3rd International Conference on Electrical and Information Technologies ICEIT’2017

Multimodel Modelization and Stabilization of a

Bioreactor
Abdelmounaim KHALLOUQ , Asma KARAMA
LAEPT-URAC 28, Faculty of Science Semlalia, P.B. 2390
University Cadi Ayyad, Marrakech, Morocco
karama@uca.ac.ma

Abstract—This paper present a modeling approach of a Lyapunov function has a non-quadratic form which is resolved
bioprocess based on a Takagi-Sugeno formulation (TS). Using too. Finally a simulation is given to prove the effectiveness of
the resulted TS model, we construct a Parallel Distributed the method.
Compensation control law (PDC) to ensure the convergence of
the state to a specific equilibrium points. The performance of
this approach is illustrated by means of a simple example of a II. TAKAGI -S UGENO F UZZY M ODELS
biomass growth process. The T–S fuzzy approach main is to describe the comport-
Index Terms—Takagi-Sugeno systems, parallel distributed ment of a nonlinear system by a finite number of local linear
compensation (PDC), bioprocesses, constraints subsystems inside different operating regions. And it’s de-
scribed by “IF-THEN” rules which represent local input/output
relations of a nonlinear system. There is three methods to
I. I NTRODUCTION
construct T-S fuzzy models [6]:
Living cells and their components play a key role within • Black box identification.
biotechnology industry. And the modeling of the bio-processes • Linearization method.
is a delicate and a fundamental step in any control law in • Nonlinearity sector methods.
order to ensure optimal working of such bioprocesses. The We are interested by the third one. We consider a nonlinear
lack of appropriate sensors has motivated the elaboration of system where the main feature is that the control input u(k)
mathematical models for the estimation of various process occurs linearly in the equation relating inputs to states. It can
parameters and states and may be helpful to describe the be represented by an equation of the form:
biological phenomena which intervene in these process [1].
Therefore the obtained models are highly nonlinear and very ẋ(t) = f (x(t)) + g(x(t))u(t) (1)
complex and the task of designing control laws and diagnostic f (.) and g(.) are two nonlinear functions.
algorithms become more difficult to achieve. In the recent To transform the model (1) to a T–S fuzzy model, the
years, the Takagi-Sugeno (T-S) structure, introduced in [2], following rules are:
provides a good way to represent nonlinear systems thanks Rule i :
to its reduced mathematical complexity and its capabilities IF z1 (t) is M i1 and . . . and zp (t). is Mip ,
to describe a large class of nonlinear systems. This paper 
aims is to design a stable nonlinear controller based on fuzzy ẋ(t) = Ai x(t) + Bi u(t)
controllers for a bioprocess. For the control of these processes T HEN (2)
y(t) = Ci x(t) i = 1, 2, . . . , r
the authors, in [3] used the identification to find the Takagi-
Sugeno model to approximate the nonlinear system in [4] the Here, Mij is the fuzzy set and r is the number of model rules;
neural network was used for the identification of the nonlinear x(t) ∈ Rn is the state vector, u(t) ∈ Rm is the input vector,
system, in [5] the linearization method is used to find the T- y(t) ∈ Rq is the output vector, Ai ∈ Rnxn , Bi ∈ Rnxm , and
S model in this work the Takagi-Sugeno approach is used Ci ∈ Rqxn ; z1 (t), . . . , zp (t) are known premise variables
to approximate the nonlinear system of a biomass growth that may be functions of the state variables. Let note:
process. Once the fuzzy model is obtained, the control design 
is carried out via the PDC scheme ensuring stabilization at ωi (z(t)) = Πpj=1 Mij (zj (t)), i = 1, 2, . . . , r
a selected equilibrium point and taking into account of the , t ≥ 0 (3)
ωi (z(t)) ≥ 0
process constraints. The problem we deal with it her is that
the direct uses of the lyapunov function to find the Controller The global model is given by:
gains lead to a 0 as equilibrium points which mean the leaching 
r

of the bio-reactor, in the literature to counter this problem ẋ(t) = hi (z(t))(Ai x(t) + Bi u(t)) (4)
we transform the stabilization problem to a tracking control. i=1

Her we use a transformation to resolve the problem and get where

the stabilization around a specific equilibrium points, another ωi (z(t))
problem due to the transformation is that the derivative of the hi (z(t)) = r (5)
i=1 ωi (z(t))

978-1-5386-1516-4/17/$31.00 2017
c IEEE
 3rd International Conference on Electrical and Information Technologies ICEIT’2017

paramètres valeur unité

where μmax h−1
 0.38
r
hi (z(t)) = 1 Ks 5 g/l
r
i=1
(6) k1 1/0.07
0 ≤ i=1 hi (z(t)) ≤ 1 Deq2 0.3 h−1
Sineq 120 g/l
The membership function hi (z(t)) indicates the activation max
Sin 200 g/l
degree of the ith associated local model. Xeq2 7.0875 g/l
Seq2 18.75 g/l

III. B IOPROCESS MODEL Table I: Simulation parameters

The dynamic of the biological process is established from
the mass-balance and can be described by the following Xeq and Seq are given by resolving these equations:
equations: 
dX (μ − D)Xeq = 0
= μ(.)X − DX (11)
dt −k1 μXeq + D(Sin − Seq ) = 0
dS (7)
= −k1 μ(.)X + D(Sin − S) The resolution of these equations lead to tow equilibrium
dt
points given by :
where : X and S are the concentrations in biomass and
• Solution 1 :
substrates and are the state variables, μ(.) is the specific 
growth rate, k1 is the pseudo stoichiometric coefficient and Xeq1 = 0
(12)
D and Sin are respectively the dilution rate and the influent Seq1 = Sin
concentrations of S and which are the system inputs
This solution corresponds to the wash of the bioreactor in other
It can clearly be seen from equations (7) that the specific words the interruption of the microbial growth. Thus it’s not
growth rate μ(.) is a key parameter for the description of the good equilibrium points.
biomass growth and substrate consumption . Many empirical • Solution 2:

expressions have been proposed to model μ(.). for simplicity Seq2 = μmaxKs D
−D
a "Monod law", which expresses the dependence of μ(.) on S −S (13)
Xeq2 = in k1 eq2
the substrate concentration S is chosen and is described as
follows: Thus, for every doublet D, Sin we have a set of equilibrium
S points. To avoid that the control law stabilyse the system(7)
μ(S) = μmax (8) arround the first equilibrium, we make a variable change as
Ks + S
bellow:
where μmax is the maximum specific growth rate; Ks is the ⎧
saturation constant. ⎪
⎨Xt = X − Xeq2
The biomass growth model (7) can be rewritten like an St = S − Seq2 (14)
⎪
⎩
affine model (1) using the following transformation: If we ut = u − ueq2
assumed that: 
D1
where u = , and replacing (13) into (7), we obtain the
D = D1 + D2 D2
(9) following system (15):
Sin = D2 max
D1+D2 Sin

where D1 (t) is a water dilution rate and D2 (t) is the ẋt (t) = f (xt ) + g(xt )ut (t) + G(t) (15)
substrate dilution rate, the affine form (1)can be obtained by where the added term G(t) is a result of the transformation
replacing D(t) and Sin (t) in (7) by their expressions (9): and :
⎡ dXt ⎤ ⎡ ⎤
⎡ dX ⎤ ⎡ ⎤ ⎡ ⎤⎡ ⎤ D1 − D1eq2
−X −X dxt dt
dt μ(S)X D1 =⎣ ⎦ ut = ⎣ ⎦
⎣ ⎦=⎣ ⎦+⎣ ⎦⎣ ⎦ (10) dt dSt
D2 − D2eq2
dt
dS
−k1 μ(S)X −S max
Sin −S D2
dt 
μ(S) − Deq2
f (xt ) =
−k1 μ(S) − Deq2
IV. T RANSFORMATION O F THE N ONLINEAR S YSTEM INTO
⎡ ⎤
A TS M ODEL −(Xt + Xeq2 ) −(Xt + Xeq2 )
A. Equilibrium points g(xt ) = ⎣ ⎦
Before transforming the nonlinear model, we need to com- −(St + Seq2 ) max
Sin − (St + Seq2 )
pute the equilibrium points of this model because in the ⎡ ⎤
following section to control the system, we need to use the (μ(S) − Deq2) ) ∗ Xeq2
Lyaponuv theory which ensure the convergence to a null G=⎣ ⎦
equilibrium point. (−k1 ∗ μ(S) ∗ Xeq2 + Sin
max
∗ D2eq − Seq2 ∗ Deq2 )
 3rd International Conference on Electrical and Information Technologies ICEIT’2017

B. The TS model design

The non-linearities considered in the model (15) are:

z1 (x) = μ(S) (16)

z2 (x) = −Xt (17)
z3 (x) = −St (18)

The global model is given by :


r
ẋt (t) = hi (z(t))(Ai xt (t) + Bi ut (t) + Gi ) (19)
i=1

To determine the subsystems(Ai , Bi , Gi ) the following rules

are used :

• Model Rule i:
IF z1 (t)is Mi1 and . . . and zp (t) is Mip , Figure 1: Validation test
then:

ẋt (t) = Ai xt (t) + Bi ut (t) + Gi i = 1, 2, ..., 8 (20) V. S YNTHESIS O F TS C ONTROL L AW

The next step is to design a stabilizing time-varying state
Here, Mi,j is the fuzzy set described in section II feedback controller ensuring the stability of the system(15).
Using the parameters summarized in Table I, we obtain the The output of the PDC control is the following form:
following results:

r
⎡ ⎤ ut = hi (z(t))Ki xt (t) (21)
0.0800 0 i=1
A1 = A2 = A3 = A4 = ⎣ ⎦
−5.4286 −0.3000 Using the quadratic Lyapunov function:
⎡ ⎤
−0.2393 0 V (t) = xTt P xt , P = P T > 0 (22)
A 5 = A6 = A7 = A8 = ⎣ ⎦
−0.8671 −0.3000 Lemma 1. Consider two matrices X and Y with appropriate
⎡ ⎤ dimensions and σ a symmetric positive definite matrix. The
−9.1804 −9.1804
B1 = B5 = ⎣ ⎦ following property is verified [7]:
−140.0000 60.0000 X T Y + XY T ≤ X T σX + Y T σ −1 Y (23)
⎡ ⎤
−9.1804 −9.1804
B 2 = B6 = ⎣ ⎦ Theorem 1. There exists a time-varying state feedback con-
−0.9506 199.0494 troller 21 for system 15 ensuring that the system state con-
verges toward an origin-centred ball of radius bounded by β if
there exists P = P T > 0, Qi,j and a scalars σi > 0 solutions
⎡ ⎤
−3.2106 −3.2106 of the following optimization problem [7]:
B3 = B7 = ⎣ ⎦
min β
−140.0000 60.0000 P,Qi,j ,σi
⎡ ⎤
−3.2106 −3.2106 ⎡ ⎤
B4 = B8 = ⎣ ⎦ Qi,j I
−0.9506 199.0494 ⎣ ⎦<0 (24)
⎡ ⎤
0.5670 I −β
G1 = G2 = G3 = G4 = ⎣ ⎦
with
−8.1000 ⎡ ⎤
⎡ ⎤ P −1 Fi,j
T
+ Fi,j P −1 I
−1.6960
⎣ ⎦ Qi,j = ⎣ ⎦ (25)
G5 = G6 = G7 = G8 =
I −σi
24.2291
and
The behavior of the developed TS model is compared to the Gi (t)T σi Gi (t) < β (26)
bioreactor model and the result is shown on the below figure:
for i, j = 1, 2, ..., r
 3rd International Conference on Electrical and Information Technologies ICEIT’2017

Proof: A. Constraints On Control Input

By derivation of 22 and by substituting 21 in 15: We note here that we deal with an real process and we have
a constraint on the control law D < μmax and Sin ≤ Sin max
.
V (t) = xTt P xt To ensure that, we use the following theorem [8] :
˙ = x˙t T P xt + xT P x˙t
V (t)
˙ = r hi r hj [(Fi,j xt (t) + Gi )T P xt
t
Theorem 2. Assume that the initial condition x(0) is known.
V (t) i=1 j=1 The constraint ||u(t)||2 ≤ d is enforced at all times t ≥ 0 if
+xTt P (Fi,j xt (t) + Gi )]
the LMIs:
≤ xTt (Fi,j
T
P + P Fi,j )xt + GTi P xt + xTt P Gi ) ⎧⎡ ⎤
⎪
⎪ 1 xTt0
⎪
⎪ ⎢ ⎥
with :Fi,j = Ai + Bi Kj ⎪
⎪ ⎣ ⎦≥0
⎪
⎪
⎪
⎪ 2
⎪ i
⎪ Y d I
Here as we said that derivative of the Lyapunov function has a ⎪
⎨
non-quadratic form this to resolve this problem we use lemma ⎡ ⎤ (34)
⎪
⎪ P 1 Y T
1: ⎪
⎪⎢
i
⎥
⎪
⎪ ⎣ ⎦≥0
⎪
⎪
˙ ≤ xT (F T P + P Fi,j + P σ −1 P )xt + GT σi Gi ⎪
⎪
V (t) t i,j i i (27) ⎪ xt0 P 1
⎪
⎪
⎩
Let us define
hold, where P 1 = P −1 and Yi = Ki P 1.
⎧
−1
⎪
⎨Qi,j = Fi,j P + P Fi,j + P σi P
T

 = minλ(−Qi, j) (28)
⎪
⎩
δ = maxGTi σi Gi

using 28 the expression 27 can be increased and we have:

V̇ (t) < −||x||2 + δ . It follows that V̇ (t) < 0 for:

Qi,j < 0
(29)
−||x||2 + δ < 0

So we guarantee that the state is bounded and converges to a Figure 2: TS fuzzy control bloc diagram

small origin-centered ball of radius δ .
Qi,j < 0 is equivalent to VI. SIMULATION RESULTS
Programming theses LMIs with x0 = (5 9) and d = 0.5
P −1 Fi,j
T
+ Fi,j P −1 + σi−1 < 0 (30) we obtain:

Remark 1. Using the variable changes β =

2868.6
⎡ ⎤
 40.7 315.7
P 1 = P −1 P1 = ⎣ ⎦
(31)
Y i = Ki P 1 315.7 5801.3
⎡ ⎤
Condition 30 can be easily linearized. Using Schur’s com- 0.0635 −0.0016
plement to 30, it becomes K1 = ⎣ ⎦
0.1093 −0.0052
⎡ ⎤
P −1 Fi,j
T
+ Fi,j P −1 I ⎡ ⎤
⎣ ⎦<0 (32) 0.0636 −0.0016
I −σ K2 = ⎣ ⎦
0.1094 −0.0052
If 32 is satisfied for all i = 1, ..., r, j = 1, ..., r
and −||x||2 + δ < 0 then V (t)˙ <0 K 3 = K4 = K5 = K6 = K7 = K8 = K2
From 24,26 with a Schur complement, it follows that
The stabilization around the equilibrium points is obtained and
1 the constraints on the control law are respected as shown in
−Qi,j > I, i = 1, ..., r, j = 1, ..., r (33) figures 3, 4.
β
Both the biomass and the substrate concentrations join the
1 1 desired equilibrium states and the obtained control variables
Then all the eigenvaluues of −Qi,j are great then . So <  are accepted since the constraints on both the dilution rate D
β β
hold and the radius is bounded by β and the influent concentration Sin are respected.
 3rd International Conference on Electrical and Information Technologies ICEIT’2017

[5] K. Turki and O. Boubaker. Modelisation Multimodele D’un Procede De

Depollution. International Conference On Sciences an Techniques of
Automatic control and computer engineering,2003
[6] A. Benzaouia and A. El Hajjaji. Advanced Takagi–Sugeno Fuzzy Sys-
tems, Springer, 2014.
[7] S. Bezzaoucha, B. Marx, D. Maquin, J. Ragot. Stabilization of Nonlinear
Systems Subject to Actuator Saturation, IEEE International Conference
on Fuzzy Systems (FUZZ), 2013
[8] K. Tanaka and H.O. Wang. Fuzzy Control Systems Design And Analysis,
John Willey And Sons, INC., 2001.

Figure 3: System States

Figure 4: Control Law Signal

VII. CONCLUSIONS
In this paper, we deal with the stabilization problem of a
bioprocess. The process is described by a nonlinear equations
obtained using the mass-balance. This model is transformed
to a fuzzy T-S model. The stabilization deals are guaranteed
by a PDC control law respecting the inputs constraints.

R EFERENCES
[1] Bastin G. and Dochain D. (1990). On-line estimation and adaptive
control of bioreactors. Elsevier, Amsterdam.
[2] T. Takagi and M. Sugeno. Fuzzy identification of systems and its
applications to modeling and control. IEEE Transactions on Systems,
Man, and Cybernetics, 15:116–132, 1985.
[3] V. H. Grisales, A. Gauthier, and G. Roux.Fuzzy Optimal Control
Design for Discrete Affine Takagi-Sugeno Fuzzy Models: Application
to a Biotechnological Process. IEEE International Conference on Fuzzy
Systems 2006
[4] A. Karama. Contribution à la modélisation et la commande par réseaux
de neurones des bioprocédés, thesis,Université Cadi Ayyad, Faculté des
Sciences - Semlalia, Marrakech, 2004