900

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 56, NO. 3, MARCH 2009

From PID to Active Disturbance Rejection Control

Jingqing Han

AbstractActive disturbance rejection control (ADRC) can be

summarized as follows: it inherits from proportionalintegral
derivative (PID) the quality that makes it such a success: the
error driven, rather than model-based, control law; it takes from
modern control theory its best offering: the state observer; it
embraces the power of nonlinear feedback and puts it to full
use; it is a useful digital control technology developed out of an
experimental platform rooted in computer simulations. ADRC
is made possible only when control is taken as an experimental
science, instead of a mathematical one. It is motivated by the
ever increasing demands from industry that requires the control
technology to move beyond PID, which has dominated the practice
for over 80 years. Specifically, there are four areas of weakness
in PID that we strive to address: 1) the error computation;
2) noise degradation in the derivative control; 3) oversimplification
and the loss of performance in the control law in the form of a
linear weighted sum; and 4) complications brought by the integral control. Correspondingly, we propose four distinct measures:
1) a simple differential equation as a transient trajectory generator; 2) a noise-tolerant tracking differentiator; 3) the nonlinear
control laws; and 4) the concept and method of total disturbance
estimation and rejection. Together, they form a new set of tools and
a new way of control design. Times and again in experiments and
on factory floors, ADRC proves to be a capable replacement of PID
with unmistakable advantage in performance and practicality,
providing solutions to pressing engineering problems of today.
With the new outlook and possibilities that ADRC represents, we
further believe that control engineering may very well break the
hold of classical PID and enter a new era, an era that brings back
the spirit of innovations.
Index TermsActive disturbance rejection control (ADRC),
extended state observer (ESO), nonlinear proportionalintegral
derivative (PID), tracking differentiator.

I. I NTRODUCTION

HE BIRTH and large-scale deployments of the powerful

yet primitive proportionalintegralderivative (PID) control law dates back to the period of the 1920s1940s in the
last century, in response to the pressing demands of industrial
automation before, during, and particularly after World War II.
Its role in the explosive growth in the postwar manufacturing
industry is unmistakable; its dominance is evident even today
across various sectors of the entire industry. It is at the same
time undeniable that PID is increasingly overwhelmed by the
new demands in this era of modern industries where the unending pursuit of efficiency and the lack and cost of skilled
labor put a high premium on feedback control technologies.
Its merit of simplicity in the analog electronics era has turned
Manuscript received March 20, 2008; revised September 23, 2008. Current
version published February 27, 2009.
J. Han, deceased, was with the Institute of Systems Science, Academy of
Mathematics and Systems Science, Chinese Academy of Sciences, Beijing
100190, China.
Digital Object Identifier 10.1109/TIE.2008.2011621

into a liability in the digital control one as it cannot fully take

advantage of the new compact and powerful digital processors.
It appears that, as any technology, PID will eventually outlive its
usefulness, if it has not already done so. The question is, what
will replace this hugely successful control mechanism in the
21st century, retaining its basic soundness and, at the same time,
shedding its limitations? It is doubtful that such question was
even entertained systematically, let alone answered, in the past.
We believe that the answer lies in our understanding of both
the characteristics of PID and the challenges it faces. It is such
understanding that will lead us to propose further developments
in the PID framework and, perhaps, even a drastic innovation
toward a new generation of digital control solutions. In this
paper, we suggest that there are four fundamental technical
limitations in the existing PID framework, and we proceed to
propose the corresponding technical and conceptual solutions,
including the following: 1) a simple differential equation to be
used as a transient profile generator; 2) a noise-tolerant tracking
differentiator; 3) the power of nonlinear control feedback; and
4) the total disturbance estimation and rejection. Together, these
new tools combine to form the backbone of a new synthesis
of digital control law that is not predicated on an accurate and
detailed dynamic model of the plant and is extremely tolerant of
uncertainties and simple to use. Moreover, we denote this new
synthesis active disturbance rejection control or ADRC.
ADRC has been a work in progress for almost two decades
[1][7], with its ideas and applications appearing in the English
literature, amid some questions and confusions, sporadically
only in recent years; see, for example, [8][14]. In ADRC, we
see a paradigmatic change in feedback control that was first
systematically introduced in English in 2001 [8]. The conception of active disturbance rejection was further elaborated in [9].
However, even though much success has been achieved in practical applications of ADRC, it appears that this new paradigm
has not been well understood and there is a need for a paper that
provides a full account of ADRC to the English audience [13].
Such need is unmistakable in the recently proposed terminologies such as equivalent input disturbance [14] and disturbance
input decoupling [15], all of which can be seen as a special case
of ADRC where only the external disturbance was considered.
It is primarily for this reason that this paper is written.
In this paper, we start with, in Section II, classical PID, a
dominant technology in industry today, and discuss its characteristics and weaknesses, followed by the proposed remedies
in Section III and the resulting ADRC control scheme in
Section IV. How this new framework is used to solve various
kinds of control problems is shown in Section V. To further
help users master this new methodology, some key points in the
application of ADRC are presented in Section VI, followed by
the concluding remarks in Section VII.

0278-0046/$25.00 2009 IEEE

Authorized licensed use limited to: Cleveland State University. Downloaded on March 1, 2009 at 09:06 from IEEE Xplore. Restrictions apply.

HAN: FROM PID TO ACTIVE DISTURBANCE REJECTION CONTROL

Fig. 1.

901

PID control topology.

II. C LASSICAL PID

Classical PID is a particular rather primitive and simplified
implementation of the basic principle in error-based feedback
control, as shown in Fig. 1, where the error between the
setpoint v = const and plant y, i.e.,e = v y, as well as its
t
differentiation de/dt and integration 0 e d are used in a linear
combination to produce the control law
t
u = k0

e d + k1 e + k2

de
.
dt

III. I NDIVIDUAL R EMEDIES

(1)

This is where PID takes its name: proportionalintegral

derivative. It found widespread applications in industry, even
when there is little or no information available regarding plant
dynamics. To help users understand this, we use the following
second-order system equation:
 x = x
1

x 2 = a1 x1 + a2 x2 + bu
y = x1

(2)

commonly found in practice, such as motion control, to illustrate why is it that a PID can be easily configured and tuned to
do its job.
Let e = v y = v x1 = e1 , e 1 = x 1 = e2 , and e =

x1 , and the error dynamics can be seen as


e 1 = e2
(3)
e 2 = a1 e1 + a2 e2 a1 v bu.
t
Denoting e0 = 0 e d , then e 0 = e = e1 , and (1) becomes
u = k0 e0 + k1 e1 + k2 e2 .

(4)

Together with (3), the error equation can be rewritten as

e 0 = e1
e 1 = e2

(5)
e = bk e + a1 v +(a bk )e +(a bk )e
2
0
0
1
1 1
2
2 2
bk0
which is asymptotically stable if

bk0 > 0, (bk1 a1 ) > 0, (bk2 a2 ) > 0
(bk1 a1 )(bk2 a2 ) > bk0 .

obvious in the presence of ever more demanding control system

performance.
To be more specific, we believe that there are four fundamental issues to be addressed in the PID framework. 1) Setpoint
is often given as a step function, not appropriate for most
dynamics systems because it amounts to asking the output and,
therefore, the control signal, to make a sudden jump. 2) PID
is often implemented without the D part because of the noise
sensitivity. 3) The weight sum of the three terms in (4), while
simple, may not be the best control law based on the current
and the past of the error and its rate of change. 4) The integral
term, while critical to rid of steady-state error, introduces other
problems such as saturation and reduced stability margin due to
phase lag. It is these fundamental limitations of PID that prompt
us to offer the following remedies.

(6)

That is, the design objective e1 = e = v x1 0, or x1

v, is met if the gains k0 , k1 , and k2 are selected to satisfy (6),
for the given range of a1 , a2 , b = 0. It is rather obvious that,
for most of the plants of the form of (2), a set of PID gains can
be easily found, analytically when the model is given or by trial
and error when it is not. Although such simplicity and the ease
of tuning could very well be behind the popularity of PID, they
also mark its fundamental limitations, which are made glaringly

We now offer the following practical solutions to address

the aforementioned issues. They are easy to implement and understand, without introducing undue complexity to the existing
controller.
A. Setpoint Jump
To avoid setpoint jump, it is necessary to construct a transient
profile that the output of the plant can reasonably follow. While
this need is mostly ignored in a typical control textbook, engineers have devised different motion profiles in servo systems.
A simple easy-to-use solution is offered in this paper.
It is well known that, for a double integral plant

x 1 = x2
(7)
x 2 = u
with |u| r and v is the desired value for x1 , the time-optimal
solution is


x2 |x2 |
u = r sign x1 v +
.
(8)
2r
Using this principle, a desired transient profile is obtained by
solving the following differential equation:

v 1 = v2

|v2 |
(9)
v 2 = r sign v1 v + v22r
where v1 is the desired trajectory and v2 is its derivative. Note
that, depending on the physical limitations in each application,
the parameter r can be selected accordingly to speed up or slow
down the transient profile. In addition, it is well known that this
continuous-time time-optimal solution in (8) could introduce
considerable numerical errors in a discrete-time implementation. To address this difficulty, a discrete-time solution for a
discrete double integral plant

v1 = v1 + hv2
(10)
v2 = v2 + hu, |u| r
was obtained as
u = f han(v1 v, v2 , r0 , h0 )

Authorized licensed use limited to: Cleveland State University. Downloaded on March 1, 2009 at 09:06 from IEEE Xplore. Restrictions apply.

(11)

902

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 56, NO. 3, MARCH 2009

where h is the sampling period, r0 and h0 are controllers

parameters, and f han(v1 , v2 , r0 , h0 ) is
d = h0 r02 ,
a0 = h0 v2 ,
y = v1 + a0

a1 = d (d + 8|y|)
a2 = a0 + sign(y)(a1 d)/2
sy = (sign(y + d) sign(y d)) /2
a = (a0 + y a2 )sy + a2
sa = (sign(a + d) sign(a d)) /2

a
sign(a) sa r0 sign(a).
f han = r0
d
Note that (11) is a time-optimal solution that guarantees the
fastest convergence from v1 to v without any overshoot and r0
and h0 are set equal to r and h, respectively. However, when
(10) and (11) are used for the purpose of defining a transient
profile, r0 and h0 can be adjusted individually according to the
desired speed and smoothness.
B. Tracking Differentiator
It is common in PID implementation that a differentiation of
a signal v is obtained approximately as
s
y=
v
s + 1
which can be rewritten as
y=

(12)

1
s + 1

C. Nonlinear Feedback Combination

PID, as a control law, employs a linear combination of
present, accumulative, and predictive forms of the tracking
error and has, for a long time, ignored other possibilities of
this combination that are potentially much more effective. As
an alternative, we propose the following nonlinear function:

f al(e, , ) =

e
1 ,

|e| sign(e),

|x|
|x|

(13)

Such approximation is still quite sensitive to the noise in

v because it is amplified by a factor of 1/ . That is, if v(t)
contains noise n(t), then v(t)

contains n(t)/ in its first term

in (13). We therefore conclude that (12) is not a good way
of approximating v(t).

Instead, we proposed the following

approximation:
v(t 1 ) v(t 2 )
2 1

(14)

which can be implemented approximately using the secondorder transfer function



1
1
1
w1 (s) =

,
2 > 1 > 0.
2 1 1 s + 1 2 s + 1
(15)
Moreover, as verified in simulations, this resolves the aforementioned problem of noise amplification. A particular second
approximation of a differentiator is s/( s + 1)2 , which corresponds to the differential equation with r = 1/
y = r2 (y v(t)) 2ry

(18)

that sometimes provides surprisingly better results in practice.

For example, with linear feedback, the tracking error, at best,
approaches zero in infinite time; with nonlinear feedback of
the form

1
(v(t) v(t )) v(t).

v(t)

provides the fastest tracking of v(t) and its derivative subject to

the acceleration limit of r. It is in this sense that (17) is denoted
as the tracking differentiator of v(t). In practical implementations, however, we again advise that its discrete version in (10)
and (11) be used to avoid unnecessary oscillations, where r0
and h0 are adjusted accordingly as filter coefficients.

f han(x1 , x2 , r, h0 )

or in the time domain as

y(t) =

fastest tracking that leads us to the tracking differentiator as

shown in the following.
Consider v(t) as the input signal to be differentiated; (9),
rewritten here as

x 1 = x2

|x2 |
(17)
x 2 = r sign x1 v(t) + x22r

(16)

u = |e| sign(e)
the error can reach zero much more quickly in finite time,
with < 1. Such can also help reduce steady state error
significantly, to the extent that an integral control, together
with its downfalls, can be avoided. An extreme case is = 0,
i.e., bang-bang control that can bring with it zero steady
state error without the I term in PID. It is because of such
efficacy and unique characteristics of nonlinear feedback that
we propose a systematic and experimental investigation. Such
nonlinear feedback functions in the forms of f al and f han play
an important role in the newly proposed control framework,
ADRC, as will be presented later in this paper.
D. Total Disturbance Estimation and Rejection via ESO
In this section, we introduce a new concept: total disturbance
and its estimation and rejection. Although such concept is, in
general, applicable to most nonlinear multi-inputmultioutput
(MIMO) time varying systems, we use a second-order singleinputsingle-output (SISO) example for the sake of simplicity
and clarity. Consider
 x = x
1

where y(t) tracks v(t), y(t)

tracks v(t),

and r determines the

speed. It is the desire to find a differentiator that yields the

x 2 = f (x1 , x2 , w(t), t) + bu
y = x1

Authorized licensed use limited to: Cleveland State University. Downloaded on March 1, 2009 at 09:06 from IEEE Xplore. Restrictions apply.

(19)

HAN: FROM PID TO ACTIVE DISTURBANCE REJECTION CONTROL

903

where y is the output, measured and to be controlled, u is the

input, and f (x1 , x2 , w(t), t) is a multivariable function of both
the states and external disturbances, as well as time.
The objective here is to make y behave as desired using
u as the manipulative variable, and for this purpose, unlike
any mathematical analysis of (19), f (x1 , x2 , w(t), t) does not
need to be expressively known. In fact, in the context of
feedback control, F (t) = f (x1 (t), x2 (t), w(t), t) is something
to be overcome by the control signal, and it is therefore denoted
as the total disturbance. At this point, we have transformed
a problem that traditionally belongs to system identification to
that of disturbance rejection, and its consequence is enormous.
Treating F (t) as an additional state variable, x3 = F (t), and
let F (t) = G(t), with G(t) unknown, the original plant in (19)
is now described as

x 1 = x2

x 2 = x3 + bu
(20)

x 3 = G(t)
y = x1
which is always observable. Now, we construct a state observer,
denoted as the extended state observer (ESO), in the form of
e = z y
1

f e = f al(e, 0.5, ),
z1 = z2 01 e

z
2 = z3 + bu 02 f e

z3 = 03 f e1

(21)

e = z y
1

f e1 = f al(e, 0.25, )
f e = f al(e, 0.5, ),
z1 = z1 + hz2 01 e

z2 = z2 + h(z3 + bu) 02 f e
z3 = z3 03 f e1 .

(22)

There are many ways to select the observer gains 01 , 02 ,

and 03 for a particular problem. For example, the observer
gains in (22) can be selected as
02 =

1
2h0.5

03 =

2
52 h1.2

(23)

for F (t) = sign(sin((0.001/h)t)). Simulation result shows

that the observer performs very well for a wide range of h, from
0.0001 to 1000.
For both continuous- and discrete-time forms of ESO in (21)
and (22), the observer gains can be made linear for the sake of
simplicity in implementation, replacing both f e and f e1 with
e. In this case, the corresponding gains are
01 = 1

02 =

allows the control law (u0 F (t))/b to reduce the plant in (19)
to a cascade integral form of
 x = x
1

x 2 = u0
y = x1

(25)

which can be easily controlled by making u0 a function of the

tracking error and its derivative, i.e., a PD controller. That is,
the control problem is transformed to that of estimation and
rejection of the total disturbance, and it is greatly simplified.
Numerous successful applications of this principle have demonstrated the validity of such an approach.
IV. P UTTING IT A LL T OGETHER : ADRC

f e1 = f al(e, 0.25, )

which is implemented in discrete form, with a sampling period

of h, as

01 = 1

Fig. 2. ADRC topology.

1
3h

03 =

2
.
82 h2

(24)

Also, note that the inputs to ESO are the system output y
and the control signal u, and the output of the ESO provides
the important information F (t) = f (x1 (t), x2 (t), w(t), t). This

Combining the transient profile generation, the nonlinear

feedback combination, and the total disturbance estimation
and rejection, the ADRC takes the form as shown in Fig. 2,
with the corresponding control algorithm in (26). Moreover,
the observer gains are given in (23), leaving only four tuning
parameters; among which, r is the amplification coefficient
that corresponds to the limit of acceleration, c is a damping
coefficient to be adjusted in the neighborhood of unity, h1 is the
precision coefficient that determines the aggressiveness of the
control loop and it is usually a multiple of the sampling period
h by a factor of at least four, and b0 is a rough approximation
of the coefficient b in the plant within a 50% range. Therefore, the main tuning parameter that caters the controller to a
particular application is h1 , with c functioning as a fine-tuning
adjustment

f v = f han(v1 v, v2 , r0 , h)

v = v + hv

1
1
2

v = v + hf v

2
2

e = z1 y

f e = f al(e, 0.5, h),

f e1 = f al(e, 0.25, h)
(26)
=
z
+
hz

e
z

1
1
2
01

z2 = z2 + h(z3 + b0 u) 02 f e

z3 = z3 03 f e1

e
e2 = v2 z2

1 = v1 z 1

u = f han(e1 ,ce2 ,r,h1 )+z3 .

b0
The last equation in (26) can be, of course, chosen differently
as a control law. Two alternatives are given as

u = 1 e1 +b02 e2 z3
u=

1 f al(e1 ,1 ,)+2 f al(e2 ,2 ,)z3

,
b0

Authorized licensed use limited to: Cleveland State University. Downloaded on March 1, 2009 at 09:06 from IEEE Xplore. Restrictions apply.

0 < 1 < 1 < 2

(27)

904

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 56, NO. 3, MARCH 2009

which amount to a linear and a nonlinear PD controller, respectively. In fact, by using different linear or nonlinear gain
combinations in the ESO and the feedback, one can easily find
over 100 different controllers in the same ADRC structure.
Regardless of which one of these control laws is chosen, we
would like to point out that the controller coefficients are
not dependent on the mathematical model of the plant, thus
making ADRC largely model independent. These coefficients
are primarily functions of the time scale, i.e., how fast the
plant changes. That is, the controller only needs to act as fast
as the plant can react, and in the previously given formulation,
this is implicitly represented by the choice of the sampling
period h.
V. A PPLICATION OF ADRC
From the aforementioned illustration, it is apparent that
ADRC has a much wider application range than PID. Even
though ADRC was presented for a relatively simple secondorder plant, the very same idea can be applied to solve
problems of much different nature and complexity. In this
section, we give a few cases that are particularly challenging
to PID and demonstrate the creative process in the applications
of ADRC.

B. Multivariable Decoupling Control

Consider the MIMO plant

x = F (x, t) + Bu
y = Cx

(32)

where D = CB is a nonsingular m m matrix. To formulate

the decoupling problem as an ADRC problem, we rewrite
(32) as
y = G + U

(33)

with the total disturbance G = CF (x, t) and the pseudo control

variable U = Du. The MIMO plant of (33) is completely
decoupled. A SISO ADRC controller is then designed for each
channel, each with yi and gi (x, t) as the output and the total
disturbance, respectively. Once such U is obtained, the real
control signal is computed as
u = D1 U.

(34)

In practice, we found that D does not need to be known

exactly, as some of its perturbation can also be rejected
in ADRC.
C. Cascade Control

A. Time Delay
Plants with time delays, such as
y = G(s)e s u

(28)

pose difficult challenges to all control design methods. Here,

we suggest the following three methods.
1) Transfer Function Approximation: There are various
methods of approximating the term e s by using linear transfer functions, for example, the Pade Approximation. However,
since ADRC is not predicated on an accurate model of the plant,
a simple approximate of e s 1/( s + 1) may prove to be
good enough. That is, (28) is approximated as
y = G(s)

1
u.
s + 1

(29)

2) Predictive Output Feedback: From (28), let

y0 = e s y = G(s)u

where the output of the second subsystem x2 is the input to

the first subsystem. Similar to the backstepping technique, we
apply ADRC to 1 using u1 = x2 as a pseudo control variable.
Moreover, once u1 is determined, it is then used as the setpoint
for 2 , where ADRC is also applied. In general, the innerloop should be designed to be faster than the outer loop, which
means that the timescale of the inner loop is made smaller than
that of the outer loop.
D. Parallel System Control

(30)

and design the control law based on the delayless y0 u

dynamics. Then, implement the control law using a predictive
output feedback e s y, which is essentially the idea of Smith
Predictor.
3) Predictive Pseudo Input Method: Let U = e s u be the
pseudo control signal; then, (28) becomes
Y = G(s)U.

Consider the two serial-connected subsystems 1 and 2

with both outputs measured, as described in

1 = f1 () + x2
1 : x
(35)
2 : x
2 = f2 () + u

y = x1

(31)

Once U , as a control law, is determined, the real control

signal can be found as u = es U . The prediction in both parts
2) and 3) can be readily approximated using the tracking
differentiator discussed previously.

In many control problems involving resonant modes, they

can be generally represented as

x
1 = 12 x1 21 1 x 1 + b1 u

x
= 22 x2 22 2 x 2 + b2 u

2
(36)
..

xm 2m m x m + bm u
m = m
x
y = c1 x1 + c2 x2 + + cm xm
which can be rewritten as in the SISO form as
y =

m

i=1

ci i2 xi 2

m

i=1

Authorized licensed use limited to: Cleveland State University. Downloaded on March 1, 2009 at 09:06 from IEEE Xplore. Restrictions apply.

ci i i x i + bu

(37)

HAN: FROM PID TO ACTIVE DISTURBANCE REJECTION CONTROL

905


with b = m
i=1 ci bi . This is clearly an ADRC problem and
should be treated accordingly if

m
m


ci i2 xi + 2
ci i i x i

i=1

i=1

is taken as a total disturbance.

Fig. 3. Relative order determination.

VI. S OME K EY P OINTS IN H OW THE ADRC I S A PPLIED

Even though the ADRC is presented for a second-order plant,
it is by no means limited to it. In fact, many complex control
systems can be reduced to first or second order, and ADRC
makes such simplification much easier by lumping many untrackable terms into total disturbance. Still, proper problem
formulation and simplification is perhaps the most crucial step
in practice, and we offer the following suggestions.
1) Identify, among many variables in a physical process,
which is the input that can be manipulated and which
is the output to be controlled. This may not be very
clear, and the choice may not be unique in digital control
systems when there are many variables being monitored
and many different types of commands can be executed.
In other words, the control problem itself sometimes is
not well defined at the beginning, and until we clearly
identify what the control problem is, including its input
and output, no amount of control theory can help us.
2) ADRCs order is chosen according to the relative degree
of the plant. In linear systems, this is easily obtained
in terms of its transfer function. For others, such as
nonlinear and time varying systems, this might not be
straightforward. From the diagram of the plant, the order
of the system can be determined simply by counting the
number of integrators in it. In the same diagram, however,
the relative degree can be found as the minimum number
of integrators from input to output through various direct
paths. For example, the plant shown in Fig. 3 is of fourth
order, but its relative degree is two.
3) The key in a successful application of ADRC is how
well one can reformulate the problem by lumping various
known and unknown quantities that affect the system
performance into total disturbance. This is a crucial
step in transforming a complex control problem into a
simple one.
4) Another effective method for problem simplification is
the intelligent use of the pseudo control variable, as
shown previously. ADRC shows an obviously quite different way of going about control design. There is really
a paradigm shift. Let us once again use the following
second-order system for illumination:
 x = x
1
2
x 2 = f (x1 , x2 , w(t)) + bu
y = x1

(38)

where w(t) is an external disturbance. As an extension

of the theory of differential equations, the problem of
concern in modern control theory may not necessarily

be the same problems faced by practitioners in industry.

The key difference may be revealed in how the term
f (x1 , x2 , w(t)) in (38) is dealt with. Given u(t), the
topological structure of the state trajectory in the state
space is completely determined by the structure and
properties of f (x1 , x2 , w(t)), and this problem is known
as the open loop analysis problem. Modern control theory
is concerned with a different problem. That is, given
f (x1 , x2 , w(t)), how do we find u(t) as a function of
the states so that the topological structure of the state
trajectory of (38) has the desired properties. This is the
perspective that is dominant in modern control, and this
dominance goes back to the 1950s and 1960s of the last
century when control theory was seen as a branch of
applied mathematics.
Things look much different from a control engineering perspective. The objective in the practice of control system design
is quite clear: Manipulate u so that the y = x1 (t) follows a
desired trajectory. In this view, f (x1 , x2 , w(t)) is a disturbance to be overcome by u in order to drive y where it
needs to be. For example, let r be the desired trajectory for
y to follow. To achieve y r, it is necessary that y r, i.e.,
f (x1 (t), x2 (t), w(t)) + bu(t) r(t) or bu(t) r(t) F (t),
F (t) = f (x1 (t), x2 (t), w(t)). In this mind set, what concerns
us is not the topological structure of the state or the property
of f (x1 (t), x2 (t), w(t)), but rather, it is the value of the latter
at t, or F (t). Moreover, as shown in the ADRC description,
f (x1 (t), x2 (t), w(t)) is dealt with as the signal to be estimated
based on the inputoutput data and compensated by the control
variable u, and it does not really matter what it is as a function
of the states. Moreover, the previously all-important distinction
between linear and nonlinear time-invariant and time-varying
systems becomes irrelevant.
VII. C ONCLUDING R EMARKS
As far as practice is concerned, PID has dominated the scene
for over 80 years, during which, few fundamental innovations
were produced. While much progress has been made in modelbased mathematical control science, it has yet to make a significant impact in industry, whose pressures force us to seek
alternatives elsewhere. This paper provides readers with such
an alternative: a new digital control technology arisen out of the
effort to address the shortcomings of PID. This paper gives an
account of each component of ADRC as well as its structure and
philosophy. This paper also demonstrates the problem solving
process and the way of thinking in how to apply ADRC to
various engineering problems.

Authorized licensed use limited to: Cleveland State University. Downloaded on March 1, 2009 at 09:06 from IEEE Xplore. Restrictions apply.

906

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 56, NO. 3, MARCH 2009

ADRC is the result of years of investigation, largely performed experimentally in computer simulations with the scientific spirit of daring imaginations, painstaking observations,
careful generalization and abstraction, and truthful verifications
of principles in real-world applications. It may help the newcomers to ADRC greatly if one abandons the initial How can
this be right? attitude and, instead, run a few simulations of the
proposed solutions and observe the results. Perhaps the facts, or
data, are more convincing than mere articulation of ideas.
If the PID was born early in the last century, where it was
instrumental in the rise of modern industry and if PID is still
dominant to this day, conceivably, a major overhaul is long
overdue and a new era could be just around the corner. Ushered
in by performance and practicality, perhaps ADRC will soon
be accepted as a viable alternative to PID, as we wait in
anticipation.
ACKNOWLEDGMENT
This paper was translated from Chinese to English by a longtime collaborator of the author, Zhiqiang Gao, at Cleveland
State University. He also revised this paper after Prof. Jingqing
Han suddenly passed away in April 2008.
R EFERENCES
[1] J. Han, Control theory: Model approach or control approach, Syst. Sci.
Math., vol. 9, no. 4, pp. 328335, 1989, (in Chinese).
[2] J. Han and W. Wang, Nonlinear tracking-differentiator, Syst. Sci. Math.,
vol. 14, no. 2, pp. 177183, 1994, (in Chinese).
[3] J. Han, Nonlinear PID controller, J. Autom., vol. 20, no. 4, pp. 487490,
1994, (in Chinese).
[4] J. Han, Extended state observer for a class of uncertain plants, Control
Decis., vol. 10, no. 1, pp. 8588, 1995, (in Chinese).
[5] J. Han, Auto disturbances rejection controller and its applications,
Control Decis., vol. 13, no. 1, pp. 1923, 1998, (in Chinese).
[6] J. Han, From PID to auto disturbances rejection control, Control Eng.,
vol. 9, no. 3, pp. 1318, 2002, (in Chinese).
[7] J. Han, Active disturbances rejection control technique, Frontier Sci.,
vol. 1, no. 1, pp. 2431, 2007, (in Chinese).
[8] Z. Gao, Y. Huang, and J. Han, An alternative paradigm for control
system design, in Proc. 40th IEEE Conf. Decis. Control, 2001, vol. 5,
pp. 45784585.
[9] Z. Gao, Active disturbance rejection control: A paradigm shift in
feedback control system design, in Proc. Amer. Control Conf., 2006,
pp. 23992405.
[10] B. Sun and Z. Gao, A DSP-based active disturbance rejection control
design for a 1-kW H-bridge DCDC power converter, IEEE Trans. Ind.
Electron., vol. 52, no. 5, pp. 12711277, Oct. 2005.
[11] Y. Su, B. Y. Duan, C. H. Zheng, Y. F. Zhang, G. D. Chen, and
J. W. Mi, Disturbance-rejection high-precision motion control of a
Stewart platform, IEEE Trans. Control Syst. Technol., vol. 12, no. 3,
pp. 364374, May 2004.
[12] Y. Su, C. Zheng, and B. Duan, Automatic disturbances rejection controller for precise motion control of permanent-magnet synchronous
motors, IEEE Trans. Ind. Electron., vol. 52, no. 3, pp. 814823,
Jun. 2005.
[13] D. Sun, Comments on active disturbance rejection control, IEEE Trans.
Ind. Electron., vol. 54, no. 6, pp. 34283429, Dec. 2007.
[14] J.-H. She, F. Mingxing, Y. Ohyama, H. Hashimoto, and M. Wu,
Improving disturbance-rejection performance based on an equivalentinput-disturbance approach, IEEE Trans. Ind. Electron., vol. 55, no. 1,
pp. 380389, Jan. 2008.
[15] M. Valenzuela, J. M. Bentley, P. C. Aguilera, and R. D. Lorenz, Improved
coordinated response and disturbance rejection in the critical sections of
paper machines, IEEE Trans. Ind. Appl., vol. 43, no. 3, pp. 857869,
May/Jun. 2007.

Jingqing Han received the B.S. degree in mathematics from Jilin University, Changchun, China, in 1958.
From 1963 to 1966, he studied as a Ph.D. student
at the Department of Mathematics and Mechanics,
Moscow University.
In 1958, he joined the Institute of Mathematics,
Chinese Academy of Sciences, Beijing, China,
where he was a Professor Emeritus with the Institute
of Systems Science, Academy of Mathematics and
Systems Science. He published six books and over
200 papers and served in many professional organizations throughout his career. In his pursuit of truth, he was never afraid of
putting his reputation on the line and challenging the establishment and the
status quo. Trained as a mathematician, he never ceased to seek solutions to the
pressing problems of the world in which we live.
Prof. Han was a leading scholar in China for over five decades, receiving
numerous prestigious awards, recognitions, and joint appointments at various
universities and research institutes.
Message from the Editor-in-Chief
Prof. Jingqing Han was a control theorist and educator who inspired generations of students and colleagues in China. He passed away on April 21,
2008 in Beijing, China, at age 71. He is survived by his wife of 46 years,
Pei Renshun, three grandchildren, and his sons, Xuefeng and Xuejun, along
with his daughter, Xuehua (Sarah Young), scattered across both sides of the
Pacific in China, Japan, and California, working in concert with his followers
to keep his legacy alive.
Prof. Han was known to be an independent thinker who made bold moves in
his illustrious career several times over, while making important contributions
in several distinct areas of study, including optimal control, game theory,
guidance and navigation, population growth, computer-aided control system
design, and, above all, active disturbance rejection control (ADRC). Perhaps,
he made his boldest move in 1989 when, upon reflecting on the several
decades of research worldwide and contemplating a seemingly unbridgeable
gap between theory and practice, he categorically rejected the basic premise of
mathematization of the world and the pure deductive reasoning in research, in
a paper titled Is it a Control Theory or Is it a Model Theory? In the following
two decades, until his death, he devoted his life to finding an alternative, and
find it he did, in ADRC, which turns modern control theory on its head. The
implication of this change of direction proved to be enormous both in theory
and practice, as hundreds of papers and a wide range of applications and patents
soon followed.
Born on November 1, 1937 to a poor peasant family of Korean decent in
Changbai, Jilin Province, a remote corner of China, Han first rose to prominence in China and beyond through his eye-opening work, in collaboration with
Jian Song, on Analysis and Synthesis of Time Optimal Control System, presented at the 3rd International Federation of Automatic Control World Congress
in Moscow in 1963. Interrupted by the Cultural Revolution, he nonetheless
returned afterward to lead a nationwide program introducing modern control
theory to a generation of scholars and educators in China whose contact with the
West had been cut off for over a decade. As a leading researcher, he continued
to make original contributions, such as the constructive method in linear
system theory, which establishes an intimate connection between the state space
representation and the polynomial matrix one, unlike anything available in
the West.
Prof. Han had a truly active and curious mind that refused to be bound by
artificial boundaries; his work transcends the distinctions between linear and
nonlinear systems, time-varying and time-invariant dynamics, modeling error
and external disturbances, etc., classifications that often box a researcher in for
life. He was a true inspiration to all who knew him, particularly his students and
long-time collaborators. With his passing, the continuation of the ADRC legacy
seems to largely fall on the shoulders of two groups, one headed by Huang Yi
in the Chinese Academy of Sciences, the other by Zhiqiang Gao at Cleveland
State University.
Prof. Bogdan M. Wilamowski, Editor-in-Chief
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

Authorized licensed use limited to: Cleveland State University. Downloaded on March 1, 2009 at 09:06 from IEEE Xplore. Restrictions apply.