12th International conference on Sciences and Techniques

of Automatic control & computer engineering

December 18-20, 2011, Sousse, Tunisia

Synthesis of observer based state

feedback control for linear time
variant systems using orthogonal
functions
Bassem Iben Warrad, Naceur Benhadj Braiek
Laboratoire d’Etude et Commande Automatique de Processus
-LECAP-
Ecole Polytechnique de Tunisie, BP. 743, 2078 La Marsa, Tunisia
e-mail: [Link]@[Link]
[Link]@[Link]

Abstract. In this work we propose a new approach to design an observer

based control law for time variant linear systems using orthogonal func-
tions. The proposed technique uses the projection of the original system
on an orthogonal functions basis, which allows converting the differential
state equation into an algebraic one. The main goal of this technique is
to shape a dynamic behavior of the considered system that is very close
to a chosen reference model.

[Link] feedback , state observer, linear time variant systems, ref-

erence model, orthogonal functions, operational matrices, shifted Legen-
dre polynomials.

1 Introduction

In process control, the conception of a state observer is useful when the state of
the considered system can’t be measured to perform a state feedback ([1]) ([2]).
Methods of state feedback with state observer were developed([3]) but they are
especially applied for stationary linear systems. When dealing with linear time
variant systems, the developed approaches of control in the literature remain
limited to particular classes of models.

In this paper, a new method of state feedback with observer using orthog-
onal functions is proposed. The improvement of the system dynamic behavior
is obtained by following the performances of a selected reference model. This
approach originized from the use of Walsh ([4]) and block-pulse ([5]) functions
was later extended to orthogonal polynomial series such as the Laguerre ([6]),
the Chebychev ([7]), the Hermite ([8]) and the Legendre polynomials ([9]). They

STA'2011-ACS-1699, pages 323-336

Academic Publication Center of Tunis, Tunisia
were also used with non linear systems ([10]). The Legendre polynomials may
have advantages over other orthogonal functions. This was shown by way of
examples ([9]) where Legendre polynomials converge to the exact solution of a
differential equation faster than the other types of orthogonal functions, as, for
example Walsh functions, Hermite and Laguerre polynomials.

This paper is organized as follows: in section 1, the orthogonal functions are

presented with interesting properties and their use for systems description. The
proposed method for linear time variant systems controller synthesis using or-
thogonal functions is derived in section 2. In the last section, an example is
presented to emphasise the effectiveness of this method.

2 Orthogonal functions for dynamic system description

2.1 Approximation using orthogonal functions
The continuous orthogonal functions have been adopted by many researchers as
a convenient and sharp tool to approximate the solution of physical systems.
The key idea of this technique is that all analytical function f (t) absolutely
integrable can be developed as follows:
∞
X
f (t) = fi ϕi (t) (1)
i=0

where
Zb
fi = 1/ri w(t)f (t)ϕi (t) (2)
a
with w(t) a positive and integrable function as the weighting function of the
scalar product.
For practical use, the development (1) is truncated up to an order N, thus giving
the following time approximation of the function f (t):
N −1
f (t) ∼
X
= fi ϕi (t) = FN φN (t) (3)
i=0
with  
FN = f0 f1 · · · fN −1
 T
ΦN (t) = ϕ0 (t) ϕ1 (t) · · · ϕN −1 (t) the vector of the orthogonal functions
basis.

The coefficients fi have the particularity to minimize the error :

Zb N −1
!2
X
ε= f (t) − fi ϕi (t) dt
a i=0
The orthogonal functions obey the orthogonality relation:
Zb
hϕi (t), ϕj (t)i = w(t)ϕi (t)ϕj (t)dt = δij r
a

where δij is the symbol of Kronecker which is equal to 1 if i = j and to 0 if i 6= j

and ri is the result of the scalar product.

2.2 Approximation of matrix functions

A time dependent matrix function A(t) ∈ Rn×m given by A(t) = [aij (t)] where
aij (t) are integrable over an interval [a, b] can be developed into orthogonal
functions series with a truncation to an order N under the following relation:
N −1
A(t) ∼
X
= AiN ϕi (t) (4)
i=0

with AiN ∈ Rn×m for i ∈ {0, 1, ..., N − 1} are matrices with constant coefficients.

2.3 operational properties

The interesting properties of an orthogonal functions basis allow to convert the
differential equations describing a dynamic system into algebraic equations lead-
ing to significant simplifications of the problems of analysis and control.

a-Operational matrix of integration

For a given basis of orthogonal functions Φ = {ϕi , i ∈ N}, the operational matrix
of integration is a constant matrix PN ∈ RN ×N such as:
Zt
ΦN (t)dt ∼
=PN ΦN (t) (5)
a

Clearly, the form of PN depends on the particular choice of the basis vector
ΦN (t).
b-Operational matrix of product
The operational vectors of product Kij have constant coefficients and verify the
property ([11]):
∀i, j ∈ {0, 1, . . . , N − 1 } , ϕi (t)ϕj (t) ∼ T
= Kij ΦN (t) (6)
From the relationship (6), we can readily get the operational matrix of product:
 T 
K0,i
MiN =  ... (7)
 

T
KN −1,i
that allows the approximation:

ϕi (t)φN (t) ∼
= MiN φN (t) (8)

3 observer based feedback for linear time variant systems

3.1 Principle of the proposed control law
Consider the system (S) described by the following state equation:

ẋ(t) = A(t)x(t) + B(t)u(t)
(9)
y(t) = C(t)x(t)
with the state vector x(t) ∈ Rn , the input u(t) ∈ R and the output y(t) ∈ R.

The system (S) is assumed to be controlable and observable and the n com-
ponentsof the state vector are not all physically measurable.
Let Ŝ a full ordre observer of the system (S), defined by the following
state equation:
˙

x̂(t) = A(t)x̂(t) + B(t)u(t) + L(t) (y(t) − C(t)x̂(t))
(10)
y(t) = C(t)x(t)

The strategy of control that we plan to develop is characterized by a law of

an observed state feedback , expressed in the following form:

u(t) = N yc (t) − K x̂(t) (11)

with:

N ∈R and K ∈ Rp×n

This control law has to provide the same behavior between the system and
a chosen reference model.

This reference model is defined by the following state equation:


ż(t) = Ez(t) + F yc (t)
(12)
yr (t) = Gz(t)

3.2 Synthesis of a state observer using orthogonal functions

One definite the observation error by:

ε(t) = x(t) − x̂(t) (13)

It comes out then:

˙
ε̇(t) = ẋ(t) − x̂(t) = (A(t) − L(t)C(t)) ε(t) (14)
 It is desired to determine the observation gain L(t) such that the observation
error has the some dynamic as a chosen observation reference model described
by a linear state equation:

ε̇ref (t) = M εref (t) (15)

Where M is a (n × n) matrix chosen such that the observer be faster than
the controlled system.

The projection of the matrices A(t), B(t), C(t) and L(t) in basis of orthogonal
functions truncated with an order N can be written as:
NP
−1 NP
−1
A(t) = AiN ϕi (t), B(t) = BiN ϕi (t)
i=0 i=0 (16)
NP−1 NP−1
C(t) = CiN ϕi (t), L(t) = LiN ϕi (t)
i=0 i=0

The integration of the equation (14) leads to::

Zt Zt
ε̇(τ )dτ = ε(t) − ε(t = 0) = [A(τ ) − L(τ )C(τ )] ε(τ )dτ (17)
0 0

The projection of the observation error ε(t) on an orthogonal basis leads to:
NP
−1 Rt
εN φN (t) − ε0,N φN (t) = AiN εN ϕi (τ )φN (τ )dτ −
i=0 0
NP
−1 NP
−1 Rt (18)
LiN CjN εN ϕi (τ )ϕj (τ )φN (τ )dτ
i=0 j=0 0

The use of the product operational matrix allows to write:

NP
−1 Rt
εN φN (t) − ε0,N φN (t) = AiN εN MiN φN (τ )dτ −
i=0 0
NP
−1 NP
−1 Rt (19)
LiN CjN εN MiN MjN φN (τ )dτ
i=0 j=0 0

The use of the integration operational matrix yields:

NP
−1
εN φN (t) − ε0,N φN (t) = AiN εN MiN PN φN (t)−
i=0
NP
−1 NP
−1 (20)
LiN CjN εN MiN MjN PN φN (t)
i=0 j=0

Making use of the vec operator, which transforms a matrix structure into a
vector one and the specific property ([12]):

vec(ABC) = C T ⊗ A vec(B)


(21)
 Simplifying by the vector φN (t) and applying the operator vec lead to:
vec(εN ) =
 N −1  −1
P T
 In×N − (MiN PN ) ⊗ AiN −  (22)
" i=0 #  × vec(ε )
 NP −1 NP
−1  0,N
T
(MiN MjN PN ) ⊗ (LiN CjN )
 
i=0 j=0

In the same way the projection of the error vector of reference observation
model, and the use of the operational matrix of integration give:

−1
vec(εN,ref ) = In×N − PNT ⊗ M

vec(ε0,N,ref ) (23)
The condition of an identical behavior of the error of reference observation
and the real error of observation can be written as follows:
εref (t) = ε(t) ⇔ εN,ref φN (t) = εN φN (t)
⇔ vec(εN,ref ) = vec(εN )
By taking into account the equations (22) and (23), one may has:
 N −1  −1
P T
 In×N − (MiN PN ) ⊗ AiN − 
" i=0 #  × vec(ε ) =
 NP −1 NP−1  0,N
(MiN MjN PN )T ⊗ (LiN CjN ) (24)
 
i=0 j=0

−1
In×N − PNT ⊗ M

× vec(ε0,N,ref )
Considering that:

vec(ε0,N,ref ) = vec(ε0,N )
It comes out:
 N −1  −1
P T
 In×N − (M iN PN ) ⊗ AiN − 
" i=0 # =
 NP
 −1 NP −1 
T
(MiN MjN PN ) ⊗ (LiN CjN ) (25)

i=0 j=0


−1
In×N − PNT ⊗ M
Or:  N −1  
P T
I
 n×N − (M P
iN N ) ⊗ AiN − 
" i=0 # =
 NP
 −1 NP−1 
T
(MiN MjN PN ) ⊗ (LiN CjN ) (26)

i=0 j=0



In×N − PNT ⊗ M
 Matrices LiN for i = 0, 1, ..., N − 1 can be determined by minimizing the
error δ which represents the difference between the two parts of the equation
(26):
 N −1  
P T
I
 n×N − (M P
iN N ) ⊗ AiN − 
" i=0 #
δ(LiN ) =  −1 NP
 NP −1 
T
(MiN MjN PN ) ⊗ (LiN CjN )

i=0 j=0


− In×N − PNT ⊗ M

3.3 Synthesis of a control law with state observer by using of the

orthogonal functions

By taking into account the equations (9) and (11), it comes out:

ẋ(t) = A(t)x(t) + B(t)N yc (t) − B(t)K x̂(t) (27)

The integration of the equation (27) on the interval [0, t] and the projection of
the state vector x(t) as well as the observed state x̂(t) and the order output yc (t)
on the basis of orthogonal functions lead to:
NP
−1 Rt
XN φN (t) = AiN XN ϕi (τ )φN (τ )dτ
i=0 0
NP
−1 Rt
+ BiN N Yc,N ϕi (τ )φN (τ )dτ (28)
i=0 0
NP−1 Rt
− BiN K X̂N ϕi (τ )φN (τ )dτ
i=0 0

The use of the product operational matrix as well as the integration opera-
tional matrix give:
NP
−1
XN φN (t) = AiN XN MiN PN φN (t)
i=0
NP
−1
+ BiN N Yc,N MiN PN φN (t) (29)
i=0
NP−1
− BiN K X̂N MiN PN φN (t)
i=0

Simplifying by the vector φN (t) and using the operator vec allow to write:
 N −1 
P T
In×N − (MiN PN ) ⊗AiN vec(XN ) =
N −1 i=0 
P T
(MiN PN ) ⊗ (BiN N ) vec(Yc,N ) (30)
i=0
N −1 
P T
− (MiN PN ) ⊗ (BiN K) vec(X̂N )
i=0
 While considering a definite observer of the form (10) and taking account of
the equality (11), it comes that:
˙
x̂(t) = (A(t) − B(t)K − L(t)C(t)) x̂(t)
(31)
+B(t)N yc (t) + L(t)C(t)x(t)
Applying the same procedure of projection of the equation (31) on the chosen
orthogonal functions basis, we get:
 N −1  
P T
 In×N − i=0 (MiN PN ) ⊗AiN + 
  
 NP −1
T


 (MiN PN ) ⊗ (BiN K) +  vec(X̂N ) =

 " i=0 # 
 N −1 N −1 
 P P T
(MiN MjN PN ) ⊗ (LiN CjN )

i=0 j=0 (32)
N −1 
P T
(MiN PN ) ⊗ (BiN N ) vec(Yc,N )+
" i=0 #
NP−1 NP
−1
T
(MiN MjN PN ) ⊗ (LiN CjN ) vec(XN )
i=0 j=0

Where X̂N is the (n × N ) matrix such that:

X̂(t) = X̂N φN (t) (33)
When merging (32) and (30), one obtains the following relation:
 N −1 
(MiN PN )T ⊗AiN vec(XN ) =
P
In×N −
i=0

N −1 
(MiN PN )T ⊗ (BiN N ) vec(Yc,N )−
P
i=0
  
NP
−1
T
 (MiN PN ) ⊗ (BiN K) × 
  i=0 N −1  −1 
 
P T
  In×N − (MiN PN ) ⊗AiN +
 

(34)
 i=0

  N −1    ×
 P T
 
 (M iN PN ) ⊗ (BiN K) +   
  i=0  
" #
 
  NP −1 NP
−1  
T
 (MiN MjN PN ) ⊗ (LiN CjN ) 
i=0 j=0

 N −1  
P T
 (MiN PN ) ⊗ (BiN N ) vec(Yc,N ) 
 " i=0 # 
 NP
 −1 NP −1 
T
(MiN MjN PN ) ⊗ (LiN CjN ) vec(XN )

i=0 j=0
 Let us write:
  
NP
−1
T
 (MiN PN ) ⊗ (BiN K) × 
  i=0 N −1  −1 
 
P T
  In×N − (MiN PN ) ⊗AiN +
 
 
 i=0 
W =   NP
  −1
 
 
T  

 (MiN PN ) ⊗ (BiN K) +  
 " i=0 #  
 N −1 N −1   
  P P (M M P )T ⊗ (L C )  
iN jN N iN jN
i=0 j=0

Then the relation (34) becomes:

 N −1  
P T
 In×N − (MiN PN ) ⊗AiN + 
 " i=0 #  vec(X ) =
 NP
−1 NP
−1  N
T
W (MiN MjN PN ) ⊗ (LiN CjN )
 
i=0 j=0
(35)
 N −1  
P T
 i=0 (M P
iN N ) ⊗ (BiN N ) − 
 
NP−1
  vec(Yc,N )
T
 
W (MiN PN ) ⊗ (BiN N )
i=0

The condition to have a similar behavior between the controlled system and
the reference model (12) can be mathematically written:

N
X −1
CiN ϕi (t)XN φN (t) = GZN φN (t) (36)
i=0

The application of the Kronecker produt and the property (21) gives the
relation:

"N −1 #
X
T
MiN ⊗ CiN vec(XN ) = (IN ⊗ G) vec(ZN ) (37)
i=0

Where vec(ZN ) is obtained by the projection of the reference model (12) on

the orthogonal functions basis as follows:

−1 T
vec(ZN ) = Ip×N − (PNT ⊗ E)

(PN ⊗ F )vec(Yc,N ) (38)

Thus the relations (35),(37) and (38) yield:

 h
−1 T
i
(IN ⊗ G) Ip×N − PNT ⊗ E PN ⊗ F vec(Yc,N ) =
N −1 
T
P
MiN ⊗ CiN ×
i=0
 N −1  −1 
P T
  In×N − (MiN PN ) ⊗AiN +  
 " i=0 

NP−1 NP
−1
#
 ×
T
 
 W (MiN MjN PN ) ⊗ (LiN CjN )
  (39)
 
 i=0 j=0 
 ×
    
  N −1 
P T
(MiN PN ) ⊗ (BiN N ) − 
 
 
  i=0 
 N −1  
 T
P  
W (MiN PN ) ⊗ (BiN N )
i=0

vec(Yc,N )

This equality must be verified for all Yc,N ∈ RN , so we have:

h
−1 T
i
(IN ⊗ G) Ip×N − PNT ⊗ E PN ⊗ F =
N −1 
T
P
MiN ⊗ CiN ×
i=0
 N −1  −1 
P T
  In×N − (MiN PN ) ⊗AiN +  
 " i=0 

NP−1 NP
−1
#
 × (40)
T
 
 W (MiN MjN PN ) ⊗ (LiN CjN )
 
 
 i=0 j=0 
 
    
  N −1 
P T
(MiN PN ) ⊗ (BiN N ) − 
 
 
  i=0 
 N −1  
 T
P  
W (MiN PN ) ⊗ (BiN N )
i=0

Finaly we obtain a nonlinear algebraic equation, which resolution by means

of an optimization technique allows the determination of the observer gain co-
efficients LiN and the control gains K and N .

4 Simulation examples

Consider a linear time variant system defined by the equations (9) with:
    
0 1 0 0 0
 0 0 1 0  0   
A(t) =   , B(t) =   , C(t) = 1 0 0 0
 0 0 0 1   0 
α1 α2 α3 α4 β1

with:
1
α1 = −1 + exp(−t), α2 = −2 + 8t+1
α3 = −3 − 0.7 cos(−0.01t), α4 = −2 + 0.5 cos(t)
β1 = 1 + 0.15 cos(1.2t − 0.5)

4.1 Synthesis of a state observer

The presented technique is applied for the synthesis of a state observer using
shifted Legendre truncated with the order N = 10 as a basis of orthogonal
functions.
With the reference model of the observation error characterized by the matrix:
 
−20 0 0 0
 0 −42 0 0 
M =  0

0 −55 0 
0 0 0 −60

We obtain the following observer gain coefficients :

 T
L0 = 9.89 10.14 10.99 6.50
 T
L1 = 10.01 10.16 10.92 7.49
 T
L2 = 10 10.08 10.48 8.73
 T
L3 = 9.1 9.98 9.91 10.27
 T
L4 = 9.99 9.95 9.77 10.64
 T
L5 = 9.99 9.95 9.76 10.67
 T
L6 = 9.99 10.03 10.24 9.32
 T
L7 = 10 10.24 11.38 6.11
 T
L8 = 9.99 10.21 11.22 6.55
 T
L9 = 10 10.34 11.96 4.52

4.2 Synthesis of the control law using a state observer

The considered reference model of the controlled system is a second order system
(12) characterized by the following parameter matrices:
   
−1.17 −0.27 0.5  
E= ,F = , G = 0 0.53
2 0 0
 Step responses of states of system considered and the those of observer
5

Amplitude
x1
0
x1observed
−5
0 1 2 3 4 5
4

Amplitude
2 x2
0
−2 x observed
2
−4
0 1 2 3 4 5
4
Amplitude

2 x3
0
−2 x observed
3
−4
0 1 2 3 4 5
2
Amplitude

x4
0
x4observed
−2
0 1 2 3 4 5
Time(sec)

Fig. 1. Step responses of the considered system state and the observer

Step Response

1.8
original system
reference model
1.6
system with state feedbook observed

1.4

1.2
Amplitude

0.8

0.6

0.4

0.2

0
0 5 10 15
Time (sec)

Fig. 2. Step responses of the original system, the reference model and the original
system in closed loop
 The proposed technique yields the following control law gains :
 
K = 100 275.8 263.5 90.6 , N = 100

The figure 2 shows the step responses of the original system, the reference
model and the system provided with the obtained control law. It appears clearly
on this figure that the proposed control law derived using the developed approach
follows perfectly the desired reference model, which proves the high availability
of the presented control technique.

5 Conclusion

In this paper, a new approach was introduced for the synthesis of an observer
based state feedback for linear time variant systems using orthogonal functions
as an approximation tool. The use of the operational matrix of integration and
operational matrix of product has allowed the transformation of differential equa-
tions into algebraic ones depending on gains of regulators. The shifted Legen-
dre polynomials have been particulary used in this work but the method stills
effective with other basis such as : Walsh and Block-pulse functions or Cheby-
chev,Laguerre, Hermite polynomials.

References

1. George, E.: Observers in Control Systems, A Practicial Guide. Academis Press, An

imprint Elsevier Science. (2002)
2. Busawon, K.K., Kabor, P.: On the design of integral and proportional integral ob-
servers. American Control Conference. volume 6 (2000) 3725-3729.
3. Ayadi, B., Benhadj Braiek, N.: State and Output feedback synthesis using orthogo-
nal functions. Conférence Internationale Francophone d´Automatique. France. (2006)
4. Chen, C., Hsiao, H.: Time-domain synthesis via walsh functions. In IEEE. volume
122 (1975) 565–570
5. Shih, L.S., Yeung, C.K., McInis, B.G.: Solution of state-space equations via Block
pulse fonctions. Int. J. Contr. volume 28 (1978) 383–392
6. King, R.E., Yeung, C.K., Paraskevopoulos, P.N.: Parametric identification of
discrete-time SISO systems. Int. J. Contr. volume 30 (1979) 1023–1029
7. Paraskevopoulos, P.N.: Chebychev Series Approach to System Identification, Anal-
ysis and Optimal Control. J. Franklin Inst. volume 316 (1983) 135–157
8. Paraskevopoulos, P.N., Kekkeris, [Link].: Hermite Series Approach to System Iden-
tification, Analysis and Optimal Control. Pro. Meas. Contr. Conf. volume 2 (1983)
146–149
9. Paraskevopoulos, P.N., Mouroutsos, S.G.: Legendre Series Approach to Identifica-
tion and Analysis of Linear Systems. Trans. IEEE. Automat. Contr. volume 30
(1985) 585–589
10. Benhadj Braiek, N.: Application des fonctions de Walsh et des fonctions modu-
latrices à la modélisation des systémes continus non linéaires. Thèse de Doctorat,
Université des Sciences et Techniques de Lille, Flandres Artois. (1990)
11. Rottela, F., Dauphin-Tanguy, G.: Non-linear systems identification and optimal
control. [Link]. volume 48 (1988) 525-544
12. Brewer, J.W.: Kronecker products and matrix calculus in systems theory. Trans.
IEEE Circ. and Syst. CAS. volume 48 (1978)