Publication [P7]

L. M. Eriksson, M. Johansson, “Simple PID Tuning Rules for Varying Time-

Delay Systems”, in Proc. The 46th IEEE Conference on Decision and Control
(CDC 2007), New Orleans, LA, USA, December 12-14, 2007, 7 p.

© 2007 IEEE. Reprinted, with permission.

This material is posted here with permission of the IEEE. Such permission of the
IEEE does not in any way imply IEEE endorsement of any of Helsinki University of
Technology's products or services. Internal or personal use of this material is permitted.
However, permission to reprint/republish this material for advertising or promotional
purposes or for creating new collective works for resale or redistribution must be
obtained from the IEEE by writing to [email protected].

By choosing to view this document, you agree to all provisions of the copyright laws
protecting it.
Simple PID Tuning Rules for Varying Time-Delay Systems
Lasse M. Eriksson and Mikael Johansson

Abstract— This paper discusses tuning of PID controllers for systems has been approached using multi-objective optimi-
varying time-delay systems. We analyze the properties of the zation to develop rules that maximize the jitter margin, i.e.
AMIGO tuning rules of Åström and Hägglund applied to vary- the maximum value of any additional varying time-delay in
ing time-delay systems and propose improved tuning rules the control system [2], [3]. This approach might result in
which increase the robustness to delay variations at the expense
complicated rule structures, which only apply for a certain
of a small degradation in nominal performance. We suggest a
tuning scheme that uses the simple AMIGO tuning on an ex- range of process parameter values. In this paper, we extend
tended plant, and define the design concepts for extending the the simple AMIGO (Approximate M-constrained Integral
plant. This approach allows treating the maximum time-delay Gain Optimization) tuning rules [4], [5] that have been de-
as a design parameter for the tuning rules. The proposed tun- veloped for PID tuning in process control, such that varying
ing rules are compared via simulations. time-delays are better taken into account while adhering to
the simple rule structures. There are AMIGO rules for non-
I. INTRODUCTION integrating and integrating processes, but we only consider

A NY practical control system suffers from delays. These

can stem from process dynamics, actuators or sam-
pling. The delays are often either assumed negligible or con-
non-integrating processes. The integrating processes are
treated in [3]. We consider extending the plant with a well
designed measurement filter before applying the AMIGO
stant, but in some cases the variance in delay times (jitter) design on the extended plant. This approach allows using
plays a significant role. There exists a variety of methods for the required jitter margin as a parameter to the tuning rules.
control of time-delay systems with constant delays, but the In the next section the basic concepts and preliminaries
toolset for dealing with varying time-delays is much more required to understand the problem and the proposed tuning
limited. While time-delay systems are abundant in practice, approach are presented. The AMIGO tuning that is fre-
our work is partly motivated by the emerging technology of quently referred to in the paper is also presented in Section
networked control. There is strong current desire to develop II. Based on the AMIGO tuning analysis presented in Sec-
technologies that allow a transition from today’s wired in- tion III, we propose an extended plant design approach and
frastructures to more flexible and cost-efficient wireless filter design concepts that better fit for varying time-delay
automation systems. This “wireless migration” requires sig- systems in Section IV. Section V compares the proposed
nificant advances in a wide range of technologies, including tuning and Section VI analyses the delay-dominant proc-
wireless communication, networking, software engineering, esses separately. Simulation results are presented in Section
and control. Our focus is on the design of practical control- VII and Section VIII states the conclusions.
lers that are robust to the varying time-delays incurred by
unreliable communications. II. PRELIMINARIES
The PID (proportional-integral-derivative) controller is
the most common controller in industrial applications today, A. PID Controller
and it is likely to be the most important controller also in The PID controller is the most common controller used in
wireless automation solutions. There are many different ar- industry. There are several versions of the basic algorithm
chitectural options for networked PID controllers, including ⎛ 1
t
de(t ) ⎞
the use of “network observers” [1] to compensate for delay, u (t ) = K ⎜ e(t ) + ∫ e(τ )dτ + Td ⎟,
⎝ Ti 0 dt ⎠ (1)
jitter, losses and other network deficiencies, or simply keep-
ing the architecture of the wired controller and modifying e(t ) = yr (t ) − y (t ),
the controller parameters. This paper takes the second path, where u is the control signal, e the error signal, yr the refer-
and develops new tuning rules that aim at providing good ence signal, y the process variable, K the gain, Ti the integra-
time-domain performance while being robust to varying tion time, and Td the derivative time. In this paper we con-
time-delays. sider the following version of the PID controller, since the
Recently, the PID tuning problem for varying time-delay “text book” version (1) is very sensitive to noise.
t

L. M. Eriksson is a Ph.D. student with the Control Engineering Labora- u (t ) = k ( byr (t ) − y f (t ) ) + ki ∫ ( yr (τ ) − y f (τ ) )dτ
tory, Helsinki University of Technology, P.O.B. 5500, Espoo, FI-02015 0
Finland (e-mail: [email protected]). (2)
M. Johansson is an associate professor with the Automatic Control,
⎛ dy (t ) dy f (t ) ⎞
+ kd ⎜ c r − ⎟
Royal Institute of Technology, Stockholm, SE-10044 Sweden (e-mail: ⎝ dt dt ⎠
[email protected]).
Here k, ki, and kd are the controller gains, b and c the set- yr yr
+ P(s) + C(s)
point weights, and yf is the filtered process variable such that _ _
1
Y f ( s ) = G f ( s )Y ( s ) = Y ( s) , (3)
(1 + sT f ) n Δ Δ
where Y(s) is the Laplace transform of the process variable C(s) P(s)
y(t). Tf is the measurement filter time-constant and n the or-
Fig. 1. Continuous-time controller (C) and plant (P) with an uncertain time-
der of the filter. Typically n equals one or two. varying delay (Δ) in the feedback loop. On the left, Δ is the controller out-
put uncertainty. On the right, Δ is the process output uncertainty.
B. Jitter margin
where Kp is the static gain, T the time-constant, and L the
There exists a variety of stability results for varying time- (constant) time-delay. The AMIGO rules were developed by
delay systems (see e.g. [6]). Most of the theorems are in time analyzing different properties (performance, robustness etc.)
domain, but there are also certain frequency domain criteria of a process test batch with over 130 processes, and the re-
such as [7]. The criteria presented in [7] are based on the sulting tuning rules are
assumption that the varying time-delay is bounded. The up-
1 ⎛ T⎞
per bound for the delay is called the jitter margin. The delay K= ⎜ 0.2 + 0.45 ⎟ ,
can be of any type (constant, time-dependent, random), but Kp ⎝ L⎠
the jitter margin determines the upper bound for it. The for- 0.4 L + 0.8T
Ti = L, (7)
mal definition of the jitter margin is given in [7], where L + 0.1T
three different controller/plant–uncertainty combinations in 0.5LT
continuous- and discrete-time are investigated. The first one Td = .
0.3L + T
is shown in Fig. 1, left, where a continuous-time plant and a
In order to use the PID controller with filtering (2), the
continuous-time controller with controller output uncertainty
rules are extended as follows [4]:
are shown. This continuous-time SISO system is stable for
⎧ ⎧0, if τ ≤ 0.5
any time-varying delays defined by ⎧k = K ⎪b = ⎨
Δ (v) = v ( t − δ (t ) ) , 0 ≤ δ (t ) ≤ δ max (4) ⎪ ⎪ ⎩1, if τ > 0.5
⎪ K ⎪
if ⎨ ki = ⎨c = 0 (8)
⎪ Ti ⎪
P ( jω )C ( jω ) 1
, ∀ω ∈ [ 0, ∞[ , ⎪ ⎪T = ⎧⎨0.05 / ω gc , if τ ≤ 0.2
< (5) ⎩kd = K ⋅ Td ⎪⎩ f ⎩ 0.1 ⋅ L, if τ > 0.2.
1 + P ( jω )C ( jω ) δ maxω
where δmax is the jitter margin. The proof of the result is Here τ refers to the relative dead time of the process
based on presenting the uncertainty (varying time-delay) L
τ= , (9)
with an operator Δ F := ( Δ − 1) D 1 s (s being the Laplace op- L +T
erator) and on the small gain theorem. which has turned out to be a significant process parameter
However, in this paper the jitter is assumed to be after the when controller tuning is considered, and ωgc is the gain
plant (e.g. sampling jitter) as depicted in Fig. 1, right. Since crossover frequency.
the signals in the control loop are all continuous, and only The development of the AMIGO rules was based on the
the plant and controller switch their positions, the small gain following robustness criterion: if the Nyquist curve of the
theorem-based stability proof, and the resulting criterion (5), loop transfer function does not intersect a circle with center
still hold for the control system in Fig. 1, right. cR and radius rR defined as
C. AMIGO tuning rules 2M 2 − 2M + 1 2M − 1
cR = − , rR = , (10)
The objective of this work is to develop simple tuning rules 2 M ( M − 1) 2 M ( M − 1)
for the PID controller in varying time-delay systems. The then the sensitivity function and the complementary sensitiv-
AMIGO tuning rules [4], [5] were selected as the point of ity function are less than M [9]. The robustness is thus cap-
comparison, since they provide good performance and are tured by one parameter only, M. The value M = 1.4 was used
robust to disturbances in control systems without varying in the AMIGO rule development, although finally the rules
time-delays. The AMIGO tuning rules are based on ap- did not quite satisfy the constraint. For the process test batch
proximating the process with the so-called KLT-process a 15 % increase of M was reported (M ≈ 1.6).
model (first order lag plus delay) possibly determined via a
simple step experiment, see [4]. The tuning rules are ob- III. ANALYSIS OF THE AMIGO TUNING
tained from the KLT-parameters. The most well-known step
In [2] the jitter margin properties of AMIGO tuning are ex-
response based tuning rules were presented by Ziegler and
amined. It is shown that the jitter margin of a pure KLT-
Nichols already in 1942 [8]. The KLT-process model is
process with AMIGO tuned controller (2) is approximately
K p − sL
PKLT ( s ) = e , (6) δmax, AMIGO ≈ 0.71⋅L. It is also suggested in [2] that the meas-
1 + sT urement filter time-constant Tf might have a great impact on
the jitter margin. To further explore the jitter margin proper- Closed-loop Bode-plots and 1/(δmax ω) for δmax = 0.5
15
ties in this case the analytical jitter margin formula is de- KLT+PID w/o filter
rived here. Consider the pure KLT-process (6) and the con- 10 1/(δmax ω)

troller (2), where the parameters k, ki and kd are tuned ac- 5

KLT+PID w. filter

cording to AMIGO tuning rules (7) and (8). The jitter mar-
0
gin δmax solved from (5) becomes

Magnitude (dB)
1 + P ( jω )C ( jω ) -5
δ max <
jω P ( jω )C ( jω ) -10

(11)
(1 + jωT ) (1 + jωT f )
n -15
1
= γ − jω L + , ∀ω ∈ [ 0, ∞[ -20
e ⎡⎣ηω 2 + jκω + λ ⎤⎦ jω
-25
where
L ( 0.4 L + 0.8T )( 0.3L + T )
-30
-4 -3 -2 -1 0 1 2
10 10 10 10 10 10 10
γ= , Frequency rad/s
0.2 + 0.45 T L
Fig. 2. Closed-loop Bode diagram of a KLT+PID system with and without
η = −0.5L2T ( 0.4 L + 0.8T ) , (12) the measurement filter.
The discussion above motivates a more thorough evalua-
κ = L ( 0.3L + T )( 0.4 L + 0.8T ) ,
tion of how the measurement filter time-constant affects the
λ = ( L + 0.1T )( 0.3L + T ) . jitter margin of AMIGO rules. Obviously, the value of Tf
It seems that (11) does not have an analytical solution, should not be dominant in the control system, i.e. it should
which could possibly give more insight on the dependency be relatively small compared to the process dynamics (time-
of jitter margin on the process parameters. Numerically, constant and delay). If the filter time-constant was dominant
though, the expression is trivial to solve if a sufficiently in the system, losses in performance would be expected. To
wide frequency range is used. better analyze the relationship between Tf and δmax, the jitter
Some remarks can be made on (11). If the measurement margins for a range of KLT-processes with AMIGO-tuned
filtering is not used, i.e. Tf = n = 0, we get PID controller and different filter time-constants Tf are cal-
culated. Kp can be omitted from the analysis, since it is can-
1 + jωT 1 celled by the controller. Fig. 3 shows the jitter margin for
δ max < γ + → 0. (13)
e − jω L
(ηω 2
+ jκω + λ ) jω ω →∞ various values of the process parameters as a function of Tf.
This means that the jitter margin is zero for a KLT- The batches of curves in Fig. 3 represent the jitter margin
process with PID controller without measurement filtering. for certain value of delay (here L = 2, 6 and 10 are used for
The reason for this is the fact that the system does not have the example). Each curve in one batch corresponds to certain
roll-off at high frequencies. This can be verified from the value of process time-constant, T. From the figure it is seen
closed-loop system’s Bode diagram, which corresponds to that the jitter margin is not greatly affected by the process
the left side of inequality (5). In Fig. 2, a KLT-process with time-constant for the relevant range of (small values of) Tf.
parameters Kp = 1, T = 3 and L = 1 in conjunction with an Another remark is that the jitter margin seems to have a cer-
AMIGO-tuned PID controller with and without the meas- tain maximum value in the range where Tf is relatively
urement filter are considered. The solid line represents the small. A closer look at the curves reveals that the value of Tf
system’s frequency response amplitude in the case without maximizing the jitter margin is nearly independent of proc-
the filter. The dotted line intersecting the solid line repre- ess time-constant T (see the circled positions in the figure).
sents the jitter margin criterion (right side of (5)) with Further analysis shows that the optimal value of Tf is linearly
δmax = 0.5. Obviously, the jitter margin criterion is not satis- dependent on L and that a very simple tuning rule for Tf
fied in this case. For comparison, the dash-dotted line shows could be derived. Using numerous values for L and T in the
the frequency response amplitude for the system with the range [0.1 10] and parameter estimation, the optimal value
measurement filter that is tuned according to the AMIGO of the measurement filter time-constant was identified as
rules. For the system with the filter, the jitter margin re- T f* ≈ 0.17 L . (14)
quirement is clearly satisfied. Although this value of Tf differs only slightly from the
The above analysis verifies the need for the measurement AMIGO tuning rule (Tf = 0.1⋅L for τ > 0.2), it has a signifi-
filter in varying time-delay systems. In addition, the filter cant effect on the jitter margin. Whereas the jitter margin for
naturally reduces the effect of measurement noise and it is AMIGO tuning is approximately 0.71⋅L, (14) gives jitter
often considered a prerequisite for using the derivative part margins 1.16⋅L…1.25⋅L depending on the value of T. Thus,
of the PID controller in practical control systems. Still it an increase of up to 75 % in jitter margin could be achieved
remains unclear if the AMIGO tuning rule for Tf is optimal by this simple modification for the pure KLT-process.
in jitter-sense.
δmax with respect to Tf for KLT+AMIGO with T = [0.1...10] the mentioned constraint. Thus the design approach needs to
14
L=2
be revised. We propose a method where the measurement
12
L=6 filter is first designed based on the required jitter margin,
L = 10
and then the plant is extended with the filter before ap-
10 proximating the KLT-model of the extended plant,
Pext(s) = Gf(s)P(s). The PID controller parameters are calcu-
8 lated using the extended plant’s KLT-approximation and the
δmax

standard AMIGO rules except for the filter time-constant

6
(which is already in the system).
Consider the KLT-process model with a PID controller.
4
As shown before, the jitter margin is hard to solve analyti-
2
cally even for this simple process model case. Nevertheless,
using well-justified approximations the analysis is signifi-
0 cantly eased. At high frequencies the following approxima-
0 5 10 15 20 25 30
Tf tions are valid.
Fig. 3. The jitter margin of KLT-process with AMIGO tuning as a function P( jω )C ( jω )
of measurement filter time-constant Tf. ≈ P ( jω )C ( jω )
1 + P ( jω )C ( jω )
This tuning rule was tested with the same process test (15)
batch that was used in the derivation of AMIGO rules (see Kp − jω L ⎛ ki ⎞ K p kd
= e ⎜k + + kd jω ⎟ → .
[4] and [5] for details). The batch consists of 133 processes 1 + jωT ⎝ jω ⎠ ω →∞ T
of nine different types. The jitter margin and the ITAE per-
formance criterion were compared to see how the change of Using the AMIGO tuning rule for kd in (15) results in
filter tuning rule affects robustness and performance. For the K p kd 0.1L + 0.225T 1
= ≈ . (16)
pure KLT-processes and those that are “close” to KLT in the T 0.3L + T 3
test batch, the tuning rule (14) gives excellent results, i.e. As seen in Fig. 2 and in (16), at high frequencies the fre-
better jitter margin without sacrificing the performance. For quency response amplitude of the KLT-model with PID con-
other processes the jitter margin remains almost equal to the troller without the filter remains approximately constant. If
values provided by the AMIGO tuning. The performance of the system has the filter, it dominates the system behavior at
the control systems remains at approximately equal level to high frequencies. In order to make the analysis general, we
AMIGO. Since this very simple modification of the AMIGO let the measurement filter order n vary and consider cases
tuning rules does not guarantee better jitter margin and per- n = 1 and n > 1 separately. Based on (16) the jitter margin of
formance for all processes, we next consider a complete the system with an nth order measurement filter becomes
measurement filter re-design to reach these objectives.
(1 + jωT )
n
f
δ max <3 . (17)
IV. EXTENDED PLANT DESIGN APPROACH jω
This section discusses a new way of designing and tuning
For n = 1, the above expression becomes
the filter for the AMIGO tuning rules. The main idea in this
approach is to allow the desired jitter margin to be an input 1 + jωT f 1
δ max < 3 = 3 T f2 + 2 . (18)
to the tuning rules, which enables adjusting the controller jω ω
performance according to the requirements posed by the The lower bound is obtained by minimizing the right side of
control system. For instance, this would be very useful for the inequality with respect to frequency. It is clearly seen
networked control systems, where often good estimates of that maximizing ω minimizes (18). Solving (18) for Tf at
the maximum delay can be given. The question is how to ω* = ∞ then gives Tf > 1/3δmax. Since small values of Tf are
tune the controller such that the criterion (5) is satisfied for preferred, we finally set the tuning rule as
the specified jitter margin. Simple tuning rules for which the 1
maximum delay can be directly assigned would be desirable T f = δ max . (19)
3
in many practical cases. In this section we propose such tun-
Similarly, for n > 1 we have
ing schemes.
3 (1 + ω 2T f2 )
n 2
In AMIGO tuning the process is approximated by the
KLT-model. In the previous section we concluded that the δ max < . (20)
ω
maximum achievable jitter margin for a pure KLT-process Minimizing (20) with respect to ω gives
with AMIGO tuning using (14) gives δmax ≈ 1.2⋅L with rea-
1
sonably small values of the filter time-constant. Especially ω* = , (21)
for lag dominated processes, for which T >> L, the require- Tf n −1
ment for the jitter margin might be considerably bigger than and the jitter margin becomes
( n −1) 2 Growth percentage of the jitter margin compared to AMIGO
⎛ n ⎞
δ max > 3 n ⎜ ⎟ Tf . (22) 300

⎝ n −1 ⎠
n=1
250 n=2
Solving (22) for Tf leads to the tuning rule n=3
(1− n ) 2

Growth of δmax w.r.t AMIGO in %

200
1 ⎛ n ⎞
Tf = ⎜ ⎟ δ max . (23)
3 n ⎝ n −1 ⎠ 150

Hence we have the design rules for the measurement filter 100
with n ≥ 1. To arrive at a complete PID-tuning, we propose
to first design the measurement filter (of desired order) us- 50

ing the rules above, and then include the filter dynamics in 0
P1 P2 P3 P4 P5 P 6 P7 P8 P 9
the plant and apply the AMIGO tuning on the extended plant
to calculate the remaining controller parameters (k, ki, kd, b -50

and c). The previously designed measurement filter is then

-100
used in the controller. The filter order may be restricted by 0 20 40 60 80 100 120 140
Process number
the process or measurement noise properties. If a high-order
Fig. 4. The jitter margin growth percentage compared to AMIGO tuning in
filter is used, losses in performance are presumable. In the the test batch.
next section, the proposed filter design guidelines are com- ITAE criterion compared to AMIGO in percentages
pared in performance and in jitter margin. 200
n=1
n=2
V. COMPARISON OF THE NEW RULES 150
n=3

ITAE cost criterion w.r.t AMIGO in %

This section compares the tuning rules presented above. We
restrict to cases where n = 1, 2, 3 to avoid unnecessary high- 100
order controllers. The rules use the extended plant approach
where the required jitter margin can be given as an input
parameter for the design. The proposed design approaches 50

are compared with respect to AMIGO tuning in Fig. 4 and

Fig. 5. We use the AMIGO process test batch and restrict to 0
P1 P2 P3 P4 P5 P 6 P7 P8 P 9
stable processes (integrating processes P6 are omitted). The
ITAE criterion is used as the performance measure through-
-50
out the study. Fig. 4 shows the relative improvement in jitter 0 20 40 60 80 100 120 140
Process number
margin compared to AMIGO tuning for the test batch. For
the extended plant approaches the objective value for the Fig. 5. The relative ITAE cost criterion with respect to the AMIGO tuning
in the test batch.
jitter margin is twice as large as the jitter margin of the
AMIGO tuned process, i.e. the dots in Fig. 4 should all be VI. DELAY-DOMINANT PROCESSES
aligned at 100 %. Fig. 5 shows the relative ITAE cost crite-
rion values of the new tuning rules with respect to those of The results in Fig. 4 indicate that the extended plant ap-
AMIGO tuning. Here the value 0 % corresponds to the per- proach might have some fundamental limitations for delay-
formance achieved with AMIGO tuning, and positive values dominant processes. The performance of the tuning rules is
indicate performance losses when applying the new rules. modest for processes for which T < L, where the variables
For most of the processes the extended plant approaches refer to the plant’s KLT-approximation parameters.
give promising results, and the improvements in jitter mar- The above presented problems gave reasons for investi-
gins are close to 100 %. Problems arise for processes with gating the extended plant approach generally for delay-
T < L, i.e. delay-dominant processes. These processes are dominant processes. The behavior of the jitter margin with
further studied in Section VI. As far as the jitter margin is respect to the filter time-constant was examined using the
considered, the third-order filter appears to give the most extended plant approach similarly as was done in Fig. 3 for
accurate results for the test batch, although the ITAE crite- the standard AMIGO tuning. This study was done with
rion increases the most for the third-order filter. In general, KLT-parameters in the ranges T = [0.02…1] and L = [1, 2].
the ITAE increases the most for slow processes with large T. Fig. 6 shows an example of the results with parameters
The increase of ITAE seems in some cases quite significant, T = 0.5 and L = 1 for n = 1, 2, 3, but consistent behavior was
but this might also stem from the time-weighting in the crite- observed for all parameter combinations tested.
rion: if the performance of a process with large T is de- As seen in Fig. 6, for n = 1, the jitter margin has an upper
creased in order to achieve bigger jitter margin, the settling bound and increasing Tf does not increase the jitter margin
time of the process increases. The longer the settling time, infinitely. The upper bound is independent of L at least
the more ITAE criterion punishes for error. Since the time is when T << L. Based on the study with a range of parameter
integrated, its effect is not linear, but rather squared. values, the upper bound for the jitter margin was identified
Jitter margin of the KLT-process, T << L, extended plant approach Growth percentage of the jitter margin compared to AMIGO
25 300
n=1 n=2
n=2 n=3
n=3 250
20

Growth of δmax w.r.t AMIGO in %

200
15
δmax

150

10
100

5
50

P1 P2 P3 P4 P5 P 6 P7 P8 P 9
0 0
0 2 4 6 8 10 12 14 16 18 20 0 20 40 60 80 100 120 140
Tf Process number

Fig. 6. The jitter margin of a delay-dominant KLT-process with extended Fig. 7. Modified extended plant approach. Growth of the jitter margin.
plant approach.
to be max(δmax,n=1) ≈ 4T, e.g. in Fig. 6 it is two seconds. We VII. SIMULATION RESULTS
can conclude that for delay-dominant processes, the first-
In this section we demonstrate the use of the proposed tun-
order filter and extended plant approach, will return very
ing rules and compare the performance and robustness to
small jitter margins compared to L (δmax ≤ 4T << L). This
varying time-delays between AMIGO and proposed tuning.
was clearly seen in Fig. 4 with process type P1. For small
We experiment the rules with both delay-dominant and lag
values of T (the values on the left) the jitter margin is very
dominated processes. The process models used in the simu-
small compared to AMIGO tuning.
lations are
For the second and third-order filters the jitter margin
1
grows linearly with Tf excluding the range where Tf is very Pexp,1 ( s ) = e− s ,
small. Thus the dependency of the jitter margin on the filter 0.3s + 1
1 (26)
time-constant may be discovered and this allows developing Pexp,2 ( s ) = .
( s + 1) ⎡⎣( 0.1s ) + 0.14s + 1⎤⎦
2
tuning rules for Tf based on the desired jitter margin. For
n = 2, the linear part of the curve in Fig. 6 was identified as
The former is a pure KLT-process and the latter is the
T f = 2.3δ max − 3L, T f ≥ 0.5 . (24)
first one of P9-type processes in the test batch. First, the
Since the jitter margin decreases rapidly in the range processes’ KLT-approximations are derived. The KLT-
0 < Tf < 0.5, the lower bound of Tf in (24) is justified. The parameters for the first process are the true process parame-
parameter fitting of (24) was based on a large number of ters (Kp = 1, T = 0.3 and L = 1) the process being delay-
process parameter combinations, not only on Fig. 6. dominant. For the other process, the KLT-parameters are
Similarly for n = 3, the relationship between filter time- Kp = 1, T = 1 and L = 0.137 the process being lag dominated.
constant, jitter margin and KLT-parameters was found to be For both processes the standard AMIGO tuning rules, and
T f = δ max − L − T , T f ≥ 0.3 . (25) for the lag dominated process the proposed tuning rules with
The tuning rules presented in this section are only used n = 1, 2, 3 were calculated. For the delay-dominant process,
for delay-dominant processes and the tuning rules (19) and only the second- and third-order filter designs were applied,
(23) for lag dominated processes (T ≥ L). In all cases, the because as explained in Section VI, the first-order filter is
filter is first designed and the extended plant design is then not applicable for delay-dominant processes. For the new
applied. If this proposed strategy is applied on the process rules, the design parameter δmax = 2δmax,AMIGO was chosen. In
test batch, the jitter margins are improved, as seen in Fig. 7. the simulations, a varying time-delay (square-wave form) of
The figure shows the results for measurement filter orders maximum amplitude 1.9δmax,AMIGO was added after the true
two and three. The first-order filter case is omitted, because plant and different controllers were experimented in this
it is not well suited for delay-dominant systems, as shown in setting. The unit step responses of the systems with various
Fig. 6. The modifications of the tuning rules for delay- controllers can be seen in Fig. 8 (delay-dominant) and Fig. 9
dominant processes are well justified and seem to give de- (lag dominated). The resulted controller parameters are
sired results. Fig. 7 should be compared with Fig. 4 that pre- shown in Table 1 and Table 2, respectively.
sents the same results without the above modifications. The The step responses of the delay-dominant process (Fig. 8)
high jitter margins for the delay-dominant processes come at show that all the systems are stable for the specific realiza-
some loss of performance, although for certain processes the tion of the varying time-delay, although the AMIGO-tuned
performance is almost equal to the AMIGO tuning. system shows signs of being close to unstable. According to
(5), the AMIGO-tuned system is not guaranteed to be stable
TABLE 1. CONTROLLER PARAMETERS FOR THE KLT-PROCESS. Pexp,2, ITAE: AMIGO: ∞, n = 1: 2.3232, n = 2: 2.0157, n = 3: 2.1043
1.5
Controller δmax k ki kd Tf ITAE
y ref
AMIGO 0.65 0.33 0.54 0.08 0.1 13.98 AMIGO
Ext. n = 2 1.33 0.56 0.43 0.27 0.5 12.17 n=1
Ext. n = 3 1.28 0.48 0.40 0.23 0.3 10.83 n=2
n=3
TABLE 2. CONTROLLER PARAMETERS FOR THE P9-PROCESS. 1
Controller δmax k ki kd Tf ITAE

Amplitude
AMIGO 0.21 3.47 7.04 0.23 0.014 ∞
Ext. n = 1 0.39 2.22 3.44 0.24 0.14 2.32
Ext. n = 2 0.42 2.00 3.04 0.24 0.071 2.02 0.5
Ext. n = 3 0.44 1.83 2.69 0.24 0.054 2.10
with the used amplitude of time-delay, but this does not
mean that the system would necessarily be unstable. The
controllers tuned using the extended plant approach are 0
0 5 10 15 20 25 30
guaranteed to be stable for this case, and they also give bet- Time (s)
ter ITAE values for the response. Fig. 9. Step responses for the P9-type process with varying time-delays.
The lag dominated system (Fig. 9) with standard AMIGO The paper suggested an extended plant approach to com-
controller is unstable, but the extended plant approach gives plement the AMIGO tuning for varying time-delay systems.
good responses. Of the extended plant designs the controller It was seen that the proposed approach results in jitter-aware
with the first-order filter gives the largest overshoot. The and well-performing controllers. The measurement filter is
controllers with second- and third-order filters give accu- first designed with the proposed tuning rules to meet a cer-
rately the desired jitter margins (as seen in Table 2), but the tain jitter margin objective. The extended plant with the fil-
requirement is not quite satisfied with the first-order filter. ter is then approximated by the simple KLT-model, and
Nevertheless, for both processes the proposed tuning rules AMIGO tuning gives the rest of the controller parameters.
give excellent results with respect to both robustness to jitter The order of the measurement filter can be chosen based on
and performance, compared to the standard AMIGO design. requirements. The third-order filter gives the most accurate
results with respect to the required jitter margin, but might
VIII. CONCLUSIONS also cause the largest degradations in closed-loop system
The paper considered the problem of tuning the PID control- performance. In addition, for all the designs, there is an ob-
ler in varying time-delay systems. The work is partly moti- vious trade-off between high jitter margin and performance.
vated by the increasing need for tuning schemes that account
for variable time-delays that appear, for example, in net- REFERENCES
worked control systems. Since the PID has many appealing [1] P. Naghshtabrizi and J. P. Hespanha, “Designing an observer-based
properties, simplicity being one of the most important, this controller for a networked control system”, In Proc. 44th Conf. on De-
cision and Control, Dec. 2005.
study concentrated on deriving new tuning rules for the con-
[2] L. M. Eriksson and M. Johansson, “PID controller tuning rules for
troller with the emphasis on jitter margin, nominal perform- varying time-delay systems,” In Proc. 2007 American Control Con-
ance and simplicity of the rules. ference, New York, USA, Jul. 2007.
Pexp,1, ITAE: AMIGO: 13.9799, n = 2: 12.1728, n = 3: 10.8327 [3] L. Eriksson and T. Oksanen, “PID controller tuning for integrating
1.5 processes: analysis and new design approach,” in Proc. Fourth Inter-
y ref national Symposium on Mechatronics and Applications (ISMA07),
AMIGO Sharjah, UAE, Mar. 2007.
n=2 [4] K. J. Åström and T. Hägglund, “Revisiting the Ziegler-Nichols step
n=3
response method for PID control,” Journal of Process Control, vol.
1 14, pp. 635-650, Sep. 2004.
[5] K. J. Åström and T. Hägglund, “Advanced PID Control,” ISA-The
Amplitude

Instrumentation, Systems and Automation Society, 2006.

[6] K. Gu, V. L. Kharitonov, and J. Chen, “Stability of Time-Delay Sys-
tems,” Birkhäuser, 2003.
0.5 [7] C.-Y. Kao and B. Lincoln, “Simple stability criteria for systems with
time-varying delays,” Automatica, vol. 40, pp. 1429-1434, Aug. 2004.
[8] J. G. Ziegler and N. B. Nichols, “Optimum settings for automatic
controllers,” Trans. ASME, vol. 64, pp. 759-768, 1942.
[9] K. J. Åström, H. Panagopoulos, and T. Hägglund, “Design of PI con-
0 trollers based on non-convex optimization,” Automatica, vol. 34, pp.
0 5 10 15 20 25 30 585-601, May 1998.
Time (s)

Fig. 8. Step responses for the KLT-process with varying time-delay.