Heat Transfer Research, 2002, Vol. 33.

7 & 8

ROBUST IDENTIFICATION OF THERMAL

PROCESSES USING ORTHOGONAL FUNCTION BASES

+
K. Bouzrara* , H. Messaoud*
* Ecole Nationale d’Ingénieurs de Monastir (ENIM), Rue Ibn El Jazzar, 5019 Monastir - Tunisia
Tel : + (216) 73 500 511 ; Fax : + (216) 73 500 514 ; email : hassani.messaoud@enim.rnu.tn

+ Author for correspondence

This paper deals with the application of bounded error identification

algorithms for updating the parameter uncertainty domain of a heating and
humidifying processes in a drying blower. The transfer function of each
process is developed on Laguerre function base and the developed algorithm
is used to determine the uncertainty set of the coefficients resulting from this
decomposition.

1. INTRODUCTION :

Parameter-bounding identification algorithms have been developed by many authors [1-4]

and they are considered as an alternative to classical identification methods based on least
squares, maximum likelihood or prediction-error method. In fact, they offer some
advantages such as directness and simplicity, few assumptions about the errors and aptness
for worst-case control design. The advantages become clearer when the model output
contains significant structural error not easily modelled as a random variable, or the available
measurement number is too small for statistical assumptions about the error to be adequately
validated, or the model is to be used for worst-case control design. When the error on the
model-output is bounded the set of parameters is a convex polyhedron (a polytope when
bounded) the element of which are consistent with the assumed model structure, the a priori
error bounds and the measurement. Several methods have been proposed in the literature for
recursively updating the polytope [6, 11] characterised either by its faces or its vertices. The
main drawback of these methods is their computation burden which disables their application
in real time control essentially when the measurement number increases that implies the
simultaneous increase of vertex and face numbers. To overcome this burden the polytope is
approximated by a simple shape region such as a parallelotope [10], an ellipsoid [1], an
orthotope.[4] or a limited complexity polytope [2].

ISSN1064-2285
2002 Begell House, Inc.

1
However all these algorithms deal with model structure such as ARX, output error or state
space models, where the transfer function parameter set has to be updated. In such a case two
major problems may arise. First the estimating of the frequency response with respect to such
structure is very often ill conditioned and its implementation may fail even for high
performance workstation [7, 8]. Second the sensitivity of the frequency response estimate is
proportional to the model order n, therefore minimising the noise induced error requires the
decrease of model order which may fail to approximate the true dynamics of the system.
To overcome these problems another model structure has to be used such that models
developed on orthogonal function bases. This representation consists on fixing poles near
where the poles of the underlying dynamics are believed to lie. Such representation is helpful
for parameter identification since the system output is a linear combination of the
decomposition coefficients.
Moreover the decomposition of a transfer function on an orthogonal base is not sensitive to
the choice of the sampling interval, it does not require a priori information about the system
delay and it results on a smaller number of parameters to estimate. For a model order
increase, the development coefficients remain unchanged. The orthogonal function bases are
mainly the FIR (Finite Impulse Response) base, the Laguerre base, the Kautz base and the
Generalised Orthogonal base. The bounding error identification algorithms update the
feasible set of coefficients resulting from these new developed models.

In this paper we focus on the determination of the exact feasible parameter set of a heating
process and a humidifying process, both in a drying blower, using Laguerre Base
representation. To do so we use bounding error identification techniques for updating the
exact feasible set (polytope) and the outer bounding ellipsoidal approximation of such set.

In the second paragraph we present the Laguerre function base where we define the Laguerre
function, its z-transform, the Laguerre pole and the Laguerre base representation. The third
paragraph will be devoted to define the Laguerre filter network which leads to a state space
representation. Since the state spaces matrices depend on the Laguerre pole we propose an
algorithm for computing the optimal value of this pole. In the fourth paragraph we apply the
bounding error identification algorithms to update the exact feasible set (polytope) and its
ellipsoidal approximation of Laguerre coefficients. In the fifth paragraph the designed
algorithms have been applied for the identification of the feasible set of a heating process and
a humidifying process in a drying blower. The output of each process is compared to
different Laguerre model outputs using the centre of the updated ellipsoid, the maximal and
the minimal values of the coefficient vector resulting from the final ellipsoid and the maximal
and the minimal values of the coefficient vector resulting from the final polytope.

2. DISCRETE LAGUERRE FUNCTION

The Laguerre polynomial has been extensively studied in [3, 9]. It can be obtained by the
following sequences:
j
 ξ 2 −1 
{∆ }=ξ ∑
min(n,k)
1
= C ξ
n n k 2k 2n j j
p k ,ξ C C  2  (1)
ξ 2k  ξ 
n n k
j=0

where ξ 〈 1 is the Laguerre pole, n is the polynomial order, ∆ is defined as ∆ f k = f k+1 - f

k-1 and C nk is the number of combinations with C nk = n ! / k ! (n-k) !

The nth discrete Laguerre function l nk, ξ is expressed as :

2
 ξk
l = 1 −ξ 2
n n
k ,ξ
p k ,ξ (2)
(−ξ)n

or equivalently

j
min(n,k)
 ξ2 −1
l = 1 − ξ ξ ( −ξ ) ∑
n 2 k n j j
C C  2 
(3)
 ξ 
k ,ξ n k
j= 0

The set of Laguerre functions l nk, ξ ( n = 0, 1, 2, ...) constitutes an orthogonal base. In fact,
any pair of these functions are orthogonal, i.e.

∀ n, m ∈ N* l k , ξ , l k , ξ = C nm d nm
n m
(4)
x , y denotes the scalar product of vectors x and y with x , y = xT y and d nm is the
kroneker delta function ( d nm = 1 if n = m and 0 elsewhere ) . If Cnn = 1 for all n = 0, 1, 2,
…, these functions are orthonormal.

The three first Laguerre functions l nk, ξ (n = 0, 1, 2) drawn in Figure 1 are given by:



l 0k , ξ = 1− ξ2 ξ k

 1   ξ 2 −1  
l k , ξ = − 1 − ξ ξ
k +1
2
1 +  2  k  (5)
   ξ  

 2   (ξ 2
)( )
−1 3 ξ 2 + 1   ξ2 −1 k 2 
2

l  1 +   k +  2 
k +2
 k,ξ = 1 − ξ 2
ξ 
   2 ξ4   ξ  2 

0.5

0.4
n=0
0.3

n=1
0.2
n=2
0.1

-0.1

-0.2

-0.3

-0.4
0 20 40 60 80 100 120 140

Figure 1 : The three first Laguerre functions

3
Let’s define L nz, ξ the z-transform of the Laguerre function l nk, ξ . It can computed recursively
as:

n
 1 − ξz  n − 1 1 − ξ 2 z  1 − ξz 
n
=   L z, ξ =   (6)
z − ξ  z − ξ 
L z, ξ
 z −ξ 

The system transfer function G(z) can be written as an infinite linear combination of the z-
transform of Laguerre functions.

∞
G( z ) = ∑ g n L n ( z , ξ) (7)
n =0

or equivalently, by using relation (6),

n
1 − ξ2 z ∞  1 − ξz 
G( z ) = ∑ g n   (8)
z − ξ n = 0  z − ξ 

Where gn (n = 0, …, ∞) are the coefficients resulting from the decomposition of the transfer
function on the Laguerre base. These coefficients can be determined as follows

g n = G ( z), L n ( z, ξ) =
1− ξ
2

G ( z −1 )
(1 − ξz ) dz n

∫ (9)
2π j T
(z − ξ) n +1
In practice the number of coefficients gn is limited to a finite number N (truncating order )
that corresponds to the minimal value for which the approximation describes correctly the
process transfer function and therefore the output of Laguerre Model feets as most as possible
the process output. The value of N depends on the laguerre pole ξ too and an optimal choice
of ξ decreases the value N.

Relation (8) is rewritten as :

n
1− ξ 2 z N  1 − ξz 
G (z ) = ∑ gn   (10)
z − ξ n= 0  z−ξ 

3. LAGUERRE FILTER NETWORK.

With respect to relation (10) the output y(k) and the input u(k) can be connected via the
network sketched in Figure 2.

By considering xi(k) ( i = 0, …, N) as state variables. The network of Figure 2 leads to a

state space representation of the system. In fact, from this network the following equations
hold:

4
 1− ξ2 z 1 − ξz 1 − ξz
u(k)
z− ξ z− ξ z− ξ

x0 (k) x1 (k) xN(k)

g0 g1 gN

∑ y(k)
∑

Figure 2: Discrete-time Laguerre network of order N

 x 0 ( k + 1) = ξ x 0 ( k ) + 1 − ξ 2 u ( k )

 x 1 ( k + 1) = (1 − ξ 2 ) x 0 ( k ) + ξ x 1 ( k) − ξ 1 − ξ 2 u( k )

 x 2 (k + 1) = − ξ ( 1 − ξ ) x 0 ( k) + ( 1 − ξ ) x 1 ( k ) + ξx 2 (k ) + ξ u( k )
2 2 2
(11)
 M

 x N ( k + 1) = ( −ξ) N−1 (1 − ξ 2 ) x 0 (k ) + .... + ( 1 − ξ 2 ) x N −1 (k ) + ξ x N ( k ) + ( −ξ) N u ( k)


The output y(k) can be written as

y ( k) = g 0 x 0 ( k ) + g 1x 1 ( k ) + L + g N x N ( k ) (12)

and the state equation is given by :

X(k + 1) = A X(k ) + B u (k )
 (13)
 y (k ) = C X(k )

with X is the (N+1)-dimensional state vector, A and B are respectively the (N+1)-square
matrix and the (N+1)-dimensional vector. The (N+1)-dimensional vector C contains the
coefficients gi (i = 0, …, N) which correspond to the coefficients of the decomposition of the
original transfer funtion on the Laguerre function base. These vectors and matrix are given
by:

5
  x0 (k)   ξ 0 0 L 0
 x (k)  
 1   1−ξ 2 ξ 0 L 0 
X(k) =  x 2 (k)  ; A =  −ξ (1 − ξ 2 ) 1− ξ 2 ξ 0 ,
   
 M   M M M O M
 x N (k)   (−ξ )N −1 (1 −ξ 2 ) N −2
( −ξ ) (1 − ξ 2 ) L ξ 
 

1 
−ξ 
 
B = 1- ξ ( −ξ ) 2  C = [g 0 L gN ]
2
; g1 g2
 
 M 
 N 
(-ξ ) 

As shown all the elements of the vector B and the matrix A depend only on the Laguerre pole
ξ. The truncating order N depends on the choice of ξ too. This order defines the dimension
of the coefficient vector C to be estimated and a large value of N may complicate the feasible
coefficient set to be updated. Thus an optimal choice of this pole is required.

Optimal choice of Laguerre pole

Any stable LTI system can be modeled by an infinite series of orthonormal functions which
involve a free parameter, closely related to the dominating pole. Theoretically, when
infinitely many parameters are employed in the expansion, the choice of dominating pole is
somewhat arbitrary. In practice, however, a truncated series expansion is used and results in
the truncation error. This error is basically a function of the model order and its dominating
pole. For a fixed model order, there exists an optimal dominating pole that minimizes the
truncation error. Hence, the choice of a dominating pole plays an important role in the quality
of an approximation of the system by Laguerre model. The selection of an optimal
dominating pole is then a crucial problem. If this parameter is chosen appropriately, the
Laguerre model can efficiently approximate a large class of linear system with less number of
estimated parameters.The estimation of the pole of Laguerre filters is based on a separable
non linear least squares methods.

Estimation of Laguerre pole of a heating process

To evaluate the optimal pole, we consider the normalized least-squares error criterion J:

∑ [y m (k) − ŷ(k, C, ξ) ]
M
2

J = 10 log 10 k =1 (14)
∑ [y m (k) ]
M
2

k =1

M is the measurement number, ym(k) is the process output at time instant k and ŷ(k, C, ξ) is
the model output computed for a given pole ? and a given coefficient vector C. This
criterion measures the normalised mean square of the error between the process output and
the laguerre model output.

6
The optimal pole corresponds to the minimal value of this criterion. In figures 3 and 4 we
draw the criterion J for different values of ? respectively for the heating process and the
humidifying process. These figures show that the optimal value of the pole ξ is ξopt = 0.92
for the heating process and ξopt = 0.9 for the humidifying process. The corresponding
truncating orders are respectively N = 4 and N = 4.

Mean quadratic error

0

-2

-4

-6
db

-8

-10

-12

-14
0 10 20 30 40 50 60 70 80 90 100
100 x opt_pole

Figure 3 : Optimal Laguerre pole, Heating Process

Mean
erreur quadraticmoyenne
quadratique error
0

-5
nc

-10

-15
0 10 20 30 40 50 60 70 80 90 100
100100*p opt
x opt_pole

Figure 4 : Optimal Laguerre pole, Humidifying process

7
 4. BOUNDING ERROR IDENTIFICATION ALGORITHM

4.1. Parameter set definition.

Once the optimal pole has been obtained, the matrices A and B are defined and the truncating
order N is determined as well. For each input u(k) the solution of the first equation of the
system (13) delivers the state vector X(k) which is used to compute the process output from
the second equation of the same system. Taking into account the error resulting from the
approximation of an infinite sum (7) by a finite one in (10), the second equation of system
(13) becomes:

y(k) = C X(k) + ε(k) (15)

with ε(k) the equation error at time instant k.

The coefficient vector C can be estimated using classical least squares or else. However no
information about the statistical properties of the error ε(k). A realistic assumption is that ε(k)
is unknown but bounded as:

ε( k ) ≤ δ (16)

The bounding error identification algorithms developed for feasible parameter set updating
can then be applied to update the feasible set of Laguerre coefficients gi, i = 1, ..., N
regrouped in the vector C.

Relations (15) and (16) define at time instant k, the set of parameters C limited by two
hyperplanes Hk1 and Hk2 orthogonal to the vector X(k) and defined as:

Hk1 = { C : C X(k) = y (k ) + δ }
 (17)
Hk2 = { C : C X(k) = y( k ) − δ }
Each hyperplane H kj (j =1, 2) generates two half spaces:

H +k1 = {C : C X(k ) ≤ y( k ) + δ }
 −
H k1 = {C : C X(k ) ≥ y( k ) + δ } (18)
 +
H k 2 = {C : C X( k ) ≥ y( k) − δ }
H −k 2
 = {C : C X( k ) ≤ y( k) − δ }

Any vector C satisfying (16) belongs to H +k1 ∩ H +k2 .

At each time instant t, the smallest set St consistent with the measurement, the model
structure and the error bounds, should satisfy the constraints resulting from all the
observations, such that

{ C : y (k ) − δ ≤ C X( k) ≤ y( k ) + δ , k = 1, ..., t }
t
St = I H+k1 ∩ H+k2 = (19)
k =1

8
St which constitutes the exact feasible set is a convex polytope the complexity of which
increases with the dimension of the parameter vector and the measurement number. In the
next paragraph we present the polytope updating algorithm. However because of the
increasing complexity of this set we present in the paragraph after the algorithm for updating
an outer bounding ellipsoid containing the exact polytope.

4.2. Exact feasible set updating

A convex polytope P is the intersection of a finite number of half spaces. It is a defined by
the set of its vertices and the set of hyperplanes adjacent to each vertex.
{
Let Pk-1 , the polytope at time instant k-1 defined by its vertex set V k−1 = v1k −1 , ..., v xk −1 and}
the set of N hyperplanes adjacent to each vertex. Polytope updating is illustrated by Figure
5. It consists in computing the polytope Pk resulting from the kth measurement acquisition.
Similarly to Pk-1 , the polytope Pk is defined by the set of its vertices Vk = {v1k , ..., vkx' } and
the set of N hyperlanes adjacent to each vertex of Vk . The kth measurement introduces two
hyperplanes Hk1 and Hk2 , each one generates two half spaces H +kj and H −kj (j =1, 2) and the
resulting polytope is Pk = Pk-1 ∩ H+k1 ∩ H+k 2 .

Hk1
Pk
Pk-1

Hk2

Figure 5 : Polytope updating

For each hyperplane Hkj (j =1, 2), the positions of all the vertices of the polytope Pk-1 , vik −1 , i
= 1, ..., x, with respect to this hyperplane are checked by computing the quantity α ij

 y( k ) + δ − XT (k ) v ik −1 if j =1
α =
i
j  i = 1, ..., x (20)
 − y (k ) + δ + X ( k ) v k−1 j= 2
T i
if

if α ij > 0, vik −1 ∈ H +kj , if zero vik −1 ∈ H kj else vik −1 ∈ H −kj . Three cases are possible

- If α ij ≥ 0 ∀ i, all the vertices of Pk-1 are in the positive half space H +kj and then the polytope
Pk-1 is contained in too. The new polytope Pk equals Pk-1 . The hyperplane Hkj is therefore
redundant.

- If α ij < 0 ∀ i, all the vertices of Pk-1 are in the negative half space H −kj and then the polytope
Pk is empty.

9
- If some but not all αij are negative, the positive half space H +kj intersects and modifies Pk-1 .
The polytope is updated as follows: the vertices of Pk-1 in the positive half space H +kj are
maintained, the vertices in the negative half spaces H −kj are rejected and new vertices are
formed at the intersection of the hyperplane Hkj with Pk-1 .

The hyperplane H kj cuts Pk-1 forming a new hyperplane H*k of the polytope Pk . H*k can be
regarded as the convex hull of its vertex set V*. Each vertex of the hyperplane H*k belongs
to Pk but not to Pk-1 .

Let ε( i, l ) the edge of Pk-1 joining the vertices vik −1 and vlk −1 . These vertices are called
adjacent ( vik −1 ↔ vlk −1 ). The edge ε( i, l ) is cut by the hyperplane H kj if its vertices satisfy
vik −1 ∈ H +kj and vlk −1 ∉ H +kj . The vertex v*i = ε( i, l ) ∩ Pk-1 is a convex combination of both
vertices vik −1 and vlk −1 .

v*i = (1- λ ) vik −1 + λ vlk −1 (21)

Since v*i ∈ H kj , it satisfies (17) and the quantity λ is given by:

XT ( k ) [ vlk −1 − v*i ]
λ = l
(22)
X ( k ) [ v k −1 − v k −1 ]
T i

The set V* of new vertices forming the hyperplane H*k is found by applying (21) - (22) to
all pairs of adjacent vertices from Vk-1 which pass the test in (20). The adjacency
( vik −1 ↔ vlk −1 ) of vertices vik −1 ∈ H+kj and vlk −1 ∉ H+kj is replaced by vik −1 ↔ v*i . All the other
(N-1) vertices adjacent to v*i are in V*. Any vertex having (N-1) hyperplanes in common
with v*i is adjacent to it.

Once the set of the new vertices is determined, we have to update the list of hyperplanes
adjacent to each vertex of the updated polytope. This list is known as the vertex- hyperplane
(v- h) list. For the maintained vertices the (v-h) list is unchanged. For the new vertices, first
we note that Hkj belongs to the (v- h) list of each vertex v*i in V* since V* ⊂ Hkj. Also since
v*i ∈ ε (i , l) it is the intersection of (N-1) hyperplanes common to (v-h) list of vik −1 and vlk −1 .
These hyperplanes with Hkj constitutes the N hyperplanes adjacent to v*i . The list of
hyperplanes and the list of vertices define together the polytope at time instant k..

4.3. Outer-bounding ellipsoid updating

An ellipsoid Ek is defined as:

Ek = {θ / [θ − θk ] ∑ −k1
T
[θ − θ k] ≤ σ 2k } (24)

Where θk is the ellipsoid centre and σk ∑k is a positive definite and symmetrical matrix
2

which defines the size and the orientation of the ellipsoid. The ellipsoid updating (Figure 4)
consists of determining at each time instant k the ellipsoid Ek which contains the intersection

10
of the anterior ellipsoid Ek-1 , obtained at time instant (k-1) and defined by its centre θk −1 and
its symmetrical matrix σ2k −1 ∑ k −1 , with the positive half space H +k1 ∩ H +k2 , generated by the
new measurement. This positive half space can be interpreted as a degenerate ellipsoid Fk
given by:

Fk = { θ / ( y(k) − C X(k)) 2
≤ δ2 } (25)

Hk2 Ek-1
E k
Ek
Hk1

Figure 6 : Ellipsoid updating

Three cases are possible depending on the position of hyperplanes Hkj (j =1, 2) with respect
to Ek-1 . This position was computed in [5] using the algebraic measure between Hkj and the
hyperplanes tangent to Ek-1 and parallel to Hkj. These conditions are based on the value of the
prediction error ε k = y( k) − C X( k) .

- If ε k > δ + X T (k ) s 2k −1 ∑ k −1 X ( k) , Hkj (j =1, 2) don’t intersect Ek-1 , the ellipsoid Ek is

then empty and the updating operation is stopped. This situation may result either form the
choice of the bounding error or form the initial ellipsoid.

- If ε k < δ − X T ( k ) s 2k −1 ∑ k−1 X (k ) , The ellipsoid Ek-1 belongs entirely to the positive half
space and the new ellipsoid Ek equals Ek-1 (Ek = Ek-1 ). The measurement yk is then
redundant.

Else hypeplanes Hkj (j =1, 2) intersect Ek-1 which should be updated to yield the ellipsoid Ek .
The updating algorithm is given in Table 1.

Ck = Ck −1 + ∑k X(k) ε k
∑ k −1 X( k ) XT ( k) ∑ k −1
∑ k = ∑ k −1 −
1 + XT ( k ) ∑ k −1 X( k )
2
εk
σ k = σk−1 + δ k −
2 2 2
T
1 + X(k ) ∑ k−1 X( k)

Table 1: Ellipsoid updating algorithm

5. IDENTIFICATION OF PHYSICAL PROCESSES

The heating process and the humidification process are a part of a drying blower which is
composed of a closed loop blower and an electric boiler that provides the blower with steam.
The blower contains a 1.1KW power centrifugal fan with a circular input and a
rectangular output which is converted to square section by a diffuser, a heating system

11
using electric resistors with total power 16 KW, a variable gear transmission, a
tranquillisation room and a trying vein used for experiments. The heat transfer from inside to
outside the blower is realised either by forced convection between the internal air and the
internal wall or by natural convection and by radiation between the ambient air and the
external wall. The boiler contains three resistors, a pressure controller, a level detector, a
safety valve, a water pump and an electromagnetic valve to enable steam injection.

The blower is therefore a bivariable process with two inputs the heating power and the steam
quantity and two outputs the temperature and the humidity. To discard the effect of humidity
on the heating process we dry the blower inside till the humidity is completely eliminated and
we have a SISO system with input the heating power and output the temperature. However
for the humidifying process the effect of the temperature couldn’t be discarded. Therefore the
experiments are done for a given temperature.

The input of the heating process is the heating power sent to the resistors. It is given by
Figure 7 which represents the percentage of the maximal power. Figure 8 draws the
humidifying process input which represents the steam quantity sent to the blower. This
quantit y is proportional to the time of electromagnetic valve opening. For each input an
interface card transforms the amplitude magnitude into a pulse with a given duration. The
pulse corresponding to the heating process temperature controls a power interface for the
specific duration so that the maximal electric electrical power is sent to the resistor.
Similarly, the pulse corresponding to the humidifying process controls the electromagnetic
valve opening for the given duration.

The error bound is assumed to be equal to 3°C for the heating process and 3 % for the
humidifying process. For the polytope updating approach as well as the ellipsoid updating
)
procedure the centre of the final set containing Laguerre coefficients C , the volume of the
final set FV and the uncertainty intervals (UI) of Laguerre coefficients are summarised in
Table 2 for both processes. It is clear that the uncertainty intervals and the final volume are
larger in case of the ellipsoid than those obtained with polytope. This is a predictable result
since the polytope is the exact parameter set.

Uncertainty set Parameters Heating Process Humidifying Process

)
C [ 0.12 0.14 0.036 -0.027 ] [ 0.106 0.035 0.048 -0.016 ]
Ellipsoid
FV 0.2443 0.3108
UI [ 1.69 1.315 1.81 1.18 ] [ 1.637 1.313 1.638 1.334 ]
)
C [ 0.30 0.014 -0.027 -0.054 ] [0.191 -0.026 -0.013 0.014]
Polytope
FV 0.019 0.0414
UI [0.499 0.722 0.683 0.75] [0.595 0.866 0.8 0.859 ]

Table2: Performances of parameter uncertainty sets.

12
 100

90

80

70

60

50

40

30

20

10

0
0 50 100 150 200 250

Figure 7 : Heating power (%)

100

90

80

70

60

50

40

30

20

10

0
0 50 100 150 200 250 300
Figure 8 : Steam Electromagnetic valve Control

In Figure 9 we draw the heating process output, the corresponding La guerre model output
computed using the final ellipsoid centre and the corresponding Laguerre model output
computed using the final polytope centre as a coefficient vector. In Figure 10 we draw the
same outputs drawn in Figure 9 for the humidifying process. It is noted for both processes
that the process output feets correctly the output of the Laguerre model computed using the
ellipsoid centre, however this is not assured when the output is computed using the polytope
centre. This result is predictable too since the ellipsoid centre is considered as the estimate of
the true parameter, however the polytope centre is the chebychef centre which may not
coincide with the real estimate. However the process output still inside the range covered by
the outputs corresponding to the minimal and the maximal values of the estimated coefficient
in the case of polytope updating as well as the ellipsoid updating and for both processes.
For each updating algorithm we evaluate the computing time and the updating rate whic h
corresponds to the number of iterations for which the algorithm is updated without
redundancy.

13
Because of its increasing complexity, the polytope updating algorithm requires a larger
computing time than the ellipsoid updating algorithm. In fact, this time is not constant from
one step to another since the number of vertices of the polytope changes. However this time
is constant for all the steps of the ellipsoid updating procedure.

100

90

80 3

70

60

50

40

30
1
20

10
2
0
0 50 100 150 200 250

Figure 9 : Comparison of different outputs of the heating Process

50

45

40

35

30

25

20
3
15
1
10

5
2
0
0 50 100 150 200 250 300

Figure 10 : Comparison of different outputs of the humidifying Process

14
 6. CONCLUSION

The bounding error identification algorithms have been applied to update the exact feasible
set and its outer bounding ellipsoid of the coefficients resulting form the decomposition of a
process transfer-function on the orthogonal Laguerre function base. The developed
algorithms have been applied to estimate the uncertainty parameter sets of a heating process
and a humidification process both in a drying blower. The determined uncertainty sets will be
used to design robust controllers for temperature and humidity regulation.

References
[1] L.V.R Arruda & G. Favier (1991), "A review and comparison of robust estimation
methods", IFAC Sy. On Identification and System Parameter Estimation, Budapest Hungaria,
pp. 1027-1032
[2] S. Maraoui & H. Messaoud (2001), " Design and comparative study of limited complexity
bounding error identification algorithms", IFAC Symposium On System Structure and
Control, Prague -Cheque Republique, 29-31 August 2001.
[3] R. Malti (1999), "Représentation de systèmes sur la base des filters orthogonaux.
Application à la modélisation de systèmes dynamiques multivariables", PhD thesis, Institut
National Polytechnique de Lorraine, Lorraine, France.
[4] H. Messaoud and G. Favier (1994), " Recursive determination of parameter uncertainty
intervals for linear models with unknown but bounded errors",10th IFAC Symposium on
System Identification, Copenhagen - Danemark, 4-6 July, pp. 365-370.
[5] H. Messaoud (1998), " Dead zone computing in Bounding Ellipsoïd algorithms",
CESA’98 IMACS Multiconferences, Hammamet-Tunisia, 1-4 April pp. 480 - 485.
[6] S. H. M0 and J.P. Norton (1990), "Fast and robust algorithm to compute exact polytope
parameter bounds", Mathematics and Computers in Simulation, Vol. 32, pp. 481-493.
[7] B. Ninness, H. Hjalmarsson and F. Gustafsson (1997), " The fundamental role of General
Orthonormal bases in system identification ", Technical Report EE9737, Centre for
Integrated Dynamics and Control (CIDAC), Department of Electrical and Computer
Engineering, University of Newcastle, Australia.
[8] B. Ninness (1998), " The utility of orthonormal bases", Technical Report EE9802,
Department of Electrical and Computer Engineering, University of Newcastle, Australia.
[9] N. Tangy (1994), " La transformation de Laguerre discrete ", Thèse de Doctorat de
l’Université de Bretagne Occidentale, Brest - France.
[10] A. Vicino and G. Zappa (1996), " Sequential approximation of feasible parameter sets
for identification with set membership uncertainty", IEEE TAC, Vol. 41, No. 6, pp. 774 -
784.
[11] E. Walter and H. Piet- Lahanier (1989), "Exact Recursive Polyhedral description of the
feasible Parameter Set for Bounded-Error Models", IEEE, Transaction on Automatic Control,
Vol. 34, N°. 8, pp. 911 – 914.

15