Professional Documents
Culture Documents
Systems and Control Laboratory, Computer and Automation Research Institute HAS, H-1518 Budapest, PO box
1 Abstract
A continuous isotherm fermentor with constant volumetric holdup and physico-chemical properties is
investigated in this paper. The fermentor is highly nonlinear due to the nonlinear reaction rate function
which has a global maximum as a function of the substrate concentration.
The nonlinear analysis shows that the fermentor is unstable near the desired operating point with the
maximum biomass production but it is locally controllable there.
Therefore various stabilizing controllers, a LQR based on the locally linearized model, feedback linearization and nonlinear feedback controllers based on the Hamiltonian description of the system have
been developed and compared for the desired operating point.
It has been shown that both a well-tuned LQR and the nonlinear feedback controller can stabilize
the system in a reasonably wide operating region near the operating point. It is, however, possible to
perform stabilization using only one of the state variables, the measurable substrate concentration by a
nonlinear feedback controller based on the Hamiltonian description.
2 Introduction
The aim beside the examination of the fermentation process model properties is to design and compare control strategies that drive and keep the system in the operating point which provides optimal
production. The dynamics of the process is given by the state space model
dX
S X XF
(1)
dt
V
(S )X
(SF S )F
dS
(2)
dt
Y
V
where
= ( )
(S ) = max
S
K2 S + S + K1
2
(3)
3 Model analysis
In this section the key dynamic properties of the open-loop system, asymptotic stability and local controllability are examined with traditional linear and nonlinear techniques.
3.1
The desired equilibrium state (from the point of view of biomass production) corresponds to the maximum
of the S function. It's easy to calculate that takes it's maximum at
( )
S0 =
K1
K2
[]
[]
[ ]
[]
[]
[ ]
[]
[]
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.5
1.5
2
Substrate concentration [g/l]
Figure 1: The
= 0 2449
2.5
function
g considering the parameters given in Table 1 (see Fig. 1)). The desired equilibrium
(S0
:
l
g and the necessary nominal input feed ow rate
value of the biomass concentration is X0
:
l
l
is F0
:
.
In
terms
of
centered
variables
the
original model can be written in the following
h
input-ane form
= 4 8776
= 3 2130
= (S + S0)(X + X0) (X +VX0)F0 (X +VX0)F
(S + S0 )(X + X0 ) (SF (S + S0 ))F0 (SF (S + S0 ))F
S_ =
+
+
X_
where
F = F
F0 ; X = X
X0 ; S = S S0
A=
@f
(0) =
@x
0
1:6065
0
0:8045
B = g(0) =
(4)
(5)
(6)
1:2194
2:4388
(7)
Stability analysis based on this local linearization shows that the problem is a critical case of asymptotic analysis since the eigenvalues of the state matrix of the linearized model have negative and zero real
parts (1
:
, 2
). Furthermore, it can be seen from open loop simulations that the desired
equilibrium point is indeed unstable. This is seen in Fig. 2 where the simulation was started from close
neighborhood of the desired equilibrium, and only numerical truncation errors acted as disturbances.
= 0 8045
=0
10
9
Biomass concentration
Substrate concentration
Concentration [g/l]
20
40
60
Time [h]
80
100
120
Figure 2: Open loop simulation started from close neighborhood of the optimal equilibrium state
3.2
In this part it is shown that nonlinear controllability analysis based on the generation of controllability
distributions is extremely helpful in identifying the singular points of the state space around which control
of the system is problematic or even impossible.
O = [B
AB ] =
1:2194
2:4388
0
0:0031
(8)
we nd that the system is controllable (in the linear sense) in the neighborhood of the required operating
point.
For this, we need to identify the vector elds f and g in equations (4)-(5) if we write them in the following
general form
x fx gxu
(9)
It's clear that
and
_ = ( )+ ( )
x = xx1 = X
S ; u = F
2
S) =
f (x) = f (X;
"
(S + S0 )(X + X0 )
(10)
+X0 )F0
(X
V
(S+S0 )(X +X0 )
(SF (S+S0 ))F0
Y
V
"
#
+X0 )
(X
V
X; S
(SF (S+S0 ))
V
g(x) = g(
) =
(11)
(12)
We follow the algorithm described for example in [5] for calculating the local controllability distribution
of the model. The initial distribution is
fgg
(13)
0
= span
and
@g
f (x)
[f; g](x) = @x
(14)
@f
g(x)
@x
(15)
2 = 1 + [f; 1] + [g; 1] = spanfg; [f; g] ; [f; [f; g]] ; [g; [f; g]]g
(16)
We do not present all the calculations with the distributions here due to space limitations, only the most
important results.
Singular points
T
g
T
SF ) all the elements of
At the point X S = X0 SF S0 (X
l , S
2 (and, of course, all the elements of
0 and
1 ) are equally zero. It means that the controllability
distribution has rank at this point. Moreover, this singular point is a steady state point in the statespace. From this it follows that if the system reaches this (undesired) point, it's impossible to drive the
process out of it by manipulating
the input feed ow rate (see Fig. 2 again).
g
If X
X0 (X
)
and
S
6 SF S0 (S 6 SF ), 2 has rank . From a practical point of view
l
g then it can't be increased by changing the
it means that if the biomass concentration decreases to
l
input ow rate.
=
=
=0
=0
Non-singular
points
T
T
At any other point in the state space including the desired operating point
=
the controllability distribution has rank 2 which means that the system is controllable
in a neighborhood of these points and we can apply state feedback controllers to stabilize the process.
X S
[0 0]
4 Controller design
In this section, the three aforementioned controllers, a LQR based on the locally linearized model, feedback
linearization and nonlinear feedback controllers based on the Hamiltonian description of the system are
compared in a qualitative way.
The presented simulation results (if otherwise not stated) correspond to the initial values X
g and S
S0 : gl respectively that represent disturbances. In open loop
: X0
:
l
(i.e. without controller) these initial values lead to the undesired singular equilibrium point X
X0
and S SF S0 discussed in the previous section.
06
4.1
= 2 9266
=
(0) = 4
(0) =
=
= 0 9796
LQ control
The starting point is the well-known LQ-controller which is designed for the locally linearized model of
the process and minimizes the cost function
J (x(t); u(t)) =
Z
0
xT (t)Qx(t) + uT (t)Ru(t) dt
(17)
where x denotes the state vector, u is the manipulable input, Q and R (the design parameters) are
positive denite weighting matrices of appropriate dimensions. The optimal input that minimizes the
above functional is in the form u
Kx. It is well known from linear system theory that only the
ratio of the elements of Q and R counts when calculating the full state feedback gain K . (Therefore the
only design parameters are Q and R if the linear model is given.) The results for two dierent weighting
matrix selections are presented below.
3.5
Input feed flow rate
4.5
3
Biomass concentration
Substrate concentration
Concentration [g/l]
3.5
2.5
2.5
1.5
1.5
0.5
10
Time [h]
15
20
10
Time [h]
15
20
Q and R
Q = I22 ; R = 1
(18)
The feedback gain (K ) and the eigenvalues of the state matrix of the linearized closed loop system (E )
were
T
K
:
:
E
:
:
(19)
= [ 0 8544 0 6905]
= [ 0 8034
2 7270]
As it is visible in Fig. 3 the linear feedback stabilizes the system quite well.
Q = 10 I22 ; R = 1
(20)
The feedback gain and the eigenvalues of the state matrix of the linearized closed loop system were
K=[
1:8971 2:5866]
E=[
0:8035
8:6225]T
(21)
The results for this case can be seen in Fig. 4. As it is visible there is no signicant dierence in the
performance of the two LQ-controllers but the energy of the necessary input signal is higher in the second
case.
4.2
A nonlinear technique, feedback linearization is applied here for changing the system dynamics into linear
one.
We chose the substrate concentration S as output variable i.e.
y = h(x) = x2 = S
(22)
To linearize the input-output behavior of the system, the following state feedback can be applied (see
e.g. [5])
u x xv
(23)
= ( )+ ( )
5
4
Input feed flow rate
4.5
3
4
2
Biomass concentration
Substrate concentration
Concentration [g/l]
3.5
2.5
1.5
1
2
0.5
10
Time [h]
15
20
10
Time [h]
15
20
where
(x)
(x)
Lf h(x)
Lg h(x)
(24)
1
Lg h(x)
(25)
Lf h(x) =
1] f (x) = f2(x)
1] g(x) = g2(x)
(26)
(27)
It's easy to check that with this full state feedback the system becomes an integrator from an input-output
point of view i.e.
y v
(28)
_=
And from this point we can apply any well-known linear controller for stabilizing the system. To show
the most important dynamic properties of the closed loop system we chose the following simple pole
placement controller
v
ky
(29)
Two simulation results are presented for feedback linearization control: the rst can be seen in Fig. 5
and corresponds to k
while the second is in Fig. 6 with a value of k
. As it is visible the chosen
output (the substrate concentration) shows a simple rst order response with a time constant determined
by k , and the biomass concentration also remains stable. Here we can arbitrarily prescribe the dynamics
of the input-output behavior of the closed loop system but a major drawback of the method is the lack
of robustness (because of the sensitivity of the functions and ).
=1
4.3
= 10
In this approach we design a nonlinear full state feedback that shapes the generalized energy function of
the process.
3.5
Input feed flow rate
4.5
Biomass concentration
Substrate concentration
Concentration [g/l]
3.5
2.5
2.5
1.5
1.5
0.5
10
Time [h]
15
20
10
Time [h]
15
20
4
Input feed flow rate
4.5
3
4
Biomass concentration
Substrate concentration
Concentration [g/l]
3.5
2.5
2
0
1.5
1
1
0.5
10
Time [h]
15
20
10
Time [h]
15
20
@H (p; q; u)
;
i = 1; : : : ; n;
(30)
@pi
@H (p; q; u)
p_i =
;
i = 1; : : : ; n:
(31)
@qi
is called a Hamiltonian control system ([6]) where q is the vector of generalized coordinates (state variables)
and the components of p are the generalized momenta (co-state variables).
q_i
In some of the cases the Hamiltonian function is in the special, input ane form
m
X
H q; p; u H0 q; p
Hj q; p uj
j =1
where
H0
tonians.
)= ( )
( )
(32)
yj =
@H
(q; p; u) = Hj (q; p) ; j = 1; : : : ; m
@uj
(33)
Notice that in the original denition of Hamiltonian control system there were no output variables present.
The natural outputs are articially dened new output variables to the system, even their number is
changed to be equal to the number of input variables m.
With the above denitions the nonlinear state-space model of an ane Hamiltonian input-output
system is in the form:
m
X
@H0
@Hj
qi
q; p
q; p uj ; i ; : : : ; n
(34)
@pi
@pi
j =1
m
X
@H0
@Hj
pi
q; p
q; p uj ; i ; : : : ; n
(35)
@qi
@qi
j =1
yj
Hj q; p ; j ; : : : ; m
(36)
_ =
( )
_ =
( )
( )+
( )
=1
( )
=1
=1
Note again, that these dentions are extension to the original mechanically motivated ones.
A simple Hamiltonian system is an ane Hamiltonian system where the functions H0 , H1 , . . . , Hm
are of the following special form:
H q; p
pT G q p V q
(37)
0
with
( ) = 12
() + ()
(38)
With the above requirements on the internal and coupling Hamiltonian the nonlinear state-space model
q_i
p_i
yj
@H0
(q; p) = G(q)p ; i = 1; : : : ; n
@pi
m
X
@H0
@Hj
(
q; p) +
(q) uj
@qi
@qi
j =1
Hj (q) ; j = 1; : : : ; m
(39)
(40)
(41)
The above simple Hamiltonian systems are special compared to the original, mechanically motivated
one in two main points:
H0 (p; q)
Hj (q)
p = [
T (1)
]
P T (C) ]T
:::
q = [P T (1) : : :
(42)
(43)
This means that the generalized momenta are the normalized conserved extensive variables and the
generalized co-ordinates are the normalized thermodynamical driving forces (see [4] and [3] ).
In addition to the state space model equations, there is a linear static (i.e. time invariant) relationship
between the state and co-state variables of a process system in the form:
q = Qp
where
(44)
p=
p1
p2
q=
where
mX
mS
q1
q2
=
p_2 =
mX0
mS0
X
S
=
m
X
m
S
(45)
(46)
q1 )F
(SF (S0
q2 ))F
(47)
(48)
Then the coupling Hamiltonians (natural outputs) used for feedback are calculated by simple integration
@H1
@q1
@H2
@q2
= X0
= (SF (S0
q1 ) y1 = H1 (q) = X0 q1
q2 )) ) y2 = H2 (q) =
12
2 q1
(49)
(50)
In this case the feedback consists of the linear combination of the calculated natural outputs
F = k1 y1 k2 y2
(51)
=1
F = k2 y2
(52)
4.5
2
4
Biomass concentration
Substrate concentration
Concentration [g/l]
3.5
2.5
2
4
1.5
1
6
0.5
10
Time [h]
15
20
10
Time [h]
15
20
Figure 7: Typical result of Hamiltonian-control using all natural outputs: state variables and input
4
Input feed flow rate
4.5
4
2
Biomass concentration
Substrate concentration
Concentration [g/l]
3.5
2.5
1.5
2
1
3
0.5
10
Time [h]
15
20
10
Time [h]
15
20
Figure 8: Typical result of Hamiltonian-control using only one natural output: state variables and input
10
Comparison viewpoints
If we want to compare dierent approaches for controlling the continuous fermentor, the following specialities of the control task should be taken into account:
The system is highly nonlinear near the chosen operating point with a maximum type nonlinearity.
This implies that the sign of the key partial derivatives is changing with the state of the system.
The system exhibits an unusual, saddle point type unstability at the operating point. For some
disturbances it shows an asymptotically stable response, for other it does not, depending on the
kind (direction) of the disturbance.
One of the two state variables, namely the biomass concentration is not measurable on-line, therefore
some kind of observers should be developed in addition when full state feedback controllers are
applied.
The system model is only considered to be an approximate one, with high structural and parametric
uncertainty. Therefore only such controllers can be used which are robust with respect to the model
parameter uncertainties.
The above specialities imply that robust, nonlinear, possibly not full state feedback techniques are
preferred for this special task.
5.2
Qualitative comparison
Our simulation results show, and the sample results presented here indicate that all the three applied
techniques are able to stabilize the fermentor near the unstable but preferred operating point. They are,
however, much dierent from the important comparison viewpoints above.
1. Nonlinearity
Surprisingly enough, a well-tuned LQR has roughly the same performance in stabilizing the operating point as the other two special nonlinear controllers. The nonlinearity of the system is taken
into account with the nolinear transformation in case of the feedback linearization and with the
kind of nonlinear feedback for the Hamiltonian controller.
2. Operating region
The operating region, that is the region in the state space from which stabilization is possible, varies
with the type of controller and also with the tuning.
Experimental simulation evidence shows that the "cheap" tuning of the LQR controller has much
wider operation region than that of the "expensive" version. However, there were cases, i.e. starting
points, for both of the LQR controllers from which they were unable to lead the system to the desired
operating point.
Both of the Hamiltonian controllers could move the system from any technically meaningful starting
point to the desired operating point.
In conclusion we can say, that the operation region is wide enough for practical applications for
both the LQR and the Hamiltonian controllers.
3. Robustness
The robustness of the controllers with respect to structural and parametric uncertainties of the
process model is a key issue in bioreactor control. This problem needs further work for the simple
continuous fermentor since we only have a very limited number of simulation experiments about it.
Again, both the LQR and the Hamiltonian controllers seem to be robust enough for our purposes.
The lack of robustness, however, makes the application of the feedback linearization technique
unacceptable for controlling bio-processes.
11
4. Observers
In case of full state feedback controllers, like the LQR controller, the feedback linearization approach
combined by pole placement controller and the Hamiltonian controller using all natural outputs
require to use an observer because only the substrate concentration can be measured. This causes
an additional diculty in the realization of the above controllers because the theory of nonlinear
observers is far from being trivial.
In the case of Hamiltonian controllers using a single natural output only, there is no need for
applying full state feedback and therefore observers.
6 Further research
There are two main directions which call for further research.
1. Controller tuning
Our simulation experiments indicate that both the controller performance and the operating region
does depend on the controller tuning. Theoretically grounded and practically applicable tuning
methods are needed both for the LQR and the Hamiltonian controllers.
2. Robustness
We need to carry out further simulation investigations to nd out the factors aecting the robustness
of the potential controllers for the continuous fermentor. Thereafter we should nd the theoretical
tools for analyzing the robustness problem for the more general case.
Acknowledgement
This research was supported by the Hungarian Research Fund (OTKA) through grant No. T032479.
References
[1] C. I. Byrnes and A. Isidori. Asymptotic stabilization of minimum phase nonlinear systems. IEEE
Trans. Automat. Contr., 36:11221137, 1991.
[2] C. I. Byrnes, A. Isidori, and J. C. Willems. Passivity, feedback equivalence, and the global stabilization
of minimum phase nonlinear systems. IEEE Trans. Automat. Contr., 36:12281240, 1991.
[3] K. M. Hangos, A. A. Alonso, and J. Perkins. A thermodynamic approach to structural stability of
process plants. AIChE Journal, 45: , 1999.
[4] Katalin Hangos, Jzsef Bokor, and Gbor Szederknyi. Analysis and control of nonlinear process
systems. Technical report, Computer and Automation Research Institute, Hungarian Academy of
Sciences, 1999.
[5] Alberto Isidori. Nonlinear Control Systems. Springer, Berlin, 1995.
[6] Arjan van der Schaft. L2-Gain and Passivity Techniques in Nonlinear Control. Springer, Berlin, 1996.
12