Available online at www.sciencedirect.

com

ScienceDirect
IFAC PapersOnLine 53-2 (2020) 1319–1324
Half-Gain
Half-Gain Tuning
Tuning for
for
Half-Gain
Half-Gain Tuning
Tuning for
for Control
Active
Active Disturbance
Disturbance
Half-Gain Rejection
Rejection
Tuning for Control
Active
Active Disturbance
Disturbance Rejection
Rejection Control
Control
Active
Gernot
Gernot Herbst
Disturbance Rejection Control∗∗
Herbst ∗ Arne-Jens
∗
Arne-Jens Hempel
∗∗
Hempel ∗∗ Thomas Göhrt ∗∗
Thomas Göhrt
Gernot Herbst ∗∗∗ Arne-Jens Stefan Hempel
Streif ∗∗ ∗∗ Thomas Göhrt ∗∗
Gernot Herbst ∗ Arne-Jens Stefan Streif Hempel ∗∗ ∗∗
∗∗ ∗∗ Thomas Göhrt ∗∗ ∗∗
Gernot Herbst Arne-Jens Stefan Streif Hempel ∗∗
Thomas Göhrt ∗∗
Stefan Streif ∗∗ ∗∗
∗∗
∗ Siemens AG, Clemens-Winkler-Str.
∗ Stefan Streif 3,
3, 09116
09116 Chemnitz,
Chemnitz, Germany
∗ Siemens AG, Clemens-Winkler-Str. Germany
Siemens
∗∗ Technische
∗∗ AG, Clemens-Winkler-Str.
Universität Chemnitz, 3,09107
09116Chemnitz,
Chemnitz,Germany Germany
∗ Siemens
∗ Technische
AG, Universität
Clemens-Winkler-Str. Chemnitz,
∗ ∗∗ Technische Universität Chemnitz, 09107 Chemnitz, Germany
∗∗ 09107
3, 09116 Chemnitz,
Chemnitz, Germany
Germany
Siemens AG, Clemens-Winkler-Str.
∗∗ Technische Universität Chemnitz, 09107 Chemnitz, Germany 3, 09116 Chemnitz, Germany
Abstract:
∗∗
Technische Universität Chemnitz, 09107 Chemnitz, Germany
Abstract: A A new new tuning
tuning rule rule is is introduced
introduced for for linear
linear active
active disturbance
disturbance rejection rejection control control
Abstract:
(ADRC), A new
which tuning
results in rule isclosed-loop
similar introduceddynamics for linearas active the disturbance
commonly rejection
employed control
bandwidth
(ADRC), which
Abstract: A results
new tuningin similar
rule isclosed-loop
introduced dynamics
for linear as active
the commonly disturbance employed
rejection bandwidth
control
(ADRC),
parameterization
Abstract: which
A new results
design,
tuninginbutsimilar
rulewith closed-loop
lower feedback dynamics linearas
forgains. In the
this commonly
manner theemployed
noise bandwidth
sensitivity of
parameterization
(ADRC), which design,
results but
inbut withisclosed-loop
similar introduced
lower feedback gains.
dynamics Inactive
as thiscommonly
the disturbance
manner the rejection
noise
employed sensitivitycontrol
bandwidth of
parameterization
the
(ADRC),controllerwhich is design,
reduced,
results paving
inpaving with
similar the lower
way
closed-loop feedback
for using gains.
ADRC
dynamics In
asin this
more manner the
noise-affected noise sensitivity
applications. of
It
the controller
parameterization is reduced,
design, but with the way feedback
lower for using ADRC
gains. inthe
In more
this commonly
noise-affected
manner theemployed
noise bandwidth
applications.
sensitivity It
of
the
is controller
proved
parameterizationthat isthereduced,
proposed
design, paving
but tuning
with thelowerway feedback
gains, for
whileusing ADRC
rooted
gains. in in this
the
In more
analyticalnoise-affected
manner solution
the noise applications.
of an algebraic
sensitivity It
of
is proved
the that isthe
controller proposed
reduced, tuning
paving the gains,
way while
for usingrooted
ADRC in the
in analytical
more solution applications.
noise-affected of an algebraic It
is
the proved
Riccati that isthe
equation,
controller proposed
can
reduced, always tuning
paving be the gains,
obtainedway while
from
for a
usingrooted
bandwidth
ADRC in the
in analytical
parameterization
more solution
noise-affected of anby
design algebraic
applications. simply It
Riccati
is proved equation,
that the can alwaystuning
proposed be obtained gains, from arooted
while bandwidth in the parameterization
analytical solution design
of an by simply
algebraic
Riccati
halving
is proved equation,
the
thatgains.
the can
This always
proposed be obtained
establishes
tuning a link
gains, from arooted
between
while bandwidth
optimal in parameterization
control
the analytical and pole
solution design
placement
of an by simply
design.
algebraic
halving equation,
Riccati the gains. Thisalways establishes a link between optimal control and pole placement design.
halving
Riccati the gains.can
equation, Thisalways
can
be obtained
establishes
be
from a bandwidth
a link between optimal control parameterization
and pole placement design bydesign. simply
halving
Copyright the© gains.
2020 The This establishes
Authors. Thisobtained
isaanlink
open from
between a bandwidth
access optimal
article under parameterization
control
the CCand design
pole placement
BY-NC-ND license bydesign. simply
halving
Keywords: theDisturbance
gains. This establishes
(http://creativecommons.org/ rejection a linkcase);
(linear
licenses/by-nc-nd/4.0) between optimalmethods;
Lyapunov control and pole placement
Observers for linear design.
Keywords: Disturbance rejection (linear case); Lyapunov methods; Observers for linear
Keywords:
systems; Disturbance rejection (linear case); Lyapunov methods; Observers for linear
systems; Time-invariant
Keywords: Time-invariant
Disturbance rejection systems.
systems. (linear case); Lyapunov methods; Observers for linear
systems; Time-invariant
Disturbance rejection systems.
Keywords:
systems; Time-invariant systems. (linear case); Lyapunov methods; Observers for linear
systems; 1. Time-invariant systems.
INTRODUCTION by
1. INTRODUCTION
1. INTRODUCTION by Zhou
Zhou et et al.
al. (2008),
(2008), denoted
denoted as as “low
“low gaingain feedback”.
feedback”. As As aa
by Zhouofetfact,
matter al. (2008),
applying denoted as “lowdesign gain feedback”.
to ADRC As a
will
1. INTRODUCTION matter
by Zhou ofetfact,
al. applying
(2008),
α-controller
denoted as
α-controller “low design
gain to ADRCAs
feedback”. willa
1. INTRODUCTION matter
lead
by Zhou to of fact,
reduced
etfact, applying
controller/observer
al. (2008), denoted as “low design
gains
gaincomparedto ADRC
feedback”. to will
the
toAsthea
α-controller
ADRC lead
matter to ofreduced controller/observer
applying gains
design compared
to ADRC will
ADRC was was developed
developed as as aa nonlinear
nonlinear general-purpose
general-purpose con- lead to ofreduced controller/observer gains compared to the
α-controller
con- lead established
matter
established
to reducedbandwidth
fact, applying
bandwidth parameterization
parameterization
controller/observer
α-controllergains design approach,
to ADRC
approach,
compared which
will
which
to the
ADRC
troller bywasHan
by developed
(2009).asA Aalinear
nonlinear
linear variantgeneral-purpose proposedcon-
was proposed by lead established
will in turn bandwidth
reduce noise parameterization
sensitivity of the approach,
resulting which
design.
troller
ADRC by Han
wasHan (2009).
developed variant was by to reduced controller/observer gains compared to the
troller
Gao (2009).as A alinear
nonlinearvariantgeneral-purpose
was proposed con- will in turn reduce
by established bandwidth noiseparameterization
sensitivity of the approach, resulting design.
which
Gao (2003),
ADRC
troller wasHan
(2003),
by facilitating
developed
(2009).asA
facilitating stability
alinear
nonlinear
stability analysis
variant and
general-purpose
analysis significantly
and proposed
was significantlycon-by will will in turn reduce
established
The in
mainturn bandwidth
reduce
contribution
noiseparameterization
noise of
sensitivity of the approach,
sensitivity
this work ofisthe
the
resulting design.
resulting
introduction
which
design.
of
Gao (2003),
reducing bythe facilitating
number
(2009).of stability
tuning analysis
linearparameters andwith significantly
the
troller
reducing
Gao (2003), Han
the number
facilitating of Atuning
stability variant
parameters
analysis was
andwith the intro-
proposed intro-
significantly by will The in
The
a new
main
mainturn
ADRC
contribution
reduce
contribution
tuning
noise ofsensitivity
of
rule,
this workofisthe
this
which work we is
theresulting
the
will
introduction
introduction
denote as
of
design.
of
“half-
reducing
duction
Gao the
of
(2003), the number
“bandwidth
facilitating of tuning parameters
parameterization”
stability analysis andwith the
approach. intro-
significantly a
The new main ADRC tuning
contribution rule,
of which
this work we iswill
the denote as
introduction “half-
of
duction ofthe
reducing thenumber
“bandwidth of tuning parameterization”
parameters with approach.
the intro- The a new
gain ADRC
tuning”. tuning
We will rule,
show which
that weiswill denote
design asaiming
“half-
duction
reducing of the “bandwidth parameterization” approach. main contribution of this work the introduction of
ofthe
thenumber of tuning parameters with the intro- again newtuning”.
ADRCWe will show
rule, that designasaiming
α-controller
duction
The “bandwidth parameterization” approach. tuning which we will denote
α-controller “half-
The seemingly
duction seemingly unorthodox use
use of
unorthodox parameterization”
of the “bandwidth of elements
elements from from modern
modern at
approach. again
at
gain
tuning”.
closed-loop
new ADRCWe
closed-loop
tuning”.
We will show
dynamics
tuning
dynamics
will rule,
show
that
similar
which
similar
that
to
towe bandwidth
will denote
α-controller
bandwidth
design aiming
parameter-
asaiming
“half-
parameter-
design
The
controlseemingly
theory unorthodox use of elements from modern
theory for creating
creating an almost model-free controller at closed-loop dynamics similar to bandwidth parameter-
α-controller
control
The seemingly for
unorthodox anuse almost
of model-free
elements from controller
modern ization
gain
ization
at will
tuning”.
will
closed-loop always
We
always willlead
dynamics show
lead to
to exactly
that
exactly
similar to halved
halved
bandwidth
α-controller gains
design
gains for
for
parameter-the
aiming
the
control
from theory
the user’s for creating
point of anuse almost model-free controller
The
from seemingly
control the user’sfor
theory unorthodox
point
creatingof view
view
an is
isof key
almost key to
to its
elements
model-free appraisal
its from
appraisalmodern
controller as ization
controller
at
as ization
controller
willand/or
closed-loop
will
always
and/or
always
lead similar
observer.
dynamics
observer.
lead
to Therefore,
to
exactlyto
Therefore,
exactly
halved
while
bandwidth
while
halved
gains for the
grounded
parameter-
grounded
gains for
in
in
the
from
a the user’s forpoint of view is al.
keymodel-free
to its appraisal as an
a “paradigm
control
from theory
“paradigm
the change”,
change”,
user’s creating
point cf.
cf.
of Gao
Gao
view et
an almost
et
is al.
key (2001);
(2001);
to its Gao
Gao
appraisal (2006),
controller
(2006), as controller
ization
an algebraic
will
algebraic
controller
and/or
and/or
Riccati
always
Riccati
observer.
equation,
lead to
equation,
observer.
Therefore,
an
exactly
an
Therefore,
whilegains
halved
α-controller
while
grounded
design
for
design
grounded
in
for
the
for
in
afrom
and“paradigm
a the change”,
differentiator
user’s point to cf.
other
of Gao
view et
is al.
model-free
key (2001);
to its Gao
approaches,
appraisal (2006),
e. g.
as an algebraic
ADRC can canand/or Riccati
be trivially
trivially equation,
obtained an
α-controller
from bandwidth
bandwidth design
parame-for
aand a differentiator
“paradigm change”, to other model-free approaches, e. g. an controller observer. Therefore, while grounded
α-controller in
and
aas a differentiator
introduced by to cf.
otherGaomodel-free
et al. (2001); Gao (2006),
approaches, g. ADRC
e.the algebraic be
beRiccati
obtained
equation, an from parame-
design for
by Fliess and Joinet (2013). And indeed, ADRC cansupersedingtrivially obtained from bandwidth parame-
α-controller
as“paradigm
and introduced
a change”,
differentiator Fliess
to cf.
and
otherGao Join al. (2001);
(2013).
model-free AndGao
approaches,indeed,(2006),
the
e. g. terization,
an algebraic
terization,
ADRC can Riccati
superseding
be trivially the
equation,
the need
need
obtained for
an
for
fromsolving
solving
bandwidth
α-controller the
the former.
design
former.
parame-For
for
For
aseaseintroduced
aof tuning, by Fliess and Join (2013). approaches,
And indeed, e.the
and
ease
as tuning,byits
ofdifferentiator
introduced its performance
to other
performance compared
model-free
compared to traditional
to indeed,
traditional g. ADRCterization,
an
an example,
superseding
detailed
cansuperseding
be trivially the
insights need
obtained are for solving
given
from into
bandwidththe
the former.
frequency-
parame-For
ease
as of tuning,
(PID-type)
introduced control,
by itsFliess
Fliess
and Join
performance
and its
and Join
(2013). And
compared
extendability
(2013). with
And to features
traditional
indeed,
the
de-
the an
and
example,
terization,
example,
time-domain
detailed
detailed
insights
the
insights
behavior
arefor
need
are
when
given
solving
given
using
intothe
into
ADRC
theformer.
the
frequency-
frequency-
with
For
(PID-type)
ease of tuning,
(PID-type)
sirable
control,
control,
and its extendability
its performance
and compared
its extendability
with
with
features
to features
traditional de- terization,
de- an andexample, superseding
time-domain detailed behaviorthe need
insights arefor
when solving
using
given ADRC
into the with half-
theformer. For
half-
frequency-
ease
sirableof for
(PID-type) for industrial
tuning,
industrial
control,
applications
its performance
applications
and its
as
as in
extendability in Herbst
compared Herbst
with
(2016)
to features
traditional
(2016) and
and
de-
andexample,
gain
an time-domain
tuning for for
detailedthebehavior
controller
insights whenand/or
are using
given ADRC
observer.
into the with half-
frequency-
gain tuning the controller and/or observer.
sirable
Madoński
(PID-type) for etindustrial
al.
al. (2019),
control, applications
and make
its it
it an
as attractive
extendability in Herbst (2016)
alternative
with features de- and
and gain time-domain
and tuning for thebehavior controller whenand/or using ADRC with half-
observer.
Madoński
sirable
Madoński
for real-world
et
for etindustrial
al.
(2019),
(2019),
control
make
applications
make
problems, it
an
an
cf.
attractive
as attractive
in Herbst
Zheng and Gao
alternative
(2016)
alternativeand
(2018). gain time-domain
tuning for thebehavior controller whenand/or using ADRC with half-
observer.
sirable for etindustrial applications asZheng
in Herbst (2016) and gain 2. tuning for the controller and/or observer.
for real-world
Madoński
for real-world
Madoński et al.
control
al. (2019),
control
(2019),
problems,
make
problems,
make itcf.
itcf.
an attractive
anZheng
attractive
and Gao
and Gao
(2018).
alternative
(2018).
alternative 2. ACTIVE DISTURBANCE REJECTION CONTROL
ACTIVE DISTURBANCE REJECTION CONTROL
for real-world control problems, cf. Zheng and Gao (2018). 2. ACTIVE DISTURBANCE REJECTION CONTROL
The core
Thereal-world of
core of ADRC ADRC is
is formed
formed cf. by an observer,
by an observer, which
which is
is 2. ACTIVE DISTURBANCE REJECTION CONTROL
for
The control is
coreasof “extended
ADRC problems,
formedobserver”by Zheng and Gaowhich
an observer, (2018). is This 2. ACTIVE DISTURBANCE REJECTION
denoted
denoted
The core asof “extended
ADRC is state
state
formed observer”
by an (ESO)
(ESO) and
observer, and
which puts
puts is This section
section provides
provides aa very
very brief
brief overview
overview of of CONTROL
continuous-
continuous-
denoted
emphasis
The core ason
of “extended
rejecting
ADRC is state
disturbances
formed observer”
by an in (ESO)
a broader
observer, and puts
sense.
which is This section
time linear ADRCprovides and a very
the brief overview
prevalent tuning ofmethod.
continuous-For
emphasis ason“extended
denoted rejecting disturbances
state observer” in a(ESO) broader and sense.
puts time section
This linear ADRC provides anda very
the prevalent
brief tuningofmethod.
overview continuous-For
emphasis
There
denoted areason rejecting
further
“extended disturbances
possible
state extensions
observer” in a to
(ESO)broader
the sense.
observer,
and puts time
a
Thismore linear
section ADRC
detailed provides and
introduction
a the
very prevalent
we
brief refer tuning
to
overview Herbst
of method.
(2013).
continuous-For
There areonfurther
emphasis rejecting possible extensions
disturbances in ato broader
the observer,sense. a
timemore detailed
linear ADRC introduction
and the we
prevalentrefer to
tuningHerbst (2013).
method. For
There as are onfurther possible extensions ato the observer,
sense. time a more detailed
linear ADRC introduction
and the prevalentwe refer to Herbst (2013).
such as
emphasis
such
There
tracking
tracking
are further
disturbance
rejecting
disturbance
possible
derivatives
disturbances
derivatives
extensions
using
in using the
Generalized
to broader
Generalized
observer, a more detailed introduction we refer tuning to Herbst method.
(2013).For
such as tracking
Proportional
There disturbance
Integral (GPI) derivatives
observers, using
cf. Generalized
Sira-Ramı́rez a
2.1 moreIdea detailed
and introduction
Structure of the we refer
Controller to Herbst (2013).
such as are
Proportional
Proportional
further
tracking Integral possible
disturbance
Integral (GPI)
extensions
(GPI)derivatives
observers, using
observers,
to
cf. the
cf.the
observer, 2.1 Idea and Structure of the Controller
Sira-Ramı́rez
Generalized
Sira-Ramı́rez 2.1 Idea and Structure of the Controller
et
et al.
such al.as(2017).
tracking
(2017). However,
disturbance
However, we
we will focus
focus on
derivatives
will onusingthe (arguably)
Generalized
(arguably) 2.1 Idea and Structure of the Controller
Proportional
et al. (2017). Integral
However, (GPI)
we observers,
will focus on cf.the Sira-Ramı́rez
(arguably)
most common
Proportional variant of the ESO, which incorporates aa 2.1The Idea
The essence andof ofStructure
linear
linear ADRC of thecan Controller
be
be described
described as as follows:
etmost
most (2017). Integral
common
al. common variant (GPI)
However, ofwe the observers,
thewill
variant ofdisturbance
ESO,
ESO,focuswhich
whichoncf.the Sira-Ramı́rez
incorporates
(arguably)
incorporates
essence ADRC
a The essence of linear ADRC can be described as follows:
can follows:
single
et al.
single lumped However,
(2017).
lumped (“total”)
(“total”) we will
disturbance focus state
on
state modeling
the
modeling (arguably)both
both
most
single common
lumped variant
(“total”) of the
disturbanceESO, which incorporates
state modeling both a The
(1) essence
assume
(1) essence of
assume ofanlinearan linear ADRCorder
n-th order can
n-th can be
integrator
integratordescribed chain as follows:
behavior
chainas behavior
unknown
most
unknown common dynamics
dynamicsvariant and piecewise
the ESO,constant
andofdisturbance
piecewise which
constant input
input distur-
incorporates distur- a The (1) P assume nn-thADRC be described follows:
single
unknown
bance
lumped
signals
(“total”)
dynamics
of the and
plant. piecewise constant state modelinginput distur- both (s)
(s) =
= bb0an/s n for a order
for a
integrator
single-input
single-input
chain behavior
single-output
single-output (SISO)
(SISO)
single lumped
bance signals
unknown (“total”)
of the and
dynamics plant. disturbance state
piecewise constant input distur-modeling both (1) assume
P
(s) = an
0 /s
for
nn-th order
a order integrator
single-input chain
single-output behavior
(SISO)
bance signals of the plant. (1) plant
assume
P
plant ofborder
of order
an
0 /snnn-thn,
for regardless
regardless
a single-input of
of its
integrator
its actual
actual chain
single-output structure;
behavior
structure;
(SISO)
unknown
bance dynamics
signals of the and
plant. piecewise constant input distur- P (s)
plant =of b 0
0 /s
order
n,
n n, regardless of its actual structure;
In this paper,
In thissignals
paper,ofwe we will
will explore the use of
explore the use of the so-called the so-called (2) apply
(2) plant (s)
apply of
P = the
theorder
b inverted
for
0 inverted
/s a gain 1/b
single-input at
gain 1/b00ofatitsthe
n, regardless the controller
single-output
controller
actual output
(SISO)
output
structure;
bance
In this paper, thewill
we plant.explore the use of the so-called (2) apply the inverted gain 1/b at the controller output
α-controller design
design for tuning
for tuning the
the observer
observer and control
andso-called
control (2) to
plant compensate
of
to compensate
apply the order
inverted for the
regardless
for the
n, gain plant
plant
1/b 0ofgain
its
gain
at theb 0 ;
actual
; structure;
controller output
In this
α-controllerpaper, we
design will
for explore
tuning the
the use
observer of the
and control to compensate for the plant 0 gain b0 ;
0 b 0
loop
In
loopthiswithin
α-controller paper,
within linear
linear ADRC.
we will
ADRC. explore It
It was
wastheput
putuseforward
of the
forward by Buhl
so-called
bycontrol
Buhl (3)
(2) set
(3) apply
set
to up
upthe aa full-order
compensate inverted
full-order
for observer
gain
the 1/b0 gain
observer
plant atfor
forthe the
the the
the integrator
controller
; output
integrator
loop design
within design
linear for
ADRC. tuning It wasthe observer and b
the put forward andby Buhl (3) chain
set upmodel a full-order observer for the the integrator
α-controller 0
0
and
and Lohmann
loop Lohmann
within
α-controller (2009)
(2009)
linear forand
ADRC.andtuningis
isItbased
based
was on
put the
observer
on the
forwardsolution
solution by of
control
of
Buhl an
an (3) to compensate
chain
set up modela (estimated
for the
(estimated
full-order states
plant
states
observer gain
forx̂ bthe;
x̂1,...,n
0 ),
),
the extended
extended
integratorby
by
and
loop Lohmann
algebraic
withinRiccati (2009)
linear equation
ADRC.and to isItobtain
based
was on the
feedback
put forwardsolution
gains by of
leading
Buhl an (3) chain
a
set model
constant
up a (estimated
input
full-order disturbancestatesfor
observer x̂1,...,n
(estimate
the
1,...,n ),theextended) by
that
integrator
algebraic
and Lohmann Riccati equation
(2009) and tois obtain
based feedback
on the gains leading
solution of an a constant
chain model input disturbance
(estimated states (estimate
x̂ ), x̂n+1 ) that
extended
x̂ by
algebraic
to
and anLohmann Riccati
exponentially equation
(2009) decaying
and tois obtain
Lyapunov
based feedback
on the gains leading
function
solution forof the
an acaptures
chainconstant
model both input
actual disturbance
(estimated disturbances
states (estimate
1,...,n
1,...,n
and ), model
n+1
extended
x̂ ) that
uncer-
by
to an exponentially
algebraic Riccati decaying
equation to Lyapunov
obtain feedback function
gains for the
leading acaptures
constant both actual
input disturbances
disturbance and model
(estimate
x̂1,...,n x̂n+1
uncer-
) that
to an exponentially
controlled
algebraic system.
Riccati A decaying
equationsimilar to Lyapunov
approach
obtain has
feedback function
been
gains for the
proposed
leading acaptures
tainties
constant both
(“extended actual
input disturbances
state
disturbance observer”, and
(estimate ESO); model
n+1 uncer-
n+1
) that
tocontrolled system. A similar approach
an exponentially has been proposed tainties (“extended
both actualstate observer”,and ESO); x̂
system. A decaying Lyapunov hasfunction for the captures disturbances model uncer-
n+1
controlled
to an exponentially similar approach been proposed tainties (“extended
both actualstate observer”,and ESO);
controlled system. A decayingsimilar approachLyapunov hasfunction
been proposed for the captures
tainties (“extended disturbances
state observer”, ESO); model uncer-
controlled system.©A2020
2405-8963 Copyright similar approach
The Authors. Thishasis anbeen proposed
open access article under the tainties
CC BY-NC-ND(“extended license state
. observer”, ESO);
Peer review under responsibility of International Federation of Automatic Control.
10.1016/j.ifacol.2020.12.1864
1320 Gernot Herbst et al. / IFAC PapersOnLine 53-2 (2020) 1319–1324

(4) compensate the disturbance using the estimate x̂n+1 ; (2013) in placing the closed-loop observer poles at λ =
(5) design a full-order state-feedback controller for the −kESO · ωCL , with kESO being the relative factor between
remaining “pure” integrator chain 1/sn to achieve the observer and control loop bandwidth:
desired closed-loop dynamics. n+1 !   
(λ + kESO · ωCL ) = det λI − AESO − lcT ESO (6)
Control law and observer equations are illustrated in = λn+1 + l1 λn + . . . + ln λ + ln+1 .
Figure 1, and given in (1) and (2) for the n-th order case.
1     Binominal expansion of (6) yields the observer gains:
u(t) = · k1 · r(t) − kT 1 · x̂(t) (1)
b0 (n + 1)!
li = · (kESO · ωCL ) ∀i = 1, ..., n + 1. (7)
i
T
with kT = (k1 · · · kn ) , x̂ = (x̂1 · · · x̂n+1 ) (n + 1 − i)! · i!
Comparing these two tuning tasks for linear ADRC, we can
  conclude that—in both cases—only integrator chains are
˙ = AESO · x̂(t) + bESO · u(t) + l · y(t) − cT
x̂(t) ESO · x̂(t)
(2) to be handled: of order n (for the closed-loop dynamics)
  and n + 1 (for the extended state observer).
 n×1 n×n  (n−1)×1
0 I 0
with AESO = 1×n , bESO =  b0 , 3. α-CONTROLLER APPROACH
0 0
0
T T
 1×n  3.1 Brief Overview of the Tuning Method
l = (l1 · · · ln+1 ) , cESO = 1 0
d n Buhl and Lohmann (2009) introduced the so-called α-
r
k1
1 u
Plant
y
controller approach, a design method leading to an expo-
b0
− nentially decreasing Lyapunov function for the closed-loop
system. The rate of decay α is the only design parameter
l of this method:

− V̇ = −αV with α > 0 and V = xT P x. (8)
x̂
b cT
With a plant as in (3):
 T
A ∂V
  V̇ = ẋ = 2xT P (Ax + bu)
kT 1 ESO ∂x
ADRC   (9)
= xT AT P + P A x + 2xT P bu
Fig. 1. Control loop with ADRC, consisting of extended = −αxT P x.
state observer (ESO) and state-feedback controller in-
cluding disturbance compensation. Signals: controlled A suitable control law for achieving a negative V̇ in (9) is:
variable y, control action u, reference r; and possible u = −bT P x. (10)
disturbances at plant input (d) and output (n).
Combining these two equations we obtain an algebraic
2.2 Tuning by Bandwidth Parameterization Riccati equation:
 α T  α 
Assuming perfect disturbance rejection and compensation A + I P + P A + I − 2P bbT P = 0. (11)
2 2
of the plant gain b0 by the control law (1), the state-
feedback controller kT has to be designed for a “virtual” In summary, the state-feedback controller gains for the
plant in form of a pure n-th order integrator chain: α-controller approach are kT = bT P , with P being the
ẋ(t) = A · x(t) + b · u(t), y(t) = x1 (t) (3) solution of (11).
 (n−1)×1 (n−1)×(n−1)   
0 I 0 (n−1)×1
3.2 Comparison to Bandwidth Parameterization
with A = , b= .
0 01×(n−1) 1
When applying the α-controller tuning approach to a loop
The predominant controller design approach in linear
with the plant being an integrator chain, the closed-loop
ADRC is called “bandwidth parameterization”, cf. Gao
poles may be complex-valued, but will all have the same
(2003), and is using pole placement with all poles at
real part of − α2 , cf. the proof in Buhl and Lohmann (2009).
λ = −ωCL , the desired closed-loop bandwidth:
   On the other hand, using the “bandwidth parameteriza-
n !
(λ + ωCL ) = det λI − A − bkT (4) tion” pole placement design as given in (4), all closed-loop
poles will be at −ωCL and real-valued only.
= λn + kn λn−1 + . . . + k2 λ + k1 .
Being the respective counterparts of PI and PID con-
Binominal expansion of (4) leads to the controller gains: trollers, first- and second-order ADRC are the most rel-
n! evant cases in practical applications, resulting in tuning
ki = · ω n−i+1 ∀i = 1, ..., n. (5) tasks for integrator chains of order up to three. When com-
(n − i + 1)! · (i − 1)! CL
paring the α-controller tuning results with pole placement
For tuning the extended state observer (ESO) with the (bandwidth parameterization for ωCL ), two observations
bandwidth approach, we will follow the notation of Herbst can be made:
 Gernot Herbst et al. / IFAC PapersOnLine 53-2 (2020) 1319–1324 1321

(1) Selecting the tuning parameters of both methods as

α = ωCL results in similar closed-loop dynamics for 1.5i · α 1.5i · α
integrator chains of order two and above, with a
slightly underdamped response for the α-controller 1i · α 1i · α
due to the complex-valued poles. For the second- 0.5i · α 0.5i · α
and third-order case, the closed-loop step response
achieved with these two methods is being compared
in Figure 2. A first-order α-controller design would −0.5 · α −0.5 · α
−0.5i · α −0.5i · α
be necessarily slower, since only one real-valued pole
(at − α2 , therefore at half the bandwidth of ωCL ) can −1i · α −1i · α
be placed.
(2) When designing with α = ωCL , the resulting con- −1.5i · α −1.5i · α
troller gains of the α-controller approach are ex- (a) n = 2 (b) n = 3
actly half of the controller gains obtained using pole
placement with bandwidth parameterization. We will 1.5i · α 1.5i · α
prove this relation in Section 4.
1i · α 1i · α
1 0.5i · α 0.5i · α

−0.5 · α −0.5 · α
−0.5i · α −0.5i · α
0.5
−1i · α −1i · α
Bandwidth parameterization, ω = 1
α-controller design, α = 1 −1.5i · α −1.5i · α
0 (c) n = 4 (d) n = 5
0 5 10 15
Time [s] Fig. 3. Closed-loop poles resulting from α-controller design
for integrator chain plants of orders n = 2 to n = 5.
(a) With plant being integrator chain of order n = 2 Bandwidth parameterization, by contrast, will always
place all poles at −α.
1
to a pole placement design with bandwidth parameteriza-
tion. Therefore the impact of measurement noise on the
control action will be reduced, making the α-controller
design an interesting alternative for noise-affected systems,
0.5
if the underdamped behavior is tolerable.
Bandwidth parameterization, ω = 1
α-controller design, α = 1 4. THE HALF-GAIN TUNING METHOD FOR ADRC
0
0 5 10 15 As pointed out in Section 3.2, there are up to three options
Time [s] of replacing bandwidth parameterization in linear ADRC
with the α-controller approach: (1) only for the controller
(b) With plant being integrator chain of order n = 3
(using α = ωCL , for ADRC of order n ≥ 2); (2) only for
Fig. 2. Comparison of the normalized closed-loop step the observer (using α = kESO · ωCL , possible for any order
responses using bandwidth parameterization (pole n ≥ 1); or (3) for both controller and observer (for n ≥ 2).
placement) and α-controller design for full-order Applying α-controller tuning to ADRC results in halved
state-feedback control of integrator chain systems of controller and/or observer gains, while maintaining similar
order n = 2 and n = 3. (albeit slightly underdamped) dynamics for the control
The closed-loop pole configurations of α-controller designs loop and/or the extended state observer. We will therefore
are presented in Figure 3, with the most important cases denote this design approach for ADRC as the “half-gain
being: tuning” method.

1 1
 This is the main result of this article, which will be proved
• λ1/2 = − ± i ·α for the second-order integra- in the following: To obtain the α-controller gains, the
2 2 Riccati equation (11) does not have to be solved. The
tor chain, and  √  gains can simply be obtained from the straightforward
1 1 3 bandwidth tuning rules, i. e. equations (5) (controller) or
• λ1 = − · α, λ2/3 = − ± i · α for the
2 2 2 (7) (observer), by halving these gains for a bandwidth
third-order integrator chain. ωCL = α (controller) or kESO · ωCL = α (observer).
Concluding this comparison: The α-controller design leads Theorem 1. For plants as given in (3), the controller gains
T
to similar closed-loop dynamics for systems of order two kBW obtained via bandwidth parameterization in (4) are
T
and above, but with only half the controller gain compared related to the α-controller gains kα from (10), (11) by
1322 Gernot Herbst et al. / IFAC PapersOnLine 53-2 (2020) 1319–1324

T
an exact factor of two, if kBW has been designed for a 5. EXAMPLES
bandwidth ωCL = α:
T 1 T 1 Aim of this section is to provide visual insights into an
kα = · kBW = · (kBW,1 · · · kBW,n ) . (12) ADRC-based control loop when using half-gain tuning for
2 2
the controller, the extended state observer, or both. For
Proof. We start by rewriting (11) as follows, this purpose we can restrict ourselves to a second-order
  plant with normalized gain and eigenfrequency:
αP = − AT P + P A + S, (13)
1
where, using (10) and (12), P (s) = 2 . (23)
s + 2s + 1
1 T 1
S = kBW kBW = · (kBW,1 kBW · · · kBW,n kBW ) . (14)
2 2 Since ADRC is almost insensitive to the damping ratio,
especially of underdamped systems, cf. Herbst (2013), the
Since A is an upper shift matrix, P A will result in P ’s informative value of our example will not be compromised
columns pi being shifted: by the particular choice of critical damping in P (s).
P A = (0 p1 · · · pn−1 ) . (15)
Bandwidth parameterization is applied to a second-order
From the first column of (13) we obtain p1 , and, as an ADRC (n = 2) using ωCL = 1 rad/s, kESO = 10, and
abbreviation, introduce Φ: b0 = 1. Four cases are being compared: (1) unmodified
kBW,1 bandwidth tuning, (2) applying half-gain tuning only to
αp1 = −AT p1 + kBW the outer control loop (“K/2 controller”), (3) applying
2 half-gain tuning only to the ESO (“L/2 observer”), and
  −1 kBW,1 kBW,1
p1 = αI + AT · kBW = Φ−1 · kBW . (4) half-gain tuning for both controller and observer.
2 2
(16)
5.1 Impact on Open-Loop Characteristics
For all other columns (i = 2, . . . , n):
For stability and dynamics, the feedback controller part of
kBW,i an ADRC control loop is essential. In Figure 4 the transfer
αpi = −AT pi − pi−1 + kBW
2 functions from controlled variable y to control signal u are
pi = −Φ−1 · pi−1 + Φ−1 ·
kBW,i
kBW . (17) compared for the four possible cases. Additionally, the loop
2 gain transfer functions are being compared in Figure 5.
We now recursively expand (17) for the final (n-th) col- |CFB | [dB]
umn: 50
n
kBW,i Original K/2 controller
pn = (−1)(n−i) · Φ−(n−i+1) · kBW . (18) 40
2 L/2 observer K/2 and L/2
i=1
30
pT
n is the gain vector of the α-controller, since, recalling
T
(10) with (3), kα = bT P = bT P T = pT n . Multiplying (18) 20
with Φn oneobtains: 
n
(n−i) (i−1) 1 10 −2
Φ · pn =
n
(−1) ·Φ · kBW,i · kBW . (19) 10 10−1 100 101 102
2
i=1  CFB [deg]
The characteristic polynomial of Φ is: 90
  
det (λI − Φ) = det λI − αI + AT = (λ − α)n . (20) 45
Comparing (20) with (4) and (5) when α = ωCL we find 0
the characteristic polynomial to be:
n
−45
det (λI − Φ) = λn − (−1)(n−i) · kBW,i · λi−1 . (21)
i=1 −90
10−2 10−1 100 101 102
This allows us to
apply the Cayley–Hamilton theorem to
(19), with Φn = i=1 (−1)(n−i) · Φ(i−1) · kBW,i we finally
n ω [rad/s]
obtain:
1 Fig. 4. Comparison of the feedback controller transfer
pn = kα = kBW . (22)
2 functions with or without half-gain tuning.
This concludes the proof. As the analytical solution of the The most interesting result might be that half-gain ob-
algebraic Riccati equation (11), it provides a link between server tuning (“L/2” case) provides significantly improved
optimal control and pole placement for linear ADRC.  high-frequency damping while having almost no impact
Remark 2. Due to the duality of the design problem, on the lower frequencies up to and including the crossover
a proof of the half-gain relation for the extended state frequency. On the other hand one has to expect some
observer design (with kESO · ωCL = α) can be constructed low-frequency performance penalty when (additionally or
in the same manner. solely) applying half-gain controller tuning (“K/2” cases).
 Gernot Herbst et al. / IFAC PapersOnLine 53-2 (2020) 1319–1324 1323

|G0 | [dB] halving the gains, as proved in this paper, establishing

50 a link between pole placement and optimal control.

ACKNOWLEDGEMENTS
0
Gernot Herbst would like to thank Michael Buhl for
−50 drawing his attention to the α-controller approach.
Original K/2 controller
L/2 observer K/2 and L/2 REFERENCES
−100
0 1 2
10−2 10−1
10 10 10 Åström, K.J. and Murray, R.M. (2008). Feedback Systems:
 G0 [deg] An Introduction for Scientists and Engineers. Princeton
−45 University Press.
Buhl, M. and Lohmann, B. (2009). Control with ex-
−90 ponentially decaying Lyapunov functions and its use
−135 for systems with input saturation. In 2009 Eu-
ropean Control Conference (ECC), 3148–3153. doi:
−180 10.23919/ECC.2009.7074889.
−225 Fliess, M. and Join, C. (2013). Model-free control. In-
ternational Journal of Control, 86(12), 2228–2252. doi:
−270 10.1080/00207179.2013.810345.
10−2 10−1 100 101 102 Gao, Z. (2003). Scaling and bandwidth-parameterization
ω [rad/s] based controller tuning. In Proceedings of the
2003 American Control Conference, 4989–4996. doi:
Fig. 5. Comparison of the open-loop gain transfer function 10.1109/ACC.2003.1242516.
with or without half-gain tuning. Gao, Z. (2006). Active disturbance rejection control: A
paradigm shift in feedback control system design. In
5.2 Impact on Closed-Loop Characteristics Proceedings of the 2006 American Control Conference,
2399–2405. doi:10.1109/ACC.2006.1656579.
With the control loop signals denoted as in Figure 1, the Gao, Z., Huang, Y., and Han, J. (2001). An alternative
“gang of six” transfer functions are defined as paradigm for control system design. In Proceedings of
     r(s)  the 40th IEEE Conference on Decision and Control. doi:
y(s) Gyr (s) Gyd (s) Gyn (s) 10.1109/CDC.2001.980926.
= · d(s) , (24)
u(s) Gur (s) Gud (s) Gun (s) Han, J. (2009). From PID to active disturbance rejection
n(s)
control. IEEE Transactions on Industrial Electronics,
providing frequency-domain insights for a two-degrees-of- 56(3), 900–906. doi:10.1109/TIE.2008.2011621.
freedom control loop as is the case with ADRC, cf. Åström Herbst, G. (2013). A simulative study on active dis-
and Murray (2008). For the four cases in our example, they turbance rejection control (ADRC) as a control tool
are presented and discussed in Figure 6 and Figure 7. for practitioners. Electronics, 2(3), 246–279. doi:
While not shown here for brevity, a discrete-time imple- 10.3390/electronics2030246.
mentation of “K/2” and “L/2” design was successfully Herbst, G. (2016). Practical active disturbance re-
tested as well, exhibiting the desired noise reduction in jection control: Bumpless transfer, rate limitation,
the control signal u. and incremental algorithm. IEEE Transactions
on Industrial Electronics, 63(3), 1754–1762. doi:
10.1109/TIE.2015.2499168.
6. CONCLUSION Madoński, R., Shao, S., Zhang, H., Gao, Z., Yang, J.,
and Li, S. (2019). General error-based active distur-
A new “half-gain tuning” rule for linear active distur- bance rejection control for swift industrial implementa-
bance rejection control (ADRC) based on the so-called tions. Control Engineering Practice, 84, 218–229. doi:
α-controller design was introduced. Compared to the 10.1016/j.conengprac.2018.11.021.
common “bandwidth parameterization” approach, similar Sira-Ramı́rez, H., Luviano-Juárez, A., Ramı́rez-Neria, M.,
closed-loop dynamics can be achieved with lower (halved) and Zurita-Bustamante, E.W. (2017). Active Distur-
feedback gains, therefore reducing the noise sensitivity of bance Rejection Control of Dynamic Systems: A Flat-
ADRC. ness Based Approach. Butterworth-Heinemann.
In view of the examples presented in Section 5, a rec- Zheng, Q. and Gao, Z. (2018). Active disturbance rejection
ommendation emerges to start with half-gain tuning for control: Some recent experimental and industrial case
the observer. This has the least impact on the closed-loop studies. Control Theory and Technology, 16(4), 301–313.
dynamics compared to bandwidth parameterization, while doi:10.1007/s11768-018-8142-x.
already providing a significant reduction of control signal Zhou, B., Duan, G., and Lin, Z. (2008). A parametric
sensitivity to measurement noise. Lyapunov equation approach to the design of low gain
feedback. IEEE Transactions on Automatic Control,
While being the analytical solution of an algebraic Riccati 53(6), 1548–1554. doi:10.1109/TAC.2008.921036.
equation, the proposed feedback gains can simply be
obtained from a bandwidth parameterization design by
1324 Gernot Herbst et al. / IFAC PapersOnLine 53-2 (2020) 1319–1324

|Gyr | [dB] |Gyd | [dB] |Gyn | [dB]

0 0 20
−20 −20 0
−40
Original −40 −20
−60 K/2 controller
−80 L/2 observer −60 −40
K/2 and L/2
−100 −80 −60
10−2 10−1 100 101 102 10−2 10−1 100 101 102 10−2 10−1 100 101 102
|Gur | [dB] |Gud | [dB] |Gun | [dB]
5 0 40

−20 30
0
−40 20
−5
−60 10

−10 −80 0
10−2 10−1 100 101 102 10−2 10−1 100 101 102 10−2 10−1 100 101 102
ω [rad/s] ω [rad/s] ω [rad/s]

Fig. 6. Gang-of-six comparison (frequency domain) with or without half-gain tuning for controller and/or observer
within ADRC. As predicted in Section 5.1, the half-gain observer (“L/2”) case provides enhanced high-frequency
damping in Gun (jω) almost without any side-effects on other performance criteria. The “K/2” cases, on the other
hand, will—while still yielding some additional high-frequency damping in Gun —involve slower reaction to reference
signal changes and more overshoot induced by disturbances at the plant input.

Step response of Gyr Step response of Gyd Step response of Gyn

0.2 1
1
0.8
0.1 0.6
Original
0.5
K/2 controller 0.4
L/2 observer 0 0.2
K/2 and L/2
0 0
0 2 4 6 8 10 0 2 4 6 8 10 0 1 2 3 4 5
Step response of Gur Step response of Gud Step response of Gun
1.2 0 20

1
0
−0.5
0.8
−20
0.6
−1
0.4 −40
0 2 4 6 8 10 0 2 4 6 8 10 0 1 2 3 4 5
Time [s] Time [s] Time [s]

Fig. 7. To ease and support the interpretability of Figure 6, a time-domain perspective is given in this figure with the
step responses of the gang-of-six transfer functions with or without half-gain tuning.