s characteristics
weighting matrices. In the classical LQG/LTR design, the (time domain and frequency domain). Elicited by this fact,
weighting matrices are usually chosen on trial-and-error an optimizing strategy based on the referenced model which
method to get good responses and robustness. can describe these quantitative performance indexes
conveniently, is given as bellow and expressed as Figure 2.

614
In the figure 3, TFL diagram is the transfer feedback
loop which has been optimal designed by DE algorithm
above mentioned. DE diagram is the DE algorithm which
Ym regulates the parameters of the weighting matrices Q2 , R2 .
R +
Kf φ(s) C
− ω is the frequency input. The cost function J 2 is selected as
Y follow:
n

solve Riccati equation E J 2 = M = ∑ [ΔG 2 ( jωi ) + K ΔΦ 2 ( jωi )] (13)

i =1
Q1 , R1 Where ΔG ( jωi ) = GTFL ( jωi ) − G ( jωi )Gc ( jωi ) and
ΔΦ = ΦTFL ( jωi ) − Φ ( jωi )Φ c ( jωi ) respectively represents
the amplitude error and the phase error between the TFL
and the return ratio G ( s)Gc ( s) . n is the number of selected
Figure 2. The principle of the TFL optimal design
frequency points and usually n = 20 . ωi are frequency points
In the figure 2, Ym is the output of the reference model, which chosen by 20 equal division according to logarithm
Y is the real output signal of the TFL, DE diagram is the DE coordinate within the range [0.1,10]rad/s. K is usually
algorithm which regulate the parameters of the weighting K = 0.0175 .
matrices Q1 , R1 . R is the input of the system. E is the When J 2 is larger than a given tolerance, the DE
absolute value of error between the output of the reference algorithm will regulate the parameters of the weighting
model and the TFL’s real output. The cost function J1 is matrices Q2 , R2 and then the optimal state feedback gain
selected as follow: matrix K c and the return ratio G ( s )Gc ( s ) will change at
J1 = ∫ E dt = ∫ Ym − Y dt (12) same time according to the value of J 2 .When J 2 is less
When J1 is larger than a given tolerance, the DE will than the given tolerance or the iteration number is achieved ,
regulate the parameters of the weighting matrices Q1 , R1 and we can consider that the Q2 , R2 are the optimal weighting
then the Kalman filter gain K f according to the value of J1 matrices and then the optimal state feedback gain matrix K c
can obtain and the return ratio can converge sufficiently
until J1 is less than the given tolerance. Then the TFL has closed to the TFL in frequency domain.
the approximative characteristics with the reference model In the LQG/LTR controller optimal design process, the
in the time domain and the interesting frequency range. In parameters of the matrices Q1 , R1 , Q2 , R2 are taken as the
this situation, the TFL can be regarded as a system that has
individuals of the DE algorithm and then we can use DE
the desired performance and robustness.
algorithm to obtain the optimal weighting
B. The Principle of LTR Optimal Design matrices Q1 , R1 , Q2 , R2 . As soon as we get the optimal
In LTR optimal design , we can by means the principle weighting matrices, the optimal Kalman filter and the
of aircraft equivalent system analog-match to optimize the optimal state feedback gain matrix are also obtained. Up to
optimal state feedback gain matrix K c so that the return now, the LQG/LTR optimal design is completed.
ratio G ( s )Gc ( s ) converge sufficiently closed to the TFL in IV. SIMULATION RESULTS
frequency domain.
To validate the effect of the LQG/LTR optimal design
The principle of the LTR optimal design is shown in
method based on DE algorithm mentioned in the section III,
figure 3.
a simulation, which is based on a kind of fighter aircraft’s
longitudinal attitude LQG/LTR flight controller, is made.
The flight state of the aircraft is M = 0.8 ， H＝0km .
The state space linear model of the aircraft is given in
ω G ( s )Gc ( s )
− +
equation (1) and x = [V α q θ ] , u = δ e . The control
T

objective is to design a LQG/LTR controller that the pitch

angle θ could follow a desired command θ c = 20D .The
Kc tracking error is eθ = θ c − θ and then the dynamics of the
M
solve Riccati equation
Q2 , R2 aircraft are augmented by integrator ε θ which εθ = eθ .The
full state vector of the augmented controlled system is
x = [V α q θ eθ ] .
T

Figure 3. The principle of the LTR optimal design

615
In the optimal design of the target feedback loop, the Figure 5. Step response of the closed loop system
θ 30
reference model is = which satisfies the
θ c s 2 + 13s + 30 25

level 1 flying quality.

After the optimal design of the target feedback loop 20

based on DE algorithm, the optimal matrices Q1 , R1 are ：

Q1 = diag [1.317 0.92188 18.901 1.0285 1810.9]
15

θ (deg)
R1 = 63.774 . 10

The optimal Kalman filter gain is: 5

K f = [1.6532 3.6321 2.7497 -0.8731 -5.3288] .

T
A B +20%
Reference Model
A B -20%

After the LTR optimal design based on DE algorithm,

0
0 1 2 3 4 5 6 7 8 9 10
t (sec)

the optimal matrices Q2 , R2 are shown as

follows: Figure 6. Step response of pitch angle with the ±20% perturbation of A, B
Q2 = diag[ 4.255e-06 6.6679e-06 9.9256e-06 4.1144e −03 1.3093e−01]
R2 = 7.2125e-06 V. CONCLUSION
The optimal state feedback gain matrix K c is: The LQG/LTR flight controller optimal design method
K c = [ 0.8963 0.3721 -1.1731 1.5915 134.9722] . based on DE algorithm is proposed in this paper. In the
optimal design process, at first, the Kalman filter of the
The singular value plots of the TFL and return ratio is target feedback loop was optimal designed by optimizing
shown in figure 4. The step response of the closed loop the weighting matrices based on DE algorithm and reference
system and the reference model is shown in figure 5. The model which satisfied the level 1 flying quality, so the
step response of the pitch angle with the ±20% perturbation optimal target feedback loop was obtained. Secondly, in the
of A, B is shown in figure 6. From figure 5,6, the time LTR optimal design process, the optimal state feedback gain
response curve of θ is almost same with the reference matrix was obtained by optimizing the weighting matrices
model output, it means that the aircraft with the optimized based on DE algorithm. The LQG/LTR flight controller
LQG/LTR controller matches the requirements of the flying optimal design method based on DE algorithm is more
quality criteria and the robustness. convenient in practice to design the LQG/LTR controller
Singular Values
than the conventional trial-and-error method by the
simulation results.
300
Return Ratio
TFL

200

REFERENCES
100
Singular Values (dB)

0
[1] G. Stein and M. Athans, "The LQG/LTR procedure for multivariable
feedback control design," IEEE Trans. on Automatic Control, Vol.
32(2),pp.105–114, February,1987.
-100

-200 [2] Mei. T. X and Goodall. R.M A, “LQG and GA solutions for active
steering of railway vehicles,” IEE Proceeding-Control Theory and
-300
10
-4 -3
10 10
-2
10
-1
10
0 1
10 10
2
10
3 4
10 Applications vol. 147(1), pp. 111–117, January,1955..
Frequency (rad/sec)

[3] Storn R and Price K, “Differential evolution–a simple and efficient

Figure 4. Singular value plots of the return ratio and the TFL adaptive scheme for global optimization over continuous spaces,”
Technical Report International Computer Science Institute, pp. 22–
25,August,1995.
25 10

20
8 [4] B.Walker and A.Saraf, “Robust flight control for coordinated
15
6
turns,”AIAA Guidance, Navigation, and Control Conference,11-14
θ (deg)

α (deg)

August,2003.
4
10
2

5
closed loop system
reference model
0
[5] David Haes, “Selection of LQG/LTR Weighting Matrices through
Constrained optimization,” Proc of American Control Conference
0 -2
0 5 10 15 20 0 5 10 15 20
t (sec) t (sec)

20 1
Seattle, Washington, pp. 458–460, June, 1995
15
0
[6] Prakash R and Rao S.V, “LQG/LTR controller design using a reduced
-1
order model,” IEEE International Conference on System Engineering,
q (deg/s)

10
δ (deg)

-2

pp.343–346, August, 1989.

e

5
-3

[7] Kevin A. Wise, “A trade study on missile autopilot design using

0
-4

-5 -5
0 5 10
t (sec)
15 20 0 5 10
t (sec)
15 20
optimal,” AIAA Guidance, Navigation, and Control Conference,20-
23, August,2007.

616