F'mceedings of the 40th IEEE

Conference on Decision and Control

Orlando, Florida USA,k e m b e r 2001 FrPO1-1
An Alternative Paradigm For Control System Design
Zhiqiang Gao 'I Yi Huang ') Jingqing Han
1) Dept. of Electrical and Computer Engineering.. Cleveland State Univ., Cleveland, Ohio 441 15 z.gao@csuohio.edu
2) Institute of Systems Science, AMSS, CAS, Beijing, 100080, P.R.China. yhuang@staff.iss.ac.cn

Abstract An alternative framework for control design is now. It is known as proponional-integral-derivative (PID)
presented in this paper. It compliments the existing theory in control, which dates back to 1922 [1,5,61, well before classical
that: 1) it actively and systematically explores the use of and modern control theory were bom. Today, PID remains as the
nonlinear control mechanisms for better performance, even for tool of choice in over 90% of industrial applications [7]. We
linear plants; 2) it represents a control strategy that is rather believe that there i s an underlining principle behind PID that
independent of mathematical models of the plants, thus achieving makes it effective.
inherent rohusmess and reducing design complexity. An
overview of this design philosophy and associated algorithms are At the same time, in working with industry on control technology
first inlmduced, followed by the summaries of preliminary issues, it appears to us that nonlinear control means are more
analysis results and practical applications. It is evident that the powerful than the linear one and. perhaps more importantly, the
proposed framework lends itself well in providing innovative availability of an accurate mathematical model should not be a
solutions to practical problems while maintaining the simplicity precondition of a good control design.
and the intuitiveness of the existing technology, namely PID.
The motivation of OUT work was to find an alternative research
1. lntmductian direction and methodology that are more in line with how control
is practiced. Philosophically, the research was carried out on the
Feedback control is a well-established engineering discipline that premises of Han's vision on control theory, introduced in two
has enjoyed a tremendous growth in both theoretical development seminal papers, first in 1989 [SI and then in 1999 [9]. This led to
and applications in the last few decades. A historical perspective a slew of promising control technjques that are free of a few
can be found in [I], and a comprehensive account of classical and fundamental limitations, such as linearity, time invariance,
modern control theory in [2]. The frequency domain based accurate mathematical representation of plant, etc.
classical control theory provided engineers with great insight on
how a single-input and single-output control (SISO) system In section 2, the design philosophy and the main components of
works, as well as a set of analysis and design tools such as Bode the new paradigm are introduced. The applications of the
and Nyquist plot, mot-locus, etc. On the other hand, the progress resulting new techniques are summarized in section 3. New
made in state-space based optimal control and estimation theory, challenges and questions to basic concepts, together with the
particularly the Kalman Filter theory, lead to the binh of modem preliminary analytical studies, are discussed in section 4. Finally,
control theory. Using singular value plot as a bridge, the concluding remarks and comments for future research are given
frequency domain insight of classical cuntml was incorporated in section 5.
into the framework of modem control. which lead to the well-
known H2 and H- synthesis methods. This provides a modem 2. The New Paradigm
control paradigm where both performance and robustness
specifications are brought into a common mathematical In his seminal paper published in 1989 181, Han contended that
framework for control synthesis of linear time invariant systems the modern control theory, originated from Kalman Filter, could
PI. be alternatively viewed as "model theory'' since modeling and
model based analysis and synthesis are its main thrusts. While it
Nonlinear control is a more challenging but less mature has undoubtedly brought significant advances to the science of
discipline. The main design methods include feedback control, "model theory" also carried with it fundamental
linearization, Lyapunov method, sliding mode, back stepping, limitations, most notably the robustness issue (or the reliance on
adaptive conml. etc. 13.41. Like its linear counter pan. nonlinear accurate mathematical representation of plant). These limitations
control law is designed based on the mathematical model of the could he the primary reasons that the practicing engineers have
plants. Progress has also been made in addressing the robustness been slow to embrace the more advanced methods. Following the
issue in nonlinear control. famous Gadel's "Incompleteness Theorem", Han suggested that
the solutions to the limitations of model theory lie outside its
In the meanwhile, the use of feedback control in industry, mostly framework [91.
known as industrial automation, has grown into a billion dollar
business. For example, it is not uncommon in manufacturing In panicular. even though the state space model (mostly the
industry to see a single manufacturing cell that uses over a (A,B,C) matrices) made it possible to discover great insight in
hundred servo loops. What was used to be a simple analog basic understanding of feedback control, such as controllability
controller has now been almost exclusively converted to digital and observability, it may not be the best vehicle lo carry out the
forms. With the rapid advance in computer hardware, new control design in practice because:
computational and network capabilities bring vast opportunities 1. The model -is not easily available in many engineering
to implement advanced control slrategies that t ~ l bring
y marked problem;
benefits to the performance. 2. Even if it is available, the resulting control law could be too
dependent on the accuracy of the model paameters and
Although there are many examples of successful applicaljons of suffers poor robustness:
modern control theory in industry, the core control technology
used in industry has remained unchanged for several decades

0-7803-7061-9/01/$10.00Q 2001 IEEE 4578

Considering marvelous feedback control mechanisms found in The above discussion regarding the nature of PID brings US to the
Nature, Han concluded that there must exist better control laws second question (QZ) and the task of systematically exploring
that are beyond the restrictions of the current control theory. more effective feedback mechanisms. To this end, several key
Taking cues from classical contml theory and engineering developments are introduced in the following sections, including
practice. as well as the long established methodology in the use of nonlinear feedback; the nonlinear differentiator; the
Guidance, Navigation and Control (GNC) in aeronautics and nonlinear PID; and the extended state observer. It will be shown
aerospace applications, Han proposed a new framework in that significant progress has been made both in developing new
seeking more effective control laws [8.10,11.12]. The key issues synthesis methods and in the analysis of the new control
to be addressed are: mechanisms. Furthermore, the new methods have made impacts
1. What do we have to know about the d a n t in order to control in solving challenging engineering problems and contributed in
it? and (Q1) new technology development [31-34).
2. How do we go about finding effective control mechanisms? (QZ)
Remark:
7he Theme Problem: To illustrate the new concepts, a GNC
example is used here where the motion of an object is described 1. Han’s methods to control system design should be viewed as
as complimentary, rather than in conflict. to modem contml theory
in that they address the issues that are not addressed within the

I i2= a(t)t U ( [ )
Y = XI
Here, XI and x2 and n(f) are position, velocity and “inherent
acceleration”, respectively, and u(f) is the control signal. In many
(1)
existing theory.

2. During the long history of control theory evolvement, there

were many ideas proposed in term of modifications of control
law, such as the ones in [13,14,37-391 and the references therein.
While we believe the work presented here was a more systematic
applications. nfrj is usually a nonlinear function of many
effort in such endeavor, which results in many unique algorithms
variables, i.e.
and practical applications, a complete comparison with past
4 f J =f(t.xi(l).x,flJ.wffJJ. (2) results is beyond the scope of this paper.
where w(f) represents internal and external disturbances. Wan
suggested that, instead of depending solely on the mathematical
2.1 Exploring Nonlinear Feedback
model of the plant. or the explicit function in (2).effective control
could be achieved based on general characteristics of the physical
The extension of his work in canonical forms of linear system led
process and certain information of the process that can be
Han to the fundamental question of linear and nonlinear systems
obtained in real time. In fact, engineers have long used an error-
[9,101. The notion of linear and nonlinear system originated from
based method known as proponional-integral-deri”ative (PID)
the classical mechanics theory where the systems of interests
controller to control such processes:
have no input and output, just states. The autonomous linear and
nonlinear systems with no input and output have distincl
topologies, which are not interchangeable. The feedback control
system, however, has an open topology that can be changed at
where the control action is bared on the current (P), past (I) and
will by employing feedback. With feedback, a linear system can
future (D) of the error: e(t) = y&) - y(t). Hcre y,(t) is the desired
become a nonlinear one, and vice versa. In other words, feedback
output and Kp,K,, and KO are controller gains that can be
control break down the boundary between linear and nonlinear
intuitively tuned. Intuitive and effective, this design approach
system. Instead of using state feedback to place desired poles, we
remains the tool of choice for engineers.
can think of placing desired nonlinearity inlo the system. Alone
this line of thoughts. Han proposed a new thought: discard the
In addition to PID, classical control theory provided additional
traditional thinking of linear and nonlinear systems and fully
control blocks such as lead and lag compensators that further
explore the potentials of feedback. especially the nonlinear
enhanced the performance of this error-based control law. m e
feedback [IO. I1,12], as shown in the following examples.
practice of control engineering demonstrates that :
1. An accurate mathematical model. such as the function
Examole 1: Consider a single integrator system with disturbance
~ f , ~ , ( f J , ~ ~ ( f j , ~is( lnot
j j , a precondition to devisc an effective
WO:
control law. In fact, model based design is inherently
susceptible to the accuracy of the model and the dynamic e=w,tu (4)
change in the plant; the objective is to design a control law such that the closed-loop
2. “Conuol” refers to a process of driving a physical variable to system satisfies e(t) + 0, as t --f -. Comparing a standard linear
its destination. Model based design invariably links the control proportional controller, U = -Ke,to a nonlinear one,
law to the global characrerisfics of the plant, represented by
its model. Analysis and synthesis of the global characteristics U = er .sign(e) (5)
of the open and closed loop system provide insight to such a for the same gain K=10, when 1 w 0 I<I,vt, the steady state
process, but they are not the objective themselves. The error is less than 0.1 for the linear controller and 0.01 for the
engineering practice s e e m to suggest, that the control be nonlinear controller with a = 112. This bound will be further
respondent to the local characteristics, such as the behavior of reduced to 0.001 with a = 113. In general, as a-0, this bound
the error at a particular operating point. Furthermore, a will approach zero, which is the case in the well-known variable
properly designed control U ( I ) in (3) can very well overcome st(Ucture control. The reason the nonlinear controller performs
the unknown inherent acceleration oft).
better is that, when k a < 1, it provides higher gain when error is
small and lower gain when error is large. It completely agrees

4579
with the intuition obtained from working with practical problems.
As a matter of fact, many fuzzy logic controllers and gain
scheduling controllers exhibit this kind of characteristics on its
error surface. Of course the fuzzy controller is much more
complicated to implement.

Examde 2: Once the unknown function aft) in (2) is estimated -M a l d , ( n l < d

fsr = (9)
and cancelled using the control signal U, the remaining double -M sign ( a ) , lal>d
integrator control problem can be conveniently solved hy using
the well known time optimal control method, which yields Note that this is the discrete time solution of the famous time
optimal control problem from the 60s.
u = - a ( t ) - M s i g f l ( x ~ - 2M
v(t)+~)
The impact of TD is profound. First, as a noise filter, it blocks
The new closed loop system is described by any pan of the signal with the acceleration exceeding M. In
( x , = x, practice, we often know the physical boundary of a signal in
terms of its acceleration rate. This knowledge can be
1 2 I xi I) (7) conveniently incorporated into TD to reject noises based on the
x2 = -Msign(x, - v ( t )+ - understanding of the physics of the plant. On the other hand, the
2M
where the output y = xI tracks the reference input v(t) in shonest traditional linear filter can only attenuate noises based on its
time possible without overshoot. Here M represents the frequency contents.
maximum acceleration this system can reach and is a function of
the maximum actuation available in the system. This is a typical Secondly, TD has a very desirable frequency response
example of nonlinearity placement in Han's work where characteristic as a filter. In panicular, it has a much smaller phase
nonlinear mechanism was pu'posely introduced into the control shift compared to linear filters, while maintaining an extremely
law. flat gain within the bandwidth. This research is still on-going.

Note that other researchers. for example, [13,14] and the Finally, perhaps the most imponant role of TD is its ability to
references therein, also noticed the advantages of using certain obtain the derivative of a noisy signal with a good signal to noise
type of nonlinearities in feedback similar to ( 5 ) . What's unique ratio. It is well known that a pure differentiator is not physically
here is that Han proposed means to systematically seek out these implemeotable. The error is often not differentiable in practice
nonlinear mechanisms in control design. It results in the hinh of a due to the noises in the feedback and the discontinuities in the
class of new controllers with unparalleled performance. It pointed reference signal. This explains why the PID controller is used
to a new direction in research for future control technologies primarily as a PI controller in most applications. The use of the
development. " D pan has been quite limited due to the extreme amplification
of noises by differentiation, or it's approximations. This noise
2.2 A Nonllear Dinerentiator problem is resolved in TD because x2 is obtained via integration.
This idea of using integration to obtain differentiation goes back
Han proposed another use of equation (7) as a nonlinear observer, to 1920s when N. Wiener proposed the definition of "fractional
where x, and x2 are the states of the observer that track the input differentiation" based on integration[28]. It led to the concepts of
signal v(t) and its differentiation v(t). respectively. I t was shown generalized function and generalized derivative, which were used
[ I l l that V E > O and T > 0 , 3 M O> O such that if . widely in the theory of panial differential equations.

M >M , JorI x,(t) - v ( f ) I dr < E . Here, the only design 2.3 A Generic Nonlinear PID Control Scheme

parameter of the filter i s the gain M, which corresponds to the Besides the usual problem with the differentiation. classical PID
upper bound of acceleration, if x I can be viewed as a position is also limited by its simple weighed sum of the current (e). past
signal. x, can track u(t) arbitrcly fast as long as M can be chosen
arbitrarily large. x, is the differentiation of xi and therefore it (Id), and future (de/dt) errors. as shown in (3). This very
approximates the generalized differentiation of "(1). when "(1) is
not differentiable. This nonlinear filter is denoted as Tracking simplicity. which makes lhe controller atlractive, also becomes a
Differentiator (TD)[15]. liability when it comes to performance. Since the digital control
era began, various attempts have been made to enhance it with
To improve the numerical properties and avoid high frequency methods like gain scheduling, fuzzy logic and other nonlinear
oscillations, a discrete time realization of TD was derived [I61 means. But the results often turned out to be quite problem
,\.
f u ( f +. .., (t> +
h> = v.~,., . ..
h IL
. 2 ~( t. ,j deDendent and not easily repeatable for different problems. Han
pmposed a fundament change in the way PID is d;signedll7):
(8)
(;,(f + h ) = v 2 ( f )+ h fsr(v,(t)- v f t ) , v , ( t ) , M , h )
U = K , I e p' sign(e) + K, 1 e, 10.) sign(e,) + K,] I e, sign(e,)
where Y , and v2 a x the state variables, "(1) is the input signal. h is
the step size and the function/si(v,,v,,M,h) i s defined as:
(10)
d = M h ; d , = d h : y=v,+h v 2 ;

a, =,-/ ; The benefits of the nonlinear proportion term K , I e lyp siznfe) ,

was discussed earlier in Example 1. A similar argument can be
made for the integral term as well. Note that an alternative

4580
implementation of the integral term is to bring the nonlinear term. desired trajectories for y(t) and y ( t ) . respectively. Instead of
le,l=', inside the integral. which will help with the integral windup
comparing the output to the setpoint directly to generate the error.
problem often encountered in practice. By choosing a, <I, it it was proven in practice that it is more beneficial 10 compare the
reduces the integral action when the error is large, which helps to output to a predetermined desired trajectory, r(t), which is
prevent the integral term from being saturated. Regarding the commonly known as motion profile in the motion control
differential term, it is advisable to make aD.1. This is because industry. From engineering practice, people found that this setup
this term plays its role of preventing output from oversbwting provides better control of not only the position, y(t), but also its
primady during the transient period when the error is relatively first, second and third derivatives, known as velocity,
large. When the output is near steady state, the differentiation acceleration, and jerk (the rate of acceleration), respectively. The
signal possesses little information and contains mostly noises. By selection of r(t), r ( f ) , ... elc. is an area of research by itself [18].
choosing a, > I , it makes the differential gain small when the
error is small. Note that a TD can also be used as a reference generator as well.

To enhance the numerical properlies of this algorithm, Han 2.4 Extended State Observer and Active Disturbance
explored various nonlinear mechanism and proposed to use the Rejection Control
function fal(x,a.S) to replace 1 e r sign(e), where Any experienced control engineer knows the impact of the
integral control on suppressing steady state error, on how fast the
output enters steady state, and on disturbance rejection. On the
other hand, the integral control brings inevitable lag into the
The graphical interpretation seen in Figure I is that the new system, which could even destabilize the closed loop system. The
function fal(x,a,b) introduces a small linear region in the gain question is that is there an alternative in dealing with the steady
function. The purpose of it is 10 prevent excessive gain when state error and disturbances?
error is small, which was known to cause high frequency
chattering in some simulation studies. Recall the Theme Problem in (1). The reason the integral control
i s needed in practice is due to the presence of a(t), which
represents both the dynamics of the physical system and possible
external disturbances. Obviously the key here is the observer that
tracks the value of a(t) closely, preferably without depending on
the mathematical model of the plant. To this end. a unique
nonlinear observer form was developed by Han's group and is
described below.

Extended State Observer

Treating the unknown dynamics, a(t), as an exlended s a t e and
expand ( I ) into

Figure 1 Comparison of linear and nonlinear gains

Finally, combining the differentiator and the nonlinear feedback

discussed above, a novel nonlinear PID (WID) control
architecture is born, as shown in Figure 2. I .it= x 2
X2 = x 3 + u ,
X3 = h(t),
Y = XI
x , ( t ) = a ( f )-
h(t) = i ( f )
~ff(x,,xd

Han's solution to this problem i s to design a nonlinear observer of

the form
(12)

I 2,
2,
=z,-g1(2,-y(f))
= z3 - g,(z, - y ( t ) ) +
i, = -g, - At))
(2,
where &(a), i=1,2,3, are appropriate nonlinear functions such as
the function fal(.) in (1 I), 2,. zI is the estimate of the state I,.xI.
z3 is the estimate of the extended state n(t).
(13)

This observn is denoted as the Extended State Observer (ESO).

Figure 2 The Nonlinear PID Configuration Interestingly, if we choose g, ( e ) = &e(i =?)
, the ESO takes
the form of the classical Luenberger observer. On the other hand,
The roles of the Tracking Differentiator are two folds: 1) to filter
out the noises in the measurement of y(t) and 2) to determine if g , ( e ) = B a , ( e + k * , s i g n ( e ) ) ,then it is consistent with a form
y ( t ) via integration. as discussed earlier. of the variable stmcture observer[251.

The Active Disturbance Reiection Control Method

The nonlinear combination (NC) is implemented as shown in For the sake of simplicity, we use the theme example (12) to
(IO), which can be improved using (11). r(t) and i ( t ) are the illustrate a new control method. Auulying the ESO given above
10 (12) with g,(-) chosen as the tal fun&n-in (I I), wkhave:

4581
 2, = I , -
-polfar(I, y(t),a,,S) 4.ADRC represents a new control concept. Its application in real
2, = z , - B o 2 f d ( z , - y ( i ) , a , , S ) + u (14) world problems can t&e on many different forms, as shown in
section 3.
z, = - p o 3 f 4 z , - y ( i ) , a , . S )
where z,. z2. and z, are the estimates of xI, x,. and x,=a(t), Extension to MlMO Case:
respectively. The plant can now be dynamically compensated
with The idea of ESO and ADRC can be easily generalized to multi-
U = ug- 22 (15) input and multi-output (MIMO) systems. As an extension of the
and the difficult control oroblem is now simolified to a double Theme Problem. consider the following dynamic equation for a
integrator control problem of multi-joint manipulator:

L=,
x, = x2
x, = U o

where U,, is the control input.

(16)

Since a(t) is estimated and cancelled via ESO, there is no need for
7 = hf(@)6+c(@, 6 )+ G ( @ ) W(t),@€ R” , + (18)
where @(i) is vector of the position of the joints, M is the n X n
inertia matrix, C is the vector of Coriolis and centrifugal forces, G
is the vector of gravity terms, W(t) is the vector of external
disturbances and K is the joint torque vector. Assuming that M is
invertible, define
integral control. Using the profile generator to generate desired F ( r ) = -M-’(e)(C(Q,B)+G(o)+W(r))
state trajectory, v, and v2. we proposed a nonlinear PD controller
of the form: +(M-’(O)-M:(Q))s
U&) = K,fal(E,,a,,S,)+ K,fal(e,,a,,S,) (17) where the nonsingular matrix MO(@) is an estimation of
where E ~ = v ~ -and
z ~ E ~ = V ~are
- Z error
~ and its differentiation, M(@).Then the system (14) is equivalent to
respectively. Kp and KD are obviously the controller gains. We
denote this type of PD control law ”Generalized P D (FPD),
o=F(i)+M,’(@)r. (19)
because higher order differentiation can be used as needed.
Viewing F(i) as the “extended state’’ of the system. a gmup of
Combining equation (14) to (17). together with a profile generator.
the ADRC configuration is now complete, as shown in Figure 3. n third-order ESOs can be used to estimate the state, 0 and
7he name of ADRC comes from its ability to actively detect and 6, as well as F(r) for each joint. These ESOs will have the same
compensate the “total disturbance” a(t), which lumps together the
structure and similar parameter selections.
effects of “internal disturbance” and the “external disturbance”.
The former comes from system dynamics can be assumed

1:
2, = Z * - P , , ( Z , - O ( t ) )
unknown. The latter represents the external forces applied to the
system that needs to be compensated by the control signal. Z, = Z , - B , f n l ( Z , - O ( r ) , a , . 6 ) + U
WIII I Z, = -Po,fnl(Z,- O(r),a,,6)

ffll(Z,-@,a.6)=......
fnf(z,, - O , , a J )

fal(z,.-e,, ,a,6)
Z,= [z,,,.... 2,” ]’ , Q = [O ,.....0” 1’
1,
The outputs U t ) . Zdr) and Z,(i) will approach 0 0, and F(r), .
Figure 3 ADRC Configuration respectively. Once again, the plant can be reduced to a group of
double integrators by applying the control law
Remarks; 7=Mg(@)U, U =-Z,(f)+U0 (20)
1. The key in this design approach is ESO. It was shown in to the system (19). which results in 6 = U. [ZO].
simulation and practical applications that, with appropriate
selections of the nonlinear functions and gains, a single fixed
ESO can tracks the states rather closely for a large class of Both the n(rJ in SlSO systems, and the F ( f ) in MlMO systems
control problems. can also be regarded as a “total disturbance”. Therefore, ESO can
be seen as an uncertain input observer (UIO) OF disturbance
2. Once the ESO is properly tuned. the GPD tuning is usually
observer.
straightforward since it has the same linear PD control intuition.
3. This control method proved to be very effective because it does Disturbance rejection is an old but key problem for high
not overly depend on mathematical model of the plant and it performance control design. A great deal of effort has been
compensates for the internal and external disturbances devoted to this topic, see 137.391 and the references therein.
dynamically. This is perhaps the most important characteristics However. these methods usually assume the knowledge of the
of ADRC. It brings new meanings LO the notion of robustness disturbance model andlor the plant model. And usually, a higher
and disturbance rejection. order observer or derivatives of the measured signal are used. The

4582
breakthrough brought by ESO is that it regards all factors In the web tension application, the process changes fmm a highly
affecting the plant. including the nonlinear dynamics, over-damped plant (<= .9) to a highly under-damped one (<= .2),
uncenainties, the coupling effects and the external disturbances, together with a bandwidth drop of over a decade and the gain
as a “total disturbance” (extended state) to be observed. This new change of 40%. A fixed parameter ADRC, with the sampling rate
vision facilitates solution for a series of challenging control of 1 kHz, produced an almost identical control performance
problems, such as disturbance rejection, dynamic linearization during the entire operation, as though there was no change in the
and decoupling control. in an ingenious way. process.

3. Industrial Applications 4. New Challenges and Questions

The NPlD and ADRC methods have been applied as a generic The proposed model-independent design philosophy and
solution to industrial control problems in several major industries. nonlinear algorithms shown above pose many new challenges and
including the Motion Control, Tension Regulation in Web interesting questions to researchers. Many traditional control
Transpon System. Truck Anti-Lock Brake Systems (ABS), and concepts can now be reexamined in a new perspective. Interesting
Computer Numeric Control (CNC). The reasons for these observations have been made in the process of applying the
problems to be selected are I ) the process exhibits significant proposed methods in solving engineering problem. Some initial
dynamic changes during operation andlor considerable external results in analytical study of the new control systems are also
disturbances are present; 2) the existing technology is inadequate, briefly introduced.
or limited, and improvements are badly needed; 3) the new
solution will have a significant impact on industry. Much effon 4.1 Redefining the Control Problem and Robustness
have been made in approaching the problem so that the successful
results are not limited to one particular machine or process, but In existing control theory, the control design problem was given
rather applicable to a large class of problems. in terms of closed-loop characteristics, such as desired pole
locations or a cost function to he minimized. that can only he
After preliminary simulation study for the proof of concepts, the captured with an accurate mathematical model. Thus it needs to
new methods were eventually evaluated either on a realistic he redefined in the new paradigm. For the sake of simplicity, the
industrial simulator, or in hardware. Details can he found in [31- following assumptions were made in order to illustrate the
361. concept.

The Desien Process m: Here, the class of plants of interests is described as

1. Understanding the problem and formulating i t properly. For x = f (x,x,t) + u ( t ) , y(r)=x(r). Let F he the set of all possible
examde. in the Truck ABS oroblem. we had to characterize the
problkm as a cascade loop feedback system, before any f (x,X,O.
systematic feedback control method can be applied.
2. Evaluating the problem and selecting the right tool. For Reference Inout: Let the desired output be v(r) which is obtained
example, a Nonlinear PD controller is quite adequate in handling dynamically from v = g ( v , v ) . Let V he the set of all possible
the nonlinearity in the truck pneumatic brake system while the
severe dynamic change in web tension system calls for the more v(r)-
capable ADRC controller.
3. Modeling and Simulation. NPlD and ADRC designs do not
Controller: A dynamic system with y(r) and v(r) as inputs and u(rJ
as output, represented by C(p). where p is the controller
require an explicit mathematical model of the process. but a
simulation model of the process, not necessarily very accurate, parameter vector.
helps greatly in understanding the process and determining the
conUoller parameters. The essential information needed is how Control oroblem: Designing the controller, C(p), such that for
quickly the process responds to a control signal and the v
b’f(x, X, 1 ) E F and v V ( r ) E , its output u(r) forces the
limitations on the sampling frequency. output of the plant, yfr), to follow the reference input, v(rj. In
4. Implementation. W I D and ADRC can he easily digitized and short, we denote ( F , V) as the “control problem” and C(p) its
implemented in a high level language, such as C. The simple solution.
Euler’s method is usually good enough in digitization because of
the inherent robustness of the controller. Remarks

Summaw of the Results 1.11seems that the above definition better describes a real control
Across the hoard, W I D and ADRC exhibit superior robustness problem with a large amount of uncenainties and the lack of
and disturbance rejection in industrial applications unmatched by precise mathematical representation. The boundaly of F can
any fixed gain controllers. In motion application hardware tests, usually be characterized using the physical limitations of the
ADRC, with its parameters unchanged, overcomes significant plant, while Vis determined based on the control specifications.
inerlia change (IOOW), torque disturbance (30%). friction
changes, etc. and produced far much performance than the current 2. The new definition calls for a reevaluation of the robustness
technology In the truck ABS application, W I D was tested concept. For example, given a control problem ( F , V) and its
against PID and a 61h order loopshaping controller. The results solution. C(po),wherepo is the controller parameter vector, define
showed that not only does NPlD provides the better performance. another set P which is the neighborhood of po and satisfies that
it is also easy to tune and retains the same intuition of PID tuning.
v p € P , C(p) is also the solution of { F , V), then one measure

4503
of robustness of C(p$ is the size of P. Another measure of We can show the concept of SSR via an inherent property of the
robustness of C(po) is its operating range. or the size of ( F , V ) . Theme Problem (I).

4.2 A Generic Time Constant Concept Assuming " ( 1 ) in ( I ) is uncertain. It's clear that all convergent
trajectories can't converge to the origin along any direction in the
The generic nonlinear PID control introduced in section 2.3 and first and the third quadrants. Assuming that 1) G is the union of
the active disturbance rejection control introduced in section 2.4 any two regions, which lie in the second and the forth quadrants
are all model independent. Their inherent robustness allows them
to be used for a large class of problems. But obviously there will respectively, and have only one contact point (0.0); 2) e
be a limit of applicability and one controller cannot control all the excludes the axes except the origin, then, if there is a trajectory
processes. The questions of how to determine such limit for a (x,(t),x,(t)) staying in G after certain time 7'. this trajectory
controller and how to adjust controller parameters for a different can only converge to the origin[22,231. This property is an
class of problems prompted the research on a generic time inherent property, which is independent on the function
constant concept, also known as the rime rcale[29,30]. It can also
a(r)(= f ( t , x , , x , , w ) )and the control input u ( t ) . In this sense,
be viewed as an attempt to characterize the essential knowledge
of the plant needed for conlrol design. we call G a Self-Stable Region, or SSR, for the system in (I).

Definition: Consider the problem defined in 4.1. let Then the purpose of control is to force all trajectories out of G
entering the interior of G and forbid the trajectories inside G
coming out G.

where GI and G2 define the operating region. Then, we define It's clear that, all stable sliding modes for the system ( I ) must lie
in the second and the fonh quadrants. Therefore the SSR concept
p = 1 1 6 as the time scale of the control problem { F , VI. reflects the essence of the sliding plane. Moreover, the restriction
of the sliding mode being a super-plane is relaxed by replacing
In the view of W I D or ADRC. regardless of specific forms of the sliding mode with the SSR, based on which a continuous
f(x,x,f) and g ( v , V ) ,pcharacterizes the problem ( F , V ) as nonlinear (non-smwth) feedback law can be constructed and the
chattering in sliding mode control can be easily avoided.
needed to determine the control parameter p. Different
applications with the same time scale can now be controlled with Via the concept of SSR, [241 analyzed the convergence and the
the same controller. The exploration of effective control laws is estimation error of ESO (13) in the second order case. 1241
no longer restricted by the traditional system classifications such
showed that the state of ESO does not convergence to a certain
as linearlnonlinear. time varyingltime invariant. etc. sliding mode but to a SSR,determined by g,C). The principle of
choosing si(') to guarantee the state trajectories of ESO
The concept of time scale also makes the controller "portable" in converging to the SSR is obtained. Funhemore, the steady
the sense that the control parameters can be converted easily estimation error can be bounded by the structure of the SSR.
between problems of different time scales. Using the ADRC in Based on this work, a Lyapunov function is constructed for
Section 2.4 as an example. assume bo,,,,P , ,bo,,,k,, ,k,, are stability proof and tracking error analysis of the ADRC designed
well tuned parameters far a process in ( I ) of the time scale pI for aircraft attitude control 1261.
with the sampling period of hl. It was shown that the
corresponding ADRC parameters for the system with the time In summary, the results presented here are merely indicative of
scale p and the sampling period of h to achieve a similar another perspective in control research. They are by no means
performance is: complete or mature. Even though there are perhaps more
questions than answers at this stage, we believe that they point to
Bo, =fi&,,.so,=(%*&,.
so,=(~,,w,,. a promising direction.
P P P

5. Concluding Remarks

More details can be found in 1301 and more research is still being An alternative paradigm for control system design developed on a
carried out. novel design concept is introduced. A survey of various control
algorithms, the initial analysis and practical applications
4.3 Convergence and Stability Analysis of ADRC employing the proposed framework is given. The design methods
discussed here compliment the existing knowledge in that it is
Due to the lack of analysis tools, introducing nonlinearity into developed closely along the line of how conlrol has been
control algorithms, as shown above, brings difficulties to practiced in industry. Therefore. it provides a vehicle for
convergence and stability analysis. While the new concepts and improving control algorithms that are being used in practical
methods are being developed, Han's group also has been applications.
persistent in pursuing the analysis of the new approaches.
Progresses are being made in several aspects, as briefly discussed The proposed work opens a wide array of research directions to
as follows. I ) systematically explore the use of other possible nonlinear
feedback mechanisms; 2) develop analysis tools for the new
The self-stable region (SSR) approach is a constructive design; 3) expand the applications of the new methods to other
continuous non-smooth synthesis method based on the industrial applications, etc. Much work is still ahead.
improvements of the variable structure control method 122,231.

4584
Acknowledgement: The second author wishes to acknowledge I231 Y. Huang and J. Han, “The self -stable region approach for
the suppon by Natural Science Foundation of China (69904010) second order nonlinear uncenain systems”, Proc. 1999 IFAC
that facilitated this work. The authors would also like to thank World Congress, 1999, Vol. E: 135-140.
anonymous reviewers for their suggestions and comments. 1241 Y. Huane- and 1. Han, “Analysis and desien for nonlinear
~~
I

continuous Extended state Observer”, Chinese Science Bulletin,

REFERENCES 45.1373-1379.2000.
I251 V. I. Utkin, Sliding Modes in Cintrol and Optimization,
[l] F.L. Lewis, Applied Optimnl Conrrol and Esri,nntion, Berlin, Springer-Verlag, 1992.
Prentice Hall, 1992. 1261 Y. Huang. K, Xu, I. Han and J . Lam, Flight control design
P I The Control Handbook, Editor W.S. Levine, pp. 198, CRC using Extended State Observer and non-smooth feedback, IEEE
Press and IEEE Press, 1996. CDC conference. December, 2001.
[3] J.E. Slotine and W. Li, “Applied Nonlinear Control”, [271 A.F.Filippov, “Differential Equations with Discontinuous
Prentice Hall Inc. 1991. Righthand Sides”, Kluwer Academic Publishers, 1988.
[4] M. Krstic, 1. Kanellakopoulos, and P. Kokotovic, Nonlinear [281 N.Wiener, The operational calculus, Math. Ann., 95, 557-
and Adaptive Control Design, John Wiley and Sons Inc., 1995. 584, 1926.
151 N. Minorsky, “Directional Stability and Automatically [291 I. Han and X. Wang. “Time Scale and Nonlinear PID
Steered Bodies”. J. Am. Soc. Nav. Eng., vol. 34. p. 280, 1922. Controller”, Proceedings of [he Chinese Control Conference,
161 J.G. Ziegler and N.B. Nichols, “Optimal Settings for pp.314-321, Chinese Science and Technology Press, Beijing,
Automatic Controllers”, Trans. ASME, vol64, pp759-768, 1942. 1994.(In Chinese)
171 K. J. Astrom and T. Hagglund, “PID Control”, The Control I301 W. Zhang. Time Scale and Auot-DisNrbances-Rejection-
..
Handbook. Editor W.S. Levine. DO. 198. CRC Press and IEEE
Press, 1996.
controller, Ph.D. Disstenation, Institute of System Science,
AMSS, Chinese Academy of Science, 1999.
.181. J. Han. “Control Theorv: Is i t a Theorv of Model or I311 Z. Gao, S. Hu, and F. Jiang, “A Novel Motion Control
ControlY, Systems Science and Mathematical Sciences, Vo1.9, Design Approach Rased On Active Disturbance Rejection”, IEEE
No.4. pp.328-335, 1989.(In Chinese) CDC conference. December, 2001.
[91 1. Han, “Robustness of Control System and the Godel’s (321 F. Jiang and Z. Gao, “An Adaptive Nonlinear Filter
Incomplete Theorem”, Control Theory and Its Applications, Approach to Vehicle Velocity Estimation for ABS”, IEEE
IMsuppl.), 149-155, 1999.(In Chinese) CCNCACSD conference. September, 2000.
1101 J. Han, “The Structure of Linear Systems and The Feedback 1331 F. Jiang and Z. Gao, “An Application of Nonlinear PID
Systems Calculation”, Proceedings of National Control Theory Control to a Class of ABS Problems”, IEEE CDC conference,
and Application Conference, Science Press, Beijing, 1981. (In December, 2001.
Chinese) [341 Y. Hou, Z. Gao, F. Jiang, and B. Boulter, “Active
1111 J. Han, “Linearity and Nonlinearity in Feedback Systems”, Disturbance Rejection Control for Web Tension Regulation”,
Control and Decision. Vo1.3, No.2, pp.27-32, 1988.(In Chinese) IEEE CDC conference, December, 2001.
[I21 J. Han, “Nonlinear Design Methods for Control Systems”. [35] R. Mahawan. Z. Luo, J. Ha” and S. Nakajima, High-speed
The Proc. OfThe 14‘hlFACWorld Congress, Beijing, 1999. high precision motion control of robot using Extended State
[I31 V. Haimo, “Finite time controller”, SIAM J. of Contr Observer, Journal of the Robotics Society of Japan, 18(2), 86-93.
Optim. pp.760-770, 1986. 2000. (In Japanese)
[I41 S . Bhat and D. Bernstein, “Continuous finite time [361 Z. Hou, Y.Huang and J. Han, Active noise cancellation with
stabilization of the translational and rotational double a nonlinear mechanism, Proc. Of the IEEE Int. Conf. On Decision
integrators”, IEEE Trans Automat Control, Vol. 43, No. 5 , and Control, 1577-1578, Tampa, UAS, Dec. 1998.
pp.678-682. 1998. [37lC.D.Johnson, A family of linear, time-invariant universal
Cl51 I. Han and W. Wang, “Nonlinear Tracking Differentiatot”, adaptive controllers for linear and non-linear plants. Int. J.
Systems Science and Mathematical Sciences, Vo1.14. No. 2, Control, Vol. 49, No.4, 1217-1233, 1989.
pp.177-183, 1994.(In Chinese) [38]M. Corless, Tu. 1, Statdinput estimation for a class of
1161 1. Han and L. Yuan, “The Discrete Form of the Tracking uncenain systems, Automatic, 34(6), 757-764, 1998.
Differentiatot”, Systems Science and Mathematical Sciences, [391T.murakami and K. Ohnishi. Dynamic identification method
Vo1.19, No.3. pp.268-273, 1999. (In Chinese) of multi-degree of freedom robot based on disturbance observer,
Cl71 J. Han, “A New Type of Controller: NLPID, Control and Journal of the Robotica Society of Japan, 11(1), 131-139, 1993.
Decision, Vo1.9, No.6, pp.401-407. 1994. (In Chinese) (In Japanese).
[I81 R. Nonon, Design of Machinery;, McGraw-Hill, Inc. 1992.
I191 1. Han, “A Class of Extended State Observers for Uncenain
Systems”, Control and Decision, Vol.10, No.1, pp.85-88,
1995.(In Chinese)
I201 Y. Huang, Z. Luo and S. Hosoe, Active Disturbance
Rejection Control of robot systems based on Extended State
Observer, the 19Ih Conference of Robot Society of Japan,
September, 2001
1211 I. Han, ”Auto-Disturbance Rejection Control and its
Applications”, Control and Decision, Vo1.13, No. I.’pp. 19-23.
1998.(In Chinese)
I221 Y. Huang and J. Han, “A new synthesis method for
uncertain systems- the self-stable region approach. Int. J.
Systems Sci., 1999, 30(1):33-39.

4505