ISSN 0280-5316

ISRN LUTFD2/TFRT--5728--SE

Autotuning of a PID-controller

Camilla Andersson
Mirjam Lindberg

Department of Automatic Control

Lund Institute of Technology
October 2004
 Document name
Department of Automatic Control MASTER THESIS
Lund Institute of Technology Date of issue
Box 118 October 2004
SE-221 00 Lund Sweden Document Number
ISRNLUTFD2/TFRT--5728--SE
Author(s) Supervisor
Camilla Andersson and Mirjam Lindberg Mattias Grundelius TAC, Malmö
Tore Hägglund LTH, Lund

Sponsoring organization

Title and subtitle

Autotuning of a PID-controller. (Automatisk inställning av PID-regulatorer)

Abstract
This master´s thesis has been performed in cooperation with TAC in Malmö. The TAC
group makes commercial buildings smarter by integrating and automating the
technical systems required to run them. TAC:s control systems use PID-controllers
to control processes such as heating and ventilation. The PID-controllers are
often badly tuned, since it is too timeconsuming to calculate good PID-parameters
at the time of deployment. A simple way of finding PID-parameters that give
faster control loops is needed. To solve this problem the thesis proposes an
autotuner based on the areamethod Method of Moments and the AMIGO tuning rules.
The implementation of the autotuner using IEC 61131 is described. The resulting
autotuner is tested on simulated processes and gives satisfactory results. The
thesis also includes practical insights on the use of the autotuner.

Keywords

Classification system and/or index terms (if any)

Supplementary bibliographical information

ISSN and key title ISBN

0280-5316
Language Number of pages Recipient’s notes
English 38
Security classification

The report may be ordered from the Department of Automatic Control or borrowed through: University Library Box 3, SE-221 00 Lund, Sweden. Fax +46
46 222 42 43
Contents
1 Introduction 3
1.1 About TAC . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.1 The Concept of Autotuning . . . . . . . . . . . . . . . 3
1.2.2 Autotuning at TAC . . . . . . . . . . . . . . . . . . . 4
2 Methods 6
2.1 A Step Response Method . . . . . . . . . . . . . . . . . . . . 6
2.1.1 Process Identication: the Three-Parameter Model . . 6
2.1.2 Design Method . . . . . . . . . . . . . . . . . . . . . . 9
2.1.3 The Method of Moments in Practice . . . . . . . . . . 9
2.2 The Relay Feedback Method . . . . . . . . . . . . . . . . . . . 11
2.2.1 Process Identication . . . . . . . . . . . . . . . . . . . 11
2.2.2 Design Method . . . . . . . . . . . . . . . . . . . . . . 12
2.2.3 The Relay Feedback Method in Practice . . . . . . . . 13
2.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3 Implementation 15
3.1 IEC 61131 and PLC programming in CoDeSys . . . . . . . . 15
3.2 Overview of the Implementation . . . . . . . . . . . . . . . . 16
3.3 Detailed Implementation . . . . . . . . . . . . . . . . . . . . . 17
3.3.1 Initial Phase . . . . . . . . . . . . . . . . . . . . . . . 17
3.3.2 Closed Loop Phase . . . . . . . . . . . . . . . . . . . . 18
3.3.3 Open Loop Phase . . . . . . . . . . . . . . . . . . . . . 20
3.3.4 PID-design Phase . . . . . . . . . . . . . . . . . . . . . 20
3.3.5 Disturbance Detection . . . . . . . . . . . . . . . . . . 20
3.3.6 Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4 Test Results 24
4.1 Lag Dominated Process . . . . . . . . . . . . . . . . . . . . . 24
4.2 Delay Dominated Process . . . . . . . . . . . . . . . . . . . . 26
4.3 Balanced Process . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.4 Process with Noise and Dead Zone . . . . . . . . . . . . . . . 28
1
CONTENTS 2
4.5 Testing on an Air Handling Unit . . . . . . . . . . . . . . . . 30
5 Conclusions 31
5.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.2 User Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.3 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
A Figures 34
A.1 FBD  Overview of the Communication Parameters . . . . . 34
A.2 SFC  Overview of the Autotuner Algorithm . . . . . . . . . 35
Chapter 1

Introduction
1.1 About TAC
TAC designs and manufactures systems for Building IT. This covers func-
tions such as access control, heating and cooling, ventilation and lighting.
With TACs integrated systems, functions such as these can be controlled
and monitored from a computer terminal.
Heating, cooling and ventilation functions include controlling entities such
as temperature and airow. This means using feedback control loops.
A typical TAC product is the Xenta R 300. It is a programmable controller
with functionality including control loops, alarm handling and time control.
The minimum sample time is 1 second. When control loops are implemented,
dierent PID-controllers may be used. The autotuner is tested with the
PIDP-block, which uses a discrete time positional PID algorithm. The PIDP-
block has dead zone capability which means that when the control error is
smaller than the dead zone the controller output remains unchanged. The
size of the dead zone is related to the accuracy of the measuring instrument.
1.2 Background
1.2.1 The Concept of Autotuning
An autotuner is a device that automatically computes the parameters of a
controller. The goal is to achieve the best control possible given the tuning
objectives. The goal is not to replace a human control engineer. The auto-
tuner should rather be seen as an aid to improvement. Many PID-controllers
in the eld use the default parameter settings, since it is too time-consuming
3
CHAPTER 1. INTRODUCTION 4
to tune them manually and the systems seem to work anyway. However,
badly tuned controllers cannot only lead to less ecient control loops, they
can also lead to unpredictable system behavior when control loops are hier-
archically dependent. An autotuner can be of great help in spotting these
kinds of problems as well as remedying them [1].
It is important to distinguish between autotuning and adaptive control. An
autotuning operation is run on the user's initiative, and takes place during
a limited time. When autotuning has been performed the control loop re-
sumes normal operation with the new parameters. An adaptive controller
continually changes its parameters during operation.
An autotuner follows the same approach that a control engineer would when
tuning a controller:
1. Process Model  Choose a model for the process
2. Process Identication  Collect process data and t it to the model
3. Design Method  Compute the new parameters
The tuning objectives of the autotuner are process specic and must be
determined before implementation. Some objectives could be low overshoot,
fast setpoint tracking and good disturbance rejection. Sometimes trade-os
must be made. Another important feature of an autotuner is usability. If
the autotuner is dicult to use, it will probably not be used. The autotuner
should not require a lot of user input to achieve good tuning. In the ideal
case a start button should be enough.
1.2.2 Autotuning at TAC
Most of TACs control systems use PID-controllers. It is important that the
PID-parameters are correctly tuned to achieve the high energy conservation
made possible by TACs integrated systems. Today not much time is spent
netuning the controllers when a system is deployed. Usually "safe" param-
eters will be set which ensure that the system is stable. These parameters
mostly result in slow and inecient control. Calculating ecient parameters
manually is too timeconsuming. An autotuner will make the deployment
process run smoother and give TACs customers better results.
Autotuning has been used at TAC before. It was implemented in the early
90's and based on a patented method developed at the Dept. of Auto-
matic Control, Lund Institute of Technology. Together with the autotuner
a controller was implemented with an algorithm dierent from the standard
PID-algorithm. This autotuner does not only give the controller parameters
but also a recommended sample interval based on the process dynamic.
CHAPTER 1. INTRODUCTION 5
The new autotuner diers from the old autotuner since it is detached and
communicates with the controller through a network. The basic idea is
that the new autotuner will be able to communicate with any of the today
existing PID-controllers at TAC. It is also dierent in the way it identies
the unknown process.
Chapter 2

Methods
In this chapter two dierent process identication methods with correspond-
ing design methods are described. Based on MATLAB testing, one of the
methods is chosen for the autotuner.
2.1 A Step Response Method
2.1.1 Process Identication: the Three-Parameter Model
Using a step response, a process can be approximated by a three-parameter
model:
Gp =
Kp −Ls
1 + Ts
e (2.1)
where K is the static gain, T the time constant and L the time delay. The
response time T = L + T for stable systems can be measured as the time
p

when the step response has reached 63% of its steady state value, see Figure
63

2.1.

6
CHAPTER 2. METHODS 7

Figure 2.1: The response time T 63 = T + L.

There are several dierent ways to determine K , T and L; three of them

are described below.
p

Estimating the delay

One way is to estimate the delay L by measuring when the measurement

signal is starting to increase. From the response time T and the estimated
delay L, the time constant T can be calculated as T = T − L. The static
63

gain K is calculated as:

63
p

Kp =
yf inal − ybegin
uf inal − ubegin
(2.2)
where u is the control signal and y is the measurement signal
Ziegler-Nichols step response method

Another method is the where the delay

L can be estimated by using the maximum slope tangent of the step response.
Ziegler-Nichols step response method

Using this tangent it is possible to nd two points, one on the horizontal
axis and one on the vertical axis. The delay is the time from the start of the
step to the point where the tangent is crossing the horizontal axis. A new
parameter, a, is introduced as the dierence between the initial value and
the tangent crossing the vertical axis. From these two parameters together
with the static gain the relationship
a=
Kp L
T
(2.3)
CHAPTER 2. METHODS 8
can be used to nd T . K is calculated as in Equation 2.2.
p

Method of Moments

A third way to determine K , T and L is the described

in [2]. The area A is the integrated distance between the stationary end
p Method of Moments

level of y and the measurement signal y in an open loop step. The area
0

A /k , also called T , can be used as a good estimate of T for the three-

parameter model, see article [3]. The area A beneath the step response is
0 ar 63

also integrated. The areas are shown in Figure 2.2.

1

Figure 2.2: A step response in open loop with the areas used in Method of Mo-
ments.

The area A is used to calculate the time constant:

1

T =
eA1
hKp
(2.4)
and the area A is used to calculate the delay:
0

L = Tar − T =
A0
−
eA1
hKp hKp
(2.5)
h is the step amplitude of the control signal and K is the static gain from
Equation 2.2.
p

The approximated processes yielded by these three dierent methods were

simulated in MATLAB and were compared with the original process. The
conclusion was that the Method of Moments gives the best approximation.
CHAPTER 2. METHODS 9
2.1.2 Design Method
When the parameters K , T and L are known a design method can be chosen
to transform them into PI-parameters or PID-parameters. One such method
p

is the well-known Ziegler-Nichols' set of tuning rules. A modern method is

the AMIGO tuning rules proposed in [3]. It uses the following equations:
1
K = (0.2 + 0.45T /L)
Kp

Ti =
0.4L + 0.8T
L + 0.1T
L (2.6)
0.5LT
Td =
0.3L + T

Instead of only using two parameters, a and L like the Ziegler-Nichols tuning
rules, the AMIGO tuning rules are based on three parameters, K , T and
L. The AMIGO method has been developed with a robustness constraint
p

and it uses a dependency on the normalized dead time τ = L/(L + T ) unlike

the Ziegler-Nichols tuning rules. The AMIGO method suits a large variation
of processes and seems to be more universal than the Ziegler-Nichols tuning
rules.
The AMIGO design method recommends a setpoint weighting which depends
on τ , this is not used since the setpoint weighting is not changeable in the
PIDP-block. The PIDP-block has its own setpoint weighting denition: b =
0 for PI and PID-control and b = 1 for P and PD-control.

2.1.3 The Method of Moments in Practice

A variation of the Method of Moments described in [2], which is further
developed in [4] is used.
First a step change is applied in open or closed loop and T is determined
by integrating the area between the normalized control and measurement
ar

signals, see Figure 2.3.

tf tf
(2.7)
Z Z
A0 = Tar = u(t)dt − y(t)dt = L + T
tb tb

where t is the time at the beginning of the step change and t is the time
at the end of the step change. The stationary gain is used as an estimate of
b f

the process gain K .

p
CHAPTER 2. METHODS 10

Figure 2.3: A step response in closed loop with the normalized control and mea-
surement signal.

A second step change in opposite direction is then applied in open loop and
the area A of the step response is calculated during the time T . The time
constant T is then calculated as in Equation 2.4 and the delay L is equal to
1 ar

T − T.
ar

The Method of Moments can be executed in both open loop and closed loop.
A closed loop is known to take care of process disturbances but requires
some safe default values for the PID-parameters of the controller. When
using closed loop, one must be aware that a high gain can cause instability
problems when the process has long delays. With an open loop, it is dicult
to be sure that the step amplitude is appropriate. An unknown high gain
can lead to unforeseen consequences in the process output.
This area-method is easy to automate since calculation of the areas through
integration can be done online. It is also robust to high-frequency distur-
bances since erroneous measurements during the integration process should
cancel each other out due to the low pass nature of integration.
A lot of tests have been performed on dierent types of processes. The
major test objective was to see if the new controller parameters improved
the performance of the control loop. The range of tested processes includes
rst-order and higher-order processes as well as processes with small and
large time constants and delays.
The tests were performed in SIMULINK. The Method of Moments was au-
tomated using an S-function. The AMIGO design method mentioned above
in Equation 2.6 was used with the approximated parameters K , T and L. A
step response of each process controlled with safe default PI-parameters and
p

a step response of the same process controlled with the new PID-parameters
were compared. The new controller parameters give a fast step response
with an acceptably small overshoot.
CHAPTER 2. METHODS 11
2.2 The Relay Feedback Method

Figure 2.4: Relay feedback loop.

2.2.1 Process Identication
The relay feedback method focuses on identifying one point on the Nyquist
curve of the process. The relay feedback causes the process to enter a limit
cycle, i.e. a stable oscillation, see Figure 2.5. The frequency of the limit
cycle approximates the ultimate frequency, ω , which is where the process
has a phase lag of 180 . The process gain at the frequency ω is called the
u
◦

ultimate gain, K , and can be approximated by

u
u

Ku = |G(iωu )| =
πa
4d
(2.8)
where d is the relay amplitude and a is the limit cycle amplitude.

Figure 2.5: Relay feedback limit cycle.

An explanation for Equation 2.8 lies in describing function theory. A relay
is a static nonlinearity and can be described by the gain N (a) where a is
CHAPTER 2. METHODS 12
the input amplitude. N (a) is called the describing function, see [2]. The
condition for oscillation is given by N (a)G(iω) = −1 which means that
G(iω) = −1/N (a), see Figure 2.6.

Figure 2.6: N (a) is the describing function of the relay.

The describing function of the relay nonlinearity is N (a) = .
4d
πa

To avoid random switching during noisy conditions a relay with hysteresis

can be used. The signal must then be larger than the hysteresis width  to
cause a switch. The negative inverse of the describing function for this kind
of relay is somewhat more complicated:
−
1
N (a)
=−
πp 2
4d
a − 2 − i
π
4d
(2.9)
2.2.2 Design Method
In the article [6] this design method is recommended for most kinds of pro-
cesses. The controlled system is specied as:
G(iω)GP ID (iω) = 0.5e−i
135π
180 (2.10)
The PID-parameters must be chosen so that the original point on the Nyquist
curve is moved to this new point. The controller has to advance the process
phase by 45 since the original phase is −180 . The PID-parameters can be
◦ ◦

calculated as:
CHAPTER 2. METHODS 13

0.5 cos 45 ◦
K =
Ku
q
tan 45 ◦ +
(2.11)
4
α + tan2 45 ◦
Td =
2ωu

Ti = αTd

where α = 6.25 according to [6].

2.2.3 The Relay Feedback Method in Practice
The relay method as described above is easy to automate. The output from
the process is measured and when a stable limit cycle is detected its period
time and amplitude are measured and averaged over several cycles. The
period time gives the ultimate frequency ω and the amplitude a is used to
calculate K as in Equation 2.8.
u
u

The method has been tested on a range of processes similar to the ones
tested in part 2.1.3. The design method described in Equation 2.11 was used.
Tests have been performed using SIMULINK and the process identication
and design are both implemented in an S-function. A step response of each
process controlled with safe default PI-parameters and a step response of
the same process controlled with the new PID-parameters were compared.
The immediate conclusion is that the design method does not give very good
results for most of the tested processes. For processes including a substantial
delay the control is very slow and inecient and processes with large time
constants receive a very large K .
The process delay could be approximated during the experiment and an
appropriate design method chosen according to the result. In article [6] a
PI-design is recommended for processes with delays.
2.3 Discussion
The step response method, represented by the Method of Moments, and the
relay feedback method are both easy to automate. Positive aspects of the
Method of Moments are that it is robust against high frequency disturbances
and safe if the rst step change is made in a closed loop. If the rst step
change is made in open loop it means taking a risk with the system. The
relay feedback method is executed in a closed loop and is also safe for the
system.
CHAPTER 2. METHODS 14
The decision of which method to choose is based on the test results with the
new PID-parameters. The parameters generated with Method of Moments
give fast step responses with a small overshoot. The parameters generated
with the relay feedback method do not give so good step responses. Depend-
ing on the process characteristics the step responses are either very slow or
have large overshoots. To get better results, a step change could be added
to the relay feedback method to nd the static gain of the unknown process.
Then the PID-design could be based on three parameters instead of only
two, K and ω . The process delay L could also be approximated. Based
on L a decision could be made in the design phase on whether to use PID or
u u

PI-control. Since the Method of Moments together with the AMIGO design
already gives good results this will be the method used in the autotuner.
Chapter 3

Implementation
3.1 IEC 61131 and PLC programming in CoDeSys
The autotuner is implemented in the IEC 61131 standard. IEC 61131 has
been developed for PLC application programming in the automation indus-
try. It contains several languages as follows:
• Instruction List (IL)
• Structured Text (ST)
• Sequential Function Charts (SFC)
• Function Block Diagram (FBD)
• Ladder Diagram (LD)
The rst two languages, IL and ST, are text based. IL is an assembly-like
language while ST resembles Pascal. The other languages are graphical lan-
guages, which means that applications are described using dierent blocks
connected to each other. An IEC 61131 application can be created using
any combination of these languages. The concept of hierarchy is well devel-
oped which means that an application can be highly structured. Blocks can
contain applications written in any of the languages mentioned above.
The most specialised language in IEC 61131 is SFC, which is used to describe
sequential behaviour. This is very useful for control applications since they
are often time- and/or event-driven. A SFC application is built using steps
and transitions. The steps represent the states in the control ow. The
transitions are conditions that allow state change. A transition can only
occur if the step immediately before the transition is active. Concurrent as
well as alternative sequences of steps are allowed.

15
CHAPTER 3. IMPLEMENTATION 16
The IEC 61131 programming tool used for implementation of the autotuner
is called CoDeSys, Controller Development System, by Smart Software So-
lutions.
3.2 Overview of the Implementation

Figure 3.1: The phases in the autotuner experiment.

The autotuner algorithm is an implementation of the Method of Moments
described in section 2.1.3. The implementation is divided into four phases,
Initial phase, Closed Loop phase, Open Loop phase and PID-design phase:
1. Initial Phase  check noise levels, ensure stationarity
2. Closed Loop Phase  Calculate T ar

3. Open Loop Phase  Calculate A 1

4. PID-design Phase  Calculate new PID-parameters

The PID-design phase occurs after the experiment.
The autotuner communicates with the user by receiving the size of the set-
point change and the size of the deadzone, if one exists. When the tuning is
nished the new PID-parameters are displayed.
There is also communication with the process that is tuned. The measure-
ment and control signals from the process are received by the autotuner.
The autotuner sends the new setpoint to the process during the Closed Loop
phase and the slavecontrol signal during the Open Loop phase.
The phases of the algorithm are implemented in a Sequential Function Chart.
Communication is handled in a Function Block Diagram. These can be found
in the appendix.
CHAPTER 3. IMPLEMENTATION 17
3.3 Detailed Implementation
3.3.1 Initial Phase
The Initial phase contains four steps, Go, NoiseDetect, SteadyStateDetect
and Initialization, see Figure 3.2. During this phase the controller is active
using the existing parameters. The step Go is used to set some constants
and to have somewhere to jump back to when the algorithm is restarted.

Figure 3.2: The steps in the Initial phase.

The two following steps run in parallel. In NoiseDetect the noise level of
the measurement and the control signal is measured. The noise level is the
basis of the tolerance level that will be used throughout the experiment
to determine if the signal is stationary, has reached a certain level or has
experienced a load disturbance. The maximum and minimum values of the
signal are picked during several 15-second intervals and the average max
value, max, and min value, mın, is computed. The length of the intervals
poses a problem if the noise has periodic components. The noise level is
calculated as |max − mın|. After each interval a rough estimate of the
tolerance is calculated for use in the parallel step SteadyStateDetect. The
nal tolerance will be the noise level multiplied by 1.2 to ensure that it is not
too sensitive. If a dead zone is present the noise level will be multiplied by 1.5
instead. Since the noise level is measured while the system is stationary the
control signal's inuence will be lesser than when a setpoint change occurs
and thus the noise will appear lesser. When the tolerance has been calculated
the step height set by the user, i.e. the size of the setpoint change, is checked.
It should be at least 10 times larger than the tolerance. Otherwise a warning
is issued since the results of the tuning cannot be guaranteed.
The SteadyStateDetect step is necessary because the signal needs to be sta-
tionary while the noise level is detected to get accurate results. It tests that
CHAPTER 3. IMPLEMENTATION 18
the measurement signal is stationary for 60 seconds by checking if it stays at
the same level, which means that it does not change more than the tolerance
level allows. 60 seconds is an arbitrarily chosen time period which seems to
work. It means that the NoiseDetect step has four intervals to compute the
noise level average from. The measurement signal is averaged online and the
latest sample is compared with the average. If the dierence is larger than
the tolerance level for more than two samples in a row both the NoiseDe-
tect and SteadyStateDetect steps are restarted. Allowing divergence for two
samples takes care of sporadic outliers so that unnecessary restarts will not
occur too often.
The last step in the Initial phase is called Initialization. Its main purpose is
to determine good values for the initial levels of the control and measurement
signals. The initial levels will be used later in the experiment. The same
test as in SteadyStateDetect is performed to see if the signals are station-
ary, and the tolerance now has the value determined in NoiseDetect. When
stationarity has lasted for half the time compared to SteadyStateDetect, i.e.
30 seconds, the averages of the control and measurement signals are deemed
to be stable enough. If stationarity is interrupted, the procedure will be
repeated including the two previous steps.
3.3.2 Closed Loop Phase

Figure 3.3: The initial and end levels for the control and measurement signals.
A step change is made with the controller active using existing parameters.
The dead zone parameter, described in section 1.1, will contribute to a sta-
CHAPTER 3. IMPLEMENTATION 19
tionary level that diers slightly from the setpoint. The stationary level
depends on the controller PI-parameters, slow parameters get a lower level
and fast parameters get a higher level than the setpoint. For reaching the
setpoint with slow PI-parameters the dead zone is added to the step size
when computing the setpoint.
The step direction is set by the user in the beginning of the experiment with
the step size. A negative step size gives a closed step in downward direction
and a positive step size gives a closed step in upward direction.
The areas of the measurement and control signals are computed as Riemann-
sums, using the sample time.
The delay is measured as the time until the signal passes 1% of the step
change and is later used in the PID-design phase.
T , the time it takes to reach 63% of the step change, is measured and twice
this time gives a time-limit for how long the system should remain stationary
63

after reaching its new level. This time-limit should not be too short since
slow step responses risk being interrupted too early, before the level is really
reached. The time-limit should not be too long either, considering that the
disturbance probability will increase. When a dead zone exists the param-
eter T is badly approximated, but still gives a long enough time-limit for
reaching stationarity. When the measurement signal reaches T a boolean
63

variable is set true and when 1/4 of the time-limit has passed the algorithm
63

allows two samples in sequence to be out of range for the tolerance level.
These samples are excluded when computing the average for the end level of
the step response.
When the condition for stationarity fails the algorithm will check for load dis-
turbances, see section 3.3.5. If a load disturbance has occurred the algorithm
will be terminated. The algorithm will also be terminated if the worst-case
time for the step change has passed. A normal Closed Loop phase is esti-
mated to take less than 2 · T + time-limit seconds, the worst-case time for
the closed loop is set to twice the normal phase time. To check the control
63

signal the only thing that is known in this phase is the initial level. If the
control signal has opposite direction to the step change and has passed its
initial level the algorithm will terminate.
Finally, when the stationarity condition is fullled, T is computed by taking
the dierence of the normalized signal areas. T is used in the next phase as
ar

the expected response time for an open loop step change, which corresponds
ar

to T for the closed loop step change. Checks to see that the control signal
does not saturate are also performed. The level of the actual step change is
63

compared with the one set by the user in the beginning of the experiment
and checked against the noise amplitude.
CHAPTER 3. IMPLEMENTATION 20
3.3.3 Open Loop Phase
The controller is deactivated and the algorithm sets the control signal to its
initial level. This will result in an open loop step in the opposite direction
to the closed loop step.
The area of the measurement signal is computed until the response time is
reached, see Figure 3.6. The total area of the signal is also computed, and
will be compared with T from the previous phase.
ar

The delay in this phase is estimated as the time until the measurement signal
changes direction from the stationary level. This delay will later be compared
with the approximated delay in the PID-design phase.
A new time-limit for the measurement signal being stationary when the
initial level has been reached is calculated as twice the response time. An
average of the measurement signal is computed when the signal is stationary,
in the same way as in the other phases. This average will be compared to the
initial level computed in the Initial phase, and will be used as an accuracy
parameter for the experiment in the PID-design phase.
If the stationary condition depending on the tolerance level fails, the algo-
rithm will check if a load disturbance has occurred, see section 3.3.5.
3.3.4 PID-design Phase
The approximations of K , T and L are calculated using the areas computed
in the Closed Loop and Open Loop phases. Using the approximations the
p

new PID-parameters are calculated by using the AMIGO tuning rules de-
scribed in the Equation 2.6. Then the credibility of the parameters is judged.
For now the only check performed is that they should be non-negative.
3.3.5 Disturbance Detection
Initial Phase

A load disturbance in the Initial phase makes the measurement signal loose
contact with the setpoint during the time it takes for the control signal to
compensate for the error. When the measurement signal is stabilized the
control signal gets a new initial level to keep the measurement signal at the
setpoint. If the control signal does not saturate the experiment will go on
without inuencing the result.
A pulse disturbance will make the measurement signal loose its setpoint for
a while and then the signals will return to their initial levels.
CHAPTER 3. IMPLEMENTATION 21
When a load or pulse disturbance occurs in the Initial phase, no specic dis-
turbance detection algorithm is used. Instead the steady state detection is
restarted and a limitation of the remaining experiment time is used to avoid
innite loop.
Closed Loop Phase

Disturbances in this phase will destroy the integration of the signal areas.
The dierence between the normalized signal areas, T , is later used in the
Open Loop phase as an estimate for the response time. The approximation of
ar

the three-parameter model, described in Equation 2.1, depends on a correctly

estimated response time for computing the area A in the Open Loop phase.
1

To detect a disturbance the algorithm checks for a change in the measurement

signal's direction against the step direction. To determine what direction the
measurement signal has the dierence between two samples and the tolerance
level is used.
dir_value :=(current_signal - old_signal);
IF dir_value <= -TOL THEN
dir := -1;
ELSIF dir_value >= TOL THEN
dir := 1;
ELSE
dir := step_dir;
END_IF

Figure 3.4: Code for deciding the direction of the signal.

If dir has the opposite direction as step_dir the level of the current_signal
is checked against its initial level and the setpoint, see Figure 3.5. And if
the current_signal is in the gray area number one or two in Figure 3.5 the
program will be terminated.
CHAPTER 3. IMPLEMENTATION 22

Figure 3.5: Closed loop step with step_dir = 1

If a change of direction has occurred in the gray area number three in Figure
3.5 the algorithm ignores it as an overshoot and then it will expect the
measurement signal to reach stationarity. If the measurement signal changes
direction again it will be considered as a disturbance though it might depend
on too fast PI-parameters.
If the measurement signal becomes stationary in the gray area number two
in Figure 3.5 the only possibility would be that the control signal has been
saturated. This will not be detected until the Closed Loop phase is nished
and the real step size is checked against the tolerance level and the user's
desired step size.
When dir has the same direction as step_dir the only check carried out is
if current_signal has reached a level larger than 2·(step size + dead zone).
Open Loop Phase

A disturbance in this phase will be considered in a dierent way than in

the previous phase. The phase is divided into two parts; before passing
the response time, called the , and after the response time, see
Figure 3.6. The area which the approximated three-parameter model is
critical part

based on is computed in the critical part. If a disturbance occurs in this

part the algorithm will be terminated. After the critical part a special timer
will end the program and allow the experiment to conclude. If a small
disturbance has happened and not been detected the computed area between
the measurement signal and the setpoint diers from the computed area in
the Closed Loop phase. In this case the algorithm will terminate because
it can not guarantee that the disturbance has occurred outside the critical
part.
CHAPTER 3. IMPLEMENTATION 23

Figure 3.6: Open Loop Step with step_dir = 1, the critical part is dened as the
time until response_time.

The detection is based on what direction the measurement signal has in

relation to the step_dir of the phase. The same code as in the disturbance
detection for the Closed Loop phase, see Figure 3.4, is used to get the signal's
direction.
Disturbances in the same direction as the step_dir are complicated to detect
in the critical part because of the dierent characteristics of the unknown
processes. This makes it impossible to detect a too fast falling/raising step
response. Instead a check is performed to see if the signal has passed 63%
of the step change before leaving the critical part.
A disturbance in the opposite direction of the step_dir will be detected in
the gray areas in Figure 3.6 but will be handled in dierent ways depending
on what part of the Open Loop phase the disturbance has happened in.
3.3.6 Filtering
When the process is supposed to be stationary, as in the Initial phase and
at the end of the Closed Loop and Open Loop phases, the measurement and
control signals are averaged using a method that is unbiased for converging
averages. The recursive average m is formulated like this, where N is the
number of samples and a is the latest signal value:
N

mN +1 =
N
N +1
mN +
1
N +1
aN +1 . (3.1)
Chapter 4

Test Results
The autotuner has been tested on three rst-order processes with dierent
characteristics, one lag dominated, one delay dominated and one balanced
process. Lag dominated means that the process time constant T is much
larger than the process delay L and τ is small. Delay dominated means that
L > T and τ is large. τ is calculated as L
and used as a basis for the
AMIGO tuning rules, see section 2.1.2. L+T

To show the further capabilities of the autotuner, it has also been tested on
a rst-order balanced process with measure noise and dead zone.
The tests have been performed using processes simulated with TAC Menta,
a FBD programming tool used by TAC to implement control applications.
"Original" PI-parameters that give a slow step response without overshoots
are chosen for each process. New PID-parameters are generated with the
autotuner and tested with a step response. Then T , the time it takes to
reach 63% of the step height, for the new parameters is compared with T
63

for the "original" parameters.

63

4.1 Lag Dominated Process

The process characteristics are as follows:
Static gain K = 1
Time constant T = 20
p

Delay L = 1
The real L is larger than 1, probably due to network delays. The τ measured
by the autotuner will therefore be larger than the theoretical τ .
Theoretical τ = 0.05 Measured τ = 0.16
24
CHAPTER 4. TEST RESULTS 25
The original and new PID-parameters are compared as well as T . 63

Original PI-parameters: New PID-parameters:

K = 0.5 K = 2.58
T = 15.0 T = 10.40
= 0 = 1.63
i i
T d T d

T 63 = 49s T = 18s
63

The new parameters give a fast step response with a small overshoot, ap-
proximately 5%.

Figure 4.1: Lag dominated process. The autotuning experiment is shown.

Figure 4.2: Lag dominated process. A step response with the new PID-parameters
is shown.
CHAPTER 4. TEST RESULTS 26
4.2 Delay Dominated Process
The process characteristics are as follows:
Static gain K = 1
Time constant T = 10
p

Delay L = 10
This is theoretically a balanced process but due to delays in the experiment
setup the τ measured by the autotuner indicates a delay dominated process.
Since this is what the controller will see, the process is classied as delay
dominated.
Theoretical τ = 0.5 Measured τ = 0.7
The original and new PID-parameters are compared as well as T . 63

Original PI-parameters: New PID-parameters:

K = 0.5 K = 0.39
T = 15.0 T = 10.21
= 0 = 4.21
i i
T
d T
d

T
63 = 46s T
63 = 42s
The new parameters give a step response that is only slightly faster than
before.

Figure 4.3: Delay dominated process. The autotuning experiment is shown.

CHAPTER 4. TEST RESULTS 27

Figure 4.4: Delay dominated process. A step response with the new PID-
parameters is shown.

4.3 Balanced Process

The process characteristics are as follows:
Static gain K = 1
Time constant T = 10
p

Delay L = 3
This is theoretically a lag dominated process but due to delays in the exper-
iment setup the τ measured by the autotuner indicates a balanced process.
Since this is what the controller will see, the process is classied as balanced.
Theoretical τ = 0.23 Measured τ = 0.43
The original and new PID-parameters are compared as well as T . 63

Original PI-parameters: New PID-parameters:

K = 0.5 K = 0.8
T = 15.0 T = 7.5
= 0 = 2.4
i i
T d T d

T 63 = 46s T 63 = 20s
The new parameters give a fast step response with a small overshoot, less
than 5%.
CHAPTER 4. TEST RESULTS 28

Figure 4.5: Balanced process. The autotuning experiment is shown.

Figure 4.6: Balanced process. A step response with the new PID-parameters is
shown.

4.4 Process with Noise and Dead Zone

This process is the same as the balanced process in the previous section, 4.3.
The process is aected by measurement noise and the controller has a dead
zone = 0.5.
The original and new PID-parameters are compared as well as T . 63
CHAPTER 4. TEST RESULTS 29
Original PI-parameters: New PID-parameters:
K = 0.5 K = 0.51
T = 15.0 T = 6.41
= 0 = 2.51
i i
T d T d

T 63 = 45s T 63= 23s

The new parameters give a fast step response with an overshoot, approxi-
mately 5%. The signal levels dier from the setpoint because of the dead
zone.

Figure 4.7: Balanced process with noise and dead zone. The autotuning experi-
ment is shown.

Figure 4.8: Balanced process with noise and dead zone. A step response with the
new PID-parameters is shown.
CHAPTER 4. TEST RESULTS 30
4.5 Testing on an Air Handling Unit
The autotuner has also been tried with the heating process of an air han-
dling unit. It was found that it is hard to bring the process to stationarity,
which is a prerequisite for the tuning experiment. Another problem is that
the outdoor temperature can change during the experiment and disturb the
results.
The autotuner makes one upwards and one downwards step change. It is
possible that the process has dierent dynamics going up or down. This
problem could be avoided by making both step changes in the same direction.
Real processes are not linear and time invariant as the simulated processes
are. This leads to a lot of interesting problems when trying to apply the
autotuner. These problems will have to be taken care of in future versions.
Chapter 5

Conclusions
5.1 Summary
An autotuner based on the Method of Moments combined with the AMIGO
tuning rules has been built. It has been tested on various processes and
the generated PID-parameters give fast step responses with acceptable over-
shoots. It is hard to tell if the parameters are optimal but at least they give
good results.
During testing we have noticed that the approximations of the process pa-
rameters T and L are often far from the original parameters. However, T +L
is close to the mark. Even though the approximations are bad, the resulting
PID-parameters give good step responses.
The total time of an experiment depends on the controller parameters used
and the nature of the process that is tuned. The autotuner uses two steps
compared to manual tuning which might need several steps to achieve good
controller parameters. This makes the autotuner time-economical when new
parameters are needed.
In some experiments the controller parameters given by the autotuner result
in an oscillating step response. This could be because some undetected
disturbance has occured or it could depend on communication problems.
The user should therefore assess the parameters before deploying them.
5.2 User Notes
Some basic knowledge about the process that will be tuned and a bit of
control theory will still be needed for the user of the autotuner.
31
CHAPTER 5. CONCLUSIONS 32
The autotuner should be connected to one PID-controller at the time, not
to entire cascaded control loops.
Before running the autotuner the user should make sure that the system is
controlled with slow and safe PID-parameters, giving a step response without
overshoot. The system should also be in steady state.
The step size of the experiment should be suciently large, at least 20 times
the approximated noise amplitude. If a dead zone is present in the process,
the user must state its size.
The experiment will take at least as much time as two setpoint changes.
If the experiment is interrupted an error message will be displayed with
information about the possible cause. Load disturbances that occur during
the experiment will invalidate the results and the autotuner therefore tries
to detect them and interrupt the experiment.
When the experiment concludes, the new PID-parameters are displayed.
They can then be manually transferred to the PID-controller.
Most of TACs control systems use only the proportional and integral parts
of the PID-controllers. The autotuner will in most cases propose that the
controller uses the derivative part as well. When the derivative part is used
the system will become more sensitive to high frequency disturbances and
therefore a lowpass lter should be used.
The autotuner will give optimal parameters, but the parameters will only
remain optimal as long as system conditions remain unchanged. A tuning of
a temperature control system that takes place during winter will probably
not give parameters suitable for summer control.
5.3 Future Work
• Better identication and ltering of the measurement noise.
• Check the generated PID-parameters to see if they seem probable.
Checks can be added, such as comparing the dierent T :s calculated
in the Closed Loop and Open Loop phases, comparing the measured L
ar

with the calculated L and checking what non-fatal errors have occurred.
• Use τ to make further adjustments to the PID-parameters as described
in article [3].
• Use τ to adjust the setpoint weighting of the controller, according to
[3].
CHAPTER 5. CONCLUSIONS 33
• Save the information from previous tuning experiments and use it to
get better process identication.
• Use the autotuner to generate several parameter sets for use in gain
scheduling.
• Further development and testing to make the autotuner t for real
systems.
Appendix A

Figures
A.1 FBD  Overview of the Communication Pa-
rameters

Figure A.1: The Function Block Diagram handles input and output.

34
APPENDIX A. FIGURES 35
A.2 SFC  Overview of the Autotuner Algorithm

Figure A.2: The Sequential Function Chart of the autotuner algorithm.

Bibliography
[1] A. Leva, C. Cox, A. Ruano.
IFAC Professional Brief 2002.
Hands-on PID autotuning: a guide to better

utilisation.

[2] K.J. Åström, T. Hägglund.

Instrument Society of America, Research Triangle Park, NC 1995.
PID controllers: Theory, Design and Tun-

ing.

[3] K.J. Åström, T. Hägglund.

Journal of Process Control 14 (2004) 635-650.
Revisiting the Ziegler-Nichols step response

method for PID control.

[4] A. Ingimundarson.
Department of Automatic Control, Lund
Dead-Time Compensation and Performance Mon-

Institute of Technology, Lund 2003.

itoring in Process Control.

[5] A. Robertsson.
Department of Automatic Control, Lund Institute of Technology, Lund
lecture slides for Nonlinear Control and Servosystems.

2004.
[6] T. Hägglund, K.J.Åström.
Automatica vol.27 no.4 (1991) 599-609.
Industrial Adaptive Controllers Based on

Frequency Response Techniques.

36