Control Engineering Practice 29 (2014) 5060

Contents lists available at ScienceDirect

Control Engineering Practice

journal homepage: www.elsevier.com/locate/conengprac

Nonlinear thermal system identication using fractional

Volterra series$
Asma Maachou a, Rachid Malti a,n, Pierre Melchior a, Jean-Luc Battaglia b, Alain Oustaloup a,
Bruno Hay c
a
Universit de Bordeaux, Institut Polytechnique de Bordeaux, IMS UMR 5218 CNRS, 351 cours de la Libration, Bt A4, 33405 Talence Cedex, France
b
Universit de Bordeaux, I2M UMR 5295 CNRS, Esplanade des Arts et Mtiers, 33405 Talence Cedex, France
c
LNE, Laboratoire National de mtrologie et d'Essais, 29 avenue Roger Hennequin, 78197 Trappes Cedex, France

art ic l e i nf o a b s t r a c t

Article history: Linear fractional differentiation models have already proven their efcacy in modeling thermal diffusive
Received 27 September 2013 phenomena for small temperature variations involving constant thermal parameters such as thermal
Accepted 28 February 2014 diffusivity and thermal conductivity. However, for large temperature variations, encountered in plasma torch
Available online 6 May 2014
or in machining in severe conditions, the thermal parameters are no longer constant, but vary along with the
Keywords: temperature. In such a context, thermal diffusive phenomena can no longer be modeled by linear fractional
System identication models. In this paper, a new class of nonlinear fractional models based on the Volterra series is proposed for
Fractional model modeling such nonlinear diffusive phenomena. More specically, Volterra series are extended to fractional
Fractional calculus derivatives, and fractional orthogonal generating functions are used as Volterra kernels. The linear
Nonlinear model
coefcients are estimated along with nonlinear fractional parameters of the Volterra kernels by nonlinear
Volterra series
programming techniques. The fractional Volterra series are rst used to identify thermal diffusion in an iron
Large temperature variation
sample with data generated using the nite element method and temperature variations up to 700 K. For
that purpose, the thermal properties of the iron sample have been characterized. Then, the fractional
Volterra series are used to identify the thermal diffusion with experimental data obtained by injecting a heat
ux generated by a 200 W laser beam in the iron sample with temperature variations of 150 K. It is shown
that the identied model is always more accurate than the nite element model because it allows, in a single
experiment, to take into account system uncertainties.
& 2014 Elsevier Ltd. All rights reserved.

1. Introduction Simulating a nite element model requires knowing all the

This work is a part of a project which aims at developing a new
device for heat ux estimation in tools during severe machining or
in high enthalpy plasma ow. This estimation is based on
temperature measurements near the heated area and on the
inversion of a model which links the measured temperature to
the heat ux. This paper focuses on the formulation of a reliable
model consistent with large temperature variations leading to a
nonlinear behavior due to temperature dependent thermal prop-
erties of the studied system.
A rst preliminary paper was presented at the 4th IFAC Workshop on Fractional
Differentiation and Its Applications, Badajoz, Spain, 2011 (Maachou et al., 2010),
and a second one at the 18th World IFAC Congress, Milan, Italy, 2011 (Maachou et
al., 2011). Here is a revisited and enhanced version of the two preliminary papers.
n
Corresponding author.
E-mail addresses: asma.maachou@ims-bordeaux.fr (A. Maachou),
rachid.malti@ims-bordeaux.fr (R. Malti),
pierre.melchior@ims-bordeaux.fr (P. Melchior),
jean-luc.battaglia@bordeaux.ensam.fr (J.-L. Battaglia), bruno.hay@lne.fr (B. Hay).
thermal characteristics (specic heat, thermal conductivity, den-
sity) of the studied thermal system, and hence requires many prior
experiments for thermal characterization. In spite of this knowl-
edge, a nite element model is generally less accurate than an
identied model, because it is difcult to take into account the
following:
(i) the exact thermocouple location due to assemblage accuracy,
(ii) the exact geometry of the contact surface between the sensor
and the sample,
(iii) boundary between the thermocouple and the sample which is
never ideal because the thermocouple is usually glued to
the material and there might be an air gap between both
components,
(iv) the nonlinear dynamical behavior of the sensor.
The identied model includes all the system uncertainties men-
tioned above and may be obtained in a single experiment. It is thus
expected to be more accurate.

http://dx.doi.org/10.1016/j.conengprac.2014.02.023
0967-0661/& 2014 Elsevier Ltd. All rights reserved.
 A. Maachou et al. / Control Engineering Practice 29 (2014) 5060
For short time experiments, thermal systems are regarded as
semi-innite media. Therefore, a fractional model appears natu-
rally as an exact solution to express the temperature at a given
point in the system versus the heat ux (Battaglia, 2008; Oldham
& Spanier, 1972). Different system identication methods based on
fractional linear models, developed in the literature, are summar-
ized in the state of the art (Malti, Victor, & Oustaloup, 2008). For
instance, Cois, Oustaloup, Battaglia, and Battaglia (2000) extended
the output error to fractional models by using a modal representa-
tion. Cois, Oustaloup, Poinot, and Battaglia (2001) proposed to use
a prediction error method combined to instrumental variable and
state variable lters. This method was enhanced in Victor, Malti,
Garnier, and Oustaloup (2013) by choosing optimal instrumental
variables. Battaglia and Kusiak (2005) implemented recursive least
squares for online parameters estimation. Malti, Aoun, Levron, and
Oustaloup (2005), Aoun, Malti, Levron, and Oustaloup (2007), and
Akay (2008) used fractional orthogonal basis for system approx-
imation. Malti, Rassi, Thomassin, and Khemane (2010) and
Khemane, Malti, Rassi, and Moreau (2012) proposed a set-
membership approach for system identication using fractional
models using frequency-domain data. Fractional multi-models
have been proposed for modeling gastrocnemius frog muscle in
Sommacal et al. (2007, 2008).
For signicant temperature variations, linear models are not
accurate enough (Maachou et al., 2012; Gabano, Poinot, & Kanoun,
2010). Recently, Gabano et al. (2010) proposed the identication of
an ARMCO (American Rolling Mill COmpany) iron sample using a
fractional linear parameter varying (LPV) model. In a direct
relationship with the results of this paper, a theoretical study
has recently been carried out for determining analytical expres-
sions of the Volterra series in Battaglia, Maachou, Malti, and
Melchior (2013). In this paper, Volterra series are extended to
fractional derivatives with kernels generated using fractional
orthogonal generating functions. Two reasons motivate this
choice. First, nonlinear model decomposition into Volterra series
allows us to separate the contribution of the linear and the
nonlinear parts. Secondly, the Volterra series may be seen as a
generalization of linear models. Volterra series have already been
used, in the literature, for the approximation and control of eddy
current brake in Simeu and Georges (1996) or recently for modeling
a greenhouse temperature in Gruber et al. (2011). The Hammerstein
model based on Volterra series helped Angerer, Hintz, and Schrder
(2004) identifying a nonlinear mechatronic system. Bibes (2004)
modeled the process of water supplies using Volterra series devel-
oped on orthonormal basis functions. Casenave (2011) used implicit
Volterra model form for the time-local formulation. This paper deals
with the extension of Volterra series to fractional derivatives and its
application in thermal modeling.
Heat transfer in a semi-innite medium is modeled in Section 2.
It is shown how linear fractional differentiation models are
obtained by solving the heat equation under some initial and
boundary conditions. Then, the limits of linear models are pointed
out for large temperature variations which justies the use of
nonlinear fractional models. Then, in Section 3, a mathematical
background is presented regarding fractional calculus and the
Volterra series are presented as a generalization of linear models.
In Section 4, a new class of models is presented as fractional
Volterra series. Fractional generating orthogonal functions are
used as Volterra kernels. In a system identication context,
parameter estimation is developed in Section 5. Either linear
coefcients of the Volterra series are estimated by least squares,
or the linear coefcients are estimated along with nonlinear
fractional parameters of the Volterra kernels by nonlinear pro-
gramming techniques. In Section 6, the fractional Volterra series
are used to model temperature variations versus heat ux for a
monovariable system using two sets of data. This system is
51
constituted of a simple thick iron disk, with one face subject to a
uniform heat ux and with a thermocouple embedded close to the
heated area. The rst set of data is generated using the nite
element method with the temperature dependent thermal proper-
ties of the iron sample characterized, for the aim of the simulation,
by the Laboratoire National de mtrologie et d'Essais (LNE), the
French National Metrology Institute (NMI). The second set of data
is obtained with an experimental set-up on the same thermal
system. The fractional Volterra series are used to model tempera-
ture variations for the two sets of data. Moreover, the measured
output, the nite element model output and the Volterra model
output are compared to each other. It is shown that the identied
Volterra model is more accurate than the nite element model
because it allows, in a single experiment, to take into account
system uncertainties (i)(iv) described in Section 1. Finally in the
conclusion, the main results of the paper are recalled.
2. Fractional calculus for thermal modeling: from small to
large temperature variations
Fractional calculus has witnessed a growing interest in various
elds as described in the old and the recent history of fractional
calculus by Machado, Kiryakova, and Mainardi (2010, 2011).
On the basis of the heat equation characterizing the heat
transfer in a semi-innite medium (Fig. 1), Battaglia, Le Lay,
Batsale, Oustaloup, and Cois (2000) showed that a fractional
model is pertinent in modeling heat transfer at a distance x from
the heated surface. The diffusion phenomenon is governed by the
linear heat diffusion equation
Tx; t 2 Tx; t
for 0 o x o 1; t 4 0 1
t x2
where is the thermal diffusivity. For convenience, isothermal and
null states are chosen
Tx; t 0; for 0 r x o 1; at t 0 2
The boundary conditions are
8
<  Tx; t t for x 0; at t 4 0
x 3
: Tx; t 0 for x-1; at t 40:
where is the thermal conductivity of a material and t is the
heat ux applied on its surface. Knowing that
; 4
C p
semi-infinite
medium
Heat flux
x
Sensor T(t)
Fig. 1. Heat ux applied on a surface S of a semi-nite medium.
52 A. Maachou et al. / Control Engineering Practice 29 (2014) 5060

with the density and Cp the specic heat, the expression of the frequency of 31.4 rad s  1 for a sampling period of T s 0:1 s.
temperature in the Laplace domain is Hence, at short distance x a Pad approximation of order 1 or
 r 2 of the theoretical model (5) is accurate enough.
1 s
x; s pp exp  x s; 5 The obtained fractional models remain linear as long as the
s C p
thermal parameters are constant, i.e. for small temperature varia-
with LT and L, which clearly shows that the tem- tions around the initial state. However, for high temperature
perature is linked to the ux though a transfer function which variations, the thermal parameters vary along with the tempera-
p ture and the linearity assumption is no longer valid. The Labor-
involves a half order integral (1= s) and an exponential term. The
closer to the heated front end, the smaller x, and the more atoire National de mtrologie et d'Essais (LNE) has characterized
negligible the exponential term. Evaluating the Pth order Pad the thermophysical properties of the ARMCO iron sample utilized
approximation of the exponential term yields in the real experimental set-up. As shown in Fig. 3, both thermal
conductivity and specic heat vary along with the temperature.
x; s 1 P an sn=2 p 2P  n! Moreover, around the Curie point1 at 1040 K, the thermal varia-
p P n 0 n 1=2 with an  x= n
s C p n 0 jan js n!P  n! tions are more pronounced, although the material remains solid.
6 Hence, for large temperature variations, linearity hypothesis is no
longer veried and a nonlinear model is required.
where the exponents of s are multiples of the commensurate order
0.5. The quality of the approximation can be assessed by compar-
ing the frequency response of the theoretical model (5) with its 3. Mathematical background: fractional models and Volterra
approximation (6). As shown in Malti, Sabatier, and Akay (2009), series
the order of the Pad approximation, P, depends on the distance x
and the Nyquist frequency (and hence the sampling period) till 3.1. Fractional linear models
which the approximation needs to be valid. For instance, when a
sensor is placed at x 2 mm from a heat surface of an ARMCO A fractional model is generally based on a fractional differential
semi-innite medium, the Pad approximations of order 0, 1, and equation
2 are respectively given by
yt a1 D1 yt aN DN yt
1
H~ 0 s pp ; 7 b0 D0 ut b1 D1 ut bM DM ut; 10
s C p S
where u(t) is the system input, y(t) is the noise-free system output,
xn aj ; bi A R2 , D d=dt is the differential operator, and the differ-
 p s0:5 2
1 entiation orders
H~ 1 s pp n ; 8
s C p S x 0:5 1 o 2 o oN ; 0 r 0 o 1 o o M ; 11
p s 2

allowed to be non-integer positive numbers, are ordered for
xn2 6xn identiability purposes. The fractional derivative D of a contin-
s  p s0:5 12 uous function f is considered in the sense of Grnwald-Letnikov
1
H~ 2 s pp n2 : 9 (Grnwald, 1867; Miller & Ross, 1993; Podlubny, 1999):
s C p S x 6xn 0:5
s p s 12  
1 K
D f Kh lim 1k f t  kh; 12
h-0 h k 0 k
For second order approximation, with 75.3 W m  1 K  1 the
thermal conductivity, Cp 470 W kg  1 K  1 the specic heat, where k stands for the binomial coefcients. When the function f
7970 kg m  3 the density and 20.09  10  6 m2 s  1 the dif- is discretized at a sampling period h, then the limit in (12) is
fusivity of ARMCO at 300 K. Fig. 2 shows Bode diagrams of these dropped and the fractional derivative is approximated by
approximations and the theoretical model (6) up to the Nyquist  
1 K
D f Kh  1k f K  kh Oh: 13
h k0 k
40

60 The Laplace transform of D when x(t) is relaxed at t0

(xt 0 for all t o 0), is given by Oldham and Spanier (1974):
Gain (db)

80 Theorical model

100
0 order approximation LfD xtg s Xs: 14
1st order approximation

120
2nd order approximation
This result is coherent with the classical case when is an integer.
140
Consequently, it is easy to dene a symbolic representation of the
4 3 2 1 0 1 2
fractional dynamic system governed by (10) using the transfer
10 10 10 10 10 10 10
function:
pulsation (rad/s)
i
Bs Mi 0 bi s
10 Gs : 15
30
As 1 j 1 aj sj
N
phase (degree)

60 Moreover, if G(s) is commensurate of order , then G(s) can be

90 rewritten as
120
m ~ i
i 0 bi s
150 Gs ; 16
1 j 1 a~j sj
n
180
210
4 3 2 1 0 1 2
10 10 10 10 10 10 10
1
In physics and materials science, the Curie temperature (Tc), or Curie point, is
Fig. 2. Physical model x; s= s in (5) and its Pad approximations H~ 0 s, H~ 1 s the temperature at which a ferromagnetic or a ferrimagnetic material becomes
and H~ 2 s. The vertical line corresponds to the Nyquist frequency. paramagnetic.
 A. Maachou et al. / Control Engineering Practice 29 (2014) 5060 53

1200 120
Curie temperature
1100 110

1000 100

Specific heat Cp (J.Kg1.K 1 )

Thermal conductivity
900 Temperature range for the finite elements modeling 90

(W.m1.K 1 )
800 80

700 70

600 Temperature 60
range for the
500 experimental setup 50

400 40

300 30

200 20
300 400 500 600 700 800 900 1000 1100
Temperature (K)

Fig. 3. Thermal characteristics of ARMCO iron sample for different temperature levels, each square corresponding to a measurement point.

where m M = and n N = are integers and 8 i0 A f0; 1; ; mg, For a nonlinear system, a model structure of Volterra series is
8 j0 A f1; ; ng written as the innite sum
8
> ~ 0 1
> bi0 bi if ( i A f0; 1; ; Mg such that i i
>
> yt yk t y1 t y2 t yk t ; 19
>
< b~ 0 0 otherwise k1
i
17
> a~j0 aj if ( j A f1; ; Ng such that j0 j
>
>
> where yk, the output of the kth Volterra kernel, writes as the
>
: a~j0 0 otherwise: generalization of the convolution product (18) to the kth order:
Z 1 Z 1
In rational transfer functions, equals 1 and usually numerator N k
yk t hk 1 ; ; k ut  i di : 20
and denominator M orders are both xed, then all coefcients 1 1 i1
bi ; i 0; ; M, and aj ; j 1; ; N, are estimated. Generally, no care
is taken to check whether any intermediate coefcient, as in (17), For example, y2, the output of the quadratic kernel, is
equals zero. Z 1 Z 1
Time-domain simulation of fractional systems is an extensively y2 t h2 1 ; 2 ut  1 ut  2 d1 d2 : 21
1 1
studied topic in the literature (Aoun, Malti, Levron, & Oustaloup,
2004; Cai & Liu, 2007 and all references therein). The simulation Therefore, by combining Eqs. (19) and (20), the output y(t) of a
algorithm used in this paper is based on the discrete Grnwald nonlinear system may be expressed as
Letnikov denition (13) of fractional derivatives. It has the advan- Z Z
1 1 1 k
tage of allowing the quantication of the simulation error while yt hk 1 ; ; k ut  i di ; 22
computing the derivative (the error is proportional to the sam- k1 1 1 i1
pling period Oh). Its major drawback is that the simulation
requires for each step the computation of sums of increasing where hk is the kth order Volterra kernel and its Laplace transform
dimension with time. Other simulation methods exist, such as the H k s1 ; ; sk is given by (Crum & Heinen, 1974)
one based on recursive poles and zeros of a fractional derivative/ Z 1 Z 1
integral as proposed by Oustaloup (1995). However, the time- H k s1 ; ; sk hk 1 ; ; k e  s1 1  sk k d1 dk : 23
1 1
domain simulation error in the latter method is more difcult to
quantify. All in all, system identication algorithms proposed in Moreover, the Laplace transform of the convolution integral (20)
this paper may be used with any time-domain simulation algo- can be written as
rithm of fractional systems, at the discretion of the user. Fractional
Y k s1 ; ; sk H k s1 ; ; sk Us1 Usk : 24
systems should however be correctly simulated with negligible
simulation errors, in order to be able to consistently estimate the Therefore, the Laplace transform of the output signal y(t), in (19),
fractional model parameters. is

Ys1 ; ; sk ; Y 1 s1 Y 2 s1 ; s2 Y k s1 ; ; sk ; 25
3.2. From linear models to nonlinear models using Volterra series
where
Volterra (1959) was the rst to introduce Volterra series; see 8
Rugh (1981) for the mathematical aspect and Helie and Roze >
> Y 1 s1 H 1 s1 Us1
>
>
(2008) for an application to a nonlinear acoustic system. Volterra < Y 2 s1 ; s2 H 2 s1 ; s2 Us1 Us2
series are relevant for slightly nonlinear time invariant systems 26
>
>
>
>
(Bard, 2005) and systems with fading memories (Boyd & Chua, : Y s1 ; ; s H s1 ; ; s Us1 Us :
k k k k k
1985) can be expressed in the time-domain as well as in the
frequency-domain. Obviously, the rst order kernel H 1 s1 corresponds to the linear
Given a linear system characterized by its impulse response part of the system. All other kernels contain system nonlinearities.
h1 t and an input signal u(t), then the output y1 t is expressed as For more simplicity, all kernels are supposed to be symmetric:
the convolution product:
Z 1 hk 1 ; 2 ; ; k hk i1 ; i2 ; ; ik ; 27
y1 t h1 ut  d: 18
1 where i1 ; i2 ; ; ik is any permutation of 1; 2; ; k.
54 A. Maachou et al. / Control Engineering Practice 29 (2014) 5060

4. Volterra series with fractional differentiation kernels with the generating functions fn dened by their Laplace transform
Fn as
As the behavior of a system modeled by Volterra series is 8
> s Re 1
described by its kernels hk in (22), an expression for the kernels hk >
> F 1 s
>
> s 1 s 2
is rst chosen and then all coefcients are estimated. Each kernel >
>
< I m 2
hk may be decomposed on fractional orthogonal generating func- F 2 s 34
>
> s 1 s 2
tions Fn. >
>
>
> 1
In Aoun et al. (2007) and Malti et al. (2005), orthonormal base >
: F 3 s F 2 s:
functions are extended to fractional differentiation orders by s 3
applying a Gram Schmidt orthogonalization procedure on a set
As for monovariable functions, the multivariable functions f m1 ;;mk
of generating functions Fn. As a consequence their inverse Laplace
transforms fn form a basis in L2 0; 1 (Aoun et al., 2007; Akay, k
f m1 ;;mk f mi ; 35
2008). Hence, the linear part of the Volterra series h1 t A L2 0; 1 i1
can always be approximated using functions fn
form a basis in L2 0; 1k . Hence, the Volterra kernels can always
M be written as
h1 t  bn f n t; 28
n1 1 1
hk t 1 ; ; t k bm1 ;;mi f m1 ;;mk t 1 ; ; t k ; 36
m1 1 mk 1
where bn, n 1; ; M, are the coefcients associated with the
decomposition of h1 on fn, and M the truncation order. For with bm1 ;;mi the coefcients associated with the decomposition,
example, the fractional Laguerre basis is characterized by the provided they belong to the L2 0; 1k space. Then, following the
presence of a single s pole at  (Aoun et al., 2007) with the developments in Bibes (2004) for rational orthonormal basis, yk in
Laplace transform of the generating functions given by (20) can be approximated by

1 M mk  1
F n s ; 29 yk t cm1 ;;mk I m1 I mk 37
s n m1 1 mk 1

with the fractional order. For thermal systems, Eq. (6) allows to where
set to 0.5. Z t
When system s poles are real and/or complex conjugate, I mk f mk k ut  k dk ; 38
fractional generalized orthonormal basis (Malti et al., 2005) may 0

be used. As compared to (29), the generalized basis allows us to M is the number of fractional orthogonal generating functions
choose any combination of poles. Their generating functions are used to approximate the kernels, and mk is the kernel order. The
iteratively dened kth order kernel is presented in Fig. 4.
for n 4 1: For instance, the quadratic kernel output developed using three
8 generating functions, i.e. M 3 and k 2, may be written as
> 1
>
> F s if n is real or
> n
>
> s n n  1 3 m1
>
<
y2 t cm1 ;m2 I m1 I m2 39
s Re n m1 1 m2 1
F 0n s and 30
>
> s n s n n  1
>
> Z
>
> I m n t
>
: F n s if n is complex with I m1 f m1 1 ut  1 d1 40
s n s n n  1
0

with Z t
and I m2 f m2 2 ut  2 d2 41
( 0
n  1 F n  1 if n 41 and n  1 is real or
31
n  1 F 0n  1 or F n  1 if n 41 and n  1 is complex

and for n 1:
8
> 1
>
> F s if 1 is real or
> 1
> s 1
>
>
<
s Re 1
F 01 s and 32
>
> s n s 1
>
>
>
> I m 1
>
: F 1 s if 1 is complex:
s 1 s 1

Fractional Laguerre basis is a particular case of the fractional

generalized orthonormal basis with all the s poles chosen to be
alike and real. For example, a linear system characterized by three
s poles of the fractional generalized orthogonal basis (GOB) at
 1  2 complex conjugate and  3 real can be written
using three generating functions:

3
h1 t bn f n t; 33
n1 Fig. 4. kth order Volterra kernel developed using M generating functions.
 A. Maachou et al. / Control Engineering Practice 29 (2014) 5060 55

or in an expanded form Then, the number Pc of parameters of an Nth order Volterra

1 2 3 model is
y2 t c1;m2 I 1 I m2 c2;m2 I 2 I m2 c3;m2 I 3 I m2 42
m2 1 m2 1 m2 1
N N M k  1!
Pc Pk ; 52
k1 k1 M  1!k!
y2 t c1;1 I 1 I 1 c2;1 I 2 I 1 c2;2 I 2 I 2 c3;1 I 3 I 1 c3;2 I 3 I 2 c3;3 I 3 I 3 43
Consequently, the second order Volterra series model output y  for the vector , the number of unknown nonlinear parameters
(t), developed on three fractional generating functions, is con- 1 ; ; M and , corresponding to the M s poles associated
stituted of the rst order kernel output y1 t (linear part) and the with generating functions and the commensurate order, is
second order kernel output y2 t:
P M 1: 53
yt y1 t y2 t with

8
> 3 Obviously, the number of parameters increases considerably with
>
>
> y1 t cm1 xm1
> the truncation order M of the generating functions (30) and with
>
> m1 1
>
> the truncation order N of Volterra kernels. For example, if the
>
< where xm L  1 fF m s1 Us1 g
1 1
44 number of generating functions, M, and the number of Volterra
>
>
3 m1
kernels, N, are both set to 3, then the number of linear coefcients
>
> y2 t cm1 ;m2 xm1 xm2
>
> equals 19, according to (52) and the number of nonlinear para-
>
> m1 1m2 1
>
>
: where xm L  1 fF m s2 Us2 g meters equals 4 (3 s poles and a commensurate order), which
2 2
yields a total number of 23 parameters. In case of using 3 generat-
where F1, F2 and F3 are dened in (34). Eq. (44) can also be ing functions and 2 Volterra kernels the total number of para-
formulated in a vector form meters reduces to 9 linear coefcients and 4 nonlinear
yt XC; 45 parameters:
T
with C c1 c3 c1;1 c3;3  and X xm1 ; ; xm3 ; xm1 xm1 ; ; 1 2 3 T : 54
xm3 xm3 . According to prior knowledge, a linear or nonlinear program-
ming method can be used for parameter estimation. If the
5. Parameter estimation fractional order and the poles are chosen according to some prior
knowledge, then only the coefcients cm1 mk are estimated by
Consider input u(t) and output yn t data collected at regular least squares. Otherwise, nonlinear programming (NLP) is
time-instants from t0 to t T final . The measured output yn t is required. Optimal parameters are computed by minimizing the
supposed to be corrupted by an additive measurement noise t quadratic norm of the output error

yn t yt t; 46
1 n
Kk ^
0 yid kh  y id kh
2
J id  100 55
 1 yn kh2
Kk
where y(t) is the hypothetical noise-free deterministic output. The 0 id

input u(t) and the output y(t) are supposed to be related by the with ynid the noisy system output and y^ id the estimated Volterra
fractional Volterra series (22). model output (the subscript id stands for identication data). For
Given the model comparison purposes the following validation criterion is com-
M mi  1
puted on validation data ynval and y^ val :
yt cm1 ;;mk L  1 fF m1 s1 Us1 gL  1 fF mk sk Usk g; 1 n ^ 2
m1 1 mk 1
Kk 0 yval kh  y val kh
J val  100: 56
 1 yn kh2
Kk
47 0 val

In case only the linear coefcients C in (49) are estimated by least

with F mi as in (30), the parameter vector squares, then
" #
C b X T X  1 X T Y;
C 57
48
b is estimated by
and the covariance matrix associated with C
bC s
P b 2 X T X  1 ; 58
is composed of a vector C of linear coefcients
where s
2
b is an empirical estimation of the noise variance.
C c1 ; ; cM ; c1;1 ; ; cM;M ; c1;;1 ; ; cM;;M T 49
In case the linear coefcients C in (49) and the nonlinear
and additionally a vector of nonlinear parameters: fractional parameters in (50) are estimated, then a nonlinear
programming (NLP) method is used. In such a case, the NLP
1 ; ; M ; T : 50
method is combined to linear programming as sketched in Fig. 5.
The number of unknown parameters depends on the truncation The rst step consists in initializing in (54). The initialization
order M of the orthogonal generating functions and the number N of and the s poles ( 1 ; 2 ; ; M ) can be done by using a rough
of Volterra kernels. Hence, estimation of a linear fractional Output Error (OE) model (Malti et
al., 2008; Victor et al., 2013). As far as thermal systems are
 for the vector C, the number Pk of unknown linear coefcients considered a good initial hit for the commensurate order is
cm1 ;;mi associated with the kth order kernel equals k-combina- 0:5 as illustrated using Pad approximations in Section 2. A
tions with repetitions among M generating functions: linear OE model can always be written using the generalized
orthogonal functions:
M k  1! Ys M
Pk : 51 Gs bm F m s 59
M  1!k! Us m 1
56 A. Maachou et al. / Control Engineering Practice 29 (2014) 5060

As in the linear models, the number of generating functions M is An estimation of the covariance matrix may be computed from
chosen when the identication criterion does not improve sig- the approximated Hessian with
nicantly and/or the validation criterion increases. Some more
sophisticated criteria can also be used (e.g. Akaike, 1974; Young, V s
b 2 J  1 ; 65
where s
2
1989). Moreover, one should keep in mind that the number of b is an estimation of the noise variance obtained from the
coefcients of the Volterra series increases signicantly with M as residual error.
indicated in (52). Secondly, the coefcient vector C0 in (49) is
computed by least square. Then, the vector 0 is built:
6. Results and discussion
0 C 0 ; 0  60

In step three, the vector from (49) is estimated using nonlinear 6.1. Identication of thermal diffusion in the ARMCO sample from
programming. The LevenbergMarquardt method, implemented simulated data
in the MATLAB function lsqnonlin of the optimization toolbox, is
used and the parameter vector k is computed iteratively by An ARMCO iron sample is considered having the shape of a
thick disc (radius R 10 mm and thickness e 19.37 mm). One face
1 0
k 1 k  fJ I J g k 61 of the disc is submitted to a time varying heat ux, uniformly
distributed on a disc of radius r 0 1:25 mm at the surface.
with
8 A thermocouple of type K (diameter 0.5 mm) is embedded at a
K 1
>
> 0 distance d 2 mm from the heated surface (see Fig. 6). The
> J  2 t k St k ;
>
>
> k0 simulated sensor works in a linear domain. Hence the nonlinea-
>
>
>
> K 1 rities are due to the temperature dependent thermal properties of
<
J  2 St k ; ST t k ;
62 the material plotted in Fig. 3. Temperature of the sensor is
>
>
k0
>
> yt k ; simulated using a nite element method in 3D conguration with
>
> ;
> k
> St quadratic Lagrange elements using the Comsol multiphysics soft-
>
>
: ware. Therefore data are generated by solving the heat equation
nonnegative damping factor;
numerically. The sampling time is set to h 0.01 s.
where is the residual error Two input signals are generated, one for system identication
and the other one for model validation, with a magnitude of 0 to 1 
id t ynid t  y^ id t: 63
108 W m  2 and a total duration of 25 s as shown in Figs. 7 and 8.
The third step is iterated till convergence when To test the identication algorithm, a white noise t of 20 dB
  is added to the Comsol noise-free output y(t):
 l;k  1 
max l;k  o;
 64 yn t yt t: 66
l l;k

where l;k stands for the lth element of the vector estimated at To help setting up the truncation order and for comparison
the kth iteration, or till the maximum number of iterations, maxiter, purposes, the system is rst identied using a linear model. As far
is reached. As for all gradient based algorithms, only convergence
to a local minimum is guaranteed.

19.37
40

1.25
10

Fig. 6. ARMCO sample pattern. All dimensions are in mm.

Input Signal
x 107
Heat flux (W/m)

10

0
0 5 10 15 20 25

Output signal
800
variations (K)
Temperature

600
400
200
0
0 5 10 15 20 25
Time (s)

Fig. 5. Parameters estimation method. Fig. 7. Identication data (ambient temperature of 300 K removed).
 A. Maachou et al. / Control Engineering Practice 29 (2014) 5060 57

as the physical model (5) contains a pure integrator of order 0.5, it where ynval and y^ val t stand respectively for simulated output and
is preferable to integrate the input signal the estimated model output on validation data is plotted in Fig. 10.
Moreover, contribution of each kernel of the Volterra series is
Us
U~ s 0:5 ; 67 highlighted which allows us to check the contribution of the linear
s
part and the contribution of the nonlinear part of the model. The
and to identify a model between the output and the integrated linear model is also plotted for comparison purposes.
input. All coefcients of the linear model As expected, the Volterra model ts better system output as
Ys M compared to the linear model. This is further conrmed by the FIT
Gs bm F m s 68 criterion used on identication and validation data:
Us m 1
 
are estimated for different values of M (number of poles) and the b 2
yn  y
FIT 100 1  : 71
quadratic norms of the output error, computed on identication yn meanyn 2
and validation data, are reported in Table 1. A truncation at M 2
appears to be a good choice, because for M Z 2 the identication As shown in Table 3, this criterion rises from 74.1% to 84.7% on
and validation criteria in Table 1 do not improve signicantly. validation data conrming that the Volterra model is more
Therefore taking a greater number of poles will not improve accurate than the linear model. Moreover, the estimated para-
signicantly the accuracy of the linear model used in the initi- meters of the linear model are biased due to the presence of
alization of the NLP algorithm. nonlinearities which are better captured in the second order
The two estimated s poles in the linear model and the Volterra series model. This rst numerical experiment allows us
commensurate order 0:5 are used to initialize in (54): to conclude on the reliability of the method for identifying
accurately the nonlinear system.
0  0:96  1:91i;  0:96 1:91i; 0:5T : 69

A Monte Carlo simulation of 300 runs with different realiza- Volterra model output
700
tions of noise with a SNR of 20 dB is carried out. The vector in first kernel output
quadratic kernel output
(48) converges after optimization to the values in Table 2. Temperature variations (K) 600 system output
linear model output
System and Volterra model outputs are plotted on validation 500
data in Fig. 9 and the difference between each model output and
400
the system output val t
300
val t ynval t  y^ val t; 70
200

100
x 107 Input Signal
Heat flux (W/m)

10
0
0 5 10 15 20 25
5
Time (s)
0
0 5 10 15 20 25 Fig. 9. Comparison between system output, linear model output, Volterra series
model output and its rst two kernels on validation data. Ambient temperature of
Output signal
300 K removed.
variations (K)

800
Temperature

600
400 80
200
0 60
0 5 10 15 20 25
40
Time (s)
20
Fig. 8. Validation data (ambient temperature of 300 K removed).
0

20

Table 1 40
Identication and validation criteria for a linear model versus the number of poles.
60 Difference between Volterra model output and system output
Difference between linear model output and system output
M 1 2 3 4 80
0 5 10 15 20 25
Jid 1.73 0.40 0.40 0.40 Time (s)
Jval 1.94 0.74 0.74 0.72
Fig. 10. Difference between the system and the model outputs.

Table 2
Estimated parameters of the Volterra series model, 1 2 .

c1 c2 c1;1 c1;2

6 6  10
Estimation 4.28  10  4.08  10  0.68  10 3.7  10  10
Standard deviation 0.71  10  6 0.19  10  6 0.41  10  10 4.6  10  10
c2,2 Re 1 Im 1

Estimation  0.07  10  10  0.96  1.18 0.35

Standard deviation 0.11  10  10 0.05 0.23 0.03
58 A. Maachou et al. / Control Engineering Practice 29 (2014) 5060

6.2. Identication of thermal diffusion in the ARMCO sample from an x 107 Input signal
4

Heat flux (W/m)

experimental set-up
3
2
An experimental set-up has been realized for the system
1
described in the previous section and presented in Fig. 12. A heat
ux generated, by a 200 W laser beam, with a wavelength of 0
0 5 10 15 20 25
960 nm and controlled by an internal function generator modu-
Output signal
lated with a maximum frequency of 10 kHz, is applied at the 150

variations (K)
surface of the ARMCO iron sample. A thermocouple is located

Temperature
100
inside that sample, through a previously drilled hole as in Fig. 11,
at a distance d 2 mm from the heated surface. The excited surface 50
S is a 1.25 mm radius disc. The aim is to provide enough energy to
0
induce large temperature variations. This experimental set-up 0 5 10 15 20 25
allows temperature variations up to 150 K. As the thermal char- Time (s)
acteristic variations are more important around the Curie point
Fig. 13. Identication data (initial condition temperature removed).
(1040 K), the ARMCO sample is heated to 940 K by an electrical
heater rolled around it.
The power of the laser beam (varying from 0 to 200 W) is x 107 Input signal
4

Heat flux (W/m)

controlled proportionally by a voltage generator varying from 0 to
3
5 V. The resulting heat ux density, considered as the model input,
2
is plotted in Figs. 13 and 14 for identication and validation data,
1
0
0 5 10 15 20 25
Table 3
Output signal
Identication and validation criteria according to the model for simulated data. variations (K) 200
Temperature

150
Model Linear (%) 2nd order Volterra (%)
100
FITid 80.4 86.5 50
FITval 74.1 84.7
0
0 5 10 15 20 25
Time (s)

Fig. 14. Validation data (initial condition temperature removed).

Table 4
Identication and validation criteria for a linear model versus the number of
generating functions.

M 1 2 3 4 5

Jid 5.06 4.49 0.03 0.02 0.02

Jval 5.68 5.03 0.07 0.06 0.06

together with the measured temperature of the sensor. As for

Fig. 11. Thermocouple attached to the iron sample.
Fig. 12. Test bench at the TREFLE laboratory (University of Bordeaux).
simulated data, the input signal is rst integrated and the quad-
ratic norms of the output error, computed on identication and
validation data, are reported in Table 4 which shows that a
truncation at M 3 is a good choice. The three estimated poles
and the commensurate order 0:5 are used to initialize 0 in
(54) for the nonlinear model:
0 1 ; 2 ; 3 ;   0:12  1:39i;  0:15 1:39i; 17:35; 0:5T : 72
However, the number of Volterra kernels is set to N 2. The vector
of (48) converges after optimization to the values in Table 5.
System and Volterra model outputs are plotted on validation
data in Fig. 15. Moreover, contribution of each kernel of the
Volterra series is highlighted which allows us to differentiate the
contribution of the linear part and the contribution of the non-
linear part of the model. The linear model output is also plotted for
comparison purposes (Fig. 16).
The Volterra model ts better system output as compared to
the linear model. As shown in Table 6, this criterion rises from
88.8% to 93.0% on validation data conrming that the Volterra
model is more accurate than the linear model. This rst experi-
ment allows us to conclude on the reliability of the method for
identifying accurately the nonlinear system.
 A. Maachou et al. / Control Engineering Practice 29 (2014) 5060 59

Table 5 160
Estimated parameters of the Volterra series model, 1 2 .
140

Temperature variations (K)

c1 c2 c3 120

Estimation  1.13  10  5 16.92  10  5  1.49  10  5 100

Standard deviation 0.31  10  5 3.05  10  5 1.19  10  5 80
c1;1 c1;2 c1;3 60

Estimation 2.94  10  8 2.42  10  8 6.76  10  8 40 Volterra model output

system output
Standard deviation 0.92  10  8 2.07  10  8 2.73  10  8 Finite element model output
20
c2;2 c2;3 c3;3
0
0 5 10 15 20 25
8 8
Estimation  21.73  10 9.83  10 8.93  10  8
Time (s)
Standard deviation 8.16  10  8 6.24  10  8 3.28  10  8

Re 1 Im 1 3 Fig. 17. Comparison between the experimental set-up output, the Volterra model
output and the nite element model output for validation data.
Estimation 0.11  1.41 15.10 0.49
Standard deviation 0.08 0.01 2.89 0.02
20
Difference between the finite element model output and the system output
Difference between the Volterra model output and the system output

15

150
10
Temperature variations (K)

5
100
0

50 5
Volterra model output
Linear model output
System output 10
0 5 10 15 20 25
0 Time (s)
0 5 10 15 20 25
Fig. 18. Errors between the experimental set-up output and each model output on
Time (s)
validation data.
Fig. 15. Comparison between system output, linear model output, Volterra series
model output and its rst two kernels on validation data. Static temperature of
the differences between the experimental set-up output and each
about 950 K was subtracted.
model output are plotted in Fig. 18 on validation data. Both gures
show that the Volterra model output is closer to the experimental
10
Difference between Volterra model output and system output
set-up output. The FIT criterion for validation data conrms these
8 Difference between linear model output and system output results as it equals 93% for the Volterra model output against only
6
74% for the nite element model output. The Volterra model
exhibits a much better performance as compared to the nite
4
element model because
2

0  The measurement point is set on the axis of symmetry of the

2 sample, to 2 mm from the excited surface in the nite element
model. However, in the experimental set-up, the thermocouple
4
position is inaccurate because of the drilling and assemblage
6
accuracy.
8  The boundary between the thermocouple and the sample is
0 5 10 15 20 25
considered as ideal. However, in the experimental set-up, the
Time (s)
thermocouple is glued to the iron sample and there might be
Fig. 16. Difference between the system output and the model outputs. an air gap between both components. Thermal characteristics
of the air and the glue are not taken into account in the nite
element model.
Table 6
 The ARMCO iron emissivity is set to 0.6 in the nite element
Identication and validation quadratic criteria according to the model. model. However, in the experimental set-up, the ARMCO iron
emissivity depends on different parameters such as the oxida-
Model Linear (%) 2nd order Volterra (%)
tion or the polishing of the sample and is known to be between
FITid 90.9 95.4
0.5 and 0.9.
FITval 88.8 93.0  The thermocouple specic heat and thermal conductivity are
set to their values at ambient temperature. However, in the
experimental set-up, the thermocouple specic heat and con-
The nite element model is now compared to the experimental ductivity variations are unknown in the temperature variation
set-up model. Both input signals of Figs. 13 and 14 are applied to range [940,1140] K.
the nite element model and the Volterra model. The outputs of  The iron sample external boundaries are insulated and set to
both models are compared to the experimental one in Fig. 17 and 940 K. However, in the experimental set-up, the temperature
60 A. Maachou et al. / Control Engineering Practice 29 (2014) 5060
initial value is reached using a conductor thread, rolled around
the iron sample, injecting a heat ux and provoking slight
temperature growth during the experiment.
Thus, on one hand, at the opposite of nite element models
which need the characterization of thermal properties versus
temperature and the exact geometric pattern of the system, the
identication algorithm gives with only one experimental set-up a
very close to reality model and, on the other hand, the Volterra
series allows a more accurate modeling as compared to a linear
model, because it allows us to take into account the nonlinearities
brought by the large temperature variations.
7. Conclusion
Linear fractional models have already proven their efcacy in
modeling thermal diffusive phenomena for small temperature
variations. However for large temperature variations, thermal
diffusive phenomena are no longer linear. Fractional nonlinear
models based on Volterra series are developed in this paper with
Volterra kernels synthesized using orthogonal generating func-
tions. They are then applied to model heat diffusion in an ARMCO
iron sample based on experimental and simulation data. Linear
and nonlinear parameters of Volterra series are estimated. It was
shown that the fractional Volterra series slightly improve model
quality in the real experiment set-up when temperature variations
are low (  150 K) and signicantly improves the quality of the
approximation when temperature variations are high. Moreover,
Volterra series allows us to separate the linear part (rst order
kernel) from the nonlinear part (other kernels). Then, modeling a
thermal system using fractional Volterra series yields an appro-
priate model with only one experimental set-up whereas nite
element model yields less accurate results with more experiments
for the characterization of thermal properties versus temperature
and the exact geometric pattern of the system.
References
Akaike, H. (1974). A new look at the statistical model identication. IEEE Transac-
tions on Automatic Control, 19, 716723.
Akay, H. (2008). Synthesis of complete orthonormal fractional basis functions with
prescribed poles. IEEE Transactions on Signal Processing, 56(10), 47164728.
Angerer, B. T., Hintz, C., & Schrder, D. (2004). Online identication of a nonlinear
mechatronic system. Control Engineering Practice, 12, 14651478.
Aoun, M., Malti, R., Levron, F., & Oustaloup, A. (2004). Numerical simulations of
fractional systems: An overview of existing methods and improvements. Nonlinear
Dynamics, 38(14), 117131.
Aoun, M., Malti, R., Levron, F., & Oustaloup, A. (2007). Synthesis of fractional
Laguerre basis for system approximation. Automatica, 43(9), 16401648.
Bard, D. (2005). Compensation des non linarits des systmes haut-parleurs
pavillon (Ph.D. thesis). Ecole polytechnique fdrale de Lausanne.
Battaglia, J.-L. (2008). Heat transfer in materials forming processes. Great Britain:
Wiley Chippenham.
Battaglia, J.-L., Maachou, A., Malti, R., & Melchior, P. (2013). Nonlinear heat diffusion
simulation using Volterra series expansion. International Journal of Thermal
Sciences, 71, 8087.
Battaglia, J.-L., & Kusiak, A. (2005). Estimation of heat uxes during high-speed
drilling. International Journal of Advanced Manufacturing Technology, 39,
750758.
Battaglia, J.-L., Le Lay, L., Batsale, J.-C., Oustaloup, A., & Cois, O. (2000). Heat ux
estimation through inverted non integer identication models. International
Journal of Thermal Sciences, 39(3), 374389.
Bibes, G. (2004). Modlisation de procds de traitements des eaux et reconstruc-
tion de grandeurs physico-chimiques (Ph.D. thesis). Universit de Poitiers.
Boyd, S., & Chua, L. O. (1985). Fading memory and the problem of approximating
nonlinear operator with Volterra series. IEEE Transactions on Circuits and
Systems, 32(11), 11501161.
Cai, X., & Liu, F. (2007). Numerical simulation of the fractional-order control system.
Journal of Applied Mathematics and Computing, 23, 229241.
Casenave, C. (2011). Time-local formulation and identication of implicit Volterra
models by use of diffusive representation. Automatica, 47(10), 22732278.
Cois, O., Oustaloup, A., Battaglia, E., & Battaglia, J.-L. (2000). Non integer model from
modal decomposition for time domain system identication. In 12th IFAC
symposium on system identication, SYSID, Santa Barbara, USA.
Cois, O., Oustaloup, A., Poinot, T., & Battaglia, J.-L. (2001). Fractional state variable
lter for system identication by fractional model. In IEEE 6th European control
conference (ECC2001), Porto, Portugal.
Crum, L. A., & Heinen, J. A. (1974). Simultaneous reduction and expansion of
multidimensional Laplace transform kernels. SIAM Journal on Applied Mathe-
matics, 26(4), 753771.
Gabano, J.-D., Poinot, T., & Kanoun, H. (2010). Linear parameter-varying fractional
modeling of a thermal system. In 4rd IFAC workshop on fractional differentiation
and its applications (FDA).
Gruber, J. K., Guzman, J. L., Rodriguez, F., Bordons, C., Berenguel, M., & Sanchez, J. A.
(2011). Nonlinear MPC based on a Volterra series model for greenhouse
temperature control using natural ventilation. Control Engineering Practice, 19,
354366.
Grnwald, A. K. (1867). Ueber begrenzte derivationen und deren anwendung.
Zeitschrift fr Mathematik und Physik, 12, 441480.
Helie, T., & Roze, D. (2008). Sound synthesis of a nonlinear string using Volterra
series. Journal of Sound and Vibration, 314, 275306.
Khemane, F., Malti, R., Rassi, T., & Moreau, X. (2012). Robust estimation of fractional
models in the frequency domain using set membership methods. Signal
Processing, 92, 15911601.
Maachou, A., Malti, R., Melchior, P., Battaglia, J.-L., Oustaloup, A., & Hay, B. (2010).
Thermal system identication for large temperature variations using fractional
Volterra series. In 4th IFAC workshop on fractional differentiation and its
applications (FDA), Badajoz, Spain.
Maachou, A., Malti, R., Melchior, P., Battaglia, J.-L., Oustaloup, A., & Hay, B. (2011).
Application of fractional Volterra series for the identication of thermal
diffusion in an ARMCO iron sample subject to large temperature variations.
In 18th IFAC world congress (pp. 56215626), Milan, Italy.
Maachou, A., Malti, R., Melchior, P., Battaglia, J.-L., Oustaloup, A., & Hay, B. (2012).
Thermal system identication using fractional models for high temperature
levels around different operating points. Nonlinear Dynamics, 70(2), 941950.
Machado, J. T., Kiryakova, V., & Mainardi, F. (2010). A poster about the old history of
fractional calculus. Fractional Calculus and Applied Analysis, 13(4), 447454.
Machado, J. T., Kiryakova, V., & Mainardi, F. (2011). Recent history of fractional
calculus. Communications in Nonlinear Science and Numerical Simulation, 16(3),
11401153.
Malti, R., Aoun, M., Levron, F., & Oustaloup, A. (2005). Unied construction of
fractional generalized orthogonal bases. In Fractional differentiation and its
applications. Mathematical tools, geometrical and physical aspects (Vol. 1, pp.
87102). Germany: U-Books.
Malti, R., Rassi, T., Thomassin, M., & Khemane, F. (2010). Set membership
parameter estimation of fractional models based on bounded frequency
domain data. Communications in Nonlinear Science and Numerical Simulation,
15(4), 927938.
Malti, R., Sabatier, J., & Akay, H. (2009). Thermal modeling and identication of an
aluminium rod using fractional calculus. In 15th IFAC symposium on system
identication (SYSID) (pp. 958963), Saint-Malo France.
Malti, R., Victor, S., & Oustaloup, A. (2008). Advances in system identication using
fractional models. Journal of Computational and Nonlinear Dynamics, 3(2),
021401,1021401,7.
Miller, K. S., & Ross, B. (1993). An introduction to the fractional calculus and fractional
differential equations. New York, USA: A Wiley-Interscience Publication.
Oldham, K. B., & Spanier, J. (1972). A general solution of the diffusive equation for
semiinnite geometries. Journal of Mathematical Analysis and Applications, 39,
655669.
Oldham, K. B., & Spanier, J. (1974). The fractional calculustheory and applications of
differentiation and integration to arbitrary order. New York, London: Academic
Press.
Oustaloup, A. (1995). La derivation non-entire. Paris: Herms.
Podlubny, I. (1999). Fractional differential equations. San Diego: Academic Press.
Rugh, W. J. (1981). Non linear theory: The Volterra/Wiener approach. Baltimore, MD,
USA: Johns Hopkins University Press.
Simeu, E., & Georges, D. (1996). Modeling and control of an eddy current brake.
Control Engineering Practice, 4(1), 1926.
Sommacal, L., Melchior, P., Cabelguen, J.-M., Oustaloup, A., & Ijspeert, J. (2008).
Fractional multi-models of the gastrocnemius frog muscle. Journal of Vibration
and Control, 14(910), 14151430.
Sommacal, L., Melchior, P., Dossat, A., Petit, J., Cabelguen, J.-M., Oustaloup, A., et al.
(2007). Improvement of the muscle fractional multimodel for low-rate stimu-
lation. Biomedical Signal Processing and Control, 2(3), 226233.
Victor, S., Malti, R., Garnier, H., & Oustaloup, A. (2013). Parameter and differentia-
tion order estimation in fractional models. Automatica, 49(4), 926935.
Volterra, V. (1959). Theory of functionals and of integral and integro-differential
equations. New York: Dover.
Young, P. C. (1989). Recursive estimation, forecasting and adaptative control. In
Control and dynamics systems: Advances in theory and applications (Vol. 30,
pp. 119166). San Diego: Academic Press.