See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/282411487

Industrial test setup for autotuning of PID controllers in large-scale

processes: Applied to Tennessee Eastman process

Conference Paper  in  IFAC-PapersOnLine · June 2015

DOI: 10.1016/j.ifacol.2015.09.012

CITATIONS READS

10 352

3 authors, including:

Esmaeil Jahanshahi Selvanathan Sivalingam

Norwegian University of Science and Technology Norwegian University of Science and Technology
43 PUBLICATIONS   300 CITATIONS    16 PUBLICATIONS   112 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

SUBPRO - Subsea production and processing View project

All content following this page was uploaded by Esmaeil Jahanshahi on 26 March 2018.

The user has requested enhancement of the downloaded file.

9th International Symposium on Advanced Control of Chemical Processes
9th
9th
JuneInternational
International Symposium
Symposium
7-10, 2015. Whistler, on
on Advanced
British Columbia,Control
Advanced Canadaof
Control of Chemical
Chemical Processes
Processes
9th
JuneInternational
7-10, 2015. Symposium
Whistler, on
BritishAdvanced
Columbia,Control
Canada of Chemical Processes
June 7-10, 2015. Whistler, British Available
Columbia, online
Canada
June 7-10, 2015. Whistler, British Columbia, Canada
at www.sciencedirect.com

ScienceDirect
IFAC-PapersOnLine 48-8 (2015) 469–476
Industrial
Industrial test
test setup
setup for
for autotuning
autotuning of
of PID
PID
Industrial
Industrial test
test setup
setup for
for autotuning
autotuning of
of PID
PID
controllers
controllers in
in large-scale
large-scale processes:
processes: Applied
Applied
controllers
controllers in large-scale
in large-scale processes:
processes: Applied
Applied
⋆⋆
to
to Tennessee
Tennessee Eastman
Eastman process
process
to Tennessee
to Tennessee Eastman process ⋆⋆
Eastman process
Jahanshahi** , Selvanathan Sivalingam,
Esmaeil
Esmaeil
Jahanshahi
Esmaeil
Jahanshahi * , Selvanathan Sivalingam,
,, Selvanathan
Esmaeil John *B.
Jahanshahi Schofield Sivalingam,
Selvanathan
John
John B. Schofield Sivalingam,
John B. Schofield
B. Schofield
Research & Development - Subsea, Siemens AS, Bratsbergvegen 5,
Research &
Research & Development -- Subsea,
Subsea, Siemens
Siemens AS, Bratsbergvegen
Bratsbergvegen 5,
7037
7037 & Development
Trondheim,
Research Norway (E-mail:
Development
Trondheim, - Subsea,
Norway (E-mail:
(E-mail: Siemens AS,
AS, Bratsbergvegen 5,
esmaeil.jahanshahi@siemens.com).
5,
esmaeil.jahanshahi@siemens.com).
7037 Trondheim, Norway esmaeil.jahanshahi@siemens.com).
7037 Trondheim, Norway (E-mail: esmaeil.jahanshahi@siemens.com).
Abstract:
Abstract:
Abstract:
Although many PID tuning approaches are available, it is not easy to find a method that does not
Abstract:
Although many
Although many PID PID tuning approaches
approaches are are available, it it is not
not easy
easy to find find aa method
method that that does
does not
require any
Although
require anymany PID tuning
engineer/operator interference.
tuning approaches
engineer/operator interference.are available,
In this work,
available,
In this it is
work, is we
we easy to
notpresent
present to aafind
fully automated
a method
fully automated does not
thatapproach
approach not
require
for PIDany
require any engineer/operator
tuning based on relayinterference.
engineer/operator feedback. This
interference. In
In this
method
this work,
work, we present
involves
we present aa fully
sending theautomated
fully relay feedback
automated approach
approachtest
for PID
for PID tuning
tuning basedbased on on relay
relay feedback.
feedback. This This method
method involves
involves sending
sending the the relay
relay feedback
feedback test test
dataPID
for from PLCsbased
tuning (Programmable
on relay Logic Controller)
feedback. This method into ainvolves
historian, analyzing
sending the test
the relay data using
feedback test
data from
data from PLCs (Programmable
PLCs (Programmable Logic Controller)
Logic Controller) into aa historian,
into historian, analyzing
analyzing the test
the test data using
data using
a tuning
data fromapplication to generate Logic
PLCs (Programmable a tuning report that
Controller) intocontains
a historian,PIDanalyzing
parameters theand testsending
data using the
aa tuning
tuning application
application to generate
to generate aa tuning
tuning report that
report that contains
contains PID parameters
parameters and sending the
areport
tuning back to the operator
application to generatestation
a to retune
tuning reportthethatcontrollers
contains inPID
PLCs.
PID parameters and
and sending
sending the
the
report
report back
back to
to the
the operator station to retune the controllers in PLCs.
This paper
report
This paper back is
paper is the operator
to focused
focused
on the
operator
on the
the
station to
to retune
following
station
following
threethe
retune
threethe
keys
keys
controllers
steps: 1)in
controllers
steps: 1) 1)inA
APLCs.
method to identify persistent
APLCs.
method to to identify
identify persistent
persistent
This
steady-state
This paper is
is focused
conditions
focused on
in athe
on following
control
following three
loop using
three keys
routine
keys steps:
operating
steps: 1) A method
data
method because any tuning
to identify test is
persistent
steady-state conditions
steady-state conditions in in aa control
control loop
loop using
using routine
routine operating
operating data data because
because any tuning
any tuning test is
test is
performed when
steady-state the process
conditions in a is operating
control loop at steady
using state,operating
routine 2) A novel procedure
data because toanyimplement
tuning relay
test is
performed
performed when
when the
the process
process is operating at steady state, 2) A novel procedure to implement relay
based tuning
performed whentest,the3) A newis operating
model at
at steady
steady state,
identification method 2) A
A novel
2) which
novelis procedure
a combination to
to implement
of frequency-relay
based tuning
based tuning test,
test, 3)process
3) A newismodel
A new operating
model identification
identification
state,
method
method which
which isprocedure
is a combination
a combination implement
of frequency-
of
relay
frequency-
domain
based and
tuning time-domain
test, 3) A newanalysis.
model Subsequently,
identification the
method identified
which plant
is a model
combination is used
of to obtain
frequency-
domain and
domain and time-domain
time-domain analysis. analysis. Subsequently,
Subsequently, the the identified plant plant model is is used to to obtain
PID tuning
domain
PID tuning and parameters
tuning time-domain
parameters
based on IMC
analysis.
based on IMC
IMC
design.
Subsequently,
design. the identified
identified plant model model is used used to obtain
obtain
PID
The approach
PID tuning parameters
has been based
parameters based
tested on on
onIMC design.
an industrial
design. test setup in which all the control loops of the
The approach
The approach has has been testedtested on on an
an industrial test test setup in in which all all the control
control loopsloops of of the
Tennessee
The approach
Tennessee has been
Eastman
Eastman
process
been tested
process
areon
are an industrial
controlled
industrial
controlled by a test setup
by a Siemens
Siemenssetup in which
PLC.
PLC. which
The all the
The necessary the control
necessary
relay loops
relay of the
parameters,
parameters, the
Tennessee
the hysteresis
Tennessee Eastman
and relay
Eastman process
process are
amplitude,
are controlled
for the by
controlled test
by aa Siemens
are estimated
Siemens PLC.
PLC. The necessary
automatically
The necessary relay
where
relay parameters,
interference
parameters,
the hysteresis
the hysteresis and and relay
relay amplitude,
amplitude, for for the test
test areare estimated
estimated automatically
automatically where where interference
by an
the
by anan
engineerand
hysteresis
engineer or
or an
or an
an
operator
relay amplitude,
operator is not for the
is not required.
the test
not required.
required.
The
The arenew
new
method automatically
estimated
method for
for model identification
for model where interference
model identification
identification
is robust
interference
is robust
robust
by
against
by an engineer
measurement
engineer or an operator
noises. The
operator is
is notproposed
required. The
method
The new
new method
is able
method to tune
for the important
model identification controlis
is loops
robust
against
against measurement
measurement noises.
noises. The
The proposed
proposed method
method is
is able
able to
to tune
tune the
the important
important control
control loops
loops
in the Tennessee
against measurement Eastmannoises.process
The successfully.
proposed method is able to tune the important control loops
in
in the
the Tennessee Eastman process successfully.
in the Tennessee
© 2015, Tennessee Eastman process
process successfully.
EastmanFederation
IFAC (International successfully.
of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.
Keywords: Auto-tuning, process performance, industrial process control
Keywords: Auto-tuning,
Keywords: Auto-tuning,
Auto-tuning, process process performance,
process performance, industrial process
performance, industrial
industrial process control
control
Keywords: process control
1. INTRODUCTION mode to obtain PID tuning values. Earlier authors pro-
1. INTRODUCTION
1. INTRODUCTION mode
mode to
to obtain PID tuning values. Earlier authors pro-
1. INTRODUCTION mode
posed to obtain
posed different
different
PID
PID tuning
obtainautotuning
autotuning tuning values.
methods
values.which
methods
Earlier
Earlier
which
have
have
authors
authors
great
pro-
great prac-
pro-
prac-
A typical industrial plant has hundreds of control loops posed posed
tical different
values.
different autotuning
However,
autotuning they methods
all
methods which
suffer
which have
from
have great
some
great prac-
major
prac-
A typical
typical industrial
industrial plant plant hashas hundreds
hundreds of of control
control loopsloops tical tical values.
values. However,
However, they
they all suffer from some major
A
where
A 90%industrial
typical of the loops plantarehascontrolled
hundreds by ofPID loops limitations
controllers
control tical values.that
limitations However,
that are they all
are explained
explained
suffer
allwell by Hang
suffer
well by
from
from et
Hang
some
some
et al.
major
al. (2002).
major
(2002).
where
where 90%
90% of
of the
the loops
loops are
are controlled
controlled by
by PID
PID controllers
controllers limitations
For example, thatthe are explained
Cohen-Coon well
methodby Hang
(Cohen et al.
and (2002).
Coon,
(Desborough
where 90% of and
the Miller,
loops are 2002). These
controlled by controllers
PID have
controllers limitations
For example, thatthe are explained
Cohen-Coon well
methodby Hang
(Cohen et al.
and (2002).
Coon,
(Desborough and and Miller,
Miller, 2002).
2002). These
These controllers
controllers have have For 1953) example,
requiresthe Cohen-Coon
an open-loop testmethod (Cohen
on the process and Coon,
(Desborough
to be tuned individually
(Desborough and Miller, to match
2002). Theseprocess dynamics
controllers in For
have example, the Cohen-Coon method (Cohen and and is thus
Coon,
to be
to be tuned
tuned individually
individually to
to match
match process
process dynamics
dynamics in 1953)
in 1953) requires
requires
inconvenient to
an
an open-loop
open-loop
apply. The
test
test on
on
disadvantage
the
the process
process
of the
and
and is
is thus
thus
closed-loop
order
to be to provide
tuned good control
individually to performance.
match process Although
dynamics the
in 1953) requires
inconvenient toan open-loop
apply. The test on
disadvantage the process
of the and is thus
closed-loop
order
order to
to provide
provide good
good control
control performance.
performance. Although
Although the
the inconvenient
step method to apply.
by apply.
Yuwana The
and disadvantage
Seborg (1982) of the
the
and closed-loop
the Bristol
heuristic
order to approaches
provide good by control
control engineers
performance. have been
Although proven
the inconvenient
step method to
by Yuwana The
and disadvantage
Seborg (1982) of and closed-loop
the Bristol
heuristic approaches
approaches by by control
control engineers
engineers have have been
been proven
proven step method by
heuristic
adequate
heuristic for a large by
approaches number
control ofengineers
control loops, proven step
the manual
have been method
method methodis the
is by Yuwana
the
need of and
Yuwana
need of
largeSeborg
and
large
setpoint
Seborg
setpoint
(1982)
(1982) and
and the
change
change
to Bristol
the trigger
to Bristol
trigger
adequate
adequate for a large
for aa large
large number
number of control
of control
control loops,
loops, the
the manual
manual method method
the tuning is the
which need may of large
drive setpoint
the process change
away to trigger
from the
tuning
adequate methods
for are very cumbersome
number of and timethe
loops, consuming
manual the tuning is the
which need may of large
drive setpoint
the process change
away to trigger
from the
tuning
tuning methods
methods are
are very
very cumbersome
cumbersome and
and time
time consuming
consuming the tuning
operating which
point. To may drive
overcome the
these process away
disadvantages, from the
Åström
in particular,
tuning methods forare
those
very plants
cumbersomewith slow and responses.
time consumingAlso, the tuning
operating which
point. To may drive
overcome the
these process away
disadvantages, from the
Åström
in
in particular,
particular, for
for those
those plants
plants with
with slow
slow responses.
responses. Also,
Also, operating
and point.(1984)
Hägglund To overcome
overcome
proposed these
an disadvantages,
automatic tuningÅström
con-
the
in improvement
particular, for in
thosecontrol
plants loop
with performance
slow mainly
responses. de-
Also, operating
and point.
Hägglund To
(1984) proposed these
an disadvantages,
automatic tuningÅström
con-
the improvement
improvement in in control
control looploop performance
performance mainly mainly de- de- and Hägglund (1984)
basedproposed an
the
pends
the on the experience
improvement in control and looptheperformance
process knowledgemainly de- of and troller
troller
that was
Hägglund
that was (1984)
based
on the relay
proposed
on the an automatic
relay
feedbacktuning
automatic
feedback tuning con-
technique.
con-
technique.
pends
pends on on the
on the
the experience
experience and
and factthe process
the process
process knowledge
knowledge of troller
This that
method wassoon basedbecame on the
a
of troller that was based on the relay feedback technique. relay
superior feedback
alternativetechnique.
to the
personnel.
pends It isexperience
a well-known and the that many industrial
knowledge of
personnel. It It isis aa well-known
well-known fact fact that
that many industrial This
many industrial This method
method soon
soon became aa superior alternative to the
personnel.
control loops
personnel. It is area poorly
well-knowntunedfact by trial
that and manyerror where This
industrial conventional
conventional soon became a superior alternative to the
tuning.
method tuning. became superior alternative to the
control loops
control loops
loopsofare are
are poorly
poorly tuned
tuned by
by trial and
trialtaken error
and error
error where conventional tuning.
where conventional tuning.
performance
control thepoorly
controltunedloopsby is not
trial and into consid-
where A huge progress has been made in the last three decades
performance of
performance of the
the control
control loops
loops is is not
not taken
taken into
into consid-
consid- A huge progress has been made in the last three decades
eration.
performanceHence, of the tuning methods without
takenhuman inter- A
eration. Hence,
eration. Hence, the control
the tuning
loops is
tuning methods
methods not
without
without
into consid-
human
human inter- in
inter- A huge
in huge
progress
the area
progress
the area
area
of auto
of auto
has
has been
auto
tuning
been
tuning
made
made of in
PID
of in
PID
the last
last three
three decades
thecontrollers.
controllers.
Major-
decades
Major-
ference draw
eration. Hence, more
the and
tuningmore attention
methods of thehuman
without researchers
inter- in
ity the
of the of
progress istuning
in lineof PID
with orcontrollers.
variations Major-
of the
ference
ference draw
draw more more
more and more
and more attention
more attention
attention of of the
of the researchers
the researchers
researchers ity in the
of area
the of
progressauto istuning
in line of PID
with orcontrollers.
variations Major-
of the
and
ferencepractising
draw engineers.
and ity
methodof the progress
suggested is in
byisÅström line with
and or
Hägglundvariations
(1984).of the
and
and practising
practising engineers.
engineers. ity
methodof the progress
suggested by in line
Åström with
and or
Hägglundvariations
(1984). ofLuy-
the
Luy-
and practising engineers.
Industrial experience has clearly indicated that it is highly method method
ben (2002) suggested
summarized
suggested by
by Åström
the
Åström and Hägglund
applications
and Hägglundand (1984).
extensions
(1984). Luy-
of
Luy-
Industrial experience
Industrial experience has has clearly
clearly indicated
indicated that that itit is highly ben
is highly ben (2002)
(2002)
auto-tuning
summarized
summarized
method and
the
the applications
applications
proposed the
and
and extensions
extensions
Auto-Tuning
of
of
Vari-
desirable to
Industrial
desirable to
have an has
experience
have an
push-button
push-button
option that
clearly indicated
option
on the
on the highly ben
it isHuman
Human
(2002) summarized
auto-tuning method and the applications
proposed the and extensions
Auto-Tuning of
Vari-
desirable
Machine to have
Interface an push-button
(HMI) to put aoption
control on the
loop Human
in tune auto-tuning
ation (ATV)
auto-tuning method
method.
method and
andIn proposed
the
proposedATV the
the Auto-Tuning
method, the
Auto-Tuning Vari-
ultimate
Vari-
desirable to
Machine Interface have
Interface (HMI)an push-button
(HMI) to to put option
put aa control on
control loop the
loop in Human
in tune
tune ation ation (ATV) method. In the ATV method, the ultimate
Machine frequency
ation (ATV)
(ATV) andmethod.
the ultimate
method. In
In the
thegainATV
ATV whichmethod, the
represent
method, the theultimate
most
ultimate
Machine Interface (HMI) to put a control loop in tune frequency and the ultimate gain which represent the most
⋆ Funding for this research was provided by SIEMENS AS frequency
important
frequency and
and the
process
the ultimate
information
ultimate gain
gain which
can
which be represent
directly
represent the most
extracted
the most
⋆
⋆ Funding for this research was provided by SIEMENS AS important
important process
process information
information can
can be
be directly
directly extracted
extracted
* Funding for this
⋆ Corresponding author
* Funding for this research was provided by
research was provided by SIEMENS AS SIEMENS AS important process information can be directly extracted
* Corresponding
Corresponding author
* Corresponding author
author
Copyright 2015 IFAC
2405-8963 © 2015, 469 Hosting by Elsevier Ltd. All rights reserved.
IFAC (International Federation of Automatic Control)
Copyright
Peer review©
Copyright © 2015
2015 IFAC
IFAC 469
469
Copyright ©under
2015 responsibility
IFAC of International Federation of Automatic
469Control.
10.1016/j.ifacol.2015.09.012
IFAC ADCHEM 2015
June 7-10, 2015. Whistler, BC, Canada Esmaeil Jahanshahi et al. / IFAC-PapersOnLine 48-8 (2015) 469–476
470
using a describing function. The recent studies on auto-
tuning of PID controllers are reported by Leva (2007).
The auto-tuning method based on relay feedback is used
widely in process industries due to its ease of implementa-
tion (Blevins and Nixon, 2011). This technique has several
advantages over other methods, for instance it is time-
saving and easy to use. The method is carried out under
closed-loop control and, with an appropriate choice of the
relay parameters, the process output can be kept close to
the set point. This maintains the operating point in the lin-
ear region where the frequency response is useful, hence the
method works well on highly nonlinear processes (Åström
and Hägglund (1988)). The method is also extended to be
applicable in presence of disturbances in the process (Hang
et al. (1993)).
The relay-based tuning method provides two pieces of
information, namely ultimate gain and ultimate period
which are used by for Ziegler-Nichols PI and PID tuning
rules . However, these tuning rules do not provide a good
trade-off between robustness and performance of control
loops. On the other hand, PI and PID tuning rules based
on IMC (Internal Model Control) design are preferred
because of their Pareto-optimality between performance
and robustness. The tuning test based on relay feedback
can easily be automated which we use for the model
identification.
Although relay based auto-tuning has so many advantages,
estimating the relay parameters to initiate the tuning pro-
cedure in plants is not straightforward. Moreover, tuning
procedure based on relay method needs to be initiated
in a control loop at steady state conditions to obtain ulti-
mate frequency and ultimate gain accurately. Generally, in
plants, visual inspection of trend plots of controller output
and process output is a way of judging whether the control
loop is in steady state or not. This is not easy when the
process variables are affected by too much measurement
noise or other sources. These two issues are addressed
in this article. We have developed a method to find the
presence of consistent steady state conditions in control
loops. In addition, we have fully automated relay based
tuning procedure by developing methods to estimate the
necessary parameters.
We have implemented our Relay Tuning Function Block
(RTFB) in the Simatic PCS7 environment. RTFB detects
consistent steady-state, estimates the required relay am-
plitude and hysteresis and performs the relay-feedback
test by a command from HMI. Simba Profibus is used for
communication between the PLC and the Tennessee East-
man process model (Ricker (2002)) running in MATLAB.
Aspen InfoPlus21 is used as the historian in the PIMAQ R
(Plant Information Management and Data Acquistion)
framework, and the PIMAQ tuning application analyses
the test data to obtain the tuning values. The PIMAQ
framework is one of the in-house products of Siemens Oil
and Gas Solutions.
This paper is organized as the following. The technical
description of RTFB and PIMAQ framework are described
in Section 2. The identification method and the tuning
rules are introduced in Section 3. The Tennessee Eastman
process is briefly described in Section 4. Results are
470
presented and discussed in Section 5, and concluding
remarks are given in Section 6.
2. INDUSTRIAL AUTO-TUNING SETUP
In order to develop a fully automatic tuning procedure
based on relay method, we need the following as has been
explained earlier.
(1) a method to detect an existence of a consistent steady
state behaviour in a loop
(2) a procedure to estimate necessary relay parameters
(3) an automatic way of selecting the controller (PID or
PI) and the tuning method (e.g. Ziegler-Nichols, IMC
based or SIMC method)
All the three issues are addressed below.
2.1 Implementation of RTFB
RTFB (Relay Tuning Function Block) is implemented in
the PCS7 Simatic environment (using SCL programming
language). RTFB is connected in series with the existing
PID controller (CA Function Block). All function blocks
are compiled, then binaries are executed in PLCs. The
main algorithms implemented in the RTFB are as follows.
Detection of a consistent steady state behavior: In gen-
eral, the visual inspection of process variables in trend
plots is a way to identify the presence of steady state
conditions. However, the steady state detection in control
loops of process plants can be automated by comparing
process output with set-point based on the student’s t-test
(Narasimhan and Jordache (1999)).
A simple algorithm is proposed in this paper to detect
the presence of a consistent steady state without using the
set-point information. The proposed algorithm is based
on the basic definition of the steady state: Steady state is
defined as the state of a system when it becomes settled
(If derivative of a quantity with respect to time is zero,
the quantity is said to be at steady state). However, in the
reality it is not possible to have a constant process variable
in a control loop due to noise.
The steady state detection algorithm in this paper is based
on three indices, I1 , I2 , and I3 which are estimated using
process output when the control loop is in the routine
operation. These indices must satisfy specific conditions
in order to conclude that the control loop is in consistent
steady state. The consistent steady state behaviour here
refers to the situation in which the process remains in the
steady state for a considerable time duration (e.g at least
30 sec in chemical process plants).
The process output is denoted by Y here. The main steps
involved in the method are as follows.
(1) Estimate recursive mean of Y (low pass filer), Yµ
(2) Estimate windowed mean of Y, Ywin
(3) Calculate VY = (Y − Ywin )2
(4) Estimate recursive mean of VY, σY2
(5) Calculate standard deviation of Y, σY
(6) Calculate upper and lower limits of Y using
YUL = Yµ + 0.5σY
YLL = Yµ − 0.5σY
 IFAC ADCHEM 2015
June 7-10, 2015. Whistler, BC, Canada Esmaeil Jahanshahi et al. / IFAC-PapersOnLine 48-8 (2015) 469–476 471
(7) If Ywin is between YUL and YLL , generate an index,
I1 = 1 , else I1 = 0 (This means that the process has
not reached steady state) .
(8) Generate second index I2 by calculating the absolute
value of the time derivative of I1 ,
I2 = |∆I1 /∆t| = |(I1,k − I1,k−1 )/∆t|
where t is time.
(9) Generate third index I3 by calculating the windowed
integral of I2 (window length Nw )
k Communication
I3,k = I2,j
j=k−Nw
(10) Consistent steady state exists if I3 = 0 and I1 = 1.
The advantages of the method are as follows: (1) it does
not require set point information, (2) the method works
well even in the presence of offset where the process
output never reaches the set-point, (3) the method detects
consistent steady state behaviour, (4) since the method
does not use setpoint, it can be used for open-loop systems
as well and (5) it is mathematically simple.
Relay hysteresis (PHYS): The relay test is sensitive to
the measurement noise, and may relay chattering happen
because of the noise. The easiest way to reduce the
influence of the noise is to use relay with a hysteresis.
Although, a small value for the hysteresis is chosen so that
it does not affect the process response, the hysteresis value
should be related to amount of the noise.
In the steady-state detection algorithm in above, we esti-
mate standard deviation of the process output σY . When
the process is at the steady-sate, σY is the noise standard
deviation. We set the relay hysteresis as
P HY S = 5σY (1)
Relay amplitude (PRAMP): Tiny oscillations in the
process output are expected when performing the relay
test. More precisely, the process output y should not vary
more than 10% of yss where yss is the steady-state value
just before staring the relay test. In addition, the relay
output cannot go beyond the controller saturation values
(umax and umin ). Therefore, the relay amplitude should
not be too large. On the other hand, the relay amplitude
must be specified large enough such that the process
output not be trapped inside the hysteresis, especially for
the processes with small gain or long time constant. For
this, the inverse gain of the process (G−1 ) is estimated
by RTFB and used to calculate a relay amplitude that
produces not larger than 10% variation in process output.
However, the variation is needed to be larger than the noise
level of the process output.
PRAMP is estimated using the following expression:
P RAM P = min(∆r1 , ∆r2 ) (2)
where
∆r1 = 0.9 × min(uss − umin , umax − uss )
−1
∆r2 = 0.1 × yss G (4)
Here, uss and yss are the controller and the process outputs
at steady state respectively. umax and umin are maximum
and minimum controller outputs set in PID algorithm (also
known as input constraints).
TE Process
Simba Profibus
(Matlab model)
PLC Operator Station
(PID, RTFB) (HMI)
Gateway (CG)
PIMAQ Tuning
PIMAQ
Historian
Application
Fig. 1. Information flow in the test setup
Disturbance rejection during the relay test: Although
relay feedback test is less sensitive to an external distur-
bance, the errors in the estimates of ultimate gain and
ultimate frequency grow as the magnitude of the load
change (disturbance) increases (Yu (2006)). In order to
reject load disturbances during the relay-test, we add an
integral action to the relay output. The integral time
constant for this integral is specified in number of the relay
cycles. This parameter (PTINT) specified based on the
user experience; we have used PTINT = 2000.
2.2 PIMAQ Framework
PIMAQ� R
(Plant Information Management and Data Ac-
quisition) developed by Siemens, Norway is a part of the
industrial IT solutions delivered for oil & gas production.
Since it logs all data from the PLC’s and HMI screens,
it can be used for analysis, production reporting and
condition monitoring events. It is also a central component
of Integrated Operations solutions.
We have developed our auto-tuning setup based on this
framework; the main components of the setup and the
information flow are illustrated in Fig. 1. The main steps
are as the following:
(1) The necessary relay parameters are either calculated
by the logic or set by operator.
(2) The chosen control loop is put into tune mode after
necessary conditions (e.g. steady-state) are verified
by the function block (RTFB). The procedure will be
stopped automatically if (a) controller constraints are
violated (b) experiment time exceeds the maximum
time set for tuning and (c) the operator sends a
command to stop the tuning.
(3) The communication gateway (CG) finds the signals
that are needed to be logged and sends them to the
PIMAQ historian.
(3) (4) The tuning application developed in PIMAQ frame-
work retrieves the data form the historian and eval-
uates tuning parameters based on the specified con-
troller structure and the tuning method. The default
controller structure varies based on type of the control
loop (i.e. temperature, level, flow or pressure con-
trol). If it is a temperature control loop the default
471
IFAC ADCHEM 2015
June 7-10, 2015. Whistler, BC, Canada Esmaeil Jahanshahi et al. / IFAC-PapersOnLine 48-8 (2015) 469–476
472

controller structure is set as PID. For other types of Kp e−θs

control loops, the default controller structure is PI. F OP T D : G1 (s) = . (7)
τ1 s + 1
(5) A tuning report that contains calculated tuning val-
For the PID tuning it is beneficial to use a second order
ues is sent back to the historian from where it is sent
plus time delay model,
to the HMI. The tuning values can now be tested in
function block that contains PID algorithm in test Kp e−θs
SOP T D : G2 (s) = 2 2 . (8)
mode for some time and can then be accepted by the τ s + 2ζτ s + 1
operator if satisfied. Many processes in the chemical industry and the oil
In the test setup presented in this paper, the PLC (where production such as liquid level in a tank are ,
PID controller and RTFB are running) communicates with Kp e−θs
the Tennessee Eastman process I/Os using Simba profibus. IP T D : G3 (s) = . (9)
s
Simba Profibus is a hardware interface that simulates
inputs and outputs for the PLC and supports up to 125 where Kp is static gain of the process, τ is the time
profibus slaves in real-time. In the automation industry, constant, θ is the time delay and ζ is the damping ratio
various protocols such as HART, Foundation Fieldbus or of the process response. Typically, response of a SOPTD
PROFINET may be used for the communication between process is classified into critical damping, over-damping or
field devices and PLCs. under-damping based on the ζ value.

Estimation of plant gain (Kp ): Kp is estimated by using

3. IDENTIFICATION AND TUNING cross spectrum and power spectrum estimates in frequency
domain.
3.1 Identification Pu,y (ω0 )
Kp = (10)
Pu,u (ω0 )
Mainly, two pieces of information, ultimate period and ul- where Pu,y (ω) is cross spectrum between controller output
timate gain, are directly estimated from relay feedback test (u) and process output (y), Pu,u (ω0 ) is power spectrum of
data. Conventionally, the ultimate period Pu is found by u and ω0 is zero. The estimation of Kp is independent of
finding crossings of the process output and the setpoint in model order and structure. The method has been tested for
the time domain. Then, the ultimate or critical frequency different processes with known gains where good estimates
ωc is found as of the process gains were achieved even in presence of
2π
ωc = . (5) measurement noise. Kp can be estimated for integrating
Pu systems easily by using the expression in Equation (11).
Instead, we use a frequency domain analysis which is 2π
robust against measurement noises. We exploit the fact Kp = (11)
that ωc is the fundamental frequency of the signal which Ku Pu
contains the peak power. We use Welch’s averaged, mod- Estimation of time constant (τ ): The time constant of
ified periodogram method (Welch, 1967) to estimate the the process is calculated similar to the approach by Wang
Power Spectral Density (PSD) of the relay test data. Then, et al. (2007) with a slight modification.
we take the frequency with the peak power as ωc . The The time constant for FOPTD: The time constant for
Power Spectrum method requires enough number of relay FOPTD when controlled by relay controller is derived as
oscillations (approximately 10 to 15 cycles ) to produce an 
accurate estimate of the ultimate frequency. Kp2 k12 − 1
τ= (12)
The ultimate gain is calculated by ωc
4h which is the equation originally used to estimate the time
Ku = , (6) constant (Yu, 2006) with Ku replaced by k1 . The constant
πa
where h is the relay amplitude and a is the response k1 is calculated by Wang et al. (2007) as follows.
amplitude. Instead of h and a, we use square root of
average powers of the input and output. 1
k2l−1 = | | (13)
G[j(2l − 1)ωc ]
The ultimate values (Ku and ωu ) are enough for tuning the
4h
PID controllers using Ziegler-Nichols or modified Ziegler- k2l−1 =  (14)
Nichols equations. In addition, the relay test data can ωc (2l − 1)| T
y(t) exp(−j(2l − 1)ωc t)dt|
be used for closed-loop process model identification. The Computation of integral over one period in equation (14)
model identification from the relay feedback has attracted is not accurate due to measurement noise. Hence, it is
significantly increasing attentions in the process control proposed to estimate the integral over the whole data
community. The model identification procedure used in length and then divide the integral by number of the
this work is described in the following. periods to minimize the effect of noise in the data.
The time constants (τ1 and τ2 ) for SOPTD are
Model Structures Most of the processes in chemical estimated by calculating k1 and k3 from equation (14) and
plants are approximated by first order plus time delay then solving the following system of equations.
(FOPTD), second order plus time delay (SOPTD) and
integrating plus time delay (IPTD) models. The obvious
Kp2 k12 − (1 + τ1 )(1 + τ2 ) = 0 (15)
choice for the PI tuning purpose for stable processes is a
first order plus time delay model Kp2 k32 − (1 + 9τ1 )(1 + 9τ2 ) = 0 (16)

472
 IFAC ADCHEM 2015
June 7-10, 2015. Whistler, BC, Canada Esmaeil Jahanshahi et al. / IFAC-PapersOnLine 48-8 (2015) 469–476 473

Table 1. SIMC-PI tuning rules 4. TENNESSEE EASTMAN PROCESS PLANT

Structure G(s) Kc Ti
Kp e−θs 1 τ1
FOPTD τ1 s+1 Kp τc +θ
min {τ1 , 4(τc + θ)}
Kp e−θs 1 τ1
SOPTD (τ1 s+1)(τ2 s+1) Kp τc +θ
min {τ1 , 4(τc + θ)}
IPTD
Kp e−θs 1 1
min {τ1 , 4(τc + θ)}
The Tennessee Eastman (TE) process model which is a
s Kp τc +θ
standard benchmark in the process control community is
Two solutions for τ1 and two solutions for τ2 are found by used in the test setup. The TE problem was introduced
solving this system of equations. For τ1 the larger solution, by Downs and Vogel (1993) in 1990. The process has eight
and for τ2 the smaller solution are the correct solutions. components, including four reactants (A, C, D, and E),
two products (G and H), an inert (B), and a byproduct
Estimation of time-delay (θ): The time delay between u (F).
and y is estimated by using the famous cross correlation
The reactions are
method with a slight modification. We use u̇, the derivative
A(g) + C(g) + D(g) −→ G(I) (product)
of u, which gives series of spikes for positive and negative
direction of relay (−h and +h). The lag at which cross
A(g) + C(g) + E(g) −→ H(I) (product)
correlation between derivative u̇ and y reaches maximum
is the time delay between u and y. Using relay with
A(g) + E(g) −→ F(I) (byproduct)
hysteresis reduces relay chattering significantly during the
test. In addition, we apply an algorithm to remove any
3 D(g) −→ 2F(I) (byproduct)
possible chattering from the data. Estimation of time delay
does not differ significantly between FOPTD and SOPTD
systems for the identification based on relay feedback test. All the reactions are irreversible and exothermic. The
The modified cross-correlation method has been tested for reaction rates are a function of temperature through an
different processes with known time delays and acceptable Arrhenius expression. The reactions are approximately
estimates of the time delays were obtained even with first-order with respect to the reactant concentrations.
highly noisy measurements. Obviously, the time delay for
integrating systems is simply Pu /4. The process has five major units: a reactor, a product con-
denser, a vapor-liquid separator, a recycle compressor, and
The selection of model structure is done based on shape a product stripper ( see Fig. 2). There are 41 measurements
factor analysis developed by Luyben (2001). and 12 manipulated variables. The process is explained in
details by Downs and Vogel (1993).
3.2 Tuning Rules
The control objectives for this process are typical for a
chemical process:
We consider the following structure for the PID controller.

1 Td s
 • Maintain process variables at desired values.
KP ID (s) = Kc 1 + + , (17) • Keep process operating conditions within equipment
sTi Tf s + 1
constraints
where Tf is the time constant of the low-pass filter on the • Minimize variability of product rate and product
derivative action. It is usually chosen to be 10% of the quality during disturbances (stream 11).
derivative time Td . Here, the low-pass filter is used only • Minimize movement of valves which affect other pro-
for the purpose of reducing the measurements noise effect. cesses
And the PI controller is as the following: • Recover quickly and smoothly from disturbances,
 
1 production rate changes or product mix changes.
KP I (s) = Kc 1 + (18)
sTi From the goals listed by Ricker (1996), the following
Three types of tuning rules used in this work are SIMC-PI, process variables need to be controlled:
IMC-PID and Modified Zeigler-Nichols PI. The Modified
Zeigler-Nichols tuning rules (Yu, 2006) in equation (19) (1) Production rate
are the same regardless of the model structure. (2) Mole % G in product
(3) Reactor pressure
Kc = Ku /3, Ti = 2×Pu , (19) (4) Reactor liquid level
where Ku and Pu are the ultimate gain and the ultimate (5) Separator liquid level
period respectively. The Skogestad PI (SIMC) tuning (6) Stripper liquid level
rules (Skogestad and Grimholt, 2012) are summarized in
Table 1. We use the control structure for Self-Optimizing Control
by Larsson et al. (2001). This control structure contains 17
In order to use the SIM-PI tuning rules for the SOPTD feedback control loops. Most of the control loops have been
model, we need to use the half-rule to obtain a FOPTD tuned by the automatic tuning methodology proposed in
model (Skogestad and Grimholt, 2012). this article. The exceptions were composition control of
τ20 τ20 Mole % of C in purge gas (Xmeas31) and the reactor
τ1 = τ10 + , θ = θ0 + , (20)
2 2 pressure (Xmeas7). Due to very slow dynamics, the tuning
where τ10 , τ20 and θ0 are the parameters of the SOPTD takes more than 24 hours for these two loops, and the
model. The IMC-PID tuning rules are given in Table 2 operating point changes in this time. The results for eight
(Lee et al., 2006). important control loops are discussed in the next section.

473
IFAC ADCHEM 2015
June 7-10, 2015. Whistler, BC, Canada Esmaeil Jahanshahi et al. / IFAC-PapersOnLine 48-8 (2015) 469–476
474

Table 2. IMC-PID tuning rules

Structure G(s) Kc Ti Td
Kp e−θs Ti θ2 θ2
1
 θ

FOPTD τ1 s+1 Kp τc +θ
τ1 + 2(τc +θ) 6(τc +θ)
3− Ti
3
τ 2 − 6(τθ +θ)
Kp e−θs 1 Ti θ2 c θ2
SOPTD (τ 2 s2 +2τ ζs+1) Kp τc +θ
2ζτ1 + 2(τc +θ) Ti
+ 2(τc +θ)
Kp e−θs 2
1 1 θ
IPTD s Kp τc +θ
– 2(τc +θ)

y5 u5
FI
y1 y10
FI u3 8 u6
FI
1 CWS 9
A JI Compressor Purge
y22 (y11)
y2 TI y20 y13
u11 PI (y ) y29
u12 6 XA
FI u1 y12 (y5)
2 Condenser
CWR LI XB y30
SC
D A XC y31
N
y3 7 A y32
y8 (y2) XD
y11 (y4) TI L
FI u2 PI y7 (y1) Y y33
LI y14 Z XE
3 5
E CWS
u7 FI E XF y34
10 R
y16
XG y35
y21 (y8) PI
(y10) XH y36
y23 XA
6 TI Vap/Liq
y24 XB A u10 Separator
N
y25 XC A CWR Stripper y
L FI y 12 TI (y18) XD y37
A
y26 XD Y y6 TI (y9) 9 y19 N
Z 3 XE y38
E A
y27 XE u9 FI L
R Reactor Y XF y39
LI Stm
y28 XF Z XG y40
y15 E
(y7) Cond R y41
y4 y17 XH

FI u4 u8
FI
4
C Product
11

Fig. 2. Tennessee Eastman process

Relay output (Process input) Relay output (Process input)

100 80
valve opening [scaled]

valve opening [scaled]

80 60

60 40

40 20

20 0
0 5 10 15 20 25 30 35 40 0 500 1000 1500
time (s) time [s]

D Feed rate (Process Output) Reactor temperature (Process Output)

39 62.5
Reactor temperature [scaled]

38
D Feed rate [scaled]

62

37
61.5
36
61
35

34 60.5
0 5 10 15 20 25 30 35 40 0 500 1000 1500
time [s]
time (s)

Fig. 3. Tuning data for D feed rate Fig. 4. Tuning data for Reactor temperature

474
 IFAC ADCHEM 2015
June 7-10, 2015. Whistler, BC, Canada Esmaeil Jahanshahi et al. / IFAC-PapersOnLine 48-8 (2015) 469–476 475

Table 3. Control loops chosen for tuning Reactor temperature (Process Output)
63.2
CVs Scaling factor MVs SIMC−PI
63 IMC−PID
A feed (xmeas1) 100 xmv3 ZN−PI
D feed (xmeas2) 0.01 xmv1 62.8

Reactor temperature [scaled]

E feed (xmeas3) 0.01 xmv2
C feed (xmeas4) 1 xmv4 62.6

Stripper underflow (xmeas17) 1 xmv7 62.4

Production rate (xmeas17) 1 Fp
Reactor temperature (xmeas9) 0.5 xmv10 62.2
Reactor level (xmeas8) 1 SP17
62

5. RESULTS AND DISCUSSION 61.8

61.6
We choose the important control loops in the Tennessee
Eastman process for the tuning test. These controlled 61.4
0 500 1000 1500 2000 2500 3000 3500
variables and the related manipulated variables are listed time [s]

in Table 3. We have scaled some of the controlled variables

such that their values become between 0 and 100. This Fig. 5. Response of reactor temperature to a step in
scaling is necessary for the analogue input drivers in the setpoint with different controllers given in Table 4
control system. Also, by this scaling, very small or very
IMC−PID, D feed rate
large numbers for the tuning values are avoided. 40

Fig. 3 shows the tuning data from the flow control loop

y [scaled]
35
for the reactant D (xmeas2). The model identification
algorithm uses shape of the relay response to identify the 30 set−point
model structure as an integrating process, Process output
25
0 10 20 30 40 50 60 70
0.059e−0.67s time [s]
G3 (s) = . (21) SIMC−PI, D feed rate
s 40
The identified models for the selected control loops and the
tuning parameters using different tunings rules are given
y [scaled]

in Table 4. We have used τc = 2×θ for SIMC-PI and IMC- 35

PID tunings, whereas the Modified Ziegler-Nichols tuning set−point

Process output
rules do not require such a tuning parameter. 30
0 10 20 30 40 50 60 70
Fig. 4 shows the relay test data for the reactor temperature time [s]
ZN−PI, D feed rate
control loop (xmeas9). The relay response is very similar 40
to a sine wave which suggests a SOPTD model. The model
y [scaled]

and the related tunings are provided in Table 4. The
responses of the reactor temperature for a step change in
the setpoint using different tunings are shown in Fig. 5.
As shown in this figure the Ziegler-Nichols tuning gives
the largest over-shoot. This is in agreement with the
theoretical robustness measures such as sensitivity peaks
and phase-margin which are given in the results table.
Fig. 6 shows response of the flow rate of D to a step in the
setpoint. The IMC-PID tuning does not give an integral
action for the integrating processes as we see in Table 4.
This leads to an sustained offset in the step response as
shown in Fig. 6.
We recommend to use PID settings for the slow loops
(with long time delay or long time constant) such as
temperature control loops. Having a SOPTD model is
beneficial to obtain proper PID settings as we identified
for the reactor temperature. However, for the integrating
processes using the SIMC-PI tuning is recommended,
because IMC-PID settings do not give an integral action.
Any load disturbance may cause the process deviate from
the desired operating point.
6. CONCLUSIONS
An industrial test setup has been developed as part of
the work in the article which can be used for different
35
30 set−point
Process output
25
0 10 20 30 40 50 60 70
time [s]
Fig. 6. Response of D feed rate to a step in setpoint with
different controllers given in Table 4
purposes like fault diagnosis and controller performance
assessment. A novel way of detecting a consistent steady
state behaviour in control loops both for open- and closed
loop scenario has been proposed. A procedure for esti-
mating necessary relay parameters to make relay based
autotuning method fully automatic without any manual
intervention is presented. The procedure for estimating
relay parameters and the method for detecting a steady
state behaviour in control loops have been successfully
tested using Tennessee Eastman standard problem. Meth-
ods for identifying model parameters from relay test data
have been applied to Tennessee Eastman process and PID
controllers are tuned based on the estimated models.
REFERENCES
Åström, K. and Hägglund, T. (1984). Automatic tuning
of simple regulators with specifications on phase and
amplitude margins. Automatica.

475
IFAC ADCHEM 2015
June 7-10, 2015. Whistler, BC, Canada Esmaeil Jahanshahi et al. / IFAC-PapersOnLine 48-8 (2015) 469–476
476

Table 4. Identified models and PI(D) tuning values for selected control loops in TE process
A feed Tuning rule Kc Ti Td |S| |T | GM PM DM IAE
SIMC-PI 4.33 6.96 – 1.39 1.22 4.54 56.69 1.67 30.39
0.13e−0.58s IMC-PID 4.33 – 0.097 1.29 1.00 5.30 74.04 2.24 17.88
G(s) = s
ZN-PI 6.80 4.64 – 1.75 1.30 2.83 46.10 0.87 22.71
D feed Tuning rule Kc Ti Td |S| |T | GM PM DM IAE
SIMC-PI 8.84 8.05 – 1.39 1.22 4.53 56.65 1.93 40.82
0.059e −0.67s
IMC-PID 8.44 – 0.11 1.29 1.00 5.29 74.02 2.59 125.61
G(s) = s
ZN-PI 13.23 5.37 – 1.75 1.30 2.82 46.1 1.00 40.28
E feed Tuning rule Kc Ti Td |S| |T | GM PM DM IAE
SIMC-PI 4.19 11.35 – 1.39 1.21 4.53 56.67 2.72 49.23
0.084e−0.95s IMC-PID 4.19 – 0.15 1.29 1.00 5.30 74.03 3.66 28.86
G(s) = s
ZN-PI 6.58 7.57 – 1.75 1.30 2.83 46.10 1.42 36.73
C feed Tuning rule Kc Ti Td |S| |T | GM PM DM IAE
SIMC-PI 24.5 12.25 – 1.39 1.22 4.53 56.66 2.94 53.12
0.013e −1.02s
IMC-PID 24.50 – 0.17 1.29 1.00 5.30 74.02 3.95 31.13
G(s) = s
ZN-PI 38.43 8.18 – 1.75 1.3 2.83 46.1 1.53 39.66
Stripper underflow Tuning rule Kc Ti Td |S| |T | GM PM DM IAE
SIMC-PI -2.03 5.92 – 1.38 1.00 4.27 73.16 7.22 60.50
−0.4865e−2s IMC-PID -1.92 5.59 0.29 1.29 1.00 5.27 73.94 7.72 60.51
G(s) = 5.26s+1
ZN-PI -6.88 7.86 – 4.74 3.88 1.31 24.28 0.68 78.96
Fp (production index) Tuning rule Kc Ti Td |S| |T | GM PM DM IAE
SIMC-PI 5.09 72.25 – 1.39 1.22 4.54 56.69 17.36 309.70
0.01e−6.02s IMC-PID 5.09 – 1.00 1.29 1.00 5.30 74.03 23.30 181.12
G(s) = s
ZN-PI 8.00 48.18 – 1.75 1.30 2.83 46.11 9.01 231.10
Reactor temperature Tuning rule Kc Ti Td |S| |T | GM PM DM IAE
SIMC-PI -16.65 280.95 – 1.69 1.29 3.61 46.17 56.42 1323
−1.67e−20.20s IMC-PID -19.33 1950.28 22.41 1.42 1.00 3.83 72.95 74.80 606.5
G(s) = 3716s2 +1947s+1
ZN-PI -25.16 192.73 – 2.39 1.88 2.28 31.52 26.46 1172
Reactor level Tuning rule Kc Ti Td |S| |T | GM PM DM IAE
SIMC-PI 1.52 2365 – 1.37 1.00 4.47 71.33 1435 1.20 × 104
G(s) = 1.29e−403s
2231s+1
IMC-PID 1.47 2297.71 63.19 1.29 1.00 5.28 73.97 1556 1.21 × 104
ZN-PI 2.48 3542 – 1.67 1.00 2.80 64.20 803.44 1.07 × 104

Åström, K. and Hägglund, T. (1988). Automatic tuning
of pid controllers. Instrument Society of America.
Blevins, T. and Nixon, M. (2011). Control Loop Foun-
dation, Batch and Continuous Processes. International
Society of Automation, Durham, NC.
Cohen, G. and Coon, G. (1953). Theoretical consideration
of retarded control. Trans. ASME, 75.
Desborough, L. and Miller, R. (2002). Increasing cus-
tomer value of industrial control performance monitor-
ing: Honeywell’s experience. Proc. AIChE Symp. Ser.,
98, 153–186.
Downs, J. and Vogel, E. (1993). A plant-wide industrial
process control problem. Computers Chemical Engineer-
ing.
Hang, C., Åström, K., and Ho, W. (1993). Relay auto-
tuning in the presence of static load disturbance. Auto-
matica.
Hang, C., Åström, K., and Wang, Q. (2002). Relay
feedback auto-tuning of process controllers — a tutorial
review. Process Control.
Larsson, T., Hestetun, K., Hovland, E., and Skogestad, S.
(2001). Self-optimizing control of a large-scale plant:
The tennessee eastman process. Industrial & Engineer-
ing Chemistry Research.
Lee, Y., Park, S., and Lee, M. (2006). Consider the
generalized imc-pid method for pid controller tuning of
time-delay processes. Hydrocarbon Engineering.
Leva, A. (2007). Comparative study of model-based pi(d)
autotuning methods. In American Control Conference.
Luyben, W. (2001). Getting more information from relay
feedback tests. Industrial & Engineering Chemistry
Research.
Luyben, W. (2002). Derivation of transfer functions for
highly nonlinear distillation columns. Process Control.
Narasimhan, S. and Jordache, C. (1999). Data reconcil-
iation and Gross Error Detection: An intelligent use
of process data. Data reconciliation and Gross Error
Detection: An intelligent use of process data.
Ricker, N.L. (1996). Decentralized control of the tennessee
eastman challenge process. Process Control.
Ricker, N.L. (2002). Tennessee eastman challenge archive.
URL http://depts.washington.edu/control
/LARRY/TE/download.html.
Skogestad, S. and Grimholt, C. (2012). The simc method
for smooth pid controller tuning. In PID Control in the
Third Millennium. Springer London.
Wang, P., Gu, D., and Zhang, W. (2007). Modified relay
feedback identification based on describing function
analysis. Industrial & Engineering Chemistry Research.
Welch, P.D. (1967). The use of fast fourier transform for
the estimation of power spectra: A method based on
time averaging over short, modified periodograms. Au-
dio and Electroacoustics, IEEE Transactions on, 15(2),
70–73.
Yu, C.C. (2006). Autotuning of PID Controllers - A Relay
Feedback Approach. Springer.
Yuwana, M. and Seborg, D.E. (1982). A new method for
on-line controller tuning. AIChE Journal, 28(3), 434–
440.

476
View publication stats