Journal of Process Control 21 (2011) 840–851

Contents lists available at ScienceDirect

Journal of Process Control

journal homepage: www.elsevier.com/locate/jprocont

PI and PID auto-tuning procedure based on simplified single parameter

optimization夽
Julio Ariel Romero ∗ , Roberto Sanchis, Pedro Balaguer
Departament d’Enginyeria de Sistemes Industrials i Disseny, Universitat Jaume I, Campus de Riu Sec, 12071, Castelló, Spain

a r t i c l e i n f o a b s t r a c t

Article history: In this paper a new auto-tuning algorithm for PI and PID controllers based on relay experiments is
Received 7 October 2010 proposed to minimize the load disturbance integral error (IE) by maximizing the integral gain, sub-
Received in revised form 24 February 2011 ject to a desired phase margin, and a minimum required gain margin constraint. The main advantage
Accepted 10 April 2011
of the proposed auto-tuning algorithm with respect to previous works is that it leads, for most of the
Available online 20 May 2011
processes, to PID tuning with close loop performance similar to PID designed using off-line numerical
optimization. Moreover the algorithm is applicable to any linear model structure, including dead time
Keywords:
and non-minimum phase systems.
PID
PI © 2011 Elsevier Ltd. All rights reserved.
Auto-tuning
Relay experiment

1. Introduction maximizing the integral gain, subject to a minimum required phase

Nowadays most of the commercially available PID controllers
include auto-tuning capabilities to simplify the task of determining
the controller parameters. In general auto-tuning methods can be
classified in two groups: model-based and relay feedback methods.
In the former, the tuning is based on simple parametric models of
the processes which are obtained from the input/output response.
On the other hand, the methods based on relay feedback typically
use the information of only one point of the system frequency
response, normally where the phase angle is − radians.
The advantages of the relay feedback auto-tuning over model-
based algorithms can be summarized as follows: (1) During the
experiment the process is under close loop control, therefore the
process does not drift away from the nominal operating point. (2)
It identifies process frequency response around the ultimate fre-
quency (where the phase angle is − radians) and no previous
information about the process model is necessary to get a success-
ful tuning of controllers. Due to this advantages, many authors have
been focusing in the improvement of the original Åström and Häg-
glund method presented in [14]. In this paper a new auto-tuning
algorithm for PI and PID controllers based on relay experiments
is proposed to minimize the load disturbance integral error (IE) by
夽 This work has been supported by the Fundación Caixa Castelló-Bancaixa and the
University Jaume I through the research project P11A2010-16.
∗ Corresponding author. Fax: +34 964 728170.
E-mail address: romeroj@esid.uji.es (J.A. Romero).
margin, and a minimum required gain margin constraints.
Regarding the design of PID controller minimizing load distur-
bance, this can be performed either off-line or on-line. The problem
of off-line tuning of PID controllers to minimize load disturbance
has been widely addressed in the literature. There are some previ-
ous works based on the maximization of the integral gain, among
these, in [15,8] a direct numerical optimization is proposed, subject
to a given value of the maximum of the sensitivity function, while
in [16] an approximation of the process to a first-order plus time
delay (FOPDT) model is proposed to derive simple tuning rules.
Things are different with auto-tuning of PID controllers. Most of
the articles propose design methods to fulfil robustness conditions
and/or desired closed-loop bandwidth, see for example [5,6,2]. Only
few articles deal with the load disturbance minimization, and the
results are quite limited. One of the main causes of this is that the
calculation of PID parameters involve a multi-parametric optimiza-
tion problem with constraints, which can not be easily solved in
computers with limited resources, such as the case of embedded
controllers where auto-tuning algorithms are implemented. This
hurdle was overcome partially in [13] by obtaining formulae for
PID and PI parameter calculation to minimize the integral square
time error (ISTE) index, using as input data the critical frequency
and critical gain measured from a relay feedback experiment. The
formulae, however, were obtained for a FOPDT plant model, so
the behavior of the process must be approximated by those sim-
ple models, and this is not always accurate, leading to a wrong
performance.
Recently, in [9], the authors presented a simple procedure to
off-line tuning PID controllers through maximizing the integral

0959-1524/$ – see front matter © 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jprocont.2011.04.003
 J.A. Romero et al. / Journal of Process Control 21 (2011) 840–851 841

While the transfer function of the PID controller with high fre-
quency filter in the derivative term is assumed to be
 Td s 1
 (1 + (s/zi ))(1 + (s/zd ))
C(s) = Kp 1 + + = Ki
1 + (Td s/N  ) Ti s s(1 + (s/zd N))
(3)

where the controller zeros have been assumed to be real and

Fig. 1. SISO PID control loop. z + z 1

d i
Kp = Ki − (4a)
zd zi zd N
gain, subject to a minimum required phase margin, and a mini- 1 1
mum required gain margin constraints. The method is based on Td = − (4b)
zd + zi − (zi /N) zd N
the use of a single tuning parameter, defined as the quotient
between the final crossover frequency and the zero of the con- zd + zi 1
Ti = − (4c)
troller. The main advantage of this method respect to previous zd zi zd N
works is that, even though the maximization of the controller zd N
gain is straightforward, since there is only one parameter to be N = −1 (4d)
zd + zi − (zi /N)
tuned, for most of the processes, the solution is close to the opti-
mal tuning obtained with direct numerical optimization methods, The process is assumed to be linear, or, if it is non linear, it is
as in [15,8]. Moreover the method is applicable to any linear assumed to stay in the vicinity of an operating point where it can
model structure, including dead time and non-minimum phase be approximated by a linear model. The model of the process in
systems. that operating point is assumed to be unknown.
The auto-tuning algorithm proposed in this paper is based in The objective of the PID controller to be calculated by the auto-
[9], and permits, from relay experiments, the tuning of PI and PID matic tuning algorithm is to reach an adequate performance and
controllers minimizing the IE with guaranteed minimum gain and robustness of the controlled system, expressed in the following
phase margins. An ad hoc algorithm is proposed to efficiently solve design specifications:
the involved optimization problem taking into account some prop-
erties of the tuning method that are presented in this paper. The • Required phase margin: m,r . The phase margin (m ) should be
procedure performs relay experiments to estimate two or, if nec- larger than or equal to this value (m ≥ m,r ).
essary, three points of the systems frequency response. The points • Minimum required gain margin:  m,r . The gain margin ( m )
are used to build a straight line approximation of the frequency should be larger than or equal to this value ( m ≥  m,r ).
response. It is shown that, despite of the simplicity of the model, • Good disturbance rejection. The step disturbance IAE should be
it is accurate enough for the interesting frequencies for control. In as low as possible.
fact the IE obtained is very similar to the one calculated using the
full model off-line numerical optimization as proposed in [9,15,8]. The proposed design methods for PI and PID controllers follow
The procedure has been tested for the model batch presented the same strategy and, therefore, a unified description of the two
in [15,8] showing its applicability in all cases. Moreover, an experi- controllers will be developed.
mental case study shows the application of the proposed approach
to a thermal system. The overall algorithm is simple enough to be
3. Properties of tuning method
implemented in a low cost embedded system.
The structure of the paper is as follows. In Section 2, the problem
As aforementioned in Section 1, the auto-tuning algorithm pro-
is stated. Section 3 shows some features of the tuning method pro-
posed in this paper is based upon the tuning method presented
posed by the authors in [9], which will be considered when deriving
by the authors in [9]. In this section new properties of the tuning
the auto-tuning algorithm. In Section 4, the auto-tuning procedure
method are presented, which will be considered when deriving the
is detailed, including the set of experiments to be performed and
auto-tuning algorithm. First, the tuning method is shortly recalled
the algorithm to automatically calculate the controller parameters.
and then the properties will be presented.
In Section 5, some simulated examples over a well known batch
of processes illustrate the validity of the approach for different
3.1. PI and PID tuning method
process dynamics, while an experimental example show the prac-
tical application of the procedure in Section 6. Finally, the main
The tuning method presented by the authors in [9] can be
conclusions are summarized.
summarized as a simplified constrained optimization approach in
which the proposed PID parameters are those that maximize the
2. Problem statement controller gain Ki , subject to the following constraints: m ≥ m,r ,
 m ≥  m,r , zd = zi . The required information about the process is the
The SISO PID control loop considered in this paper is shown in frequency response data, as a list of points defined by frequency,
Fig. 1, where yr is the reference signal to be tracked, u is the control magnitude and phase. The controller zeros are imposed to be equal
action, y is the controlled output, and d is an input disturbance. in order to derive a very simple tuning procedure, however the lost
The transfer function of the PI controller is assumed to be of performance due to this restriction is not important for most
 1
 1 + (s/zi )
processes.
C(s) = Kp 1 + = Ki (1) The key to solve the previous constraint maximization in a
Ti s s straightforward way is to introduce a tuning parameter, that is the
relation between the final gain crossover frequency of the process
where
with controller and the controller zero
Ki 1 ωcg
Kp = ; Ti = (2) a= ; z c = zi , z d (5)
zi zi zc
842 J.A. Romero et al. / Journal of Process Control 21 (2011) 840–851

where ωcg is the frequency where the magnitude of the open loop
frequency response is 1, i.e. where the phase margin is measured.
For a given value of this parameter a, the phase of the controller
at the final crossover frequency ωcg depends only on that value, as
shown by the following equations.
3.2. Method properties: shapes of functions Ki (a) and  m (a)
As aforementioned, the tuning method looks for the value of
parameter a that maximize Ki (a) fulfilling the robustness condi-
tions  m (a) ≥  m,r . In this section the functions Ki (a) and  m (a) are


1 studied and some of their properties are revealed which will be
PI : arg(C(jωcg )) = arctan(a) − = − arctan (6) used in the auto-tuning algorithm. For this study, the tuning pro-
2 a
cedure is applied to design controllers for the following test batch
a  of models proposed in [15,8]. For the PID controllers the test batch
PID : arg(C(jωcg )) = 2 arctan(a) − arctan − (7)
N 2 of models is:
1
Therefore, for a given value of parameter a, the calculation of G1 (s) = (12)
the controller is automatic, following two steps: s(s + 1)3

e−5s
1. The final crossover frequency (ωcg ) is obtained as the frequency G2 (s) = (13)
(s + 1)3
where the phase of the system fulfills the phase margin equation,
i.e. where the phase of the system is 1
G3 (s) = (14)
(s + 1)(1 + 0.2s)(1 + 0.22 s)(1 + 0.23 s)
arg(G(jωcg )) = − + m,r − arg(C(jωcg )) (8)
1
G4,5,6,7 (s) = ˛ = 4, 5, 6, 7 (15)
Then, the value of the zero (zi = zd ) is calculated as (s + 1)˛

ωcg −2(s − 2)
zi = zd = G8 (s) = (16)
a (s + 1)3
In the case of PI controllers the considered models are:
2. The value of Ki is then calculated from the condition of unit
magnitude at the final crossover frequency 1
G9 (s) = (17)
(s + 1)3
|C(jωcg )G(jωcg )| = 1 → Ki (9)
1
G10 (s) = (18)
The resulting equations are: (s + 1)(1 + 0.2s)(1 + 0.22 s)(1 + 0.23 s)
ωcg e−15s
PI : Ki (a) =  (10) G11 (s) = (19)
|G(jωcg )| 1 + a2 (s + 1)3
 1
ωcg 1 + (a2 /N 2 ) G12 (s) = (20)
PID : Ki (a) = (11) s(s + 1)2
|G(jωcg )|(1 + a2 )
1 − 2s
3. Once the controller parameters are defined, the final gain margin, G13 (s) = (21)
 m , is calculated. (s + 1)3
9
G14 (s) = (22)
Applying the previous steps for different values of a, two graphs (s + 1)(s2 + 2s + 9)
relating a with Ki and  m can be easily built, and the value of a
These models capture the typical dynamics encountered in con-
where Ki reaches a maximum value, constrained to  m ≥  m,r can
trol applications: multiple poles, different poles, integrator, time
be obtained straightforwardly. Once the values of Ki and zc are
delay, non minimum phase, therefore the conclusions of the next
obtained, the parameters of the controller (Kp , Td and Ti ) are cal-
study may be assumed to apply for most of the industrial process
culated applying Eqs. (2) or (4).
models.
Remark 1. In the case of the PID controller, the two zeros are The shapes of functions Ki (a) and  m (a) obtained when design-
imposed to be equal, i.e. zi = zd , in order to simplify the design ing PID and PI controllers for models in the test batch are shown in
method. This restriction could result in a significant loss of perfor- Figs. 2 and 3 respectively. It can be noted that in all cases Ki (a) is a
mance for some particular processes. However, in [4] it is proved smooth function with only one maximum in aKmax . In most cases,
that for most of the industrial processes dynamics this limitation the function  m (a) has also a maximum which is always located
produce a minor deterioration of the closed loop performance com- in values of a smaller than aKmax , more precisely, the maximum
pared with the optimal design obtained with independently fixed of  m (a) lies in the interval [0, 0.5]. Therefore, the function  m (a)
zeros. For the batch of models proposed in [15,8], that are rep- increases as a decreases from aKmax until the maximum of  m is
resentative of most of the industrial processes dynamics, the loss reached.
of performance due to this restriction is not important. Interested The features of functions Ki (a) and  m (a) pointed out in this
readers are referenced to [9], where a detailed comparison between section will be used to develop an ad hoc algorithm to automatically
the above described method and the MIGO (M-constrained inte- solve the optimization problem involved in the design procedure.
gral gain optimization) methods presented in [15,8] for PID and PI
controllers respectively, was made. The MIGO methods are based 4. Auto-tuning algorithm
on a pure numerical optimization approach where the zeros are
independently fixed, leading to complex zeros most of the times. In this section the auto-tuning algorithm which automati-
Despite the restriction in the zeros placement, both methods lead cally tunes the PI and PID controllers according to the procedure
to controllers with similar closed loop performance, measured in described in Section 3 is presented. Since the simplified optimiza-
terms of disturbance IAE, when similar robustness conditions are tion tuning procedure needs the frequency response of the process,
fixed. and this is unknown, a straight line approximated model of the
 J.A. Romero et al. / Journal of Process Control 21 (2011) 840–851 843

0.04 40 0.2 4
0.03 G 30 0.15 G 3
1 2
0.02 20 0.1 2
0.01 10 0.05 1
0 0 0 0
0 1 2 3 4 5 0 1 2 3 4 5
a a
20 500 0.5 10
G3 0.4 G 8
4
0.3 6
0.2 4
0.1 2
0 0 0 0
0 1 2 3 4 5 0 1 2 3 4 5
a a
0.5 5 0.5 5
G5 G6

0 0 0 0
0 1 2 3 4 5 0 1 2 3 4 5
a a
0.4 4 0.4 4
0.3 G 3 0.3 G8 3
7
0.2 2 0.2 2
0.1 1 0.1 1
0 0 0 0
0 1 2 3 4 5 0 1 2 3 4 5
a a

Fig. 2. Controller gain Ki (solid lines, left axis) and gain margin  m (dashed lines, right axis) variations as a function of parameter a (N = 10, m,r = 60◦ ). PID design.

frequency response is proposed to be obtained by estimating two
points of the Nyquist plot. The first point should be near the critical
point (where the phase of the process is around −180◦ ), and the
second point at a lower frequency. These points can be obtained by
performing two modified relay feedback experiments. In [11] some
alternatives can be found. For example, one of them consists of
adding a delay in the loop (the phase of the estimated point depends
on the delay). Once two points of the frequency response have been
obtained (M1 ej1 at frequency ω1 and M2 ej2 at frequency ω2 ), the
magnitude and phase are approximated by the straight line defined

1 10 3 60
2.5 50
G 2 40
9 G10
0.5 5 1.5 30
1 20
0.5 10
0 0 0 0
0 1 2 3 4 5 0 −3
1 2 3 4 5
a x 10 a
0.06 4 6 300
0.05 G11 3.5 5 250
0.04 3 4 G12 200
0.03 2.5 3 150
0.02 2 2 100
0.01 1.5 1 50
0 1 0 0
0 1 2 3 4 5 0 5 10
a a
0.2 4 1.5 6
G G
13 14
1 4
0.1 2
0.5 2

0 0 0 0
0 1 2 3 4 5 0 1 2 3 4 5
a a

Fig. 3. Controller gain Ki (solid lines, left axis) and gain margin  m (dashed lines, right axis) variations as a function of parameter a (N = 10, m,r = 60◦ ). PI design.
844 J.A. Romero et al. / Journal of Process Control 21 (2011) 840–851

by those points in a logarithmic scale, i.e.,

10

|G(jω)|db = c − d log(ω) (23) 5

0
with
−5
20 log(M1 /M2 )
d= ; c = 20 log(M1 ) + d log(ω1 ) 100
log(ω2 /ω1 )

and −100

arg(G(jω)) = e − f log(ω) (24)

−200
with
10−0.5 10−0.4 10−0.3 10−0.2 10−0.1
1 − 2
f = ; e = 1 + f log(ω1 ) Fig. 4. Bode diagram of G8 (s) and the approximation by one or two straight lines.
log(ω2 /ω1 )
The circles are the points obtained by the relay feedback experiments.
With this straight line approximated frequency response model,
the PID is tuned automatically using the procedure described in 4. If log (ω2 /ωgt1 ) > rlog (ω1 /ω2 ), perform a third relay feedback
ω
Section 3. If the resulting gain crossover frequency lies somehow (2 −1 ) log ω 1
near the estimated points, the error of the straight line approxi- experiment with a delay L3 =  ω1 . The result is a fre-
gt1
ωgt1 log ω
mation is assumed to be low and the tuning concludes. If, on the 2
quency ω3 , and amplitude A3 . The third point of the frequency
other hand, the resulting gain crossover frequency lies far away
response is defined by frequency
ω3
, magnitude M3 = A3 /4
from the estimated points, the error is assumed to be large, and   2
a new relay experiment is performed (with a different delay) to 4 h
and phase 3 = L3 ω3 −  − arg A3
1− A3
− j Ah .
obtain a third point of the Nyquist plot. This third point should be 3

near the gain crossover frequency of the first design (ωgt1 ), because
5. Design the PID controller using the approximation by the two
then the error in the final phase margin due to the straight line
straight lines, with specifications m,r and  m,r , and finish the
approximation will be lower. The three points then define a two
auto tuning process.
straight lines approximate frequency response model that is used
again to tune automatically the PID using the procedure described
The most adequate value of parameters  and r depend on the
in Section 3. Two parameters must be fixed to define the auto tun-
type of controller (PI or PID). For PID controller, adequate values
ing procedure. The first one is the phase  of the process where the
are  = − 130◦ , r = 1, while for PI controller, adequate values are
second point should be identified. The second parameter is the one
 = − 100◦ , r = 0.5. These numerical values have been decided by
that defines the condition to perform or not the third experiment.
a heuristic approach using the test batch of models proposed in
This parameter can be defined as the maximum relative distance
[15,8], that represent the most common dynamics encountered in
between the gain crossover frequency of the first design, ωgt1 , and
industrial processes. Decreasing the parameter r implies increas-
the interval of frequencies defined by the two known points. If this
ing the probability of requiring a third experiment, leading to an
parameter is defined as r, the third point should be calculated only
improved accuracy, but at the cost of a longer experiment phase.
if
On the other hand, increasing the parameter r implies decreasing
  ω  the probability of requiring a third experiment, reducing the length
ω2 1
log > r log . (25) of the experiment phase, but at the cost of a less accurate result.
ωgt1 ω2

To summarize, the auto tuning procedure is then defined by the 4.1. Examples
following steps, including the equations that result from using the
relay feedback with delay experiments: As an example, consider the system
1 − 2s
G8 (s) =
1. Perform a relay feedback experiment without delay, with ampli- (s + 1)3
tude  and hysteresis h. The result is a frequency ω1 , and
Fig. 4 shows the Bode diagram of the system, and the approxi-
amplitude A1 . The first point of the frequency response is defined
mation by one or two straight lines, as a result of the relay feedback
by frequency
  ω1 , magnitude M 1 = A1 /4, and phase 1 = − −
 experiments, with  = 1 and h = 0.01. If the third experiment is not
4 2
arg A1
1 − ( Ah ) − j Ah (here, the describing function taken into account, the straight line approximation defined by the
1 1
two points has an significant error especially outside the inter-
of the relay with hysteresis has been used). val defined by those points. Using the point obtained with the
2. Perform a second relay feedback experiment with a delay third experiment, the approximation by two straight lines is more
L2 = ((− 1 + )1 )/ω1 , where  is the phase of the desired accurate. Fig. 5 shows the disturbance response of the PID con-
point. The result is a frequency ω2 , and amplitude A2 . The trollers designed with the two points approximation and with the
second point of the frequency response is defined by fre- three points approximation, compared to the response of the PID
quency
 ω2, magnitude M2 = A 2 /4 and phase 2 = L2 ω2 −  −
 designed with the full frequency response data, with specifications
arg 4 2
1 − ( Ah ) − j Ah . m ≥ 60◦ ,  m ≥ 2. The use of the third point improves the distur-
A2 2 2 bance response, making it closer to the true response. The cost is,
3. Tune the PID controller using the straight line approximation, of course, the longer experiment phase. In this case, the third exper-
with specifications m,r and  m,r , and calculate the gain crossover iment would not be performed if a parameter r > 1.3 is taken. The
frequency ωgt1 . If log (ω2 /ωgt1 ) < rlog (ω1 /ω2 ) the auto tuning performance of the controller tuned after only two experiments is
ends. slightly worse but still reasonable.
 J.A. Romero et al. / Journal of Process Control 21 (2011) 840–851 845

4.2. Controller tuning using straight lines approximated model

1
Once an approximate frequency response model is available
0.8
(one straight line or two straight lines approximation), the tuning
0.6 of the controller consists of calculating the optimum value of a∗ by
Amplitude

solving the following optimization problem

0.4
argmax0<a<amax Ki (a, m ) (26a)
0.2
m ≥ m,r (26b)
0
m (a, m ) ≥ m,r (26c)
−0.2
The shapes of functions Ki (a) for a given m were investigated in
−0.4
Section 3.2. It was shown that Ki (a) is a smooth function with only
0 5 10 15 20 25 30 one maximum in aKmax . When a one straight line approximation
time (sec) of the frequency response is used as process model in the tun-
ing procedure described in Section 3, the value of a that maximize
Fig. 5. Disturbance response of system G8 (s) with PID controllers. (−) Full data. (− −)
Ki (a) does not depend on m , and can be explicitly calculated using
Two points approximation. (· · ·) Three points approximation.
Eqs. (27) and (28) for PID and PI design respectively (see proof in
Appendix A).
   
p2 3 3
0 aKmax =− + R+ Q3 + R2 + R− Q 3 + R2 (27)
3
−2 where
3p1 − p22
−4 Q =
10 −1
9

p2 p1 p0 p3
0 R= − − 2
6 2 27
−100
1
 d

−200 p2 = (N − 2) 1+ ln(10)
f 20
10 −1
p1 = (2N 2 − 1)
Fig. 6. Bode diagram of G2 (s) and the approximation by one or two straight lines.
The circles are the points obtained by the relay feedback experiments.
1
 d

p0 = (N − 2N 2 ) 1+ ln(10)
f 20
1
ln(10)
 d

0.8 aKmax = ˇ ln(10) = 1+ (28)
f 20
0.6
The result of these equations is the starting point of an algorithm
0.4 developed to solve (26), that is shown in Fig. 8. In the first step of the
Amplitude

0.2 algorithm, Eq. (27) for PID or (28) for PI is used to calculate the value
a1 = aKmax where Ki (a) has a maximum. A PID or PI controller (PID1 )
0 is designed for this value of a applying the two steps described in
−0.2 Section 3, with m = m,r and using the straight lines approximated
process model. The gain margin  m is also estimated using that
−0.4
model. If the gain margin restriction is fulfilled (condition mi >
−0.6 m,r is true), then the controller PID1 is the optimal one and the
algorithm stops, otherwise the algorithm looks for the value of a
−0.8
0 10 20 30 40 50 60 70 80 90 100 which fulfills the gain margin restriction. Taking into account that
time (sec) according to Section 3.2,  m (a) increases as a decreases from aKmax ,
the value of a∗ is expected to be at the left of a1 . Therefore a is
Fig. 7. Disturbance response of system G2 (s) with PID controllers. (−) Full data. (− −) decreased by a in each iteration of the algorithm until the first
Two points approximation. (· · ·) Three points approximation.
value of a fulfilling the gain margin restriction is found (ai such that
mi > m,r ). Then the value of a∗ is calculated by interpolating the
last two points, i.e., as the intersection point of straight lines defined
Figs. 6 and 7 show the Bode diagram and the disturbance by the points (ai−1 , mi−1 ), (ai , mi ) and (ai−1 ,  m,r ), (ai ,  m,r ), Eq.
response for the same analysis of previous example applied to the (29). The value of a is defined as a = a1 /imax , where imax is the
system G2 (s) = 1 3 e−5s . In this case the third experiment would maximum number of iterations to cover the interval ]0, a1 [.
(s+1)
not be performed if a parameter r > 1.7 is taken, but in this case,
(mi − m,r )ai−1 − (mi−1 − m,r )ai
the behaviour of the controller tuned after only two experiments is a∗ = (29)
mi − mi−1
much worse due to the high approximation error, especially in the
phase (that results in a much lower phase margin (m = 17◦ ) than While the value of a is decreased, in a steps, looking for the gain
desired). margin condition ( m (ai ) >  m,r ), two events may occur that must
846 J.A. Romero et al. / Journal of Process Control 21 (2011) 840–851

be taken into account. Both events imply that there is no solution

that fulfils the robustness condition  m (a) >  m,r . The first one is
that a value ai < 0 is reached. The second one is that  m,i <  m,i−1 ,
since this implies that the maximum of  m (a) has been reached
and it is not large enough. If one of these conditions is true, then a
more robust design is attempted by increasing m in 5◦ , till there
exists a solution for which  m (a) >  m,r . The maximum value of m
for which designs are attempted is proposed to be 70◦ . When this
value is exceeded, the algorithm stops.
In the case that the approximate model is defined by two straight
lines, the starting point defined by Eq. (27) or (28), should be cal-
culated using the straight line defined by the second and third
estimated points, because the gain crossover frequency lies in that
section of the approximate model.

4.3. Configuration parameters of the auto-tuning algorithm

The proposed autotuning algorithm has been developed to be

implemented in a stand alone PID controller with low computing
resources. The previous algorithm defined to solve the optimiza-
tion problem seems complex, but in fact requires a low computing
capacity, as will be proved in the next section (it can be easily
implemented in a low cost embedded microcontroller).
In order to completely define the algorithm to be implemented,
the following parameters must be selected:

• : the phase of the second point of the frequency response to be

estimated.
• r: the parameter that defines the condition to calculate the third
point.
• m,r : the minimum required phase margin.
•  m,r : the minimum required gain margin.
• : the amplitude of the relay.
• h: the hysteresis of the relay. It should be slightly higher than the
measurement noise amplitude to avoid wrong switchings of the
relay.
• N: the parameter of the PID high frequency filter.
• b: the weighting factor for the proportional term of the reference
signal. This parameter does not affect the disturbance response,
but only the overshoot of the step reference response. Therefore,
strictly speaking, it is not a parameter needed for the autotuning,
but a value should be assigned for the controller implementation.

If the algorithm is implemented in a commercial PID controller,

some of those parameters should be fixed by the PID manufacturer,
and only a few of them should be left to be selected by the final
PID user. Several alternatives can be considered. The simplest one
from the user’s perspective could be the manufacturer to fix all
the parameters. A possible set of parameters could be  = − 130◦ ,
r = 1, m,r = 60◦ ,  m,r = 2,  = 40%, h = 5%, N = 10 and b = 0.7. Another
possibility could be to define a set of default values, but to let the
user modify some of them before the autotuning is performed (for
example, , N or b). With respect de phase and gain margin, the
most appropriate solution could be to let the user select between
2 or 3 configurations, for example: higher robustness and slower
response (m,r = 70◦ ,  m,r = 3), lower robustness and faster response
(m,r = 50◦ ,  m,r = 2).

5. Simulation study

In order to verify the feasibility of the auto-tuning algorithm

Fig. 8. Algorithm to solve the optimization problem on Eq. (26).
proposed in the previous section, a simulation study is developed
using the models in the test batch presented in Section 3.2. For
these models, the auto-tuning algorithm is used to calculate the
controllers parameters for the following values of gain and phase
 J.A. Romero et al. / Journal of Process Control 21 (2011) 840–851 847

5 1

9
1

y,G
y,G

0 0

−5 −1
0 50 100 150 200 250 0 50 100 150
1
0.5
2
y,G

y,G10
0
−1
0 50 100 150 200 250 300 350 −0.5
0 5 10 15 20 25
0.2
2
3
y,G

11
0

y,G
−0.2
0 5 10 15 20 25 −2
1 0 100 200 300 400 500 600 700
4
y,G

0 10

y,G12
−1 0
0 50 100 150 200 250
1 −10
0 50 100 150 200 250
y,G5

0
2

13
−1 0

y,G
0 50 100 150 200 250
1
−2
0 50 100 150 200 250
6
y,G

0
1
−1
0 50 100 150 200 250
14

0
y,G

1
y,G7

0 −1
0 10 20 30 40 50
time(sec)
−1
0 50 100 150 200 250
2 Fig. 10. Relay experiments and disturbance step responses for the close-loop sys-
tems with models G9 –G14 and PI controllers.
y,G8

−2
0 50 100 150 200
time(sec)
Fig. 9. Relay experiments and disturbance step responses for the close-loop systems
with models G1 –G8 and PID controllers.
margins: m,r = 60◦ ,  m,r = 2. A relay with amplitude  = 1 and hys-
teresis h = 0.001 is used. The results of simulations for PID and PI
are presented in Figs. 9 and 10, respectively.
The oscillations at the beginning of simulations correspond to
relay experiments carried out in order to estimate points in the sys-
tems frequency responses. The vertical lines in the oscillation zones
indicate the beginning and end of experiments, each experiment
allowing the identification of one point in the frequency response.
As can be noted, for PID designs (Fig. 9) most of the models
require two experiments: only models G2 and G8 need the iden-
tification of three points of the frequency response. Model G2 has a
large time delay in comparison with the time constants, whereas G8
is a non-minimum phase system. In the case of PI designs (Fig. 10),
models with time delay (G11 ), integrator (G12 ) and non-minimum
phase (G13 ) require three experiments. For the rest of models two
experiments are enough to tune the controllers. It is worth to note
that the number of experiments is linked with the parameter r in
Eq. (25). For PID and PI designs empirical values of r = 1 and r = 0.5
have been used. These values, however, can be modified in order
to obtain more exact models or to reduce the maximum number of
experiments to two, as suggested in Section 4.
After the experiments, in Figs. 9 and 10, once the controllers are
tuned with the proposed frequency response autotuning method
(i.e. FR AUTO), the disturbance step response is obtained for the test
250 batch. The IAE and the sensitivity peak Ms obtained are presented
in Tables 1–3. These values are compared first with the tuning pro-
cedure using the full transfer function information as described in
[9], and secondly with the classical Ziegler–Nichols tuning rules,
described in [14] and the tuning method proposed in [13], which
will be referred as Zhuang–Atherton method.
Firstly, as can be seen in Tables 1–3, the IAE and the Ms obtained
by the FR AUTO are very similar to the method FR presented in [9],
despite the fact that FR method uses the full exact model and the
FR AUTO algorithm performs relay experiments to obtain only two
or, at most three, points in the frequency response. As a result it
can be concluded that the autotuning approach is able to capture
the interesting frequencies for control.
Secondly, the proposed FR AUTO method is compared with
the Ziegler–Nichols and the Zhuang–Atherton methods. The
Zhuang–Atherton method is derived for FOPDT models, so no result
can be provided for integrating models G1 and G12 . Moreover,
unstable tunings are noted by an ∞ IAE. In general, the values of IAE
obtained with the proposed auto-tuning algorithm are smaller than
those obtained with Ziegler–Nichols and Zhuang–Atherton meth-
ods, while the robustness indexes, Ms , are kept very similar. The
only exception is the Zhuang–Atherton method for model G8 which
provides an IAE slightly smaller than the one achieved with the FR
AUTO, however at the cost of reducing the robustness. The same
happens with Ziegler–Nichols PI design for G12 . In this case the
IAE obtained is smaller but again at the cost of a lower robustness
(higher Ms ).
In most of the cases, the algorithm found the solution for
m = m,r . Only in three cases (models G8 , G13 and G14 ), marked
848 J.A. Romero et al. / Journal of Process Control 21 (2011) 840–851

Table 1
Comparison of the results reached by the proposed PID auto-tuning method and
the ones in [14] (Ziegler–Nichols method) and [13] (Zhuang–Atherton method). FR
AUTO: frequency response relay experiment method, FR: frequency response model
based method.
Table 3
Comparison of the results reached by the proposed PI auto-tuning method and
the ones in [14] (Ziegler–Nichols method) and [13] (Zhuang–Atherton method). FR
AUTO: frequency response relay experiment method, FR: frequency response model
based method.

Gn Design IAE Ms Gn Design IAE Ms

G1 FR (m,r =45,  m,r =2) 21.44 1.93 G9 FR (m,r =45,  m,r =2) 1.72 1.93
FR (m,r =60,  m,r =2) 40.84 1.70 FR (m,r =60,  m,r =2) 2.20 1.63
FR AUTO(m,r =45,  m,r =2) 24.16 1.88 FR AUTO(m,r =45,  m,r =2) 1.68 1.80
FR AUTO(m,r =60,  m,r =2) 46.25 1.62 FR AUTO(m,r =60,  m,r =2) 2.71 1.52
Ziegler–Nichols 24.55 1.59 Ziegler–Nichols 11.97 1.26
Zhuang–Atherton – – Zhuang–Atherton 5.76 1.36

G2 FR (m,r =45,  m,r =2) 12.83 2.19 G10 FR (m,r =45,  m,r =2) 0.21 1.67
FR (m,r =60,  m,r =2) 8.55 2.04 FR (m,r =60,  m,r =2) 0.41 1.40
FR AUTO(m,r =45,  m,r =2) 9.13 2.27 FR AUTO(m,r =45,  m,r =2) 0.25 1.59
FR AUTO(m,r =60,  m,r =2) 8.59 1.10 FR AUTO(m,r =60,  m,r =2) 0.54 1.33
Ziegler–Nichols ∞ – Ziegler–Nichols 4.00 1.10
Zhuang–Atherton ∞ – Zhuang–Atherton 2.02 1.14

G3 FR (m,r =45,  m,r =2) 0.04 1.70 G11 FR (m,r =45,  m,r =2) 37.01 2.21
FR (m,r =60,  m,r =2) 0.09 1.46 FR (m,r =60,  m,r =2) 25.02 2.05
FR AUTO(m,r =45,  m,r =2) 0.04 1.68 FR AUTO(m,r =45,  m,r =2) 29.19 2.61
FR AUTO(m,r =60,  m,r =2) 0.09 1.45 FR AUTO(m,r =60,  m,r =2) 26.89 2.58
Ziegler–Nichols 1.70 1.21 Ziegler–Nichols 68.15 1.71
Zhuang–Atherton 1.00 1.29 Zhuang–Atherton 38.77 1.43

G4 FR (m,r =45,  m,r =2) 1.62 2.08 G12 FR (m,r =45,  m,r =2) 49.91 1.62
FR (m,r =60,  m,r =2) 2.17 1.75 FR (m,r =60,  m,r =2) 179.47 1.32
FR AUTO(m,r =45,  m,r =2) 1.62 2.09 FR AUTO(m,r =45,  m,r =2) 48.91 1.57
FR AUTO(m,r =60,  m,r =2) 2.25 1.75 FR AUTO(m,r =60,  m,r =2) 115.07 1.33
Ziegler–Nichols 7.56 1.67 Ziegler–Nichols 21.10 2.17
Zhuang–Atherton 4.35 1.95 Zhuang–Atherton ∞ –

G13 FR (m,r =45,  m,r =2) a 9.10 2.23

FR (m,r =60,  m,r =2) 6.85 2.09
Table 2
FR AUTO(m,r =45,  m,r =2) a 8.63 2.03
Comparison of the results reached by the proposed PID auto-tuning method and
FR AUTO(m,r =60,  m,r =2) 7.58 1.93
the ones in [14] (Ziegler–Nichols method) and [13] (Zhuang–Atherton method). FR
Ziegler–Nichols 24.64 1.49
AUTO: frequency response relay experiment method, FR: frequency response model
Zhuang–Atherton 7.49 1.81
based method.
G14 FR (m,r =45,  m,r =2) a 0.96 2.15
Gn Design IAE Ms
FR (m,r =60,  m,r =2) 0.90 2.07
G5 FR (m,r =45,  m,r =2) 2.67 2.35 FR AUTO(m,r =45,  m,r =2) a 1.00 2.36
FR (m,r =60,  m,r =2) 3.12 2.00 FR AUTO(m,r =60,  m,r =2) 0.95 2.37
FR AUTO(m,r =45,  m,r =2) 2.85 2.26 Ziegler–Nichols 4.66 1.56
FR AUTO(m,r =60,  m,r =2) 3.15 1.94 Zhuang–Atherton 2.14 1.76
Ziegler–Nichols 10.22 2.10 a
Cases explained in the text.
Zhuang–Atherton 5.83 2.84

G6 FR (m,r =45,  m,r =2) 4.45 2.31

FR (m,r =60,  m,r =2) 4.08 2.18
FR AUTO(m,r =45,  m,r =2) 4.79 2.33 with large time delays [10,3,12]. In addition to relay experiments,
FR AUTO(m,r =60,  m,r =2) 4.03 2.10 Phase-Locked Loop based methods are also available to estimate
Ziegler–Nichols 13.02 2.83 points of the system frequency response with a desired system
Zhuang–Atherton 7.45 4.40
phase angle [1]. The auto-tuning algorithm proposed in this paper
G7 FR (m,r =45,  m,r =2) 6.32 2.27 could be used with any of these estimation methods.
FR (m,r =60,  m,r =2) 8.50 1.55
FR AUTO(m,r =45,  m,r =2) 8.28 2.46
FR AUTO(m,r =60,  m,r =2) 5.40 2.15
5.1. Computation load study
Ziegler–Nichols 15.86 4.05
Zhuang–Atherton 10.24 7.93
In order to evaluate the implementation feasibility of the pro-
G8 FR (m,r =45,  m,r =2) a 7.16 2.26
posed algorithm for real-time execution in a low cost embedded
FR (m,r =60,  m,r =2) 6.08 2.12
FR AUTO(m,r =45,  m,r =2) a 6.98 1.97 system, the required computation load is calculated in terms of
FR AUTO(m,r =60,  m,r =2) 6.34 1.92 floating point operations per second (FLOPS). The procedure has
Ziegler–Nichols 8.72 1.79 been to measure the number of FLOPS required by the algorithm,
Zhuang–Atherton 5.71 2.12 implemented as a Matlab function, for all the batch cases previ-
a
Cases explained in the text. ously considered (with m,r = 45◦ ,  m,r = 2). The Lightspeed Matlab
toolbox [7] has been used in order to obtain accurate results. The
results obtained are summarized in Table 4. Assuming a standard
with asterisk (*) in Tables 2 and 3, m had to be increased because
architecture (floating point microcontroller or DSP) with 20 Mega-
no solution for m = 45◦ ,  m ≥ 2 could be found. In the three cases,
FLOPS, it can be seen from Table 4 that the computation time
the algorithm found the solution for a final value of m = 55◦ .
required allows the implementation of the algorithm on standard
Remark 2. In this simulation study, relay experiments as pro- low cost embedded devices. Moreover it should also be stressed
posed in [11] are used to estimate points of the system frequency that the proposed algorithm does not require the use of look-
response. There are some other versions of relay experiments up tables, and therefore, no especial memory requirements are
which improve the accuracy of estimations, especially for systems needed.
 J.A. Romero et al. / Journal of Process Control 21 (2011) 840–851 849

Table 4 68
Computation load of the algorithm (evaluated in number of FLOPS) when applied
to the test batch of models. The column TIME contain the execution time when the 66
experiment experiment
algorithm is implemented in a hardware architecture with 20 Mega-FLOPS. 64 1 2
PID PI 62

y
Models FLOPS TIME (␮s) Models FLOPS TIME (␮s) 60

G1 3635 181.8 G9 1410 70.5 58

G2 3079 153.8 G10 2225 111.3 56
G3 3635 181.8 G11 1657 82.8 0 500 1000 1500 2000 2500 3000 3500 4000 4500
G4 1892 94.5 G12 1657 82.8 t
G5 2141 107 G13 4971 248.6 10
G6 2390 119.5 G14 5253 262.6
G7 3841 192 8
G8 9237 461.4
6

u
4

0
0 500 1000 1500 2000 2500 3000 3500 4000 4500
t

Fig. 12. Output and input of the thermal process during auto-tuning and control
experiment.

|G(jω1 ) | = 0.1122. The second point was obtained introducing a

delay L2 = 33 s in the loop, resulting in an oscillation A = 0.2026 and
Tosc = 432 s (experiment 2 in Fig. 12), leading to ω2 = 0.0145 rad/s,
arg (G(jω2 )) = − 138.2◦ and |G(jω2 ) | = 0.159.

Fig. 11. Picture of the thermal process used for the experiment.
7. Conclusions

6. Experimental case
This section illustrates the applicability of the proposed PID auto
tuning method to a real experimental case. The process to be con-
trolled is a thermal system, composed by a heater resistor attached
to an iron plate inside a plastic cover. A switched amplifier drives
the resistor, while a thermocouple with an amplifier is used to mea-
sure the temperature. The system is completed by a small fan that
can be used to apply a disturbance to the plant. The auto tuning
algorithm and the PID controller are implemented by a PC with a
data acquisition card using Matlab Real Time Windows Target soft-
ware. A simple Euler discretization with period T = 1 second and
a parallel realization of the controller with anti windup is used. A
picture of the complete thermal system (without the plastic cover)
is presented in Fig. 11.
Fig. 12 shows the input and output of the process during the
auto tuning phase and during the control phase for the follow-
ing values of the parameters that define the algorithm: m,r = 60◦ ,
 m,r = 2,  = − 130◦ , r = 1,  = 1, h = 0.05, N = 10, b = 1. An output ref-
erence yref = 60◦ C is selected to perform the experiment. As the
required input for that output value is unknown, an estimated value
of uav = 3 V is used. Taking into account the relay amplitude  = 1,
the relay switched between umax = 4 V and umin = 2 V. This approxi-
mate selection of the input needed for the desired output reference
led to an asymmetrical switching, as can be seen in the figure. How-
ever, this imperfection in the relay switching symmetry was not a
problem, as the behaviour of the tuned controller was still correct.
In order to test the controller behaviour under reference and distur-
bance changes, the reference was increased 6◦ C at instant t = 2600 s,
and the fan was connected at instant t = 3400 s.
In this case, only two points were needed to approxi-
mate the frequency response. The first point, obtained without
delay, corresponds to A = 0.143 and Tosc = 326 s (experiment 1 in
Fig. 12), leading to ω1 = 0.01927 rad/s, arg (G(jω1 )) = − 159.5◦ and
In this paper, a PID auto-tuning procedure has been developed.
It is based on the use of modified relay feedback experiments
(for example with added delay) to obtain two or three points of
the frequency response of the process, and the use of straight
line approximations of the Bode diagram in order to tune the PID
through a simplified single parameter optimization method previ-
ously developed by the authors.
The auto-tuning procedure implies a low computing cost, and
the performance of the resulting control loop is somehow near
the optimum that could be obtained if the full information of the
process model were available.
The proposed algorithm has been compared with other auto-
tuning approaches and with the optimum performance, using a
batch of models widely used in the literature. The conclusion is that
previous approaches are clearly improved, and the overall perfor-
mance obtained is quite similar to the one obtained by assuming
full information of the process model.
Finally, the algorithm has been tested on a real experimental
thermal process, showing that it works well with real measurement
noise and disturbances.
Appendix A. Study of Ki (a)
Assume that a straight line approximation of the frequency
response is used for the controller design as
|G(jω)|db = c − d log(ω) (30)
arg(G(jω)) = e − f log(ω) (31)
that can be written as
c−d log(ω)
G(jω) = 10 20 ej(e−f log(ω)) (32)
850 J.A. Romero et al. / Journal of Process Control 21 (2011) 840–851

Now, consider the case of the PI controller, whose frequency Applying the phase margin equation for a given value of a
response is
  
1 + (jω/zi ) K a arg(C(jωcg )) + arg(G(jωcg )) = − + 2 arctan(a)
C(jω) = Ki = i 1+ jω 2
jω jω ωcg a
− arctan + e − f log(ωcg ) = − + m,r
whose argument at the final crossover frequency ωcg is N

 one obtains the final cross over frequency for that value of a
arg(C(jωcg )) = − + arctan(a)  +2 arctan(a)−arctan a +e−
2 (N) m,r
2 ωcg = 10 f

Applying the phase margin equation for a given value of a

Now, applying the unit magnitude equation
(| G(jωcg ) ||C(jωcg ) | = 1) at that frequency, one obtains the controller
 gain Ki as a function of a
arg(C(jωcg )) + arg(G(jωcg )) = − + arctan(a) + e − f log(ωcg )
2
= − + m,r 
ωcg 1 + (a2 /N 2 )
Ki (a) =
one obtains the final cross over frequency for that value of a |G(jωcg )|(1 + a2 )
 +arctan(a)+e−
2 m,r 
ωcg = 10 f 1 + (a2 /N 2 ) a
= 10˛+2ˇ arctan(a)−ˇ arctan( N )
(1 + a2 )
Now, applying the unit magnitude equation
(| G(jωcg ) ||C(jωcg ) | = 1) at that frequency, one obtains the controller where again
gain Ki as a function of a
1
 d
  c
˛= 1+ + e − m,r −
f 20 2 20
ωcg
Ki (a) =  and
|G(jωcg )| 1 + a2
 
 1 d
1 1 d  +e− c ˇ= 1+
=  10 f
1+ 20 ( 2 m,r +arctan(a) ) − 20
f 20
1 + a2
It can be demonstrated that the curve Ki (a) has a unique max-
that can be written in a more compact form as imum for a > 0. For that purpose, calculating the derivative with
respect to a and equating to zero one obtains:
1
Ki (a) =  10˛+ˇ arctan(a)
1 + a2 a3 + (N − 2)ˇ ln(10)a2 + (2N 2 − 1)a + (N − 2N 2 )ˇ ln(10) = 0 (34)

where that is a degree 3 polynomial with a unique real root for a > 0. In
   order to demonstrate this, if a1 , a2 , a3 are the roots of the above
1 d c polynomial, the following well known relations between the roots
˛= 1+ + e − m,r −
f 20 2 20 and the coefficients of the polynomial are true

and (N − 2)ˇ ln(10) = −(a1 + a2 + a3 )

1
 d

ˇ= 1+
f 20 (2N 2 − 1) = a1 a2 + a2 a3 + a1 a3

The derivative of the above expression with respect to a can be

easily found to be (N − 2N 2 )ˇ ln(10) = −a1 a2 a3

dKi  10˛+ˇ arctan(a) In the case that d > 0, f > 0 and N > 2, the following inequalities
= ˇ ln(10) − a fulfil
da (1 + a2 )
3/2

−(a1 + a2 + a3 ) > 0
that has a unique root at

dKi ln(10)
 d

a1 a2 + a2 a3 + a1 a3 > 0
= 0 ⇒ aKmax = ˇ ln(10) = 1+ (33)
da f 20

The second derivative is easily shown to be negative at that −a1 a2 a3 < 0

point, and therefore, the Ki (a) curve has a unique maximum at the
If two roots are complex conjugate (for example a2 and a3 ), then
above value aopt = ˇln (10).
the real root (a1 ) must be positive, since a2 a3 > 0 and −a1 a2 a3 < 0.
Consider now the case of the PID controller, whose frequency
In the case that the three roots are real, they can not be all positive,
response is
since −(a1 + a2 + a3 ) > 0. They can not be all negative, nor one nega-
(1 + (jω/zi ))
2
K (1 + (a/ωcg )jω)
2 tive and two positive, since −a1 a2 a3 < 0. Therefore, one of the roots
C(jω) = Ki = i is positive and the other two are negative, or complex conjugate.
jω(1 + (jω/Nzi )) jω 1 + (a/Nωcg )jω
Furthermore, it can be easily shown that the derivative at a = 0 is
dKi 
whose argument at the final crossover frequency ωcg is positive ( da = 10˛ ln(10)ˇ 2 − 1
N
> 0), while the derivative

a a=0
for a = ∞ is zero. This implies that Ki (a) has a unique maximum
arg(C(jωcg )) = − + 2 arctan(a) − arctan
2 N for a > 0. The value of a where Ki is maximum is the only positive
real root of polynomial (34), that can be obtained by using well
 J.A. Romero et al. / Journal of Process Control 21 (2011) 840–851 851

known formulas for degree 3 polynomials. Defining the polynomial [3] I. Kaya, D.P. Atherton, Parameter estimation from relay autotuning with asym-
coefficients as metric limit cycle data, Journal of Process Control 11 (4) (2001) 429–439, 8.

1
 d
 [4] B. Krinstianson, B. Lennartson, Robust tuning of PI and PID controllers: using
derivative action despite sensor noise, IEEE Control Systems Magazine 26 (1)
p2 = (N − 2) 1+ ln(10) (2006) 55–69.
f 20
[5] A. Leva, PID autotuning algorithm based on relay feedback, Control Theory and
Applications 140 (5) (1993) 328–338, ID: 1.
p1 = (2N 2 − 1) [6] M. Ma, X. Zhu, A simple auto-tuner in frequency domain, Computers & Chemical
Engineering 30 (4) (2006) 581–586.

1
 d
 [7] T. Minka, Lightspeed matlab toolbox, 2002, http://research.microsoft.com/en-
us/um/people/minka/software/lightspeed/.
p0 = (N − 2N 2 ) 1+ ln(10) [8] H. Panagopoulos, K.J. Åström, T. Hägglund, Design of PID controllers based on
f 20
constrained optimization., IEE Proceedings of the Control Theory Application
the optimum value of a can be obtained as: 149 (1) (2002) 32–40.
[9] R. Sanchis, J.A. Romero, P. Balaguer, Tuning of PID controllers based on sim-
3p1 − p22 plified single parameter optimization, International Journal of Control 83 (9)
Q = (2010) 1785–1798.
9 [10] K.K. Tan, T.H. Lee, S. Huang, K.Y. Chua, R. Ferdous, Improved critical point esti-
mation using a preload relay, Journal of Process Control 6 (5) (2006) 445–
p2 p1 p0 p3 455.
R= − − 2 [11] K.K. Tan, T.H. Lee, R. Ferdous, PID control, new identification and design meth-
6 2 27 ods, Automatic PID Controller Tunning-The Nonparametric Approach, Springer,
   
2005, pp. 147–182.
[12] S. Vivek, M. Chidambaram, An improved relay auto tuning of PID controllers
p2 3 3
aKmax = − + R+ Q 3 + R2 + R− Q 3 + R2 (35) for unstable FOPTD systems, Computers & Chemical Engineering 29 (10) (2005)
3 2060–2068.
[13] M. Zhuang, P. Atherton, Automatic tuning of optimum PID controllers, IEE Pro-
References ceedings D 140 (3) (1993) 216–224.
[14] K.J. Åström, T. Hägglund, Automatic tuning of simple regulators with specifica-
tions on phase and amplitude margins, Automatica 20 (5) (1984) 645–651.
[1] J. Crowe, M.A. Jonhson, PID control, new identification and design methods,
[15] K.J. Åström, H. Panagopoulos, T. Hägglund, Design of PI controllers based on
Phase-Locked Loop Methods, Springer, 2005, pp. 213–256.
non-convex optimization, Automatica 34 (5) (1998) 585–601.
[2] W.K. Ho, Y. Hong, A. Hansson, H. Hjalmarsson, J.W. Deng, Relay auto-tuning
[16] K.L. Åström, T. Hägglund, Revisiting the Ziegler–Nichols step response method
of PID controllers using iterative feedback tuning, Automatica 39 (1) (2003)
for PID control, Journal of Process Control (14) (2004) 635–650.
149–157.