Sunteți pe pagina 1din 6

Adaptive Control and On-line Controller Design for Multivariable Systems

in State-Space
Istvn Harmati

Bla Lantos

Department of Control Engineering and


Information Technology
Technical University of Budapest
H-1117 Budapest, Pzmny Pter stny 1/D.
Hungary

Department of Control Engineering and


Information Technology
Technical University of Budapest
H-1117 Budapest, Pzmny Pter stny 1/D.
Hungary

E-mail: harmati@seeger.iit.bme.hu

E-mail: lantos@iit.bme.hu

Abstract - In many practical cases, the system has unknown or


partially unknown parameters making the control more
difficult. The use of adaptive control strategies is one of the
most promising approaches to overcome this problem. The
paper deals with the indirect adaptive control of slowly
varying multivariable (MIMO) linear systems in the statespace. Our results are intended to exhibit the effectiveness of
the on-line identification and to compare the different control
algorithms. In connection with this subject, the possibilities of
real-time realization are also investigated.

decoupling of MIMO systems is a standard problem in the


control engineering. Our goal is to find a state feedback in
order to reduce the system to independent SISO channels.
The algorithm consists of different steps. First time, one
must check the Gilberts condition for decoupling and find
the unstable decoupling zeros based on the canonical form
of the system decoupled in integrator sense. Sometimes it
may be necessary to design a parallel compensation for
removing the unstable decoupling zeros in the model
and/or to perform a precompensation, if the Gilberts
condition is not satisfied. Our method is the combination
of some early results of Gilbert [3], the controllability
condition of Haustus and some new results of Lohmann
[4], extended with modifications needed for numerical
stability of the design steps.
The paper is arranged as follows. Section II gives the
survey of the concept and theory on the area of the applied
adaptive control algorithms. A similar approach can be
found in [1]. The reader may find some deeper
interpretations and a more unified approach in [5]. The
results are presented in the frame of an illustrative example
in Section III together with the demonstration of different
control strategies. One of our main intentions is to compare
their behaviors. The identification has a distinguished role
in the whole discussion because the control performance
depends above all on the precision of the model.
It is evident in the realization that the on-line concept
requires suitable devices and quick, simple numerical
algorithms to fulfill the realtime specifications. On-line
controller design methods for slowly varying MIMO
systems are implemented in MATLAB. Section IV
introduces the ACE Kit of dSPACE as a possible
alternative for the real-time implementation. The ACE
KIT is based on the board DS1102 and has a Real-Time
Interface (RTI) to assure the connection with the
MATLAB environment. More information about the
total development environment is described in [6].

I. INTRODUCTION
In the more general applications, there are not given
enough information of the process. However, some popular
techniques suggest successful treatments for this problem
including also the different adaptive methods. Indirect
adaptive control describes a body of theoretical and
practical engineering principles in which one brings into
union the model-based development of feedback
controllers with the on-line identification based on the
recursive system parameter estimation and the state
estimation. The identification is accomplished in a
common loop in every sampling instant. The identified
model may be given in the Luenbergers controllability
normal form where the structure indices can be determined
in separate off-line experiments. The estimated parameters
are connected to the corresponding system parameters and
the parameters of the innovation noise model. The
parameter and the state estimation are carried out parallel,
which rely on the state-space description of the linear
MIMO system. Optimal estimation needs the computation
of the gradient of the estimation error, which is related to a
simple matrix recursion and the computation of the
derivatives of system matrices.
After the system identification, various type of control
algorithms can be chosen. In our investigation on-line
controller design methods are developed and implemented
including pole placement, LQ optimal control, predictive
control and decoupling. In order to avoid the often
occurred numerical and stability problems in state-space
predictive control, it is advantageous to apply numerically
robust and stable predictive methods based on the results
of Rossiter, Kouvaritakis and Rice [2]. The improved
numerical conditions allow us to get a powerful tool for
many multivariable linear systems.
Each subsystem of MIMO system influences on the
other one. Of course, this phenomenon needs a more
sophisticated treatment in the control system. The

II. A SURVEY OF THE APPLIED THEORY


In this section we give the theoretical background of
the implemented methods. Subsection II-A sketches the
control system structure. The following subsections
summarize briefly the theory of identification, pole
assignment, LQ optimal control, predictive control and the
decoupling. The final subsection formulates the steps of
the controller design.

A. Control system structure

\ W +  = & [ W + 

The control system architecture is shown in Fig. 1. It


can be seen that an explicit separation between
identification and control is assumed. A known model for
the controller design is needed in the technique of adaptive
control. Since the plant parameters and the states are in fact
unknown, they are obtained using a recursive parameter
and state estimation algorithm.

The obtained results allow us to establish a model in state


space. Further on, this model can be regarded as a suitable
substitution of the real system. In our implementation, the
identification obtains the model in Luenbergers
controllability form, where known controllability indices
are supposed.

(11)

C. Pole assignment
Several algorithms are available for multivariable
feedback design based on pole assignment. One of the
methods uses the Luenbergers controllability form. The
advantage of the approach is that the identification supplies
immediately this normal form. If the characteristic
polynomial of the specified closed loop is

F = V U + D U  V U  +  + D 

(12)

then we can introduce the notations

Fig.1. The scheme of indirect adaptive control

The estimated parameters and states enable us to make


a model, which is the basis of the controller design. For
this reason, the key issue of the whole concept is the
performance of the identification. The different controller
design methods are elaborated in state-space. Additionally,
one also has to take care of the static behaviors.

$ F

B. Identification

$/8( is the matrix consisting of general rows in $

The aim of identification is to obtain the estimated


model parameters and states from the input and output of
the system. The algorithm is described in several steps.
The algorithm of recursive parameter estimation has to be
executed in each sampling instant
7

W = \ W W W  = \ W \ W

[ , + W 3 W  W ]
W ]3 W  

. W = 3 W  7 W

3 W = , . W

W = W  + . W W

(1)
(2)
(3)
(4)

then the algorithm can be continued with the steps of state


estimation as

W $%&/
G%
G/
G$

0. =
[ W +
X W +
W
G
G
G

G&
9. =
[ W
G
:. = $ / & :. + 0 . / 9.
[ W +  = $ [ W + % X W + / W
G&
9. =
[ W + 
G
7 W +  = 9. + & :. +



=


D 

(6)
(7)
(8)
(9)
(10)

(13)

%/8( is the matrix consisting of general rows in %


a
$ F is the matrix belonging to general rows of $
where $ , % are the matrices of the identified model.
In this case, the state feedback is

a
a

. /8( = %/8(
$/8( $F

(14)

D. LQ optimal control
The quadratic regulator problem is to find the control
input to minimize the cost functional

) [ =

 1
{ 4N [N  [N + 5N XN  XN
 N =

(15)

where 4  5 >  . The solution has rich and varied


literature. Detailed discussion can be found for example in
[7]. LQ optimal control is the one of most popular methods
in linear control systems.
N

(5)




 O





D L D U 

E. Predictive control in state-space


Stable generalised predictive control (SGPC) is a class
of the most powerful methods in control engineering. Our
methods are based on results of Kouvaritakis et al.[2]
where some numerically robust predictive methods

improving the earlier algorithms are developed. Let the


model be given by

xk +1 = A xk + B uk
y k = C xk

(16)

where x n ,u r , y m . It is assumed that an


integrator is included in the model so that u is a vector of
incremental input. Furthermore, define the terminal
constraint (which is one of many possible cases)

yi 0, i , ui = K q xi , i N

(17)

where u , y denote predicted values of inputs and outputs at


the current sampling instant and . T is an LQ optimal
controller. The key idea is to attain a control law, which
satisfy the terminal constraint and at the same time
minimize the cost function

J =
i =1

{ y

+ ui 1

(18)

(20)

Di := Ci Ad i B

(21)

If the Gilbert matrix ' = '7 '7 L 'P7 has


maximal rank then the Gilberts condition is satisfied and
the system can be decoupled. In this case, the following
step provides a decoupled system in integrator sense by
means of static state feedback. At this point, it is possible
to find a transformation, which gives canonical form of the
system. The canonical form manages to discover the
unstable decoupling zeros. If they exist then a parallel
compensation is desired. Let us now consider again the
Gilberts condition. If it is not satisfied then the planning
of an appropriate precompensation is unavoidable at the
current sampling instant. The additional dynamic system
(as a dynamic feedback) may cause difficulties and stateinconsistency in the on-line algorithm, since the system
order with precompensation differs from that without
precompensation.
G. Design algorithms

with respect to uk (for all k = 0,1,K ). The solution can be


easily obtained if the Lagrange multiplier is introduced
into the performance index (18). The main weakness of
this simple approach in itself is the poor numerical
conditions for unstable plant, which can cause inaccuracies
in computations. The reason is that some powers of the
matrix $ are present in the computations. For large
horizon, these powers can be small and large at the same
time. Consequently, the results may lead to the inversion of
ill-conditioned matrix. An ingenious proposal brings into
use the new control law

K q xi + ti , i = 0,K, N 1
u =
K q xi , i N

(19)

where the above mentioned . T gives a stabilizing state


feedback, i.e. = $ % . T is stable. The new form is
powerful because it implies the stability (because of the
stable ) and the remaining free parameters W are able to
assure the fulfillment of terminal constraint.
L

F. Decoupling
In order to eliminate the undesirable effects between
channels, it is necessary to apply more intricate control
laws. We refer to [3], as the foundation of the algorithm
used in our implementation. For the sake of more unified
approach, this method is extended with some new results
based on [4] and inserted in the indirect adaptive control
environment.
The first step for the controller design is to check the
Gilberts condition for decoupling. Let Ci be the ith row
of C and define

0
if
Ci B 0

di :=
k
max l n 1 : Ci A B = 0, k = 0,1,..., l 1

In spite of the use of different control methods, the


steps are similar in the general concept. This subsection is
intended to give an overview of the sequence of design
steps in the case of predictive control:
1) Choosing the specified closed-loop system
2) Recursive parameter and state estimation
3) Realization of terminal constraint
4) Prestabilizator design
4) Calculation of the state feedback
5) Introducing the reference signal (satisfying the
static requirements)
6) Obtain the controller output
7) Limit the controller output (if necessary).
8) Limit the output signal of the identified model.
9) Refine the identification (if necessary).
Except of step 3) and 4), the methodology is the same for
every method. Step 3) and 4) depend on the current type of
controller.
III. ILLUSTRATIVE EXAMPLE
This Section describes the applied 2-input 2-output
plant model and compares the results of the simulations of
different control strategies. Identification plays a
distinguished role in the whole discussion.
A. Plant model
The plant model is defined in [1] as a turbo generator
consisting of two linear MIMO model candidates, Model 1
(M1) and Model 2 (M2) . The models are described in statespace form with innovation noise model. M2 (model 2) has
a zero outside the unit circle. The system assumed to be
one of the two model candidates at a given sampling
instant. However, the sample instant where the change is
between the model candidates is unknown for the control

strategies. Two model changes are implemented in the


simulation. The switching from M1 to M2 is at the
sampling instant 300 and the second, from M2 to M1 is at
the sampling instant 700. The whole simulation contains
1000 simulation steps.
B. Identification results during LQ optimal control

models can also cause large transient errors in the


identification. The phenomenon is illustrated in Fig. 6.
3

The results of identification during LQ optimal control


(i.e. in closed loop) are illustrated in Figs. 2-6. Denote
D (z ) the transfer function having D11 ,K, D22 elements.
For the sake of more unified approach, Figs. 2 and 3.
30

10

15

20

25

30

35

40

45

50

10

15

20

25

30

35

40

45

50

1.5
1
0.5
0
-0.5

20

10

Fig. 4. The step response of the M2 (solid) and the identified (dashdot)
subsystems
0

10

15

20

25

30

35

40

45

50

' (top) and ' (bottom) for input u1

In this case, the plant changes from model M1 to model


M2 in the sampling instant 300 and the identified model
gives incorrect parameters forcing large errors within a
small interval. After this, the identification yields again a
reliable model providing acceptable control performance.

0.4
0.3
0.2
0.1
0
0

10

15

20

25

30

35

40

45

50

Fig. 2. The step response of the M1 (solid) and the identified (dashdot)
subsystems

' (top) and ' (bottom) for input u1

0.5
0
-0.5
-1

compare the step responses of the identified model with the


current model M1 in sampling instant 999, while Figs. 4-5.
relate to identified and current model M2 in sampling
instant 699. It is clear that the identification error is greater
in the case of model M1. Obviously, the main reason is
that the gain of M1 is greater, as well. The identification of
the model M2 shows more precise results as shown in Fig.
4-5. It is beyond argument that the natures of transients in
the identified and real case are very similar providing an

-1.5

10

15

20

25

30

35

40

45

50

10

15

20

25

30

35

40

45

50

2
1.5
1
0.5
0

Fig. 5. The step response of the M2 (solid) and the identified (dashdot)
subsystems

' (top) and ' (bottom) for input u2

-5
1
-10
-15

0.5

-20
-25

0
0

10

15

20

25

30

35

40

45

50

-0.5
150

0.5

200

250

300

350

400

450

200

250

300

350

400

450

0
0

10

15

20

25

30

35

40

45

50

Fig. 3. The step response of the M1 (solid) and the identified (dashdot)
subsystems

' (top) and ' (bottom) for input u2

acceptable model for control strategies. Of course, the


identification error shows great values at the beginning of
the simulation. The extremely great values are vanished
after the transients. However, the change between the two

-0.5
150

Fig. 6. The identification error during LQ control in the neighborhood of


the model switching.

An important and interesting consequence is that the fast


varying control signal may be helpful in the identification.
Consequently, the transients of the identification are faster

but their values are greater, too. The phenomenon may


easily lead to disadvantageous output signals.
Unfortunately, this case is typical in the controlled loop.
There is not any unified approach to handle the mentioned
trade-off. An idea is to try detecting the model changes.
Since, the behaviors of the identification considerably
depend on the values of the parameter vector and the
covariance matrix P , it is a useful treatment to change
these parameters, if necessary. The detection of model
switching means such a highlighted case. An important
conclusion is that identification is a very sensitive
algorithm. The successfulness depends strongly on the true
plant parameters and on the start values in the
identification algorithm.

values compared to other control methods then the control


may lead to weaker performance. The pole assignment (not
documented in the figures) falls often in this critical class.
1
0.5
0
-0.5
-1
-1.5
150

200

250

300

350

400

450

200

250

300

350

400

450

1
0.5
0
-0.5

C. Control results

-1
150

This subsection examines the pole assignment, the LQ


optimal, the stable generalised predictive control and the
decoupling in the above-mentioned illustrative example.
To compare the results of different type of control
strategies, consider the Fig. 8-12. A single peak at 300
marks the place of model switching from M1 to M2.
1

Fig. 9. The system output (solid) and the reference signal (dashdot) for
the two subsystems by predictive control
1
0.5
0
-0.5

0.5

-1
150

200

250

300

350

400

450

200

250

300

350

400

450

0
0.5
-0.5
-1
150

200

250

300

350

400

450

1
0.5

-0.5
150

Fig. 10. The control signals for the two subsystems by predictive control

-0.5
-1
150

200

250

300

350

400

450

Fig. 7. The system output (solid) and the reference signal (dashdot) for
the two subsystems by LQ optimal control.

1
0.5
0
-0.5
-1
150

200

250

300

350

400

450

200

250

300

350

400

450

1
0.5
0
-0.5
-1
150

Fig. 8. The control signals for the two subsystems by LQ optimal control

Every result is illustrated in the same time interval with the


output and the control signals. One of the main conclusions
is that the identification is the key point of the whole
control performance that has been already discussed above.
If the control signal is very inhomogeneous and has great

To reach a hopeful compromise, many useful proposals


can bring into sight. For example, if the identification error
changes in great extent that means possibly a detected
model switching then one of the most effective solutions in
our implementation is to decrease the trace of the
covariance matrix. The main disadvantages of the pole
assignment are the strong coupling (greater than 100%),
the large transients, inhomogeneous control signal and
noise sensitivity. On the other side, this method needs the
smallest time for the realization. The LQ optimum control
provides a reliable, numerically robust approach on the
area of indirect adaptive control. The transients are more
acceptable and the control signals are relative
homogeneous. The LQ control shows considerably smaller
couple effects and noise sensitivity than the pole
assignment. At the same time, the real-time realization
requires a little more computational time. It is evident to
see that the robustness of control strategies plays (because
of the identification error) a crucial role. To improve the
mentioned property, the stable generalized predictive
control approaches suggest a good alternative. Their
robustness promises powerful results in point of view of
the transients. Additionally, the most properties are similar
to LQ optimum method, since the introduced SGPC
involves an LQ optimal controller. However, time for
realization increases forcefully with respect to horizon,

which can injure the real-time specifications. Although, the


predictive algorithms can easily lead to numerical
problems, the numerically improved SGPC algorithm is
able to avoid the troubles. In order to eliminate the couple
effects, one has to apply decoupling algorithms. In our
implementation, the nominal multivariable MIMO system
has been reduced into independent SISO systems.
Unfortunately, the plant is unknown from the view of point
of the algorithm. The effectiveness of the indirect adaptive
control based on decoupling algorithm depends strongly on
the accuracy of the identified model. An accurate model
guarantees good performance in the transients, in tracking
the reference signals and in the noise sensitivity. However,
the method is very sensitive to the model switching. The
results are demonsrateded in the Fig. 11 and 12. Because
of the complexity, the decoupling algorithm can be
regarded as the computationally most complicated method
in the real time specification.
1
0.5
0
-0.5
-1
150

200

250

300

350

400

450

200

250

300

350

400

450

1
0.5
0
-0.5
-1
150

multivariable digital controllers and real-time simulations


in various fields of control engineering. The board is based
on the Texas Instruments TMS320C31 floating point
Digital Signal Processor (DSP). All of the implemented
algorithms hold the real-time specifications. The
development environment allows us to record the time
history of the variables and tune the real-time parameter. It
is especially advantageous in the indirect adaptive control,
where the theory for choosing of parameters has not been
elaborated yet. A typical development and test session for
real-time application consist of some important steps
which can be realized by using MATLAB, Simulink,
dSPACE I/O board library, and special tools of dSPACE
TRACE and COCKPIT.
VI. CONCLUSION
Indirect adaptive control strategies provide a
powerful tool for the multivariable linear system. The key
point of the whole concept is the performance of the
identification. Different type of control methods has been
examined. The pole assignment is a quick and simple
algorithm with weak performance. The LQ controller is an
acceptable compromise between the error transients and
the real time specifications. In order to increase the
robustness, the numerically robust predictive controllers
suggest an adequate alternative. The decoupling of
multivariable systems damps considerably the effects
between the subsystems. All of the algorithms satisfy the
real time requirement in the dSPACE development
environment.
VI. ACKNOWLEDGEMENT

Fig. 11. The system output (solid) and the reference signal (dashdot) for
the two subsystems by decoupling
2
1
0

Support for the research of intelligent control systems


is provided by the Hungarian National Research Programs
under grant No. OTKA T 029072 and FKFP 0417/1997
VII. REFERENCES

-1
-2
150

200

250

300

350

400

450

200

250

300

350

400

450

1
0.5
0
-0.5
-1
150

Fig. 12. The control signals for the two subsystems by decoupling

In subsection II-F was mentioned the possibility of stateinconsistency while switching in/out the precompensator.
Theoretically, it is not guaranteed to avoid the situation.
Corresponding with our experiences, if the Gilberts
condition is satisfied for the nominal plant then it is also
valid for the identified model. A further interesting
experience that the identified model often satisfies the
Gilberts condition, even if the nominal plant does not.
IV. IMPLEMENTATION CONCEPT
The dSPACE DS1102 [6] is a single board system, which
is specifically designed for development of high-speed

[1] Nazaruddin, Y.Y.: Adaptive Regelung von Ein- und


Mehrgrensystemen
auf
der
Basis
der
Zustandsraumdarstellung.
Dissertation,
Ruhr
Universitt, Bochum, 1994.
[2] Rossitter, J.A., Kouvaritakis, B., Rice, M. J.: A
numerically robust state-space approach to stablepredictive control strategies. Automatica, Vol 34, No
1, pp.65-73, 1998.
[3] Gilbert, E.G.: The decoupling of multivariable systems
by state feedback. SIAM Journal of Control, Vol 7, No
1, 1969, pp.50-63
[4]Lohmann,
B.:
Vollstndige
und
teilweise
Fhrungsentkopplung im Zustandsraum.
Fortschritt-Berichte der VDI-Zeitschriften, Reihe
8,
Nr.244,
VDI-Verlag,
Dsseldorf.
Dissertation,
Universitt Karlsruhe, 1991.
[5] Sastry, S, Bodson M.: Adaptive Control. Prentice-Hall,
Inc. Englewood Cliffs, New Jersey,1989
[6]dSPACE DS1102 Software Environment (Reference
Guide)
[7] Kailath, T., Linear Systems, Prentice-Hall, Englewood
Cliffs, New Jersey, 1980

S-ar putea să vă placă și