Professional Documents
Culture Documents
Abstract : Fictitious reference iterative tuning (FRIT) is one of the eective data-driven tuning methods for parameters
of a controller with only one-shot experiment. This paper applies FRIT to internal model control (IMC) for linear, time
invariant, stable and non-minimum phase systems, which enables us to simultaneously obtain a desired controller and
a mathematical plant model. Here, the authors consider the case where we do not have any information on the system.
To overcome the diculty of treating the non-minimum phase behaviors of the system, Laguerre expansion is used to
describe the internal model.
Key Words : data-driven approach, fictitious reference iterative tuning (FRIT), internal model control (IMC), non-
minimum phase, Laguerre expansion.
1. Introduction
It is well known that internal model control (IMC), which is
shown in Fig. 1, is useful and eective for the attainment of
a desired tracking property in many practical applications [1].
With an internal model P implemented in parallel to the ac-
tual plant P, the controller C I MC compensates the mismatch be- Fig. 1 Internal model control (IMC).
tween P and P. Ideally, if the model exactly equals to the plant,
IMC completely yields the desired tracking property. In prac-
tice, there is always a dierence between a model and the actual method, in which approximations of the gradient, the Hessian,
plant. The parameters of the controller are therefore tuned to and so on, consist of experimental data. This means that IFT re-
meet the performance requirement. In these cases, constructing quires many experiments to update parameters of a controller.
a mathematical model based on the identification is a rational Thus, it takes considerable expense and time, which are crucial
approach. However, there are also many cases where it is di- problems with respect to practical points of view. On this point,
cult to apply persistently excited signals from the viewpoints of VRFT and FRIT have a great advantage with only a one-shot
the safe operation of the plant. In addition, it is also preferable experiment, then the time and expense for obtaining optimal
to reduce time and cost from the viewpoints of the management parameters are drastically reduced. In [4], VRFT was applied
for the plant. In such cases, the direct use of the data collected to a class of linearly parameterized IMC, and hence, the flexi-
from experiments is also a useful and eective approach. More- bility and the freedom of the controller are limited. In [6][8],
over, since the data have fruitful information on the dynamics FRIT was utilized for controller parameter tuning of IMC or
of the plant, it is expected that data-driven approaches yield a the Smith predictor. The approaches proposed in [6][8] treat
more desirable IMC controller. From such points of view, there the controller whose denominator and numerator are parame-
are a number of studies on data-driven approaches to IMC in terized, which implies that more eective tuning of IMC can be
literatures, such as F. De Bruyne [2] with iterative feedback performed.
tuning (IFT, [3]), S. Formentin et al. [4] with virtual reference On the other hand, one of the features of IMC is that the con-
feedback tuning (VRFT, [5]), O. Kaneko et al. [6][8] with fic- troller contains a plant model internally, thus the data-driven
titious reference iterative tuning (FRIT, [9]) and so on. approach in IMC simultaneously yields a mathematical model
IFT, which was used in [2], is a tuning method of a param- of the plant [6][8]. From practical points of view, the obtained
eterized controller by directly minimizing the performance in- model is useful for finding out information on model uncertain-
dex which consists of the data. This minimization can be done ties, monitoring the actual status, detecting an aging variation
by a nonlinear optimization technique, e.g., the Gauss-Newton of the plant, re-designing more advanced controllers, and so
on. From theoretical points of view, it is meaningful to study
this issue since there is a crucial interplay that cannot be sepa-
Graduate School of Natural Science and Technology, Kanazawa
University, Kakuma-machi, Kanazawa, Ishikawa 9201192, rated between a mathematical model and a designed controller
Japan as stated in [10]. In fact, the reference [11] provides the itera-
Institute of Science and Engineering, Kanazawa University, tive design of a controller and a model in the IMC architecture
Kakuma-machi, Kanazawa, Ishikawa 9201192, Japan for time delay systems. However, the proposed method in [11]
E-mail: hien@moccos.ec.t.kanazawa-u.ac.jp,
o-kaneko@ec.t.kanazawa-u.ac.jp, shigeru@t.kanazawa-u.ac.jp
requires many experiments.
(Received March 9, 2012) By the way, it is seen that there are many systems with a non-
(Revised July 2, 2012) minimum phase behavior which cannot be eliminated in various
JCMSI 0001/13/06010038
c 2012 SICE
SICE JCMSI, Vol. 6, No. 1, January 2013 39
M
G(s) k Lk (s, a) (11)
k=1
= Td (8) := m n (15)
1 + PC( )
generically holds. This means that the minimization of JF () is where
deeply related to attaining the desired tracking property.
m = [0 1 . . . . . . + ] R++1 (16)
3. IMC with FRIT for Non-Minimum Phase Systems:
Laguerre Expansion Approach with constraints such that the numerator and the denominator
3.1 Laguerre Expansion for Non-Minimum Phase Sys- of Pm are Hurwitz polynomials, and
tems
Laguerre functions are bases of a complete orthonormal n = [1 . . . M a] R M+1 . (17)
set on the Lebesgue space of squared integrable functions in
The factorization (12) is a specific approach to handle the
[0, ). The continuous Laguerre functions have Laplace trans-
unstable zeros and/or time delay, which has been introduced in
forms [16]
[1] and [11]. In addition, such a factorization is also convenient
2a s a k1 for treating the reference model in the cases of non-minimum
Lk (s, a) = , a > 0, (9) phase systems, as discussed below.
s+a s+a
As stated above, the time scale a was chosen based on the
where the parameter a is called the time scale.
knowledge of the system to guarantee the precision of the ap-
It is shown in [16],[17] that any stable systems can be exactly
proximation. An appropriate choice of a enables us to decrease
represented by an infinite sum of Laguerre functions as
the term M of Laguerre expansion [18][21]. However, since
there is no any information of the system in our setting, we con-
G(s) = k Lk (s, a) (10) sider a as a parameter that is tuned together with other parame-
k=1
ters of the controller to minimize the model-reference criterion.
with suitable coecients k , k = 1, 2, , . However, in Regarding the reference model T d , since the limitation of the
practice, with a given order M of the expansion, a truncated tracking property is deeply related to the non-minimum phase
Laguerre expansion is used for the approximation, e.g. behavior of a system, T d would have the same unstable zeros
SICE JCMSI, Vol. 6, No. 1, January 2013 41
M lim T dm (s) = 1, (24)
T d (n ) = T dm k Lk (s, a). (18) s0
k=1
and
3.2 Simultaneous Attainment of a Desired Controller and
a Mathematical Plant Model 0 0. (25)
As stated earlier, since IMC involves a mathematical model Then, the steady state error of the closed loop system depicted
internally, it is expected that the achievement of the desired out- in Fig. 4 with respect to the step reference signal r is eliminated
put also yields a mathematical model of the actual plant [6] if and only if
[8],[11]. As shown in [8], if the controller C I MC is parameter-
ized as
M
a
(1)k1 k = (26)
2
C I MC () = T dm Pm (m )1 , (19) k=1
a
M1
desired output and the attainment of the plant model. For some
M = (1)k1 k (29) parameters and , the following three statements are equiv-
2 k=1
alent
for the elimination of the steady state error. Thus, we use the
notation (i) JF ( ) > JF ( ), (36)
where Pm (m ) and Pn (n ) are described as (13) and (14), re- See [8] for the more details.
spectively. Substituting r() into (6), we obtain The above equivalence also holds in this paper. In the state-
ment (ii), the relative error of the closed loop Gry () and the
JF () = y(0 ) T dm Pn (n )r()2N (32) reference model T d () is evaluated. In the statement (iii), the
relative error of the actual plant P and the internal mathematical
together with (26). model P() is evaluated. Thus, the decrease of JF () leads to
As a result of the minimization of JF (), we obtain the opti- the decreases of the involved quantities on the achievement of
mal vector of the controller the desired output and the accuracy of the mathematical model
under the influence of y(0 ) and (1 T d ()) y(0 ), respectively.
= arg min JF (). (33)
3.5 Algorithm
If the minimized value of JF is regarded as suciently small
(JF ( ) 0), it is regarded that the optimal parameters yield The algorithm of the proposed method is summarized as fol-
a controller C( ) for a desired specification, i.e., Gry ( ) = lows.
T d ( ) generically holds. Moreover, it follows from (20) that 1. Parameterize the minimum phase and the non-minimum
P( ) = P also generically holds. phase parts of a plant model as (13) and (14).
3.4 Remarks 2. The parameter vector is determined by m and n as (16)
Remark 1 The results deduced above for the continuous sys- and (30), respectively.
tem case are easily adapted to the discrete system case. Based
on the discrete Laguerre functions with z-transforms [21] 3. Set an initial parameter vector 0 and perform a one-shot
experiment to obtain the data u(0 ) and y(0 ).
k1
1 a2 1 az
Lk (z, a) = , |a| < 1, (34) 4. Calculate the fictitious reference signal r() by using (31),
z a z a
construct the cost function JF () as (32) and minimize it
the condition (26) is modified as by an o-line nonlinear optimization.
5. Obtain the optimal parameter vector = arg min JF ()
M
1 a
k = . (35) which yields both a desired controller and a mathematical
1 + a
k=1 model of the plant.
Namely, the extension to the discrete-time case can be trivially Note that, at each step in the optimization process, m is
done. checked with the constraints such that the numerator and the
denominator of Pm (m ) are Hurwitz polynomials. The process
Remark 2 For a fixed M, the number of the terms of Laguerre
will continue if the constraints are satisfied. Otherwise, it stops,
expansion, the optimal a might not yield convergent coecients
we either take the parameter vector of the previous step or ad-
for the approximation. In this case, it is expected that increasing
just the initial parameter and then repeat the procedure.
M leads to the improvement of precision of the identification.
However, increasing this term makes calculation more compli- 4. Numerical Example
cated. How to decide the term M as well as the degree and
We apply the proposed approach to an unknown plant
of Pm is a significant issue that should be addressed as impor-
tant future researches. 4(2s 3)(s 2)
P= es (39)
(s2 + 5s + 2)(s2 + 4s + 5)
Remark 3 In a practical sense, it might be dicult to achieve
JF ( ) = 0. In [8], the authors analyzed the relationship among with a time-delay and unstable zeros.
the minimization of the cost function, the achievability of the The reference model is given
SICE JCMSI, Vol. 6, No. 1, January 2013 43
Fig. 5 The reference signal r (the dot-and-dash line), the initial output Fig. 6 The reference signal r (the dot-and-dash line), the optimal output
y(0 ) (the solid line) and the desired output T dm Pn (0n )r (the dotted y( ) (the solid line) and the desired output T dm Pn (n )r (the dotted
line). line).
1
T dm = . (40)
2s + 1
Without information on the plant P, we parameterize the
plant model P with a first-order minimum phase part Pm for
simplicity
0
Pm (m ) = (41)
1 s + 1
M
Pn (n ) = k (a)Lk (s, a). (42) Fig. 7 Step responses: P (the solid line), P( ) (the dotted line).
k=1