IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 17, NO.

1, JANUARY 2009 29

Design of a Data-Driven PID Controller

Toru Yamamoto, Member, IEEE, Kenji Takao, and Takaaki Yamada

Abstract—Since most processes have nonlinearities, controller algorithms (GA) [4] have been proposed, but there are obstacles
design schemes to deal with such systems are required. On to these methods. According to these methods, the learning cost
the other hand, proportional-integral-derivative (PID) controllers is considerably large, and these PID parameters cannot be ade-
have been widely used for process systems. Therefore, in this paper,
a new design scheme of PID controllers based on a data-driven quately adjusted due to the nonlinear properties. Therefore, to
(DD) technique is proposed for nonlinear systems. According to the best of our knowledge, there are currently few PID param-
the DD technique, a suitable set of PID parameters is automati- eter tuning schemes of practical use for nonlinear systems.
cally generated based on input/output data pairs of the controlled In contrast, the development of computers has been enormous
object stored in the database. This scheme can adjust the PID in the same time period. These advances in computers enable
parameters in an online manner even if the system has nonlinear
properties and/or time-variant system parameters. Finally, the them to store, quickly retrieve, and read out an ever-increasing
effectiveness of the newly proposed control scheme is evaluated on amount of data. With these advantages in mind, the following
some simulation examples, and a pilot-scale temperature control method has been proposed based on the computer data storage
system. system: whenever new data is obtained, the data is stored. Next,
Index Terms—Database, nonlinear control, proportional-inte- similar neighbors to the information requests, called “queries,”
gral-derivative (PID) control, process control, self-tuning control. are selected from the stored data. Additionally, the local model
or the local controller is constructed using these neighbors. This
is known as a data-driven (DD) technique. The DD techniques
I. INTRODUCTION called the Just-in-Time (JIT) method [5], [6], the Lazy Learning
method [7], [8], and the Model-on-Demand (MoD) method [9]

I N RECENT years, many complicated control algorithms

such as adaptive control theory and/or robust control theory
have been proposed and implemented for real systems. Even
have received lots of attention in the last decade.
A few similar PID controllers have been already proposed
based on the JIT method [10] and the MoD method [11].
though these complicated and subtle control algorithms exist, Generally, the JIT method is used to supplement the feedback
less sophisticated proportional-integral-derivative (PID) con- controller with a PID structure, however it is problematic. The
trollers continue to be widely employed in process industries. tracking property is not reliable enough due to the nonlinearities
The reasons for the continued popularity of PID controllers are occurring in the case where reference signals are changed. In
summarized as follows: 1) the control structure is quite simple; this method, the controller does not create any integral action
2) the physical meaning of control parameters is clear; and 3) in the whole control system. On the other hand, the latter
the operators’ know-how can be easily utilized in designing method has a PID control structure. PID parameters are tuned
controllers. by operators’ skills, and they are stored in the database in
Given these reasons, it is still attractive to design PID con- advance. Also, a suitable set of PID parameters is generated
trollers, but they do have their drawbacks. Since most process using the stored data. Still, good control over the performance
systems have nonlinearities, it is difficult to obtain good control cannot be necessarily obtained in the case where nonlinearities
performances for such systems simply using the fixed PID pa- are included in the controlled object and/or system parameters
rameters. The adaptive or self-tuning PID control schemes [1], are changed. This is due to the fact that PID parameters are not
[2] have been frequently employed for systems with weak non- adjusted in an online manner corresponding to properties of the
linearity. However, since these methods have been mainly re- controlled object.
searched and discussed for use with linear systems, it is impos- In an effort to overcome the shortcomings of the JIT and
sible to obtain suitable control results for systems with strong MoD methods, this paper proposes and describes a new design
nonlinearity. Where there is strong nonlinearity, PID parame- scheme of PID controllers, based on the DD technique. Ac-
ters tuning methods using neural networks (NN) [3] and genetic cording to the proposed method, PID parameters obtained using
the DD technique are adequately updated in proportion to con-
Manuscript received April 19, 2006; revised February 04, 2007 and October trol errors, and modified PID parameters are stored in the data-
06, 2007. Manuscript received in final form January 31, 2008. First published base. The main feature of the newly proposed scheme is that
September 26, 2008; current version published December 24, 2008. Recom-
mended by Associate Editor S.-L. Jamsa-Jounela. the PID parameters are updated corresponding to the control
T. Yamamoto is with the Department of Technology and Information error in an online manner and they are stored in the database.
Education, Hiroshima University, Higashi-Hiroshima 739-8524, Japan (e-mail: Therefore, the learning effect is gradually accumulated in the
[email protected]).
K. Takao is with the Technical Headquarters, Mitsubishi Heavy Industries,
database. In order to avoid the excessive accumulation of the
Ltd., Mihara 729-0393, Japan. stored data, an algorithm is put forward which reduces the need
T. Yamada is with the Industrial Automation Company, OMRON Corpora- for memory storage and computational costs. Finally, the effec-
tion, Kusatsu 525-0035, Japan.
Color versions of one or more of the figures in this paper are available online
tiveness of the newly proposed control scheme is examined on
at http://ieeexplore.ieee.org. some simulation examples, and experimentally evaluated on a
Digital Object Identifier 10.1109/TCST.2008.921808 pilot-scale temperature control system.
1063-6536/$25.00 © 2008 IEEE

Authorized licensed use limited to: Toru Yamamoto. Downloaded on September 23, 2009 at 01:01 from IEEE Xplore. Restrictions apply.
30 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 17, NO. 1, JANUARY 2009

II. DD PID CONTROLLER DESIGN Thus, with those additional pieces of information, (6) can be
rewritten as
A. DD Technique
In order to create the DD technique, several mathematical (7)
considerations must first be undertaken. The first consideration (8)
is the following discrete-time nonlinear system: (9)
(1) where denotes a function. By substituting (7) and (8) into
(1) and (2), the following equation can be derived:
where denotes the system output and denotes the non-
linear function. In this case, is called the “information
(10)
vector,” which is defined by the following equation:

(11)
(2)
where , , and denotes a nonlinear function.
where denotes the control input. Also, and denote the Therefore, the PID parameters are related to the informa-
orders of the system output and the control input, respectively. tion vector , which is expressed as
Accordingly, (2) means that the controlled object is described
using elements included in . In addition, it is assumed (12)
that, with regard to the controlled object, the sign of the partial
derivative of with respect to is known. The sign of
(13)
the partial derivative is also the sign of the system Jacobian, so
the system Jacobian is a known variable.
Since the future output included in (13) cannot be
According to the DD technique, the data is stored in the form
obtained at , is replaced by , because the
of the information vector expressed in (2). Moreover, the
purpose of the control considered in this paper is to realize
query is required in calculating the estimate of the
. Therefore, included in (13) is newly
output , that is, after some similar neighbors to the query
rewritten as follows:
are selected from the database, the prediction of the system can
be obtained using these neighbors. The controller design is dis-
cussed next.
(14)
B. Controller Design
After making the above preparations, it is possible to design a
In this paper, the following control law with a PID structure
new PID control scheme based on the DD technique. The pro-
is considered:
cedure of the controller design is summarized as follows.
[STEP 1] Generate Initial Data-Base: The DD technique
(3) cannot work if the historical data are not saved at all, as there
(4) will be no information from which to work. It is necessary, then,
to first create a database based on historical data of the con-
where denotes the control error signal defined by trolled object. Therefore, PID parameters are initially calculated
using either the Ziegler and Nichols method [12] or the Chien,
(5) Hrones, and Reswick(CHR) method [13]. Of course, PID pa-
rameters determined through the use of the operators’ skill and/
The reference signal is denoted by . Also, , , and , or know-how can be also utilized as the initial database, that is,
respectively, denote the proportional gain, the reset time, and indicated in the following equation is generated as the ini-
the derivative time, and denotes the sampling interval. Here, tial database:
, and included in (4) are derived by the relations
, and . denotes (15)
the differencing operator defined by . However,
it is quite difficult to obtain a good control performance due where and are given by (14) and (9). Moreover,
to nonlinearities if the PID parameters ( , , and ) in denotes the number of information vectors stored in the initial
(4) are fixed. Therefore, a new control scheme is proposed that database. If one set of fixed PID parameters is chosen as being
can adjust PID parameters in an online manner corresponding typical, then all PID parameters included in the initial infor-
to properties of the system. Thus, instead of (4), the following mation vectors may be equal. Expressed numerically, that is,
PID control law with time-variant PID parameters is employed: in the initial stage.
[STEP 2] Calculate Distance and Select Neighbors: It is nec-
(6) essary though, to determine the distances between the query

Authorized licensed use limited to: Toru Yamamoto. Downloaded on September 23, 2009 at 01:01 from IEEE Xplore. Restrictions apply.
YAMAMOTO et al.: DESIGN OF A DATA-DRIVEN PID CONTROLLER 31

and the information vectors . These are cal- Thus, let be the denominator of the discrete-time ver-
culated using the following -norm with some weights: sion of the following desirable second-order continuous-time
transfer function :

(25)
(16)
In other words, is defined as follows:
where denotes the number of information vectors stored
in the database when the query is given. To further explain (26)
the variables: denotes the th element of the th infor-
mation vector, and, similarly, denotes the th element of (27)
the query at . Among the th element of all information vec-
tors ( , ) stored in the database, the max- (28)
imum element is denoted by . Similarly,
denotes the minimum element. Here, -pieces with the smallest
(29)
distances between them are chosen from all information vectors.
[STEP 3] Compute PID Parameters: Next, using -neigh- (30)
bors selected in [STEP 2], the suitable set of PID parameters
is computed around the query using the following linearly where denotes the rise-time. The damping coefficient is
weighted average (LWA) [14]: denoted by . The damping coefficient itself is adjusted by
changing . Where and , the step response of
shows the Binominal model response and the Butterworth
(17)
model response, respectively.
Moreover, each partial differential of (19) is developed as
where denotes the weight corresponding to the th informa- follows:
tion vector in the selected neighbors. This is calculated by

(18)

Using the PID parameters computed in (17), the control input

is generated, and the output is measured.
[STEP 4] PID Parameters Adjustment: In the case where
the suitable control performance cannot be obtained using the
PID parameters computed in [STEP 3], these control parame-
ters have to be updated and stored in the database, that is, it is
necessary to adjust the PID parameters so that the control error
is decreased. The following steepest descent method is utilized
in order to modify the PID parameters: (31)

(19) Note that a priori information with respect to the system Jaco-
bian is required in order to calculate (31). Here,
(20) using the relation , the system Jacobian can be
obtained by the following equation:
where denotes the learning rate, and the following
denotes the error criterion:
(32)
(21)
where , . From the assumption
(22)
mentioned above, the sign of the system Jacobian is obtainable.
The output of the reference model is denoted by , given by And by including in , the usage of the system
Jacobian can be made easy [3]. Therefore, it is assumed that the
sign of the system Jacobian is known in this paper.
(23)
[STEP 5] Remove Redundant Data: In implementing the pro-
(24) posed DD technique in real systems, the newly proposed scheme
has a constraint: the calculations between [STEP 2] and [STEP
Here, is designed based on the following two features: 4] must be finished within the sampling time. Here, storing the
1) rise-time; redundant data in the database results in the computer wasting
2) damping property. time that could be used on the process. Therefore, an algorithm

Authorized licensed use limited to: Toru Yamamoto. Downloaded on September 23, 2009 at 01:01 from IEEE Xplore. Restrictions apply.
32 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 17, NO. 1, JANUARY 2009

Fig. 1. Block diagram of the proposed data-driven PID control system.

Fig. 2. Static properties of System 1 and System 2.

to avoid excessive stored data is put forward. The procedure is
carried out in the following two steps.
[First condition] For information vectors in which
A. Ex.1 Hammerstein Model
-neighbors are excepted beforehand, the information
vectors which satisfy the following condition are ex- To begin, the following Hammerstein models are discussed
tracted: (see [15] for more information).
[System 1]
(33)

Note that the distance is computed using only input/output (35)

data included in the information vector. Meeting the first
condition means that only the information vectors with
short distance to the query , are extracted from the [System 2]
database.
[Second condition] For information vectors extracted in
the first condition, the information vectors which satisfy (36)
the following condition are extracted:
where denotes the white Gaussian noise with zero
(34) mean and a variance of . The static properties of the
System 1 and System 2 are shown in Fig. 2. On examining
Fig. 2, it is apparent that the gains of System 2 are larger
where , , and mean , , and , respec- than those of System 1 at .
tively. Using the above procedure, it is possible to ex- Here, the reference signal is given as follows:
tract and delete from the database redundant data. This is
achieved by extracting and deleting from the database in-
formation vectors with high similarity in the relationship
(37)
between the newly generated PID gains and PID gains in-
cluded in the extracted information vectors. If plural infor-
mation vectors exist which satisfy the second condition,
then only the information vector with the smallest value in The information vector is given by
the second condition is removed.
Note that the query and the corresponding updated PID pa-
rameters are always stored in the database. In practice, and (38)
should be set between 0.1 and 1.0, but some trial and error
may be necessary. The desired characteristic polynomial included in the
The above algorithm is illustrated by the block diagram in reference model was designed as follows:
Fig. 1.
(39)
III. SIMULATION EXAMPLES
where was designed by setting as 1.0 and as 0.0.
In order to evaluate the effectiveness of the newly proposed Here, is equal to 1.0. Table I shows the user-specified pa-
scheme, some simulation examples for nonlinear systems are rameters as determined in the proposed method.
considered. These include a Hammerstein model and a system For the purpose of comparing the proposed method with con-
with a hysteresis. They are discussed below. ventional schemes, the fixed PID control scheme widely used

Authorized licensed use limited to: Toru Yamamoto. Downloaded on September 23, 2009 at 01:01 from IEEE Xplore. Restrictions apply.
YAMAMOTO et al.: DESIGN OF A DATA-DRIVEN PID CONTROLLER 33

TABLE I
USER-SPECIFIED PARAMETERS INCLUDED IN THE PROPOSED METHOD
(HAMMERSTEIN MODEL)

Fig. 4. Trajectories of PID parameters corresponding to Fig. 3.

Fig. 5. Behavior of PID parameters adjustment around t = 100 [step].

The control results for System 1 [see (35)] are shown in Fig. 3
using the fixed PID controller and the proposed control scheme.
Trajectories of the PID parameters in the proposed scheme are
summarized in Fig. 4.
From Fig. 3, it is clear that the control results of the fixed PID
controller are not good. This is owing to nonlinearities of the
controlled object. The nonlinearities had the most influence in
the third step where the control responses become slow. On the
other hand, looking at Figs. 3 and 4, it is apparent that good con-
trol results can be obtained using the proposed method because
PID parameters are adequately adjusted.
The adjustment of PID parameters merits further discussion.
Fig. 3. Control results using (a) the fixed PID control and (b) the proposed Fig. 5 shows the behavior of the updating scheme set out in
method for System 1. [STEP 4], where the reference signal is changed from 0.5 thru
1.0 around 50 [step]. According to Fig. 5, it is clear that a new
set of PID parameters is gradually built up in the database, and
in industrial processes is first employed. These PID parameters, PID parameters are suitably adjusted according to the change of
tuned by the CHR method [13], are as follows: the reference signal or the operating point.
(40) Moreover, in Fig. 6 shows the point in which a new infor-
mation vector is stored in the database, that is, Fig. 6 shows that
In addition, a certain type of neural-net-based PID controller [3], the process of removing the redundant data in [STEP 5] works
called the NN-PID controller, was applied. This particular NN adequately, and the new data (i.e., the new information vector)
was used because it complements the fixed PID controller. are stored instead of the redundant data. Fig. 6 also illustrates

Authorized licensed use limited to: Toru Yamamoto. Downloaded on September 23, 2009 at 01:01 from IEEE Xplore. Restrictions apply.
34 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 17, NO. 1, JANUARY 2009

Fig. 6. Replacement behavior of the redundant data by the new data, where
shows the point in which a new information vector is stored in the database.
Fig. 8. Control result using the NN-PID controller for Hammerstein model.

Fig. 7. Error behaviors using the NN-PID controller for Hammerstein model.

that new data are stored only in the transient state where the ref-
erence signals are changed. Note that there is very little replace- Fig. 9. Control result using the fixed PID controller in the case where the
system is changed from System1 to System2.
ment of the redundant data by the new data when the system is
in a steady state (i.e., where the reference signal is constant).
As a result of using the algorithm in [STEP 5] to remove need-
The next step is to consider the case where the system changes
less data, the number of data items stored in the database was
from (35) to (36) at . First, the control results using the
quite low: 38, reduced from an original count of 206. In addition,
fixed PID controller are shown in Fig. 9, where the PID param-
using the proposed method, the error given by the following
eters used are the same as those shown in (40). Although the
criterion is 0.0417:
gain of the controlled object becomes high after , the
PID parameters are not changed. Therefore, the control perfor-
(41) mance becomes oscillatory after . On the other hand, the
proposed control scheme was also employed in this case. The
control results and trajectories of PID parameters are shown in
where denotes the number of steps per 1 [epoc], and is set Figs. 10 and 11, respectively. Figs. 10 and 11 demonstrate that,
as 200 in this case. Furthermore, because PID parameters can even if system parameters are changed, a fairly good control
be adjusted in an online manner by the proposed method, the performance can be obtained because PID parameters are ad-
number of iterations was set as 1. justed adequately using the proposed method. This result shows
Moreover, the NN-PID controller [3] was applied to this the adaptability of the proposed method remarkably well in the
system [see (35)], where the NN was utilized in order to Hammerstein model. Next, the proposed method was applied to
supplement the fixed PID controller. The trajectory of the another type of nonlinear system: a system with a hysteresis.
learning error [expressed in (41)] is shown in Fig. 7, and
control results are shown in Fig. 8. We see from Fig. 7 that B. Ex.2 System With a Hysteresis
the number of learning iterations remains 86 [epoc] until the
Next, the following model with another nonlinear type of
control results using the NN-PID controller can produce the
system was tested. In this system, like that set out by Narendra
same control performances as the proposed method. In other
and Parthasarathy [16], the following equation applies:
words, until is satisfied. Since this condition has
been satisfied, we can go ahead and examine the effectiveness
(42)
of the proposed method nonlinear systems.

Authorized licensed use limited to: Toru Yamamoto. Downloaded on September 23, 2009 at 01:01 from IEEE Xplore. Restrictions apply.
YAMAMOTO et al.: DESIGN OF A DATA-DRIVEN PID CONTROLLER 35

TABLE II
USER-SPECIFIED PARAMETERS INCLUDED IN THE PROPOSED METHOD

Fig. 10. Control result using the proposed method in the case where the system
parameters are changed.

Fig. 11. Trajectories of PID parameters corresponding to Fig. 10.

Fig. 13. Control result using the fixed PID controller for Ex.2.

Equation (38) was utilized as the information vector , and

the desired characteristic polynomial included in the
reference model was designed as

(44)

where and are set as 1.0 and 0.0, respectively, and

equals 1.0. In addition, the user-specified parameters included
in the proposed method were determined and are summarized
in Table II.
For the purpose of comparison, the fixed PID control scheme
was first employed, wherein the PID parameters were tuned by
the CHR method [13]. Then, PID parameters were calculated as
Fig. 12. Static property of the system (42).
(45)
where denotes the white Gaussian noise with zero mean and The control results are shown in Fig. 13. As seen in Fig. 13,
a variance of . The static property of this model is shown due to nonlinearities of the controlled object, the control re-
in Fig. 12 and discussed below. From Fig. 12, it is clear that this sults using the fixed PID controller include excessive oscillation
model has strong nonlinearities. In particular, there is a kind of around . The consequence is that the system falls into
hysteresis occurring between and . Here, the an unstable state.
reference signal was given by Next, the self-tuning PID control scheme was employed, and
the control results are shown in Fig. 14. The self-tuning PID
control scheme examined in this paper has been previously con-
(43) sidered in [2]. For details, see Yamamoto and Shah [2]. Thus,
the system parameters were recursively estimated and the PID

Authorized licensed use limited to: Toru Yamamoto. Downloaded on September 23, 2009 at 01:01 from IEEE Xplore. Restrictions apply.
36 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 17, NO. 1, JANUARY 2009

Fig. 14. Control result using the self-tuning PID controller for Ex.2.

Fig. 17. Photograph of the experimental temperature control system.

Fig. 15. Control result using the proposed method for Ex.2.

Fig. 18. Schematic figure of the experimental temperature control system.

IV. EXPERIMENTAL RESULTS

The proposed DD method was experimentally evaluated on a
pilot-scale temperature control system. Fig. 17 shows a photo-
graph of this equipment, and the corresponding schematic figure
of this system is illustrated in Fig. 18.
Two heaters were secured on a steel plate. These heaters
worked synchronously, corresponding to the input signal from
the computer. One thermo-couple was also prepared on the
Fig. 16. Trajectories of PID parameters corresponding to Fig. 15. steel plate, and the measured temperature of the steel plate
was sent to the computer as the system output signal. The
control objective was to regulate the temperature of the steel to
parameters were computed using the estimates based on the re- the desired reference signal by manipulating the power of the
lationship between the generalized minimum variance control heater. The static properties of the temperature control system
and PID control. Due to the strong nonlinearity, the control per- are shown in Fig. 19. Especially, a system gain of around 100
formance was inferior in the second reference signal, that is, in the temperature is about a third of the temperature around
. 60 . The system gain changes drastically during the control.
Finally, the control results using the proposed scheme are The fixed PID controller and the proposed method were em-
shown in Fig. 15, and then trajectories of PID parameters are ployed for this system. The reference signal was changed alter-
shown in Fig. 16. Figs. 15 and 16 show that the domination of nately from 60 in the temperature and 100 . The fixed PID pa-
the proposed method is quite clear. The PID adjustment in an rameters were computed using the CHR method [13] as follows:
online manner works exceptionally well and removes the ex-
cessive oscillation. (46)

Authorized licensed use limited to: Toru Yamamoto. Downloaded on September 23, 2009 at 01:01 from IEEE Xplore. Restrictions apply.
YAMAMOTO et al.: DESIGN OF A DATA-DRIVEN PID CONTROLLER 37

Fig. 19. Static properties of the temperature control system.

Fig. 21. Trajectories of PID parameters corresponding to Fig. 20.

TABLE III
USER-SPECIFIED PARAMETERS INCLUDED IN THE PROPOSED METHOD
(TEMPERATURE CONTROLLER)

Fig. 20. Control result using the fixed PID controller(dotted line) and the pro-
posed controller(solid line).

Fig. 20 shows the control results. The dotted line and the
solid line show the results using the fixed PID and the pro-
posed method, respectively. The above PID parameters were
determined using the historical data gathered when the system
output was around 80 in temperature. Because the PID param- is first put on the main steel board, and the object is processed
eters were not suitably calculated, the control results are not so while keeping the temperature constant. It moves to the next
good. In particular, due to the small integral gain, the tracking process when the processing ends, and the next new object is
property to the reference signal is considerably inferior. put on the main steel board. This procedure is repeated many
On the other hand, according to the proposed method, PID times. As a result, if the temperature recovers quickly, a lot of
parameters were appropriately adjusted in an online manner ac- objects can be processed. Therefore, the tracking properties of
cording to the reference signal as shown in Fig. 21. Here, the the PID controller are strongly demanded in industries.
user-specified parameters included in the proposed method were For the purpose of comparison, the fixed PID controller and
determined as shown in Table III, and PID parameters in the ini- the proposed method were employed for the case where the pe-
tial database were set the same as in (46). riodic disturbance is introduced. The control results are sum-
Next, the robustness and the adaptability to cyclic distur- marized in Fig. 22: the dotted line shows the result given by
bances was investigated. Key to this investigation is the solenoid the fixed PID controller and the solid line shows the proposed
coil (see Fig. 17 or Fig. 18 which is part of the apparatus in this method. Moreover, Fig. 23 shows the trajectories of PID param-
experiment). Because this solenoid works only periodically, it eters created by the proposed method. Here, the fixed PID pa-
creates a cyclic disturbance in the following way. The solenoid rameters were computed using the CHR method [13] as follows:
is attached to the second steel plate which comes into contact
with the main steel plate. It returns to its original position when (47)
the temperature is kept constant. This results in the main steel
plate acting as the periodic disturbance. In other words, when Also, the user-specified parameters included in the proposed
the solenoid and the steel plate come in contact, the temperature method utilized were the same as those given in Table III. It is
of the main steel plate falls. The temperature rises if it returns. clear that the tracking property to the reference signal is gradu-
This cycle of temperature rising and dropping was created as ally improved according to the proposed method. The behavior
an imitation of a common industrial process wherein the object is shown more clearly in Figs. 24 and 25, where these figures

Authorized licensed use limited to: Toru Yamamoto. Downloaded on September 23, 2009 at 01:01 from IEEE Xplore. Restrictions apply.
38 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 17, NO. 1, JANUARY 2009

Fig. 22. Control result using the fixed PID controller(dotted line) and the pro- Fig. 25. Enlarged figure around t = 1100 in Fig. 22.
posed control method(solid line) in the case where the disturbance is periodi-
cally added.

V. CONCLUSION
In this paper, a new design scheme of PID controllers using
the DB modeling method has been proposed. To date, many PID
controller design schemes using NNs and GAs have been pro-
posed for nonlinear systems. In employing these schemes for
real systems, however, the considerably large learning cost be-
comes a serious problem. This problem can be avoided by using
the proposed method because such computational burdens can
be effectively reduced by using the algorithm for removing the
redundant data. In addition, the effectiveness of the proposed
method has been verified by some numerical simulation exam-
ples. Given the success of this newly proposed DD technique
for PID controllers on the temperature control system, there is
good reason to continue to explore its success on other systems.
Applications of the newly proposed scheme for existing sys-
tems and the extension to multivariable cases are currently under
consideration.
Fig. 23. Trajectories of PID parameters corresponding to Fig. 22.
REFERENCES
[1] K. J. Åström and T. Hägglund, Automatic Tuning of PID Controllers.
Research Triangle Park, NC: Instrum. Soc. of America, 1988.
[2] T. Yamamoto and S. L. Shah, “Design and experimental evaluation of
a multivariable self-tuning PID controller,” IEE Proc. Control Theory
Appl., vol. 151, no. 5, pp. 645/652–645/652, 2004.
[3] S. Omatu, K. Marzuki, and Y. Rubiyah, Neuro-Control and Its Appli-
cations. London, U.K.: Springer-Verlag, 1995.
[4] B. Porter and A. H. Jones, “Genetic tuning of digital PID controllers,”
Electron. Lett., vol. 28, pp. 843/844–843/844, 1992.
[5] A. Stenman, F. Gustafsson, and L. Ljung, “Just in time models for dy-
namical systems,” in Proc. 35th IEEE Conf. Decision Control, 1996,
pp. 1115/1120–1115/1120.
[6] Q. Zheng and H. Kimura, “A new just-in-time modeling method and
its applications to rolling set-up modeling,” (in Japanese) Trans. SICE,
vol. 37, no. 7, pp. 640/646–640/646, 2001.
[7] J. Zhang, Y. Yim, and J. Yang, “Intelligent selection of instances for
prediction functions in lazy learning algorithms,” Artif. Intell. Rev., vol.
Fig. 24. Enlarged figure around t = 500 in Fig. 22. 11, pp. 175/191–175/191, 1997.
[8] G. Bontempi, M. Birattari, and H. Bersini, “Lazy learning for local
modeling and control design,” Int. J. Control, vol. 72, no. 7-8, pp. 643/
658–643/658, 1999.
show the control results around and , respec- [9] A. Stenman, “Model on demand: Algorithms, analysis and applica-
tions,” Ph.D. dissertation, Dept. Elect. Eng., Linköping University,
tively. For the disturbance, another set of PID parameters is im- Linköping, Sweden, 1990.
mediately extracted from the database corresponding to the cur- [10] Q. Zheng and H. Kimura, “Just-in-time PID control,” in Proc.
rent state. It is clear that the proposed method effectively uses 44th Japan Joint Automatic Control Conf., Tokyo, Japan, 2001, pp.
336–339.
the previously accumulated knowledge. This is one of the typ- [11] J. Ohta and S. Yamamoto, “Detabase-driven tuning of PID controllers,”
ical features of the proposed method. (in Japanese) Trans. SICE, vol. 40, no. 6, pp. 664–669, 2004.

Authorized licensed use limited to: Toru Yamamoto. Downloaded on September 23, 2009 at 01:01 from IEEE Xplore. Restrictions apply.
YAMAMOTO et al.: DESIGN OF A DATA-DRIVEN PID CONTROLLER 39

[12] J. G. Ziegler and N. B. Nichols, “Optimum settings for automatic con- Kenji Takao received the M.Eng. and D.Eng
trollers,” Trans. ASME, vol. 64, no. 8, pp. 759/768–759/768, 1942. degrees from Hiroshima University, Higashi-Hi-
[13] K. L. Chien, J. A. Hrones, and J. B. Reswick, “On the automatic con- roshima, Japan, in 2004 and 2007, respectively.
trol of generalized passive systems,” Trans. ASME, vol. 74, pp. 175/ He is currently a Research Engineer with the
185–175/185, 1972. Technical Headquarters, Mitsubishi Heavy Indus-
[14] C. G. Atkeson, A. W. Moore, and S. Schaal, “Locally weighted learning tries, Ltd., Mihara, Japan. He was a Research Fellow
for control,” Artif. Intell. Rev., vol. 11, pp. 75/114–75/114, 1997. of the Japan Society for the Promotion of Science for
[15] L. Zi-Qiang, “On identification of the controlled plants described by two years starting in April 2005. His current research
the hammerstein system,” IEEE Trans. Automat. Contr., vol. AC-39, interests are in the area of PID control, intelligent
no. 2, pp. 569–573, Apr. 1994. control, process control, and data-based modeling/
[16] K. S. Narendra and K. Parthasarathy, “Identification and control of dy- control.
namical systems using neural networks,” IEEE Trans. Neural Netw., Dr. Takao is a member of the Society of Instrumentation and Control Engi-
vol. 1, no. 1, pp. 4/27–4/27, Feb. 1990. neers in Japan and the Institute of Systems, Control and Information Engineers
in Japan.

Toru Yamamoto (M’92) received the B.Eng. and

M.Eng. degrees from the University of Tokushima, Takaaki Yamada received the B.Eng. degree from
Tokushima, Japan, in 1984 and 1987, respectively, Osaka University, Osaka, Japan, in 1985.
and the D.Eng. degree from Osaka University, He joined in the Industrial Automation Company,
Osaka, Japan, in 1994. OMRON Corporation, Kusatsu, Japan, in 1985. His
He is currently a Professor with the Department of work is mainly in the field of development of temper-
Technology and Information Education, Hiroshima ature control.
University, Higashi-Hiroshima, Japan. He was a Vis- Mr. Yamada is a member of the Society of Instru-
iting Researcher with the Department of Mathemat- mentation and Control Engineers in Japan (SICE) and
ical Engineering and Information Physics, University the Information Processing Society of Japan. He re-
of Tokyo, Tokyo, Japan, in 1991, and, during the au- ceived a license of instrument and control engineers
tumn of 1996, he was a Visiting Professor with the Department of Chemical and from SICE in 2005.
Materials Engineering, University of Alberta, Edmonton, AB, Canada. In addi-
tion, he was an Overseas Research Fellow of the Japan Society for Promotion
of Science with the University of Alberta for six months in 2006. His current
research interests are in the area of self-tuning control, intelligent control, and
process control.
Dr. Yamamoto won the Best Poster Paper Award at the 3rd Asian Control
Conference in 2000 and was the recipient of the Prize of Progress from the
Institute of Electrical Engineers in Japan and the Technology Award from the
Society of Instrumentation and Control Engineers in Japan in 2003.

Authorized licensed use limited to: Toru Yamamoto. Downloaded on September 23, 2009 at 01:01 from IEEE Xplore. Restrictions apply.