Scaling and Bandwidth-Parameterization Based specifications including disturbance rejection, noise sensitivity,

Controller Tuning stability margins, and smoothness of the control signal.

Zhiqiang Gaa
Dept. of Electrical and Computer Engineering This lack of design insight leads to the heuristic nature of the
Cleveland State University, Cleveland, Ohio 441 15 tuning methods implemented in indushy. Furthermore, the
practice of control design and tuning tended in the direction of art
Abstract: A new set of tools, including controller scaling, rather than science. This paper presents a comprehensive
controller parameterization and practical optimization, is approach that moves control design and tuning in the direction of
presented to standardize controller tuning. Controller scaling is science.
used to frequency-scale an existing controller for a large class of
plants, eliminating the repetitive controller tuning process for The paper is organized as follows. Controller scaling is
plants that differ mainly in gain and bandwidth. Controller introduced in Section II. Parameterization and optimization of
parameterization makes the controller parameters a function of a model-based controllers are discussed in Section 111. Design,
single variable, the loop-gain bandwidth, and greatly simplifies parameterization, and optimization of a model-independent
the tuning process. Practical optimization is defined by controller design method are discussed in Section IV. Finally,
maximizing the bandwidth subject to the physical constraints, some concluding remarks are given in Section V.
which determine the limiting factors in performance.
Collectively, these new tools move controller tuning in the 11. Controller Scaling
direction of science. A controller is generally not “portable”, i.e., a controller designed
for one plant is usually not applicable to another plant. The
Keywords: Tuning, PID, Scaling, Auto-Scaling, Auto-Tuning, objective of controller scaling is to make a good controller
Adaptive Self-Tuning, Gain-Scheduling, Disturbance Observer, “portable”, much like the filter design. With the bandwidth, pass
Computer Aided Controller Design band, and stop hand requirements given, the filter design is
straightfonvard. First, a unit bandwidth filter, such as an nth
1. Introduction order Chebeshev filter H(s), is found that meets the pass band and
The proportional-integral-derivative (PID) controller, first stop hand specifications; then it is frequency scaled by wo to
proposed by N. Minorsky in 1922 [I], is used in over 90% of achieve the desired bandwidth of coo. It is shown in this section
current industrial control applications [ 2 ] . In addition, the that the controller design can be performed similarly.
controller parameters are still determined by N k S of thumb, such
as look-up tables [3]. 2.1 Frequency Scale and Time Scale
Consider a unit feedback control system with the plant GJs) and
Classical control theory has successfully provided the analysis the controller G,(s), as shown in Figure 2.1. Assume that G,(s)
and design tools for single-input single-output (SISO), linear, was designed for desired command following, disturbance and
time-invariant systems, since the 1940s. The PID design noise rejection, and stability robustness. Now, consider a similar
approach moved from empirical (i.e., ad hoc tuning methods such class of plants G,(s/o,). for any given U+ Can a controller be
as Ziegler and Nichols tuning tables [3]) to analytical (i.e., pole found without a repetition of the tedious loop shaping design
placement, frequency response). In particular, the frequency process?
response-based methods (Bode and Nyquist plots, stability
V
margins, lead-lag compensators) have proved to be especially
useful in solving control problems.

Historically, determining controller parameters to meet design

specifications (w, rather than the design of the controller
Figure 2. I Feedback Control Configuration
itself, has been the main concern in industry Most industrial
plants are inherently stable and consist of SISO subsystems.
Simple PID controllers implemented in a digital form can usually Definition2.1: Denote rap as the fresuencv scale of the plant
meet the performance needs. But the problem of tuning has G,(s/u,J with respect to G,(s), and 5=l/mn. the corresponding
hardly received much attention in the existing control theory. The time scale.
variety of ad-hoc tuning algorithms in industrial control products Definition2.2: Denote k as the of the plant kC&) with
shows the lack of in-depth understanding of the problem and the respect to Gp(s).
need for further research.
The differences in many industrial control problems can be
The PID gains are commonly “tuned” on a trial-and-error basis in described in terms of the frequency and gain scales defined here,
practice. A general lack of knowledge regarding the relationship such as the temperature processes with different time constants (in
between “design objectives” and “practical performance first-order transfer functions), motion control problems with
measures” makes the use of well-known design techniques such different inertias, motor sizes and frictions.
as Root Locus @ole-placement) and linear optimal control
difficult. For example: in pole-placement design, the objective is The use of the scales allows the development of a generic solution
to place the closed-loop poles at given locations, based on the for a class of problems. Any linear time-invariant plant, strictly
understanding of how the location of poles affects the transient proper and without a finite zero, can be reduced to one of the
response of a system. Although the transient response is usually following forms
an important design consideration, it is not the only issue in pole- _ 1_ 1 1 1 1 1
placement methods with which to contend. The pole-placement
~- , ... (2.1)
s + 1 ’ s ’ $ 2 +25s+ 1 ’ ++I) ’ s* ’SI +5,s’ t & + I
method is ill-equipped to handle other common design
Proceedings of the American Control Conference
02003 IEEE
0-7803-7896-2/03/$17.00 4989 Denver, Colorado June 4-6,2003
through gain and frequency scaling. For example, the motion w S
control plant of G,(s) = (k, t k,2t kd - ) l k (2.7)
23.2 S w,
Ge(5]=-
s(st1.41) That is, the new PID gains, kp, F, and k, are obtained from the
is simply a variation of a generic motion control plant original ones as
I
Go@)=-
s(s t 1)
with gain and frequency scales of k = 11.67 and O+ = 1.41,
respectively. That is
Example 2.1 Consider one of the plants in (2.l), which has a
~-23.2 11.67
(2.2) transfer function of
s(stl.41) i I
141'1"'')
.. . .. . 1
Equation (2.1) covers the majority of industrial control plants, G,(s) =-
s'+stl
which are usually approximated by a pure first order or a second and the PID controller gains of k,=3, k,=l, and kd=2. Now,
order transfer function response. For completeness, (2.1) may be assume that the plant has changed to
appended by terms such as
1
stl sit25,s + 1 G
,
(.
, =
(2.3) ($2 t -+I
s ~ t 2 ~ s + l ' s ~ t ~ , s ' + ~"' 2 s + l ' 10
to include systems with finite zeros. Furthermore, for a particular The new PID gains determined from (2.8) are
class of plants, the scaling concept can be applied accordingly to k = 3,k, = 10,kd= . 2 . Applying a unit step function as the set
reflect the unique characteristics of the class. For example, plants
commonly seen in motion control with significant resonant point, the responses of the original controller and the scaled
problems can be modeled and scaled as: controller are illustrated in Figure 2.2, demonstrating that the new
response is exactly the same as the original scaled by 7=I/wp,
w.=IO radlsec.

1
10

(2.4)
where the resonant frequencies satisfy @,p=nwp. @,=my. Such
problems with multiple frequency scales, up,nop, and mmp3are
referred to as multi-scale oroblems.

2.2 Controller Sealing /urn.

Theorem 2.1: Assuming CJs) is a stabilizing controller for plant
Figure 2.2 Auto-Scaling of PID
G,(s), and the loop gain crossover frequency is o,,then the
controller
From the frequency response of loop gain transfer function (i.e.,
G,(s) =G,(s/o,)/k (2.9 the product of the controller and the plant in this case), it is
will stabilize the plant c p ( s ) =kC,,(s/qJ, and new loop gain determined that the gain margins of both systems are infinite, and
the phase margins are both 82.372 degrees; and the 0 dB
E(s) = G,(s)G,(s) will have a bandwidth of uc@,,
and the crossover frequency for both systems are 2.3935 and 23.935 rls,
same stability margins of L(s)= C,(s)C,(s). respectively.

111. Bandwidth-Parameterization and Optimization

Proof. The proof is obvious since E(s) = L(s/w,) , Q.E.D.
The controller scaling method that is demonstrated above
resolved the long-standing issue of portable controller design.
Note that the new closed-loop system has the same frequency Once a good controller is obtained for one plant, it is easily scaled
response shape as the original system, except that it is shifled by to control similar plants that are different only in gain and
U,, That is, all feedback control properties, such as disturbance frequency scales, thus avoiding tedious control redesign.
and noise rejection, as wcI1 as stability robustness, are retained
from the previous design, except that their frequency ranges are However, there are always differences in design specifications
all shifted by wr and constraints for different control problems. The question we
want to address now is how do we realistically "optimize"
The use of controller scaling eliminates the repetitiveness of controller design for each application, i.e., how does an engineer
control design and tuning in industry today. Applying controller get "the most performance" out of a given set of
scaling in ( 2 . 5 )to a PID controller, hardwarelsoflware?
k
G,(s)=k, t--LtkdS (2.6) For the sake of simplicity and practicality, bandwidth, denoted as
S
o,,is selected here as the measure of performance. It is well-
results in known that higher bandwidth corresponds to better command
Proceedings of the American Control Conference
4990 Denver. COloladO June 4-6.2W3
following, disturbance rejection and sensitivity to parameter
variations. On the other hand, achievable bandwidth is limited by
the presence of sensor noise and dynamic uncertainties. Most
design trade-offs are reflected in the selection of the 0,.For this
reason, a unique controller parameterization is proposed that
makes U, the only design parameter to be determined (tuned).

Definition 3.1: Bandwidth-Parameterization, also known as I I

Parameterization, refers to assigning all closed-loop poles at -w, Figure 3.1 Loop-shaping
and making all parameters of the controller a function of 0,.Here,
w, is denoted as the bandwidth of the feedback control system and Consider Figure 3.1, the desired loop-gain can be characterized as
C,(s,o,) as the controller.

The concept of we-Parameterization applies to all controller

design methods. The key is to make all controller parameters as
functions of U, Two particular design techniques are used in the where wc is the bandwidth, and
following subsections to show how this is accomplished. wI<w,, wz>w,, m 20, and ntO (3.3)
are selected to meet constraints demonstrated in Figure 3.1. Both
3.1 Parameterization of the Pole-Placement Design rn and n are integers. The default values for w, and col are
Consider the normalized plants in (2.1) and assume that the w, = wJlO and w2 =IOW, (3.4)
desired closed-loop transfer functions are which yields a phase margin of approximately 90 degrees.
(3.1) Once the appropriate loop gain constraints are derived and the
Applying the simple pole-placement design to the first and second corresponding lowest order L(s) in (3.2) is selected, the controller
order plants in (2.l), a set of w,-parameterized controllers is can he determine from
obtained as shown in Table 2.1. Similar solutions for higher order
(3.4)
plants in (2.1) can be easily obtained. Note that with 0. as the
only parameter. the tuning of the controller is greatly simplified.
An additions1 constraint on n is that

Plant - 1
s t l
1
- I 1 I
- (3.5)
G,W s s 2 + 2 5 s + l l s(s+l) sz
is proper. This design is valid only if the plant is minimum phase,
as assumed, For a non-minimum phase plant, a minimum phase
approximation of GP' (s) should be used instead.

A compromise between o1and the phase margin can be made by

adjusting wl upwards, which will increase the low frequency
3.2 Parameterization of the-Loop-Shaping Design gains at the cost of reducing the phase margin. Similar
The loop-shaping design'mcthod, also known as the frequency compromise can be made between phase margin and w2.
response-based controller'design method, is an insightful method
used by many engineers in practice. It is the only design method 3.3 Parameterization of State Feedback Controller and State
in classical control theory that comprehensively addresses Observers
multiple design concerns, such as transient response, disturbance The state feedback controller can be represented by
rejection, stability margins, and noises. Loop-shaping as a
u=r+K2 (3.6)
concept, and as a design tool, helps practicing engineers greatly in and is based on the state space model ofthe plant:
improving the PID loop performance and stability margins. For
k(f) = A x ( / ) + Bu(t)
example, a PID with a lead-lag compensator is commonly seen in (3.7)
industry today. Unfortunately, manipulating loop gain frequency y ( t ) = Cx(t)+ Du(t)
response and managing competing performance measures can be Here, U is the control signal, r is the setpoint for the output to
tedious and, in some cases, frustrating. A computer algorithm to follow, x is the state vector, and (A,B,C,D} matrices are given.
automate this process was proposed in [4-71, which not only When the state x is not accessible, a state observer (SO):
replaces the manual design but also leads to a self-tuning 2 = A ; + B~ + L ( -~j ) (3.8)
controller design and implementation.
is often used to find its estimate, i . The state feedback gain K
The loop-shaping process consists of two steps: 1) convert all and the observer gain L are determined using the standard
design specifications to loop gain constraints, as shown in Figure eigenvalue assignment technique based on the equations:
3.1; and 2) find a controller G,(jw) to meet the specifications. eig(A+BK) = Ac(s) and eig(A+LC) =A&
where h,(s) and k,(s) are polynomials of s that are chosen by the
designer. Usually K and L have many parameters and are

Proceedings of the American Control Conference

4991 Denver, Colorado June 4-6, 2003
difficult to tune. The parameterization of state feedback and The design objective is to rotate the load one revolution in one
state observer gains are achieved by making second with no overshoot. The physical characteristics of this
A<(s)=(s+oJ = eig(A+BK) and A,,(s)=(s+oJn= eig(A+LCJ (3.9) control problem are: I)lu(<3.5 volt, 2) the sampling rate is 1 H z ,
where W. and wo are the bandwidth of the state feedback system 3) the sensor noise is 0.1% white noise, 4) there could be a torque
and the state observer, respectively, and n is the order of the disturbance up to 10% of the maximum torque, and 5 ) the control
system. Now the tuning of K and L become much simpler and signal should be smooth with noise level limited to *lo0 mV.
intuitive since w, and we have explicit physical meanings. An
example will be given in a later section of this paper. The plant transfer function is
G,W= ,k = I I .67 and up=1.41
3.4 we-Optimization
-(-+I)
Consider a common controller design scenario in Figure 3.2. The 9 %
goal of controller design is to achieve maximum closed-loop A simple PD design of
bandwidth, subject to design constraints. This is denoted as the
w,-Optimization. u = k,(r - y ) + k,,-L)
with
kp = w: and kd = 20, - 1
results in a closed-loop transfer function of

-
Figure 3.2 A Common Controller Design Scenario

Desien Procedure:
I

I . From the given plant transfer function, determine its

frequency and gain scales, upand k;
2. Determine the type of controller required (either a simple
pole placement design in Table 3.1, or a more
comprehensive design in equation (3.4)) based on the design
specifications;
3. Select the G,(s, w,) corresponding to the scaled plant in the
form of(2.1);
4. Scale the controller to -1G c ( - s, q )
k "0

5.' Digitize and implement G,(sho,, wJk ;

6. Set an initial value ofw, based on the bandwidth requirement
from the transient response;
7. Gradually increase wC while performing tests on the
simulator, until either one of the followings is observed:
a. Control signal becomes excessively noisy;
b. Control signal exceeds physical limits in the
magnitude andlor the rate of change;
c. Indication of instability (oscillatory behavior)
, .~.
~ ~
I s n d l , l m 1 ,.Sb.C
Example 3.1 .. . I * n d l , m " . 2 0 ,.d/I.C
...... ,r.nr,.n, eco+,m
Consider a motion control test bed as pictured in Figure 3.3. The ..,.,., -. '...
mathematical model of the motion system was derived and I
.* ,
,
.:r,' ,
verified in hardware test, as
:V = (-I . 4 I j +23.2Td) 1 - 2 3 . 2 ~ (3.10) I
where y is thc output position, U is the control voltage sent to the
power amplifier that drives the motor, and T, is the torque
disturbance.

IV. Parameterization of Model-Independent Controllers

Figure 3.3 The Motion Control Test Plant The design techniques presented so far are based on a linear time-
invariant mathematical model of the plant. In practice, how,ever,
the physical systems are usually nonlinear and time varying. The
ProceedingsOf the American Control Conference
4992 Denver, Colorado June 4-6, 2 W 3
 lack of a descriptive mathematical model is often a limitation to which is easily controlled with a PD controller
the systematic controller design. In this section, the u,=kp(r-z,)-kdzl (4.9)
parameterization and optimization methods introduced above are
applied to a model-independent design approach proposed in [X- where I is the setpoint. Note that - k , z , , instead of k d ( t - z l ) ,
I I]. is used to avoid differentiation of the setpoint and to make the
closed-loop transfer !kction pure second order without a zero:
4.1 A Model Independent Controller Design
For the sake of simplicity, consider a second order plant c , = s2 + k,s
kP
+ k, (4.10)
y = -ay-by t w+ bu (4.1)
Here, the gains can be selected as
where y and U are output and input, respectively, and w is the
external disturbance. Here, the parameters, a and b, are both kd = 2{w, and k, = 0,' (4.11)
unknown, although we have some knowledge of b, i.e., bo = b . <
where w, and are the desired closed loop natural frequency and
Rewrite (4.1) as damping ratio. {is selected to avoid any oscillations.
j ; = -ay -by + w+(b- b&+ bou= f +b,u (4.2)
Remarks:
where f = -ay - by + w + ( b- bo). . Here, fis referred to as the 1. The disturbance observer-based PD controller achieves zero
aeneralized disturbance. or disturbance. because it revresents
~ -
steady state error without using- an integrator;
both the unknown internal dynamics, -ay - by ( b- bo)u and + 2. The design is model independent. The only parameter
the external disturbance w. needed is the approximate value of b in (4.1).
3. The combined effects of the unknown disturbance and the
internal dynamics are treated as a generalized disturbance.
The basic idea is to obtain f . an estimate of I: and use it in the Bv, auementine
- -
the observer to include an extra state. it is
actively estimated and canceled out, thereby achieving active
I

control law, U = ( - f + u , ) l b o , to reduce the plant to a unit-gain disturbance rejection.

double integrator control problem, y = (f - f) uo , with a + 4. The PD controller in (4.9) can be replaced with a more
elaborate loop-shaping design, if necessary.
disturbance (f - j ).
The above controller, (4.7) and (4.9), combines with the LESO to
The plant in (4.2) is written in state equation form actively compensate for the disturbances and is, therefore,

Ix, = x2
x1 = x, +bo.
i,=h
(y=x,

with x, = f added as an augmented state, and h = f as

(4.3)
denoted as Linear Active Disturbance Rejection Controller
(LADRC). This is a special case of the ADRC controller
originally proposed by Han in [8, 91, which uses nonlinear gains
in place of the linear ones in (4.6) and (4.9). While the nonlinear
gains may be more effective, they also produce extra complexity
in the control algorithm implementation and tuning. In addition,
stability of the system becomes more difficult to prove. The
unknown disturbance. Now f can be estimated using a state discussions in this paper are limited to the linear case.
observer based on the state space model
4.2 Stability
x = Ax+ Bu +Eh
(4.4) Let ei=x,-zi, i=l. 2, 3 , and combine equation (4.5) and (4.6) and
y=cz subtract it from (4.4).
. .. the error equation can be written as:
where e = A,e+ Eh (4.12)
where

The state space observer, denoted as the linear extended state

observer (LESO), of (4.4) is constructed as
i = Az+ Bu + L ( y - 9 )
A,=A-LC=
[::1 1
-b, 0 1

and E is defined in (4.4). Obviously, the LESO is bounded-input

bounded-output (BIBO) stable if the roots of the characteristic
9 =cz (4.5)
polynomial ofA,
and L is the observer gain vector, which can be obtained using A(5) = s3+ fi, s2 +fl, sip, (4.13)
any known method such as the pole placement technique, are all in the left half plane and h is bounded. The separation
L = @I h BIT (4.6) principle also applies to LADRC.
where [I' denotes transpose. With the state observer properly
designed, the controller is given by Theorem 4.1: The LADRC design from (4.5) to (4.8) yields a
-2, + uo BIB0 stable closed-loop system if the observer in (4.5) and (4.6)
U=-
Do
(4.7) -
and the control law (4.8) for the double internator are stable.
Proof: Equation (4.'7) and (4.9) can be combined into a state
Ignoring the estimation error in 13, the plant is reduced to a unit feedback form of u=(libo)[-kp -b-l]z=Fz, where F = (I/bo)[-kp -
gain double integrator, 4 -I]. The closed-loop system is then represented by the state-
j=(f-z,)+u, =Uo (4.8) space equation of

Proceedings 01 the American Control Conference

4993 Denver, Colorado June 4-6, 2W3
 observer a function of 0,. Here, o, is denoted as the bandwidth
of the observer.

The plant (4.4) for which LESO is designed has all three poles at
where =B/bo. (4.14) is BIB0 stable if its eigenvalues are in the the origin. Intuitively, the observer will be the least sensitive to
left half plane. By applying elementary row and column noises if the observer gains in (4.6) are the smallest for a given
ooerations. it is obvious that the closed-looo eieenvalues satis!? bandwidth of o, (a proof of this would be interesting). But the
observer gains are proportional to the distance from the poles of
the plant to those of the observer. Both this and simplicity
suggest that all three of the observer poles should be placed at -
=eig ( A + EF) ueig ( A - L C ) o., or equivalently,
=(rootsofs2+kds+kJu(roots ofsJ+pIs2 +fl~s+BJ Q.E.D. A($ = s J + p ,s’+f12s+f13=(s+0J’ (4.15)
That is,
Since r, as the reference signal, is always bounded, the only 8, =30,, B, = 3 4 , 8, = d (4.16)
nontrivial condition on the plant is that h = f is bounded. In Remarks:
I ) Equations (4.15)
. . and (4.16)
. , are easily extended to an nrh
other words, the disturbancefmust be differentiable, which is a orber LESO;
reasonable assumption.
2) This parameterization method can also be readily extended
4.3 Parameterization and a.-Optimization of LESO
to the Luenbereer- Observer for arbitrary. A.. B,. and C
matrices, using the following steps:
State observers provide information on the internal states of the -_-
plants that are otherwise unavailable. They also function as noise a. Obtain (A,B,C}as observable canonical form of
filters. The primary concern in observer design is the selection of {A,B.C);
the bandwidth. The closed-loop observer, or the correction term, b. Determine the observer gain, E , so that all the
L(y- j ) in particular, is used to accommodate the unknown initial poles of the observer are at -ma;
c. Use the inverse state transformation to obtain the
states, the uncertainties in parameters, and the distuhances. The
observer gain, L, for ( A . 5 . C ) .
ability to meet the control requirements is largely dependent on
the speed at which the observer can track the states and, in case of The parameters in L are all functions of w, which can be
LESO, the disturbance f(t,xl,x2, w). In general, observers should easily adjusted.
be made to work as fast as the measurement noise allows.
4.3.3 w,-Optimization
4.3.1 Limiting Factors In Observer Design The practical optimality of the observer is similarly defined to
There are three common limiting factors in design I ) sensor that of the controller as shown in Section 111. That is, increase the
noise; 2) sampling rate; 3) dependency on the state space model bandwidth as much as allowed by the hardware and software
of the plant. limitations, which are mainly noises and the fixed sampling rate.

The level of sensor noise is hardware-dependent but it is Definition 4.2: oiOutimization refers to maximizing the
reasonable to assume that it is a white noise with the peak value at observer bandwidth U,, subject to the condition that the
0.1% to 1% of that of the output. The observer bandwidth should sensitivity to sensor noises and the delay in sampling are
be selected so that there i s no significant oscillation in its states acceptable.
due to noises. This bandwidth will also be limited by the given
sampling rate. In general, the faster the LESO, the sooner the disturbance is
observed and cancelled by the controller. Therefore, the 0.-
The dependency of the design on the state space model limits its optimization should be applied for each application. More
application to problems where this type of model is available. importantly, the repetitive observer design and tuning is reduced
The Observer also becomes sensitive to the inaccuracies of the to the adjustment ofone parameter: 0,.
model, which cause the plant dynamic to change. These
limitations make the LESO, which, is largely model-independent Relationshio between w, and w,
as shown above, unique and appealing. A common rule of thumb is to choose
W O = 3-50, (4.17)
4.3.2 Parameterization of LESO This applies to the state feedback control system where o, is
Although the pole placement technique is used widely for the determined based on the transient response requirements,
state observer design, the location of observer poles has never particularly the settling time specification. The controller design
been systematically addressed. In practice, a compromise is made can he more aggressive by using a smooth transient profile,
between the speed at which the observer tracks the states and its instead of a step command, as the desired trajectory to allow the
sensitivity to the sensor noises. Sampling rate also limits how fast output to follow more closely. In this case, there are hvo
the observer can operate. These design issues are addressed here bandwidths to consider: the actual control loop bandwidth, o,and
using the same parametcrization and optimization method in the the equivalent bandwidth of the transient profile, Be. Since the
controller design, introduced in Section 111.
observer is evaluated on how closely it tracks the states and Bcis
Definition 4.1: o,Parameterization refers to assigning all more indicative than U, of the speed at which the plant states
observer eigenvalues at -0.and making all parameters of an move, Bcis used in place of 0,in (4.17). Furthermore, taking
other design issues, such as the sampling delay, into
Proceedings of the American Control Conference
4994 Denver, Colorado ~ u n 4
e4.2~3
consideration, a more appropriate minimum wQ is found through step torque disturbance of 10% of the maximum torque is added
simulation and experimentation to he at t=3 sec. The threshold for the ripples in the control signal is
0.= 5-1 0 D< (4.18) t100mV peak to peak, which is in line with that of the PD
controller in Example 3.1.
4.4 Optimization of LADRC
To summarize the observer and controller design methods Tuni ndODtimization
discussed in section 4.1 and 4.3, a cohesive LADRC design and Following the design procedure describe above, W. and W. are
optimization procedure is given as follows: increased together, with an initial value of 40 ragsec. When they
Step 1: Design parameterized LESO and controller where w, and reach 95 radisec, the ripple threshold for control signal is reached,
indicating that they cannot be increased funher. The
wc are design parameters;
corresponding error of r-y has a maximum value of 3.4% and
Step 2: Design a transient profile with the equivalent bandwidth
0.5% for the transient period (t from 0 to 1 sec) and the
of Dc ; disturbance rejection period (t from 3 to 3.2 sec), respectively.
Step 3: Select an w, from (4.18); The subsequent changes in w, and w, and simulations show
Step 4: Set o,=o, and simulateltest the LADRC in a realistic interesting facts:
software simulation or a hardware set-up; 1. The noise ripples in the control signals are virtually
Step 5 : Incrementally increase both by the same amount until the unchanged for o,+ w,=190 and min(w, and wo}2 40;
noise levels and/or oscillations in the control signal and 2. Further increase of w, beyond 40 only affects the disturbance
output exceed the tolerance; rejection error. The transient error in this case appears to be
Step 6: Incrementally increase or decrease wc and o. individually, solely dependent on U,;
if necessary, to make trade-offs among different design 3. Bringing the transient error and the disturbance error to the
considerations, such as the maximum error during the same level while maintaining the same noise level in the
transient period, the disturbance attenuation, and the control signal leads to w , 4 0 radisec and oc=150raasec;
magnitude and smoothness of the controller. 4. The maximum bandwidth is quickly determined using the
proposed design procedure.
Remarks: Figure 4.1 shows comparisons between the LADRC and the
1. If the simulationitest results in Step 4 are unsatisfactory, it is regular PD conholler in Example 3.1 (w,=20 radisec). The
likely that the transient design specification described by D' fairness of the comparison is based on the fact that both control
is untenable due to the noise and/or sampling limitations. In signals have the same level of noises.
this case, the control "goals" may have to be lowered by
reducing Dc and therefore ocand w.. The LADRC shows distinct advantages over the regular PD
controller because:
2 . Note that this approach can be easily extended to the
1) No detailed mathematical model is required;
Luenberg state observer- based state feedback design. 2) Zero steady state error is achieved without using the
integrator term in PID;
Example 4.1: Consider the same control problem in Example 3.1, 3) Much better command following is demonstrated during the
hut apply the LADRC in (4.5) to (4.10). Note that h=23.2 for this transient stage;
problem, but to make the design realistic, assume the best
4) The controller is extremely robust as the plant's damping
estimate of b is b,,=40. Rewrite the plant differential equation
coefficient deviates from its original value of 1.41. In fact,
(3.10)as
the output response is essentially unchanged when this value
9 = ( - 1 . 4 l y t 23.2Td)+(23.2-40)u+4Ou = / + 4 0 u changes from-30 to 30!
The LESO is

and
zI
[
i = -30;
-4 0
+ y,zi + j,and
01
-30, I 0

[0
0 30.
0 1 z + 40 36Jz
High performance is achieved through the use of a disturbance
observer and is examined in Figure 4.2. The further increase in
w, from 40 to SO radisec does not greatly improve hacking
performance, but it allows more noise into the observer states.
OYII)",

z, +/=-1.41j+23.2T,+(23.2-40)u,as f+-
The controller is defined as O I

U =a
40
and ug = k p ( r - z , )- k,z,

kd = 25uC,5 = 1, and k p =U:

.o 2

where w, is the only design parameter to be tuned. A trapezoidal

transient profile is used with a settling time of one second, or
~3~4. From (4.1 S), w, is selected to he 40 radsec.
*!me ..CO"d

Figure 4.1 Comparison of PD and LADRC Controllers

Simulation Setuw
The LADRC is simulated using Sirnulink ("sine
I , 1 ode1 with a fixed
step of 1 ms) with a white sensor noise of the peak value 0.1% of
that of the output, a 1 ms sampling period. As in Example 3.1, a
Proceedings of the American Control Conference
4995 Denver. Colorado June 4.6.2003
 . . ~ '
~I
~ ~
optimization, introduced in Section 3.3. The o,-optimization can

-' be similarly applied using

~

',
I S" t p,s"-'+... + p,-,.+p, = (s +w,y (4.28)
2 ".#&,I" 4

V. Concluding Remarks
A new set of controller design, tuning, and optimization tools are
presented, based on innovative scaling, parameterization, and
optimization concepts. These concepts prove to be applicable not
only to the existing model-based controller design methods such
as pole placement and loop shaping, but also to model-
independent design methods. These new tools make controller
Figure 4.2 Performance of the LESO design and tuning easier and more effective. For example, the
tuning of a PID controller is reduced to adjusting a single
4.5 Extensions to a Plant of an Arbitrary Order parameter, instead of three. Practical optimization is successfully
For a general nrh order plant with unknown dynamics and incorporated into the one-parameter tuning. The new concepts and
external disturbances, methods were tested successfully in software simulations, which
y(n) =J ( ~ y , + , ...,y ( n - ~ ) , u , i ,,,,u ( m - ~ ~ ,w ) + bu (4, 19) incorporated real world scenarios such as sampling, sensor noise,
disturbance, and the lack of mathematical model.
Han's observer can be similarly derived, stalting from the
REFERENCES
.i,=x* . _N. Minorskv.~. "Directional Stabiliw and Automaticallv
111
i, =x, Steered Bodies," J. Am. Soc. Nav. Eng., vol. 34, p. 280,
... 1922.
(4.20) [2] K. J. Astrom and T. Hagglund, "PID Control," The Control
in= x"+, + b,u Handbook, W.S. Levine, Ed. CRC Press and IEEE Press,
=h
in+< 1 9 9 6 , ~198.
.
[3] J.G. Ziegler and N.B. Nichols, "Optimal Settings for
.Y = 2,

'2, = z2 -P,(z, - y ( l ) ) IEEE Conference on Control Applications; Real-Tinre

2, = z3 - iadz, - ~ ( 0 ) S)stems, The International Journal of Time-Critical
... (4.22) Computing Systems, vol. I I , no. 3, Nov. 1996.
[6] Z. Gao, "An Algorithmic Approach to Loop Shaping with
2, =I.+, -fl,(z, -y(O)+b,,u Applications to Self-Tuning Control Systems," Journal of
.%.,
=-B.+,(z, - y ( t ) ) rhe Franklin Institure, vol. 332B. no.6. pp. 643-656, 1995.
..
With the gains properly selected, the observer will track the states System Design Using Matrix Interpolation", Proc. of the
and yield 33rd IEEE Conference on Decision and Control, pp. 2506-
251 1, December 1996.
z,(t) + Y(t),Z,(t) --f j ( f ) , . . . , Z " ( t )+ y ' " - " w (4.23) [8] J. Han, "A Class of Extended State Observers for Uncertain
z,+,(t) + f(t, y,+"..,y'"-'',u,i ,...U("-", w ) Systems," Control and Decision, vol. IO, no.1, pp. 85-88,
1995.(in Chinese)
The controller can also be similarly designed as in (4.7) and (4.9): [9] J. Han, "Nonlinear Design Methods for Control Systems,"
Proc. Of The 14" P A C World Congress, Beijing, 1999.
- +U0
="+I
(4.24) [IO] Z. Gao, Y. Huang, and J. Han, "An Alternative Paradigm
b" for Control System Design," Proc. of the 2001 IEEE
which reduces the plant to approximately a unit gain cascaded Conference on Decision and Conrrol, Dec. 2001.
integrator plant [ I l l Z. Gao, "From Linear to Nonlinear Control Means: A
y'"' = ( J -z,,+,)+ u0 uO (4.25) Practical Progression," /SA Transoclions, vo1.41, 110.2, pp.
177-89, April 2002.
and
uo = k,(r - 2 , ) - k,,z, - kdm-,
zn (4.26)
where the gains are selected so that the closed-loop characteristic
polynomial has n poles at -wc, i.e.,
S" + k,,-,s"-' f... + k,,s + k , = (s+w,)" (4.27)
Proceedings o f the American Control Conference
4996 Denver, Colorado June 4-6.2003