Mathematical Biosciences 216 (2008) 154–162

Contents lists available at ScienceDirect

Mathematical Biosciences
journal homepage: www.elsevier.com/locate/mbs

Structural identifiability of a model for the acetic acid fermentation process

Jorge E. Jiménez-Hornero a,*, Inés M. Santos-Dueñas b, Isidoro Garcı́a-Garcı́a b
a
Computing and Numerical Analysis Department, University of Córdoba, P.O. Box 14071, Córdoba, Spain
b
Chemical Engineering Department, University of Córdoba, P.O. Box 14071, Córdoba, Spain

a r t i c l e i n f o a b s t r a c t

Article history: Modelling has proved an essential tool for addressing research into biotechnological processes, particu-
Received 26 February 2008 larly with a view to their optimization and control. Parameter estimation via optimization approaches is
Received in revised form 15 September among the major steps in the development of biotechnology models. In fact, one of the first tasks in the
2008
development process is to determine whether the parameters concerned can be unambiguously deter-
Accepted 19 September 2008
Available online 27 September 2008
mined and provide meaningful physical conclusions as a result. The analysis process is known as ‘iden-
tifiability’ and presents two different aspects: structural or theoretical identifiability and practical
identifiability. While structural identifiability is concerned with model structure alone, practical identi-
Keywords:
Modelling
fiability takes into account both the quantity and quality of experimental data. In this work, we discuss
Structural identifiability the theoretical identifiability of a new model for the acetic acid fermentation process and review existing
Parameter estimation methods for this purpose.
Acetic acid fermentation Ó 2008 Elsevier Inc. All rights reserved.
Vinegar

1. Introduction yield coefficients in generic wastewater treatment models [10] and

Parameter identification – or parameter estimation from exper-
imental data – is one of the critical steps in developing models
[1–3]. Some biotechnological models are especially difficult to
solve because of the high-order non-linear multi-parametric equa-
tion systems and some experimental data being subject to consid-
erable errors. This entails conducting a preliminary analysis in
order to ascertain whether the specific parameters to be estimated
can be uniquely determined by using a given optimization algo-
rithm. Such an analysis must include the model structure, and also
the quality and quantity of experimental data available. This study
is known as identifiability and presents two different aspects:
(a) Structural or theoretical identifiability, which is exclusively
concerned with the complexity of the model structure for a
given input–output behaviour [4].
(b) Practical identifiability, which considers the quality and
quantity of experimental data in addition to model structure.
This paper discusses the structural identifiability of a new mod-
el for the acetic acid fermentation process. Identifiability in bio-
technological processes has been the subject of much literature.
Such processes include wastewater treatments [5–8] and denitrifi-
cation [9], which have been examined for structural identifiability.
Some authors have also examined the theoretical identifiability of
* Corresponding author.
E-mail address: jjimenez@uco.es (J.E. Jiménez-Hornero).
the structural identifiability of a model for an activated sludge bio-
reactor [11]. Others have analysed a known model for endogenous
respiration of sludge in a batch reactor under no dissolved oxygen
constraints [12]. Still others have examined several models for aer-
obic and anaerobic processes used in wastewater treatment [13].
Structural identifiability in a microbial growth model has also been
studied [14]. Also, a reparametrization technique proposed else-
where [15] was applied to two models of a high biological and
pharmacokinetic interest; the models were also dealt with in
another study [16], but using differential algebra, and the same
approach was used for immunological and epidemiological models
elsewhere [17]. Some authors have conducted theoretical identifi-
ability studies on robotic models for movement control [18], as
well as on mechanical systems in general [19]. A method for trans-
forming chromatographic columns models intended to avoid over-
parametrization was proposed in [20] and the generating series
method, which is described below, used to check for structural
identifiability in the resulting model. Finally, key issues regarding
the use of various statistical tools in modelling procedures were
discussed in [21] and a new approach to verifying model structural
identifiability involving semi-infinite programming and max–min
problems was reported.
1.1. Structural identifiability
The concept of structural identifiability has been under investi-
gation ever since it was first defined [22]. Formally, structural iden-
tifiability relies on the following definitions [23]:

0025-5564/$ - see front matter Ó 2008 Elsevier Inc. All rights reserved.
doi:10.1016/j.mbs.2008.09.004
 J.E. Jiménez-Hornero et al. / Mathematical Biosciences 216 (2008) 154–162
– Definition 1: parameter hi of a model is structurally globally
identifiable for an input class U if and only if for almost all
h 2 P (P being the parametric space),
) ...
^h 2 P
) ^hi ¼ hi
^ð^h; tÞ ¼ y
y ^ðh; tÞ; 8t > 0; 8u 2 U
– Definition 2: parameter hi of a model is structurally locally
identifiable for an input class U if and only if for almost all
h 2 P a neighbourhood V(h) exists such that
) other hand, if a countable set of solutions exists, then the model
^h 2 VðhÞ  P
) ^hi ¼ hi
y^ð^h; tÞ ¼ y
^ðh; tÞ; 8t > 0; 8u 2 U
– Definition 3: model M(h) is structurally globally identifiable if
and only if all parameters hi are structurally globally identi-
fiable. Therefore, this definition also holds with local
analyses.
While conceptually relevant, these definitions are scarcely use-
ful for studying structural identifiability in practice. Also, identifi-
ability can be examined in various ways. Thus, theoretical
identifiability in linear models can be checked with several reliable
tests including Laplace transforms, Taylor series expansion of
observations, the Markov parameter matrix approach and methods
determining the ranks of matrices A and B from the state–space
representation of the model [23–25]. Confirming identifiability in
non-linear models is much more complex and relatively few meth-
ods exist for verifying their theoretical identifiability. Also, such
methods usually require software tools affording symbolic compu-
tation [26] and substantial computational resources.
1.2. Review of existing methods
1.2.1. Analytical methods
One of the earliest methods for structural identifiability analysis
was based on series expansions of the input–output map of the
model. For a specific input function or a zero-input experiment,
this involves Taylor series expansion [27]. This approach assumes
a state–space model for the process of the following form:
dx
¼ f ðx; u; hÞ; xð0Þ ¼ x0 ðhÞ
dt
yðt; hÞ ¼ hðx; hÞ
where x 2 Rn is the state vector, h 2 X (X being an open subset of
Rq ) a constant parameter vector, u 2 U the input vector and y 2 Rr
the output vector. This formulation assumes that, for each h 2 X,
a connected open subset of Rn [denoted by M(h) and containing
x0(h)] exists such that f(, u, h) and h(, h) are analytical functions on
M(h) and x 2 M(h) for any t P 0.
Outputs y(t) are expanded as power series and derivatives are
evaluated at a specific time (usually t = 0) in order to obtain math-
ematical expressions as simple as possible:
dy
2
t2 d y
yðtÞ ¼ yð0Þ þ t ð0Þ þ ð0Þ þ   
dt 2! dt2
The derivatives in Eq. (4) can be calculated as model parameter
functions from Eq. (3):
yð0Þ ¼ c0 ðhÞ
dy
ð0Þ ¼ c1 ðhÞ
dt
...
If one assumes all derivatives in (5) to be known, then the next step
involves solving the model parameters as functions of the deriva-
tives and inputs:
155
 
h1 ¼ b1 yð0Þ; dy ð0Þ; . . . ; u
dt
 
h2 ¼ b2 yð0Þ; dy
dt
ð0Þ; . . . ; u ð6Þ
 
ð1Þ hq ¼ bq yð0Þ; dy ð0Þ; . . . ; u
dt
Analysing the solutions of (6) allows one to draw revealing conclu-
sions as regards structural identifiability. Thus, if a single solution
exists, then the model is theoretically globally identifiable. On the
is locally identifiable. Finally, with an uncountable set of solutions
ð2Þ
the model is structurally unidentifiable [23,27–29].
Despite its apparent simplicity, this method is subject to two
major problems, namely:
(a) Initially, the number of derivatives used for computation
should be at least equal to that of parameters of the model so that
equation system (6) can be solved. However, if some equations
are not independent, the number of derivatives must be even
greater; therefore, no upper limit for the number of power series
expansion coefficients of the outputs necessarily exists [30]. Also,
finding no solution does not necessarily mean that the model is
truly unidentifiable since additional independent equations
may exist in the system. An upper limit has indeed been shown
to exist for linear [29], bilinear [31] and polynomial systems
[32,33].
(b) Provided equation system (6) can be solved, working out
every parameter value can be excessively labour-intensive,
even with the aid of dedicated software for symbolic computa-
tion [8,13,23,28] such as Matlab (Mathworks, Inc.), Maple
(Maplesoft, Inc.) or Mathematica (Wolfram Research, Inc.).
Alternatively, one can test for local identifiability by using
derivatives calculated in (5) [34]. The test is based on the jacobian
matrix of the derivatives with respect to the parameters. If the rank
of this matrix is less than the number of parameters, then the mod-
el is unidentifiable.
The generating series approach is one other way of solving the
problem that shares a number of characteristics with the previous
one. This approach is based on non-linear control theory and uses
the relationship between Lie derivatives and non-linear observabil-
ð3Þ
ity [5]. The method requires the model to be linear regarding the
inputs u(t) and allows one to extend the analysis to the entire class
of bounded and measurable input functions (e.g. the set of piece-
wise continuous functions [35]):
dx Xm
¼ f0 ðx; hÞ þ ui ðtÞfi ðx; hÞ; xð0Þ ¼ x0 ðhÞyðt; hÞ ¼ hðx; hÞ ð7Þ
dt i¼1
where x is the state vector, u = {u1, u2, . . . , um} the input vector, h the
parameter vector and y the output vector. The analysis relies on the
model output functions h(x, h) and their successive Lie derivatives,
Lfj0 . . . Lfjk hðx; hÞ, as evaluated at a specific time where simple enough
mathematical expressions can obtained (usually t = 0). The Lie
derivative along a vector field fi is equivalent to
ð4Þ
X
n
o
Lfi ¼ fj;i ðx; hÞ ð8Þ
j¼1
oxj
where fj,i is the j-th component of fi.
As in the power series expansion method, all Lie derivatives at a
ð5Þ given time are assumed to be known to calculate the solution set
for the parameters. Again, any conclusions about structural identi-
fiability will rely on the number of solutions obtained [29,36]. The
mathematical expressions obtained with the generating series
method are usually simpler than those provided by previous meth-
od [37].
156 J.E. Jiménez-Hornero et al. / Mathematical Biosciences 216 (2008) 154–162
Like the generating series method, the method based on the lo-
cal state isomorphism theorem, also known as the ‘similarity trans-
formation’ method, assumes that the entire class of bounded and
measurable functions are available [30,38]. Also, the model should
be locally reduced at x0(h) for almost any h. This implies that the
system must fulfil the Controllability Rank Condition (CRC) and
Observability Rank Condition (ORC) [39,40] at x0(h). These condi-
tions determine controllability and observability in non-linear sys-
tems and can be considered an extension of those for linear
systems, which were established by Kalman [41]. If these condi-
tions are verified, the local state isomorphism theorem establishes
that, if two states x and ~
x corresponding to two different parameter
sets h and ~h, respectively, are considered, then the corresponding
models will have the same input/output behaviour for any input
u(t) and t > 0 if and only if local state isomorphism
k : V ! Rn ; x!~ x ¼ kð
xÞ [where V is a neighbourhood of  xð0Þ] ex-
ists which fulfils the following conditions:
 
ok 
rank ¼n ð9Þ
oxT x¼x
kðxð0ÞÞ ¼ ~xð0Þ ð10Þ
 braic Equations) solvers to analyse structural local identifiability.
ok 
f ðkðxÞ; ~hÞ ¼ T   f ðkðxÞ; hÞ ð11Þ
ox x¼x
 preset values for the parameters and the other considers them as
ok 
gðkðxÞ; ~hÞ ¼ T   gðkðxÞ; hÞ ð12Þ
ox x¼x
hðkðxÞ; hÞ ¼ hðx; hÞ
~ ð13Þ
 and k are obtained and,
If all solutions of the system (9)–(13) for h
for almost any ~ h, the only possible solution is 
h¼~h; kðxÞ ¼ ~
x, then
the model is structurally globally identifiable [30,42].
Overall, this method is only applicable to models of low com-
plexity. Otherwise, the mathematical expressions required to ver-
ify conditions (9)–(13) are rather complex.
Differential algebra has also been used in this context.
This method reformulates model equations as linear regressions
regarding parameters by using traditional algebraic operations
jointly with differentiation [32,43,44]. The method is based on
the Ritt–Wu algorithm [45,46], which is only applicable to polyno-
mial or rational models. Despite these restrictions, many models in
biotechnology and other areas fulfil this condition. Additional
restrictions require that the models have non-zero, differentiable
inputs. The application of this method to structural identifiability
analysis is described elsewhere [44]; also, a software implementa-
tion was developed in [47] and an improved version can be found
in [48].
Notwithstanding its elegant formulation, experience has shown
that differential algebra methods not only converge on the solution
very slowly, but also fail with fairly complex models [49]. Also,
they share some difficulties with the previous methods as regards
mathematical handling of even low-order models [44].
Finally, some methods use a combination of differential algebra
and either generating series or power series expansion methodol-
ogy [16].
A different set of approaches uses reparametrization methods
[15,23,29,50–53] to transform the original models into structurally
identifiable forms. However, these methods are not systematic and
the physical meaning of the ensuing parameters can be lost
through reparametrization. Although some authors have tried to
reduce heuristics at some stages of the process, some intuition
and experience are still needed for proper implementation.
Some reported methods are only applicable under special cir-
cumstances. Thus, some necessary and sufficient conditions for
verifying structural identifiability of uncontrolled systems with
or without restrictions have been reported in [14] and [48,54–
56], respectively. Finally, two methods for solving structural iden-
tifiability of high-order linear systems regarding states and non-
linear regarding parameters are described in [57].
1.2.2. Numerical methods
In general, applying analytical methods to practical problems is
not easy. This has fostered the development of numerical methods
to assess theoretical local identifiability and, in some cases, global
identifiability. Thus, a method for parameter estimation by use of a
global optimization algorithm based on interval analysis was pro-
posed in [58]. The method works on data previously generated by a
known model around a specific point or region of the parametric
space. This procedure allows checking whether the system has
only one possible solution and is therefore globally identifiable.
Its greatest drawback is the long computing time needed, which
is only acceptable for models with up to three dimensions. This se-
verely restricts its use with the models frequently encountered in
practice. An improved alternative was developed in [59], which
uses a forward–backward contractor to reduce the number of
bisections on interval analysis. This allows the global algorithm
to be more efficient and to deal with more complex problems.
One other numerical approach [60] uses DAE (Differential-Alge-
Models are solved in two different ways; one involves the use of
unknowns. Because the number of equations will usually differ
from that of parameters, the system is expanded adding as many
equations as necessary taking successive derivatives with respect
to time. If the system has a unique, constant solution, then it can
be deemed structurally locally identifiable; otherwise (e.g. if it
has a non-constant solution or more than one solution), then either
the model is unidentifiable or unsuitable derivatives were chosen.
Finally, one other numerical method for structural identifiabil-
ity analysis, which relies on similarities between this concept
and observability, is described in detail [61]. It considers parame-
ters as constant state variables with null derivatives and examines
their local observability together with the remainder ‘real’ state
variables. A probabilistic semi-numerical algorithm is used for
computing, within a reasonable time, locally observable variables
and the number of non-observable variables. The algorithm certi-
fies that any variables deemed observable actually are. A method
for transforming non-observable models into observable models
with an identical input/output behaviour was also developed; the
raised models are known as symmetric models.
2. Proposed model
Basically, acetic acid fermentation is an aerobic biological reac-
tion effected by acetic acid bacteria; these cause partial oxidation
of the medium, mainly by converting ethanol into acetic acid:
C2 H5 OH þ O2 ! CH3 COOH þ H2 O; DH ¼ 493 kJ ð14Þ
Vinegar production relies heavily on this stage, bacterial activity
on which is markedly dependent on the particular environmental
conditions.
The proposed model is based on previously reported models [62–
73], and also on extensive research work on process optimization.
If one assumes a thoroughly mixed liquid phase and isothermal
conditions in a semi-continuous operation mode, then the follow-
ing mass balances can be written:
dX v dV
V þ Xv ¼ Vðr Xc  r Xd Þ ð15Þ
dt dt
dX d dV
V þ Xd ¼ Vðr Xd  r lisis Þ ð16Þ
dt dt
dE dV
V þE ¼ F i  E0  V  r E ð17Þ
dt dt
 J.E. Jiménez-Hornero et al. / Mathematical Biosciences 216 (2008) 154–162 157


4
dA dV E
V þA ¼ V  rA ð18Þ fdE ¼ 1 þ ð28Þ
dt dt K mE
dO dV  4
V þO ¼ F i  O0 þ V½bðO0  OÞ  rO  ð19Þ A
dt dt fdA ¼1þ ð29Þ
K mA
dV
¼ Fi ð20Þ
dt where:
where: ld specific cell death rate (h1).
Xv 1
viable cell concentration (g L ). l0d minimum specific cell death rate (h1).
Xd non-viable cell concentration (g L1). KmE ethanol cell death constant (g ethanol L1)
E ethanol concentration (g L1). KmA acetic acid cell death constant (g acetic acid L1).
A acetic acid concentration (g L1).
O dissolved oxygen concentration (g L1). The kinetic expression for ld assumes a minimum specific cell
V volume (L). death rate ðl0d Þ due mainly to cell ageing. Such a rate is weighted
Fi volumetric feed flow rate (L h1). by fdE and fdA terms, which consider the influence of the substrate
E0 ethanol concentration in raw material (g L1). and product.
O0 dissolved oxygen concentration in equilibrium with air A first-order kinetic law exclusively dependent on the non-via-
(g L1). ble cell concentration is used to represent the cell lysis rate:
b constant including KLa factor, aeration flow and volume
(h1). rlisis ¼ l0lisis  X d ð30Þ
rXc cell growth rate (g viable cells L1 h1). where:l 0 1
lisis :
specific lysis rate (h ).
rXd cell death rate (g non-viable cells L1 h1).
The expression for the ethanol uptake rate, rE, obtained from Eq.
rlisis cell lysis rate (g lysed cells L1 h1).
(17) is as follows:
rE ethanol uptake rate (g ethanol L1 h1).
rA acetic acid formation rate (g acetic acid L1 h1). rE ¼ aE=X  r X c ð31Þ
rO dissolved oxygen uptake rate (g oxygen L1 h1).
where:
aE/X: a yield factor accounting for the amount of ethanol to be
and the following kinetic expressions can be put forward for the
consumed in order to produce the amount of energy required for
growth rate:
biomass growth (g ethanol g1 biomass). aE/X is approximately
116.96 g ethanol g1 cell.
r X c ¼ lc  X v ð21Þ The expression for the acetic acid formation rate is
lc ¼ lmax  fe  fa  fo ð22Þ
rE
rA ¼ ð32Þ
where: Y E=A
1
lc specific cell growth rate (h ).
where:
lmax maximum specific cell growth rate (h1).
fe, fa and fo terms representing the influence of ethanol, acetic acid YE/A: stoichiometric coefficient of ethanol consumption for ace-
and dissolved oxygen on cell growth. tic acid formation (0.767 g ethanol g1 acetic acid).
As regards the oxygen mass balance (19), coefficient b [75] al-
lows one to use (O0), which is constant and roughly equal to
Also, the following expressions can be used to weight the max-
7.5 103 g O2 L1.
imum specific cell growth, which ultimately represent its kinetics:
KLa
E b¼ ð33Þ
fe ¼ 2
ð23Þ K L a RT
1 þ VV H
E þ K SE þ KEIE m

Q
1 V  Vm ¼ ð34Þ
fa ¼  4 ð24Þ V
A
1þ K IA where:
O KLa global volumetric mass transfer coefficient for the liquid
fo ¼ ð25Þ phase (h1).
O þ K SO
V  Vm relation between the input air flow and volume (h1).
where R gas constant (0.082 atm L K1 mol1).
1
KSE ethanol saturation constant (g ethanol L ). T temperature (K).
KIE ethanol inhibition constant (g ethanol L1). H Henry constant.
KIA acetic acid inhibition constant (g acetic acid L1). Q input air flow rate (L h1).
KSO dissolved oxygen saturation constant (g oxygen L1). V volume (L).

The previous ethanol limitation and cell growth inhibition The dissolved oxygen uptake rate can be represented by
expression for fe was proposed by Andrews [74]. fa only considers
rE
growth inhibition by acetic acid. Finally, fo uses a simple Monod rO ¼ ð35Þ
Y E=O
structure to represent the influence of dissolved oxygen on cell
growth. where:
In addition, the following expressions are proposed to represent YE/O: ethanol/oxygen stoichiometric coefficient (1.44 g etha-
the cell death rate: nol g1 oxygen).
In summary, if all previous expressions for the proposed model
r Xd ¼ ld  X v ð26Þ
are considered, the final state–space form used for structural iden-
ld ¼ l0d  fdE  fdA ð27Þ tifiability analysis is as follows:
158 J.E. Jiménez-Hornero et al. / Mathematical Biosciences 216 (2008) 154–162

dX v E 1 O – Possible combinations of the output power series expansion

¼ lmax    4   Xv
dt E2 O þ K SO coefficients for the systems can grow exponentially; there-
E þ K SE þ 1þ K IE
A
K IA fore, if no solvable system is found, it is impossible to ascer-
" 4 # "
  4 #
E A 1 tain whether the model is unidentifiable or a different
0
 ld  1 þ  1þ  X v  X v  F i ð36Þ combination of coefficients should be tested.
K mE K mA V
"  4 # "  4 # – The computation time needed to obtain each set of power
dX d 0 E A series expansion coefficients – which sometimes entails cal-
¼ ld  1 þ  1þ  Xv
dt K mE K mA culating high-order derivatives – is substantial, even if spe-
1 cific software tools for symbolic computation are used. As
 l0lisis  X d  Xd  Fi ð37Þ a result, the computation time required to solve each equa-
V
tion system is very long.
dE F i E 1
¼ ðE0  EÞ  aE=X  lmax    4
dt V E2
E þ K SE þ K IE 1þ A
We can therefore conclude that verifying structural identifiabil-
K IA
ity in the proposed model with the Taylor series expansion method
O
  Xv ð38Þ is very difficult. Indeed, we reached no reliable conclusion from its
O þ K SO
2 3 application to several equation systems. By way of example, let us
dA 1 6 E 1 O illustrate the process for one of the many systems studied – an
7
¼  4aE=X  lmax    4   Xv5 unsolvable system, obviously. The nine parameters involved would
dt Y E=A E2 O þ K SO
E þ K SE þ K IE 1þ A
K IA require at least nine independent equations established by com-
1 puting the power series expansion coefficients. Based on the
 A  Fi ð39Þ above-described measured outputs,
V
dO F i 0 KLa y1 ¼ X v
¼ ðO  OÞ þ  ðO0  OÞ
dt V 1 þ K LQaV  RT y2 ¼ X v þ X d ð45Þ
H
2 3 y3 ¼ E
1 6 E 1 O 7 where yi are the outputs. At t = 0,
  4aE=X  lmax  2
  4   Xv5
Y E=O E þ K SE þ KEIE 1 þ A O þ K SO y1 ð0Þ ¼ X v ð0Þ
K IA

ð40Þ y2 ð0Þ ¼ X v ð0Þ þ X d ð0Þ ð46Þ

y3 ð0Þ ¼ Eð0Þ
dV
¼ Fi ð41Þ where Xv(0), Xd(0) and E(0) are initial conditions of these three state
dt
y1 ¼ X v ð42Þ variables, corresponding to the viable cell concentration, total cell
concentration and ethanol concentration, respectively. For simplic-
y2 ¼ X v þ X d ð43Þ
ity, state variables can be renamed xi and parameters pi:
y3 ¼ E ð44Þ
x1 ¼ X v ; x2 ¼ X d ; x3 ¼ E; x4 ¼ A; x5 ¼ O; x6 ¼ V ð47Þ
The parameters introduced in the model are lmax ; K SE ; K IE ; K IA ; p1 ¼ lmax ; p2 ¼ K SE ; p3 ¼ K IE ; p4 ¼ K IA ;
K SO ; l0d ; K mE ; K mA and l0lisis .
p5 ¼ K SO ; p6 ¼ l0d ; p7 ¼ K ME ; p8 ¼ K MA ; p9 ¼ l0lisis ð48Þ
Because the initial conditions are known and constant, they are
3. Structural identifiability analysis of the proposed model
independent of the model parameters. Therefore, they cannot be
used to solve for the parameters in Eqs. (5) and (6). To simplify
3.1. Analytical methods
the notation in the equations, the following can be done:
The proposed model is non-linear, with six state variables and nine c1 ¼ x1 ð0Þ; c2 ¼ x2 ð0Þ; c3 ¼ x3 ð0Þ; c4 ¼ x4 ð0Þ; c5 ¼ x5 ð0Þ;
parameters, and has the volumetric feed flow (Fi) as its sole input. Also, c6 ¼ x6 ð0Þ ð49Þ
it only uses the viable cell concentration, total cell concentration and
dyi
ethanol concentration as outputs (measured variables); neither the Therefore, appropriate expressions for dt
ð0Þ are needed which can
acetic acid concentration – which is related, though variably, to the be obtained by using Maple:
ethanol concentration – nor the volume – which represents the bal- dy1 p1 c1 c3 c5
ð0Þ ¼   
ance between the feed and product flows – are employed. Also, it dt c2 c4
c3 þ p2 þ p3 1 þ p44 ðc5 þ p5 Þ
was impossible to obtain reliable dissolved oxygen concentration 3 4
  
measures owing to the characteristics of the culture and reactor; c43 c44 c1  F i ð0Þ
 p6 1 þ 4 1 þ 4 c1  ð50Þ
therefore, this variable was excluded as a measured output. p7 p8 c6
The mathematical structure of the proposed model is amenable    
dy2 dy c4 c4 c2  F i ð0Þ
to analysis with any type of method for verifying structural identi- ð0Þ ¼ 1 ð0Þ þ p6 1 þ 34 1 þ 44 c1  p9 c2 
dt dt p7 p8 c6
fiability (Taylor series expansion, generating series, local state iso-
morphism and differential algebra). However, in order to preserve ð51Þ
the physical meaning of the parameters – if any – we chose to ex- dy3 aE=X  p1 c1 c3 c5 F i E0  c3  F i ð0Þ
ð0Þ ¼   c23
 c44
 þ
clude analytical methods involving any model transformation. dt c6
c3 þ p2 þ p3 1 þ p4 ðc5 þ p5 Þ
The Taylor series expansion approach was implemented in sev- 4

eral trial-and-error cycles that provided mathematical expressions ð52Þ

for the initial coefficients of the outputs of the power series expan- Eqs. (50)–(52) are only three of at least nine independent equations
sions until a system of nine independent equations in nine un- required to build a solvable system. Therefore, additional output
knowns (parameters) was obtained. The aim was to find a power series expansion coefficients must be obtained. We computed
solvable system. However, we encountered several problems in d2 yi 3
ð0Þ and ddty3i ð0Þ to this end, but their mathematical expressions
dt 2
the process, namely: were too complex to be shown here. Also, the ensuing system pro-
 J.E. Jiménez-Hornero et al. / Mathematical Biosciences 216 (2008) 154–162 159

vided no solution with Maple software. The Maple code used is shown Lie brackets were computed in this way:
in Appendix A. These results led us to try an alternative method.
of1 of0
Such a method was the generating series method, which is very ½f0 ðx; hÞ; f1 ðx; hÞ ¼ f0  f1 ð58Þ
ox ox
similar to the previous one but was expected to be less computa-
tionally demanding based on the ability to isolate output terms. where f0(x,h) and f1(x,h) are vector fields.
The proposed model for the acetic acid fermentation process can Verifying the controllability rank condition required building an
be rewritten as follows in terms of Eq. (7): n  n matrix (n being the model order) by computing as many Lie
brackets as needed to fill its columns. If the rank of the matrix in
dX v ¼ r  r  Xv F
Xc Xd V i the initial state and for any value of each model parameter is n,
dt
dX d ¼ r  r  X d F then the controllability condition will be fulfilled. Otherwise, either
dt Xd lisis V i the model does not satisfy this condition or a different matrix
dE ¼ r þ E0 E F needs to be built from other Lie brackets.
EX i
dt V
Similarly, verifying the observability rank condition required
dA ¼ r  A F building an n  n matrix, the rows of which would be filled with
A V i
dt ð53Þ gradients of the Lie derivatives with respect to the state vector.
dO ¼ bðO0  OÞ  r þ O0 O F
dt OE V i Likewise, if the rank of the matrix in the initial state and for any
dV ¼ F value of each model parameter is n, then the observability condi-
i
dt tion will be fulfilled.
y1 ¼ X v The difficulty involved in computing the Lie brackets and Lie
y1 ¼ X v þ X d derivatives for the proposed model, and their infinite potential
combinations, made the problem unsolvable in practice.
y1 ¼ E
In any case, we tested some controllability and observability
Therefore, matrices for the proposed model, but no solid evidence of their
2 3 2 3 ranks being 6 (=n from Eq. (53)) could be obtained.
r Xc  rXd  XVv
6 Xd 7 By way of example of the verification of the controllability rank
6r  r 7 6 V 7
6 Xd lisis 7 6 7 condition, we built the following controllability matrix from the
6 7 6 E0 E 7
6 r EX 7 6 7 model Eq. (54)
f0 ðx; hÞ ¼ 6
6r
7
7 f 1 ðx; hÞ ¼ 6 V 7
6 A 7 6A 7 ½ f0 f1 ½f0 ; f1  ½½f0 ; f1 ; f0  ½½f0 ; f1 ; f1  ½½½f0 ; f1 ; f1 ; f1   ð59Þ
6 7 6 V 7
0
4 bðO  OÞ  r OE 5 6 O0 O 7 ð54Þ
4 5
V The mathematical expressions for these Lie brackets are too com-
0 1 plex to be shown here. Their complexity precluded confirming
2 3
Xv whether the matrix rank for any parameter vector would be 6.
6 7 We used the following matrix, among others, to verify the
hðx; hÞ ¼ 4 X v þ X d 5
observability rank condition:
E
2 3
rx h1
This method exhibited the same problems as the power series 6 rx h2 7
expansion. Although Lie derivatives gave simpler mathematical 6 7
6 7
expressions, the equation system was unsolvable, so no reliable 6 rx h3 7
6 7 ð60Þ
conclusions about structural identifiability in the model could be 6 r ðL h Þ 7
6 x f0 1 7
drawn. By way of example, below is described one of many systems 6 7
4 rx ðLf0 h2 Þ 5
studied here. Because solving for the parameters required using
rx ðLf0 h3 Þ
at least nine equations, we chose to use the following Lie deriva-
tives: (evaluated at t = 0): Lf0 h1 ð0Þ; Lf0 h2 ð0Þ; Lf0 h3 ð0Þ; Lf1 Lf0 h1 ð0Þ; The only simple expressions in this case were those for rx h1, rxh2
Lf1 Lf0 h2 ð0Þ; Lf1 Lf0 h3 ð0Þ; Lf1 Lf1 Lf0 h1 ð0Þ; Lf1 Lf1 Lf0 h2 ð0Þ and Lf1 Lf1 and rxh3:
Lf0 h3 ð0Þ. We used the nomenclature introduced in Eqs. (47)–(49),
and computed the first three derivatives with Maple: rx h1 ¼ ½ 1 0 0 0 0 0 
  rx h2 ¼ ½ 1 1 0 0 0 0  ð61Þ
oh1 p1 c1 c3 c5
Lf0 h1 ð0Þ ¼  f0 ð0Þ ¼   
ox c23 c4 rx h3 ¼ ½ 0 0 1 0 0 0 
c3 þ p2 þ p3 1 þ p44 ðc5 þ p5 Þ
4
   Other gradients were mathematically more complex and are thus
c43 c44
 p6 1 þ 4 1 þ 4 c1 ð55Þ not shown. As with the controllability rank condition, we could
p p8
  7 not ascertain whether the rank of the observability matrix was 6.
oh2 p1 c1 c3 c5
Lf0 h2 ð0Þ ¼  f0 ð0Þ ¼    Therefore, this method also failed to afford clear-cut conclusions
ox c2 c4
c3 þ p2 þ p3 1 þ p44 ðc5 þ p5 Þ about structural identifiability in the proposed model.
3 4
Although the differential algebra method was applicable in the-
 p9 c2 ð56Þ
  ory, the complexity of the model required extremely elaborate
oh3 aE=X p1 c1 c3 c5 algebraic and differentiation operations involving a vast number
Lf0 h3 ð0Þ ¼  f0 ð0Þ ¼  c2
 c4
 ð57Þ
ox c3 þ p2 þ p33 1 þ p44 ðc5 þ p5 Þ of potential choices. Also, although several alternatives were used
4 to obtain a set of linear equations for the parameters, no definite
The mathematical expressions for the other derivatives were conclusions could be reached.
exceedingly complex and are thus not shown here. Using Maple
code provided no unique solution for the nine Lie derivative system. 3.2. Numerical methods
Applying the local state isomorphism theorem method required
initially verifying the controllability and observability rank condi- Based on the results of the analytical methods, only numerical
tions in the initial state. This entailed analysing the ranks of matri- methods could provide an effective way of solving the problem.
ces built from Lie brackets and Lie derivatives, respectively. Although they only afford conclusions about structural local iden-
160 J.E. Jiménez-Hornero et al. / Mathematical Biosciences 216 (2008) 154–162

tifiability, numerical methods can be applied to much more com- where x and ~ x are state vectors, p is the known parameter vector, p
plex models and with a higher computing efficiency. the unknown parameter vector and y the output vector. Eqs. (62)–
The first method studied here is described in detail elsewhere (79) contain 21 unknowns. Nevertheless, they can be reduced to 18
[60]. We used Ecosimpro software (Empresarios Agrupados, S.A., since, from Eqs. (68)–(70) and (77)–(79), it is clear that
Madrid, Spain) as the numerical solver for application to the pro- ~x1 ¼ x1 ; ~x2 ¼ x2 and ~x3 ¼ x3 . It is therefore necessary to add six
posed model. Although the computational cost was virtually equations in order to obtain a solvable system. Such equations must
negligible, non-linear algebraic loops and/or structural singulari- be obtained by differentiating with respect to time some of the Eqs.
ties often caused convergence problems. In any case, the main (71)–(75) on the assumption of constancy in the parameters.
problem was to make an appropriate choice of equations in or- According to the proponent of this method [60], the equations to
der to obtain a solvable system. If the system concerned has be differentiated should be chosen in accordance with two essential
no solution, then alternative equations (e.g. expressions contain- criteria: the equations should contain as many parameters as possi-
ing additional derivatives) must be tested. If no solution is even- ble and excessively complex, high-order derivatives – numerical
tually found, this does not necessarily mean that the model is integration of which can be rather difficult – should be avoided.
structurally unidentifiable. Application of this method to the pro- By way of example, we chose to derive Eqs. (72)–(75) once and
posed model provided no solvable system. The initial system Eq. (71) twice – using Maple in both cases. The ensuing mathemat-
was of the following form [see nomenclature in Eqs. (38 and ical expressions were too complex to be shown here. We used the
39)]: following initial conditions for integration:
  
1 x1 x3 x5
p x43 x44 x1 ð0Þ ¼ 0:15; x2 ð0Þ ¼ 0:02; x3 ð0Þ ¼ 11:6; x4 ð0Þ ¼ 105;
x_ 1 ¼     
p 6 1 þ 1 þ
x2 x4
2 þ p3 1 þ 44 ðx5 þ p
x3 þ p 5 Þ 47
p 48
p x5 ð0Þ ¼ 3  103 ; x6 ð0Þ ¼ 4 ð80Þ
3 p 4

x1 F i and the following parameter values (falling in the preset range):

x1  ð62Þ
x6
   p 2 ¼ 3:5; p
1 ¼ 0:6; p 3 ¼ 3; p 5 ¼ 3  104 ;
4 ¼ 90; p
x4 x4 x F
x_ 2 ¼p 6 1 þ 34 1 þ 44 x1  p  9 x2  2 i ð63Þ 6 ¼ 2:5  105 ; p
p 7 ¼ 12:5; p
8 ¼ 35; p
9 ¼ 0:5 ð81Þ
7
p p8 x6
aE=X p  1 x1 x3 x5 F i ðE0  x3 Þ The solver exhibited convergence problems from the first integra-
x_ 3 ¼  x23
 x44
 þ ð64Þ
2 þ p
x3 þ p 1 þ p4 ðx5 þ p 5 Þ x6 tion step and provided no solution. Alternative equation systems,
3 4 different initial conditions and trial parameter values were tested,
aE=X p 1 x1 x3 x5 x4 F i but no unique solution for pi could be found. Therefore, despite its
x_ 4 ¼  x23
 x44
  ð65Þ
2 þ p
x3 þ p 1 þ ðx þ 
p ÞY x6 simplicity and low computational cost, the method failed as a
4
p 5 5 E=A
3 4
means for verifying structural identifiability in our model.
K L aðO0  x5 Þ aE=X p 1 x1 x3 x5 Finally, we used a software tool developed as a Maple package
x_ 5 ¼   
1 þ K LQax6  RT
H 
x23 x4
x3 þ p2 þ p3 1 þ p44 ðx5 þ p5 ÞY E=O by its proponent (http://www2.lifl.fr/~sedoglav/Software/) to
4 implement the method introduced in [61]. This method, which fea-
F i ðO0  x5 Þ tures virtually negligible computation times, provided a positive
þ ð66Þ
x6 answer as regards theoretical local identifiability in the proposed
x_ 6 ¼ F i ð67Þ model. The software tool certified that all model parameters were
observable.
y1 ¼ x1 ð68Þ
y2 ¼ x1 þ x2 ð69Þ
4. Conclusions
y3 ¼ x3 ð70Þ
  
p1 ~x1 ~x3 ~x5 ~x43 ~x44 The model proposed for the acetic acid fermentation process
~x_ 1 ¼     p6 1 þ 1 þ ~x1
~x2 ~x4
~x3 þ p2 þ p3 1 þ 44 ð~x5 þ p5 Þ p47 p48 has been examined for structural identifiability by using various
p
3 4 analytical and numerical methods. The former included Taylor ser-
~x1 F i ies expansion, generating series, local state isomorphism and dif-
 ð71Þ
~x6 ferential algebra, all of which exhibited problems such as difficult
  
~x4 ~x4 ~x2 F i implementation and substantial computational costs for models
~x_ 2 ¼ p6 1 þ 34 1 þ 44 ~x1  p9 ~x2  ð72Þ
p7 p8 ~x6 of small-medium dimension and complexity frequently found in
practice.
aE=X p1 ~x1 ~x3 ~x5 F i ðE0  ~x3 Þ
~x_ 3 ¼    þ ð73Þ Numerical methods were easier to implement than the analyt-
~x2 ~x4 ~x6
~x3 þ p2 þ p 3
1 þ p4 ð~x5 þ p5 Þ
4
ical methods, but only afforded verification of local structural
3 4
identifiability.
aE=X p1 ~x1 ~x3 ~x5 ~x4 F i
~x_ 4 ¼     ð74Þ The results of our study of structural identifiability in the pro-
~x2 ~x4 ~x6
~x3 þ p2 þ p3 1 þ 44 ð~x5 þ p5 ÞY E=A posed model for the acetic acid fermentation process allow us to
3 p 4
draw the following conclusions:
K L aðO0  ~x5 Þ aE=X p1 ~x1 ~x3 ~x5
~x_ 5 ¼   
~ ~x2 ~x4
1 þ K LQax6  RT
H ~x3 þ p2 þ p3 1 þ 44 ð~x5 þ p5 ÞY E=A – No clear-cut criteria exist for choosing the best potential
3 p
4
0 method for structural identifiability verification of a
F i ðO  ~x5 Þ dynamic model. This is consistent with the variety of meth-
þ ð75Þ
~x6 ods used for this purpose to date.
~x_ 6 ¼ F i ð76Þ – The substantial computational cost involved entails the use
y1 ¼ ~x1 ð77Þ of symbolic software to analyse models for theoretical iden-
tifiability. In some cases, computational complexity is so
y2 ¼ ~x1 þ ~x2 ð78Þ
high that the problem is intractable, even with powerful
y3 ¼ ~x3 ð79Þ
software.
 J.E. Jiménez-Hornero et al. / Mathematical Biosciences 216 (2008) 154–162
Only one of the studied methods provided positive results
affording the conclusion that the proposed model is structurally lo-
cally identifiable.
While theoretical identifiability analysis is important, model
parameters can only be estimated uniquely after practical identifi-
ability has been verified. If a model is practically identifiable, then
it should be structurally identifiable as well; on the other hand, a
model may not be practically identifiable either because of the lack
of appropriate available experimental data or the model structure
itself. In this situation, it is important to also have information
about the theoretical identifiability of the model.
Acknowledgments
The authors wish to acknowledge funding of this research
(AGL2002-01712, PET2006_0827) by Spain’s Ministry of Science
and Technology (MCyT) and Ministry of Education and Science.
Also, they wish to thank the firm Vinagres y Salsas, S.A., Grupo
SOS (Córdoba, Spain), for valuable help and advice.
Appendix A
x1:¼f1(t): x2:¼f2(t): x3:¼f3(t): x4:¼f4(t): x5:¼f5(t):
x6:¼f6(t):
fe:¼x3/(x3+p2+x3̂2/p3):
fa:¼1/(1+(x4/p4) ˆ4):
fo:¼x5/(x5+p5):
rxc:¼p1*fe*fa*fo*x1:
rxd:¼p6*(1+(x3/p7) ˆ4)* (1+(x4/p8)ˆ4)*x1:
rlysis:¼p9*x2:
re:¼aex*rxc:
ra:¼re/Yea:
ro:¼re/Yeo:
bt:¼Kla/(1+RTH*x6*Kla/Q):
x6p:¼Fi:
x1p:¼rxc-rxd-x1*x6p/x6:
x2p:¼rxd-rlysis-x2*x6p/x6:
x3p:¼-re+(Fi*E0-x3*x6p)/x6:
x4p:¼ra-x4*x6p/x6:
x5p:¼bt*(O0-x5)-ro+(Fi*O0-x5*x6p)/x6:
x1pp:¼diff(x1p,t):
x2pp:¼diff(x2p,t):
x3pp:¼diff(x3p,t):
x1ppp:¼diff(x1pp,t):
x2ppp:¼diff(x2pp,t):
x3ppp:¼diff(x3pp,t):
solve({x1p=c1, x2p=c2, x3p=c3, x1pp=c4, x2pp=c5,
x3pp=c6, x1ppp=c7, x2ppp=c8, x3ppp=c9}, [p1,p2,p3,p4,
p5,p6,p7,p8,p9])
References
[1] J.R. Banga, E. Balsa-Canto, C.G. Moles, A.A. Alonso, Improving food processing
using modern optimization methods, Trends Food Sci. Technol. 14 (2003) 131.
[2] C. Gutiérrez, Optimización global de procesos de la industria alimentaria y
biotecnológica, Dept. Ingenierı́a de Sistemas y Automática, University of Vigo,
Vigo, Spain, 2003.
[3] K. Schittkowski, Parameter estimation in systems of nonlinear equations,
Numer. Math. 68 (1994) 129.
[4] E. Walter, Identifiability of Parametric Models, Pergamon, Oxford, 1987.
[5] D. Dochain, P.A. Vanrolleghem, Dynamical Modelling and Estimation in
Wastewater Treatment Processes, IWA Publishing, London, 2001.
[6] D. Dochain, P.A. Vanrolleghem, M. Vandaele, Structural identifiability of
biokinetic models of activated-sludge respiration, Water Res. 29 (1995) 2571.
[7] B. Petersen, Calibration, identifiability and optimal experimental design of
activated sludge models, University of Ghent, Ghent, Belgium, 2000.
161
[8] B. Petersen, K. Gernaey, P.A. Vanrolleghem, Practical identifiability of model
parameters by combined respirometric–titrimetric measurements, Water Sci.
Technol. 43 (2001) 347.
[9] S.V. Bourrel, J.P. Babary, S. Julien, M.T. Nihtila, D. Dochain, Modelling and
identification of a fixed-bed denitrification bioreactor, Syst. Anal. Model.
Simulat. 30 (1998) 289.
[10] L. Chen, G. Bastin, Structural identifiability of the yield coefficients in
bioprocess models when the reaction rates are unknown, Math. Biosci. 132
(1996) 35.
[11] S. Julien, J.P. Babary, P. Lessard, Identifiability and identification of an activated
sludge process model, Syst. Anal. Model. Simulat. 37 (2000) 481.
[12] K.J. Keesman, H. Spanjers, G. van Straten, Analysis of endogenous process
behaviour in activated sludge, Biotechnol. Bioeng. 57 (1998) 155.
[13] N. Noykova, T.G. Muller, M. Gyllenberg, J. Timmer, Quantitative analyses of
anaerobic wastewater treatment processes: identifiability and parameter
estimation, Biotechnol. Bioeng. 78 (2002) 89.
[14] L. Denis-Vidal, G. Joly-Blanchard, An easy to check criterion for
(un)identifiability of uncontrolled systems and its applications, IEEE Trans.
Automat. Control 45 (2000) 768.
[15] N.D. Evans, M.J. Chappell, Extensions to a procedure for generating locally
identifiable reparametrisations of unidentifiable systems, Math. Biosci. 168
(2000) 137.
[16] G. Margaria, E. Riccomagno, M.J. Chappell, H.P. Wynn, Differential algebra
methods for the study of the structural identifiability of rational function
state–space models in the biosciences, Math. Biosci. 174 (2001) 1.
[17] G. Margaria, E. Riccomagno, L.J. White, Structural identifiability analysis of
some highly structured families of state space models using differential
algebra, J. Math. Biol. 49 (2004) 433.
[18] T.J. Debus, P.E. Dupont, R.D. Howe, Distinguishability and identifiability testing
of contact state models, Adv. Robotics 19 (2005) 545.
[19] R. Serban, J.S. Freeman, Identification and identifiability of unknown
parameters in multibody dynamic systems, Multibody Syst. Dyn. 5 (2001) 335.
[20] F. Couenne, C. Jallut, M. Tayakout-Fayolle, On minimal representation of
heterogeneous mass transfer for simulation and parameter estimation:
application to breakthrough curves exploitation, Comput. Chem. Eng. 30
(2005) 42.
[21] S.P. Asprey, S. Macchietto, Statistical tools for optimal dynamic model
building, Comput. Chem. Eng. 24 (2000) 1261.
[22] R. Bellman, K.J. Aström, On structural identifiability, Math. Biosci. 7 (1970)
329.
[23] K.R. Godfrey, J.J. Distefano, Identifiability of model parameters, in: E. Walter
(Ed.), Identifiability of Parametric Models, Pergamon, Oxford, 1987, p. 1.
[24] C. Cobelli, A. Lepschy, G.R. Jacur, Identifiability of compartmental systems and
related structural properties, Math. Biosci. 44 (1979) 1.
[25] C. Cobelli, G. Toffolo, Identifiability from parameter bounds – structural and
numerical aspects, Math. Biosci. 71 (1984) 237.
[26] A. Raksanyi, Y. Lecourtier, E. Walter, A. Venot, Identifiability and
distinguishability testing via computer algebra, Math. Biosci. 77 (1985) 245.
[27] H. Pohjanpalo, System identifiability based on power-series expansion of
solution, Math. Biosci. 41 (1978) 21.
[28] M.J. Chappell, K.R. Godfrey, S. Vajda, Global identifiability of the parameters of
non-linear systems with specified inputs – a comparison of methods, Math.
Biosci. 102 (1990) 41.
[29] E. Walter, Identifiability of State Space Models, Springer, New York, 1982.
[30] S. Vajda, K.R. Godfrey, H. Rabitz, Similarity transformation approach to
identifiability analysis of nonlinear compartmental models, Math. Biosci. 93
(1989) 217.
[31] S. Vajda, Structural identifiability of dynamical systems, Int. J. Syst. Sci. 14
(1983) 1229.
[32] S. Diop, M. Fliess, On nonlinear observability. in: C. Commault (Ed.),
Proceedings of the First European Control Conference, Hermes, Grenoble,
Switzerland, 1991, p. 152.
[33] S. Vajda, Deterministic identifiability and algebraic invariants for polynomial
systems, IEEE Trans. Automat. Control 32 (1987) 182.
[34] H. Pohjanpalo, Identifiability of deterministic differential models in state
space, Technical Report 56, Technical Research Centre of Finland, Espoo, 1982.
[35] E. Walter, Y. Lecourtier, Global approaches to identifiability testing for linear
and nonlinear state–space models, Math. Comput. Simulat. 24 (1982) 472.
[36] Y. Lecourtier, F. Lamnabhi-Lagarrigue, E. Walter, Volterra and generating
power series approaches to identifiability testing, in: E. Walter (Ed.),
Identifiability of Parametric Models, Pergamon, Oxford, 1987.
[37] E. Walter, L. Pronzato, On the identifiability and distinguishability of nonlinear
parametric models, Math. Comput. Simulat. 42 (1996) 125.
[38] E.T. Tunali, T.J. Tarn, New results for identifiability of nonlinear systems, IEEE
Trans. Automat. Control 32 (1987) 146.
[39] R. Hermann, A.J. Krener, Nonlinear controllability and observability, IEEE
Trans. Automat. Control 22 (1977) 728.
[40] H.J. Sussman, Existence and uniqueness of minimal realizations of nonlinear
systems, Math. Syst. Theory 10 (1977) 263.
[41] R.E. Kalman, A new approach to linear filtering and prediction problems, Trans.
ASME – J. Basic Eng. 82 (1960) 35.
[42] S. Vajda, H. Rabitz, State isomorphism approach to global identifiability of
nonlinear systems, IEEE Trans. Automat. Control 34 (1989) 220.
[43] M. Fliess, S.T. Glad, An algebraic approach to linear and nonlinear control, in:
H.L. Trentelman, J.C. Willems (Eds.), Essays on Control: Perspectives in the
Theory and its Applications, Boston, MA, 1993, p. 223.
162 J.E. Jiménez-Hornero et al. / Mathematical Biosciences 216 (2008) 154–162
[44] L. Ljung, T. Glad, On global identifiability for arbitrary model parametrizations,
Automatica 30 (1994) 265.
[45] J.F. Ritt, Differential equations from the algebraic standpoint, American
Mathematical Society Colloquium Publications, 1932.
[46] W.J. Wu, On zeros of algebraic equations – an application of Ritt principle,
Kexue Tongbao 31 (1986) 1.
[47] D.M. Wang, An implementation of the characteristic set method in MAPLE, in:
J. Pfalzgraf, D. Wang (Eds.), Automated Practical Reasoning: Algebraic
Approaches, Springer, New York, 1995, p. 187.
[48] L. Denis-Vidal, G. Joly-Blanchard, C. Noiret, Some effective approaches to check
the identifiability of uncontrolled nonlinear systems, Math. Comput. Simulat.
57 (2001) 35.
[49] T.G. Müller, Modelling Complex Systems with Differential Equations, Albert
Ludwigs Universität, Freiburg, 2002.
[50] M.J. Chappell, R.N. Gunn, A procedure for generating locally identifiable
reparametrisations of unidentifiable non-linear systems by the similarity
transformation approach, Math. Biosci. 148 (1998) 21.
[51] J.J. Distefano, Complete parameter bounds and quasi-identifiability conditions
for a class of unidentifiable linear systems, Math. Biosci. 65 (1983) 51.
[52] R.N. Gunn, M.J. Chappell, V.J. Cunningham, Reparametrisation of unidentifiable
systems using the Taylor series approach, in: D.A. Linkens, E. Carson (Eds.),
Proceedings of the Third IFAC Symposium on Modelling and Control in
Biomedical Systems, Pergamon, Oxford, 1997, p. 247.
[53] E. Walter, Y. Lecourtier, Unidentifiable compartmental models – what to do,
Math. Biosci. 56 (1981) 1.
[54] N.D. Evans, M.J. Chapman, M.J. Chappell, K.R. Godfrey, Identifiability of
uncontrolled nonlinear rational systems, Automatica 38 (2002) 1799.
[55] N.D. Evans, R.J. Errington, M. Shelley, G.P. Feeney, M.J. Chapman, K.R. Godfrey,
P.J. Smith, M.J. Chappell, A mathematical model for the in vitro kinetics of the
anti-cancer drug topotecan, Math. Biosci. 189 (2004) 185.
[56] N.D. Evans, L.J. White, M.J. Chapman, M.J. Chappell, K.R. Godfrey, The structural
identifiability of the susceptible infected recovered model with seasonal
forcing, Math. Biosci. 194 (2005) 175.
[57] H.G.M. Dotsch, P.M.J. VandenHof, Test for local structural identifiability of
high-order non-linearly parametrized state space models, Automatica 32
(1996) 875.
[58] I. Braems, L. Jaulin, M. Kieffer, E. Walter, Guaranteed numerical alternatives to
structural identifiability testing, in: Proceedings of the 40th IEEE Conference
on Decision and Control, Orlando, CA, 2001.
[59] E. Walter, I. Braems, L. Jaulin, M. Kieffer, Guaranteed numerical computation as
an alternative to computer algebra for testing models for identifiability, in: R.
Alt (Ed.), Numerical Software with Result Verification, Springer-Verlag, Berlin,
2004, p. 124.
[60] M. Gerdin, Using DAE solvers to examine local identifiability for linear and
nonlinear systems, in: Proceedings of the 14th IFAC Symposium of System
Identification (SYSID-2006), 2006.
[61] A. Sedoglavic, A probabilistic algorithm to test local algebraic observability in
polynomial time, J. Symb. Comput. 33 (2002) 735.
[62] R. Bar, J.L. Gainer, D.J. Kirwan, An unusual pattern of product inhibition – batch
acetic acid fermentation, Biotechnol. Bioeng. 29 (1987) 796.
[63] J.M. González-Saiz, C. Pizarro, D. Garrido-Vidal, Evaluation of kinetic models
for industrial acetic acid fermentation: proposal of a new model optimized by
genetic algorithms, Biotechnol. Prog. 19 (2003) 599.
[64] T. Ito, H. Sota, H. Honda, K. Shimizu, T. Kobayashi, Efficient acetic acid
production by repeated fed-batch fermentation using 2 fermenters, Appl.
Microbiol. Biotechnol. 36 (1991) 295.
[65] R.K. Kruppa, D. Vortmeyer, Transient growth and product-formation kinetics
of acetic acid bacteria, Bioprocess Eng. 20 (1999) 545.
[66] A. Nanba, A. Tamura, S. Nagai, Synergistic effects of acetic acid and ethanol on
the growth of Acetobacter sp, J. Ferment. Technol. 62 (1984) 501.
[67] Y.S. Park, H. Ohtake, M. Fukaya, H. Okumura, Y. Kawamura, K. Toda,
Enhancement of acetic acid production in a high cell density culture of
Acetobacter aceti, J. Ferment. Bioeng. 68 (1989) 315.
[68] Y.S. Park, H. Ohtake, K. Toda, M. Fukaya, H. Okumura, Y. Kawamura, Acetic acid
production using a fermenter equipped with a hollow fiber filter module,
Biotechnol. Bioeng. 33 (1989) 918.
[69] Y.S. Park, H. Ohtake, M. Fukaya, H. Okumura, Y. Kawamura, K. Toda, Effects
of dissolved oxygen and acetic acid concentrations on acetic acid
production in continuous culture of Acetobacter aceti, J. Ferment. Bioeng.
68 (1989) 96.
[70] Y.S. Park, M. Fukaya, H. Okumura, Y. Kawamura, K. Toda, Production of acetic
acid by a repeated batch culture with cell recycle of Acetobacter aceti,
Biotechnol. Lett. 13 (1991) 271.
[71] Y.S. Park, K. Toda, M. Fukaya, H. Okumura, Y. Kawamura, Production of a
high-concentration acetic-acid by Acetobacter aceti using a repeated fed-
batch culture with cell recycling, Appl. Microbiol. Biotechnol. 35 (1991)
149.
[72] Y.S. Park, K. Toda, Multistage biofilm reactor for acetic-acid production at high
concentration, Biotechnol. Lett. 14 (1992) 609.
[73] L.E. Romero, J.M. Gomez, I. Caro, D. Cantero, Kinetic model for growth of
Acetobacter aceti in submerged culture, Chem. Eng. J. Biochem. Eng. J. 54 (1994)
B15.
[74] J.F. Andrews, A mathematical model for continuous culture of microorganisms
utilizing inhibitory substrates, Biotechnol. Bioeng. 10 (1968) 707.
[75] F.J. Nieto, Algunos aspectos de la tecnologı´a de la fermentación acética, in: C.
Llaguno, M.C. Polo (Eds.), El Vinagre de Vino, CSIC, Madrid, Spain, 1991.