Sie sind auf Seite 1von 6

SICE-ICASE International Joint Conference 2006

Oct. 18-2 1, 2006 in Bexco, Busan, Korea


Maximum Sensitivity based PID Controller Tuning:
A survey and comparison.

Richard W. Jones1 and Ming T. Tham2


' Institute of Technology & Engineering,
Massey University, Wellington, New Zealand
(Tel: +64-4-801-2794; E-mail: r.wjonesgmassey.ac.nz)
2 School of Chemical Engineering and Advanced Materials,
University of Newcastle upon Tyne, U.K.
(Tel: +44-191-222-7285; E-mail: ming.thamgncl.ac.uk)
Abstract: Maximum sensitivity, M5, is a closed-loop design specification that satisfies many of the requirements of a
useful and practical PID controller tuning method. A variety of PID tuning methods using this specification have been
developed with the majority being developed primarily for auto-tuning purposes. This contribution examines the
general applicability of a number of the tuning methods by comparing their performance on some benchmark process
plant models.
Keywords: PID, Maximum Sensitivity, Controller Tuning. Process Control.

1. INTRODUCTION
The initial part of the paper surveys the available
The process industries are dominated by PID maximum sensitivity-based PID design methods and
controllers with these comprising of at least 9500 of all characterizing them in terms of the way the controller
controllers. Despite the proliferation of countless PID parameters are determined. The majority of the
controller tuning methods, the search still continues for surveyed methods were developed with auto-tuning in
an approach that can encompass the desirable features
mind though many of the methods can also be used for
for a useful and practical design method; namely that it control design when a low-order model of the process is
should be applicable to a wide range of systems and already available. The comparison of the different
should introduce specifications that capture the essence approaches on a variety of process models completes
of real control problems. A major consideration in real the paper.
control problems is the design trade-off between
performance and robustness.
2. THE PID CONTROLLER
Maximum sensitivity, which represents the inverse of Industrial Process PID controllers include set-point
the minimum distance on the Nyquist diagram between weighting to provide flexibility in shaping the transfer
the loop transfer function and the critical point, has been function from the set-point to the output. In addition the
shown to be effective as a robust performance tuning derivative term is often implemented in the feedback
parameter, the choice of which is transparent to the signal path to reduce 'derivative kick'. Filtering of the
potential user [1]. Shinskey [2] initially suggested derivative action term is also carried out to reduce the
optimizing load disturbance rejection with constraints possible effects of measurement noise on the output
on the maximum sensitivity, Ms, and complementary signal. The Instrument Society of America (ISA) defines
maximum sensitivity, Mp. Using only maximum the following general PID controller:
sensitivity as a control design parameter was proposed
and investigated by Persson and Astrom [3] via a
dominant pole-placement control approach. U(s) = k bR(s)-Y(s)±+ 1 (R(s) -Y(s)) +
Subsequently this work became the basis of the K-l (1)
(Kappa-Tau) maximum sensitivity-based PI and PID (Td N)s+1 (cR (s) Y(s))}
control design approach [4]. More recent control
design-related work has concentrated on examining, in
further detail, the nature of both the maximum and where Y(s), R(s) and U(s) denote the Laplace transforms
complementary sensitivity constraints as well as of the process variable, set-point and manipulated
developing efficient numerical procedures to carry out variable. The proportional gain is given by kc and Ti and
control design, see for example [5, 6, 7, 8]. Because of Td are the integral and derivative time constants
these developments and its increasing importance as a respectively. The set-point weighting is provided by b
design approach, this contribution will examine while the c (=0 or =1) weighting determines if the
maximum sensitivity-based methods for tuning PID derivative action in the controller uses the error or
controllers via their application to a range of process output signal. The parameter N specifies the degree of
models. derivative action filtering. This form of PID controller

89-950038-5-5 98560/06/$10 2006 ICASE


3258
encompasses all the PID control structures used towards the desired specification. This approach
throughout the paper. usually embodies optimisation techniques and
borrows from adaptive loop-shaping ideas, see for
3. MAXIMUM SENSITIVITY example [7, 8].
* Non-convex optimisation: This is the approach
With G(s) representing the plant model and GC(s) the developed by Astrom et al to develop maximum
controller then the maximum sensitivity specification is sensitivity based designs for both PI [ 1] and PID [6]
defined as controllers.
A maximum sensitivity based design approach from
Ms=max
O<c)<oo 1+G(jo))Gc(jo))
2
(2) each of these groups will now be described. These four
approaches will be used later for the simulation studies.
where the loop transfer function GI(s)=G(s)Gp(s). The 4.1 K-T Tuning
quantity Ms is the inverse of the shortest distance from
the Nyquist curve of the loop transfer function to the The K-l approach was derived from extensive
critical point (-1, jO). Reasonable values of Ms are in the simulation investigations of the dominant pole
range 1.4 to 2.0; the lower the value the more robust the placement design method applied to a range of typical
control. The efforts to generate simple control design process models [5]. Both step response based and
relationships for this specification are hindered by the frequency response correlations have been developed.
actual geometric definition though - the condition that In the frequency response method, the process is
the loop transfer function goes through a point on the characterized by static gain Kp,, ultimate gain Ku, and the
Nyquist curve is ultimate period T,. The controller parameters are given
as functions of a gain ratio, K= 1(Krp KU).
GI(j) =-1+ e JO (3) Table 1: K-T tuning formula and coefficients for PID

where 0 is the angle that a line drawn between the Ms 1.4 2.0
critical point and the point on the Nyquist curve makes ai aO a, a2 aO a, a2
with the negative real axis. The condition for tangency k/Ku 0.33 -0.31 -1.0 0.72 -1.6 1.2
at the same point can be written as T/TU 0.76 -1.6 -0.36 0.59 -1.3 0.38
T/TU 0.17 -0.46 -2.1 0.15 -1.4 0.56
b 0.58 -1.3 3.5 0.25 0.56 -0.12
dG1( }jo) /T
arg 0 (4) Table 1 gives parameters for the function:
dco 2

4. DESIGN METHODS f (t) = aO exp (alK + a2K2) (5)


The maximum sensitivity-based PID design method
manipulates the Nyquist plot to provide desirable whereJ(t) is kIK1, T, T], T/TU or b.
closed-loop system characteristics and therefore some of
the design approaches have a lot in common with loop 4.2 Approximate analytical function
shaping design. In the last few years, a variety of M,
design approaches have been developed. These can be Shao and Wang introduced a PID control design
characterized as follows: approach that uses an approximate analytical approach
to satisfy a closed-loop maximum sensitivity
* Correlation based design: Tuning rules usually specification [5]. The basic procedure involves making
developed through extensive simulation studies to the controller zeros equal to the poles of the model to
develop correlations between some process defined produce a loop transfer function of the following form:
characteristic and the PID parameters, for example
see [4].
* Approximate Analytical approaches: Similar in G, (s)G(s)= ke-sL (6)
concept to the correlation approach except that the
relationships between the process related where L is the time-delay associated with the process
parameters and the control parameters are derived model and k is a controller associated gain term. From
from the nonlinear analytical relationship between (3) and (4) the relationship between the sensitivity
the control parameters and the maximum sensitivity, specification and the model and control parameters is
see for example [5, 9]. nonlinear is due to the geometric definition of Ms. Wang
* Iterative Tuning: Starting with an initial control and Shao were able to obtain the following approximate
design, the controller parameters are then iterated expression, through simulation studies, for the

3259
calculation of k as a function of L and Ms: Ziegler-Nichols tuning rules does provide the global
minimum via this design procedure.
1 1.508
k= 1.451 - (7) 4.4 Non-Convex Optimisation
L 9 Ms
The work of Astrom, Panagopoulos and Hagglund is
This expression makes it is easy to fimu J the the most in-depth examination of PI and PID controller
corresponding PID controller parameters. For a pirocess tuning using the maximum sensitivity based approach.
model of the form They formulated the design problem as a non-convex
optimisation problem, optimising load disturbance
-Ls rejection with a constraint on the maximum sensitivity.
e (8) This was done because in process control applications,
G (s) 2
axs +ds+y
the primary control task is to reject disturbances
whereas set-point responses are typically of secondary
importance. They consider both M, and Mp, in their
The resulting PID design equations are: optimisation procedure by utilising a circle with a centre
C and a radius R where:
kc 1.451
1.508 kc Y1.508)
1.451
-
MJs Ti L Ms M- MMpp -2MM2
s~~ p +M2p~ _1

1.508
(9) C C= 2M, -) (Ms (12)
kc Td= 1.451-
Ms +Mp -1
R=
2Ms (Mp -1)
4.3 Iterative Tuning of the Modulus Margin
When optimising the PID control parameters, the main
Garcia et al [8] developed an iterative ttuning constraint, apart from requiring the closed-loop to be
approach with a number of closed-loop specifica tions: stable, is that the Nyquist curve of the loop transfer
modulus margin (inverse of the maximum value of the function is outside the circle with centre at s=-C and
sensitivity function), phase margin and cro,ssover radius R. This constraint can be formulated as:
frequency. The first two specifications ensur*e the
stability and robustness of the closed-loop system while (13)
the third specification, the crossover freqiuency,
C + kc - -(ki - O kd )G (ic) ) R2
influences the rise time of the closed loop system. Only
the first two specifications will be considered her(e. The where k,=kJT, and kd= kCTd .The constraint that the
frequency criterion on which the iterative con-troller maximum sensitivity is smaller than M, corresponds to
tuning method is based is defined as the weighte d sum C= landR= 1M,.
of squared errors between the desired specificatiorns and
the actual values of the frequency characteristics: 5. MODELS AND CONTROLLERS

I( (10) Table 2 presents the process models used to compare


J ( a Md) +
(a- )J)
the M, design approaches. These comprise a simple
second-order plus time-delay (SOPTD) system, GC a
where Al and A2 are weighting factors, Ma and Md are non-minimum phase system, G2 a plant with complex
respectively, the actual and desired inverse of the poles, G3 and a high-order lag, G4. Their corresponding
SOPTD approximation is also shown.
infinity norm of the sensitivity function. ()a and ('b are
the actual and desired phase margins.
The controller parameters for each of the maximum
The controller parameters minimizing the criterion can sensitivity based design methods are shown in Table 3.0.
be obtained iteratively by the Gauss-Newton method: Although Astrom and co-authors do examine a whole
range of possible maximum sensitivity values, most
researchers restrict themselves to two values: M5 1.4, a
pi+, pi =
7jR- IY(pi) (11) robust choice, and M,=2.0, a more aggressive controller.
where i is the iteration number, yi the step size, R a The approximate analytical and iterative design
positive definite matrix of dimension n'p x n'p, that is approaches use a set-point weighting value of b=1
chosen to be the Hessian matrix for a fast convergence. throughout. The K-l and non-convex optimisation
The iterative algorithm only gives a local minimum of methods also use an additional set-point filter with time
the criterion so the initial choice of controller is constant Tf whenever the set-point weight b=0. All the
important. However, it has been found that using design methods use a PID with c=1 except the

3260
non-convex optimisation approach where c=0. The only for a set-point response and using a PID controller
phase margin used for the iterative approach is fixed at structure with fixed b=c=1.0 even though the K-l design
600 while both the iterative and approximate analytical includes a set-point weighting term. The studies
approaches fix the ratio of the PID parameters T, Td-4. presented here examine both set-point and disturbance
The derivative filter time constant was chosen as Td /20 responses using a general PID structure.
for each controller.
Model GC is a simple second-order plus time-delay
Table 2: Process Plant Models. system and Figs. l(a) and l(b) show the controlled
responses for M,= 1.4 and M,=2.0.
Model SOPTD
O0.2s
GI=esodel G (se +O0.2s1)2 A

1.2F

1-2s e 2?.868s
(s+) (I( 0+.4454s)
08F
e
s2 +1.5s+1I
~~~~e
011
(1+ 0.9623s)2 0.G

C e4 1.73s 04F

(s +1)5G= (1+1.8 9S)2 0.2

U_ -
Table 3: Process Models and their PID parameters. Time
10
(AA-approximate analytical, IT-iterative,
OPT-optimisation) Fig. 1 (a). Model G1: MA= 1.4,L) K-T,(... ) Analytical,
(--) Iterative.
Ms Kc Ti Td b
GI 1.4 K-l- 3.43 1.329 0.325 0.53 .4
2.0 K-l 6.77 1.067 0.269 0.26
1.4 AA 3.74 2.0 0.5 1 1.2

2.0 AA 6.970 2.0 0.5 1


1.4 IT 3.57 1.64 0.41 1
G2 1.4 K-l 0.13 1.056 0.169 2.73
2.0 K-l 0.508 1.881 0.503 0.37
1.4 AA 0.116 0.891 0.223 1 0lFA-

2.0 AA 0.217 0.891 0.223 1 04k


1.4 OPT 0.312 2.25 0.8 0.60
2.0 OPT 0.542 2.07 0.79 0
G3 1.4 K-l 4.95 1.091 0.261 0.53
2.0 K-l 10.02 0.866 0.219 0.26 n
S 1u 15
1.4 AA 5.493 1.925 0.481 1 Time

2.0 AA 10.24 1.925 0.481 1 Fig. (b). Model GI: Al,==2.0,() K-T,(...)Analytical.
1.4 IT 5.76 1.88 0.47 1
G4 1.4 K-l 0.757 3.603 0.971 0.56 For M,=1.4 the iterative approach provides a very nice
2.0 K-l 1.375 3.394 0.852 0.3 set-point response with the analytical approach being
1.4 AA 0.817 3.78 0.945 1.0 almost dead-beat in nature and the K-l being fairly
2.0 AA 1.523 3.78 0.945 1.0 damped. For M,=2.0, the analytical tuning approach
1.4 OPT 0.784 2.68 1.24 0.0 resulted in a large overshoot while the K-T set-point
2.0 OPT 1.47 2.33 1.25 0.0 response has been detuned too much by the set-point
weighting. With regard to the disturbance rejection
responses the K-l easily out performs the other two
6. SIMULATIONS approaches.
Although Astrom et al. [1, 7] carried out extensive
performance comparisons between their non-convex Model G2 is a non-minimum phase system, where the
optimisation derived PI and PID controllers, the right-half plane zero is quite close to the imaginary axis.
literature is sparse with regard to comparisons between This leads to a large time-delay when the system is
the different M, design methods. The iterative tuning approximated as a second order-plus time delay model,
approach has been compared with K-l design [7], but (see Table 2). Figs. 2(a) and 2(b) show the controlled

3261
responses for M,= 1.4 and M,=2. 0.

14
1.2

068

0.6

04

0.2

1.4

1.2

0G

086
0.4

0n2
0
0 10 20 30
Time
40 50 6;0

Fig. 2(a). Model G2: M,=1.4, (_) K-T, (..) Analytical,


(--) Optimisation.
10

0.6

0.4

0.2

1.4

12

6.8
Jf

~0
ll
Similar responses are obtained for disturbance rejection
when M,=2.0 but the K-l set-point response is under
damped and anything but smooth. The analytical
set-point response yet again demonstrates an increase in
overshoot for the larger value of Ms although not as
much as that seen in the previous simulations.

1.4

1.2

II
II=
I

2 4 6 8 10
Time
12 14 16

Fig. 3(a). Model G3: Ms= 1.4, (_) K-T, (...) Analytical,
(--) Iterative
18 20

0.2 0.6
0 10 20 30 40 50 b0
Time
04

Fig. 2(b). Model G2: M,=2.0, (_) K-l, (...) Analytical, 0.2
(--) Optimisation.
0 2 4 6 8 10 12 14 16 18 20
For Ms, 1.4 the responses are very similar being Time
differentiated slightly by their rise times to a set-point Fig. 3(b). Model G3: Ms=2.0, (_) K-T, (...) Analytical.
change. For MA=2.0, the optimisation approach uses an
additional first-order reference filter, time constant, The final model, G4, is a high order process model. Figs.
T,> 1.03, to reduce the percentage overshoot of the 4(a) and 4(b) show the controlled responses for Ml, 1.4
set-point response. This unfortunately has the effect of and M,=2.0. For M,=1.4 the analytical and K-l designs
slowing the rise time of the response as well. The both provide damped set-point responses with the
percentage overshoot of K-l response has increased optimisation approach giving an overshoot of
appreciably over that shown for Ml, 1.4 while the approximately 10% even with b=0 and an additional
analytical response has an overshoot of over 400O. reference filter of Th,=0.51. For M,=2.0, the K-l
approach again provides an under damped set-point
The model G3 has complex poles plus a small associated response coupled with a very good disturbance rejection
time-delay. Figs. 3(a) and 3(b) show the controlled response. The other two approaches both exhibit
responses for M,=1.4 and M,=2.0. The K-l approach appreciable overshoot (20% or above). The optimisation
again demonstrates excellent disturbance rejection, even approach uses b=0 and an additional set-point filter with
though the set-point response was under damped Tsv=0.72.
set-point response for Ml, 1.4. The analytical and
iterative approaches have almost exactly the same
controller parameters, hence their similar responses.
While the set-point responses are excellent with about
500 percentage overshoot, their disturbance rejection
responses are much more damped than the K-l approach.

3262
methods gave consistently good set-point responses but,
at the higher value of M5, the approximate analytical
1.2 approach consistently gives large overshoots. This
perhaps could be improved by using a PID-P structure,
see [9], which allows the calculation of a set-point
118 weighting parameter for the PID controller. Most of the
methods work best on time-delay dominant systems
016 giving good set-point tracking, apart from the
014
non-convex optimisation design on G4, and disturbance
rejection responses.
0 12
REFERENCES
0 111 20 30 40 rn 60 70D
Time [1] K. J. Astrom, H. Panagopoulos and T. Hagglund,
Fig. 4(a). Model G4: M,=1.4,( K-,(o...)Analytical, "Design of PI Controllers based on Non-Convex
(--) Optimisation. Optimization," Automatica, Vol. 34, No.5, pp.
585-601, 1998.
[2] F.G. Shinskey, "How good are our controllers in
I1.4
absolute performance and robustness?"
1.2 Measurement and Control, Vol. 23, pp. 114-121,
1990.
[3] P. Persson and K. J. Astrom, "Dominant pole
068
design-a unified view of PID controller tuning,"
IFAC Sympos.. on Adaptive Systems in Control
0.6 and Signal Processing, Grenoble, pp. 127-132,
1992.
0.4
[4] K. J. Astrom and T. Hagglund, PID Controllers
02
Theory, Design and Tuning, Instrument Society of
America, Research Triangle Park, NC, 1995.
0 10 20 30 40 50 60 70
[5] Y-G. Wang and H-H. Shao, "A new PID auto-tuner
Time based on sensitivity specification," Chemical
Fig. 4(b). Model G4: M,=2.0, (_) K-T, (..) Analytical, Engineering Research and Design, Vol. 78, No. 2,
(--) Optimisation. pp. 312-316, 2000.
[6] H. Panagopoulos, K.J. Astrom and T. Hagglund,
7. DISCUSSION "Design of PID controllers based on constrained
optimisation," IEE Proceedings - Control Theory
As well as providing insight into the range of Applications, Vol. 149, No.1, pp. 32-40, 2002.
performance, on a variety of process models, of [7] D. Garcia, A. Karimi and R. Longchamp, "Robust
different maximum sensitivity-based PID control design PID controller tuning with specification on
approaches it was hoped that the simulations might also modulus margin," Proceedings of the American
provide some insight into the use of different PID Control Conference, Boston, pp. 3297-3302, 2004.
control structures. Overall, the K-l design would seem to [8] J. Crowe and M.A. Johnson, "Automated
be the best. Excellent disturbance rejection performance maximum sensitivity and phase margin
was provided even though the set-point response was in specification attainment in PI Control," Asian
most cases over-damped (implying some modification Journal of Control, Vol. 4, No. 4, pp. 388-396,
of the b correlation could be carried out) it was 2002.
acceptable in all cases. The non-convex optimisation [9] Y-G. Wang, and W-J. Cai, "PID Tuning for
approach is a powerful off-line control design approach. Integrating Processes with Sensitivity
The fact that the set-point responses, despite the use of a Specification," 40th IEEE Conference on Decision
set-point weighting and an additional set-point filter, and Control, Orlando, Florida, 2001.
still produced poor performance in some cases was
disappointing. It implies that concentrating overly on
optimising the disturbance rejection response has been
to the detriment of the set-point tracking.

The iterative and approximate analytical approaches


consistently produce similar disturbance rejection
responses which might be due to both methods
imposing the constraint T, Td-4. The choice of a smaller
value of phase margin in the iterative method might also
lead to improved disturbance rejection. For M,=1.4,both

3263

Das könnte Ihnen auch gefallen