Control 2004, University of Bath, UK, September 2004 ID-095

PRACTICAL CONTROL LOOP TUNING USING A

MATLAB/SIMULINK TOOLBOX
A. Adgar*, C.S. Cox*

* University of Sunderland, UK.

Keywords: Computer aided control system design, control strategy. In all of the above scenarios, it is obviously
Modelling and simulation, Non–linear systems, System desirable if some kind of advice could be generated
identification and signal processing, Utilities automatically. This would be especially useful if the advice
would also embrace the modelling phase that inevitably
Abstract accompanies any design strategy. The software toolbox
described in this paper has been specifically designed to
This paper describes a suite of software which has been produce this kind of assistance.
developed at the University of Sunderland. The software
takes the form of a toolbox for the MATLAB and Simulink
environment. It assists in the many stages of designing a 2 The Software
feedback controller for a real process.
The authors have developed a comprehensive set of software
The software may be used for data collection, data pre– tools running under the MATLAB and Simulink platform to
processing, process modelling, controller design and finally assist in performing real–time modelling and control
controller evaluation. The software has been demonstrated on experiments. MATLAB and Simulink are closely related
a laboratory based process trainer and several controller computer packages which provide an environment that allows
architectures have been designed and evaluated. the integration of computation, visualisation, and
programming in an easy–to–use environment. Typical uses
The software presented here is thought to provide several include: mathematics and computation; algorithm
advantages over some of the commercial loop–tuning and development; modelling, simulation, and prototyping; data
controller design software. These are discussed within the analysis, exploration, and visualisation; scientific and
paper. engineering graphics; and application development, including
graphical user interface building.

1 Introduction The name MATLAB stands for matrix laboratory, since most
of the functionality of the software comes from matrix
It is well known that many complex processes are often manipulation. It has many uses therefore in mathematics,
satisfactorily controlled by means of simple three term (PID) engineering, and science. In this case the software has been
controllers. In fact, many industrial solutions are realised utilised for teaching process modelling and control systems
using only the P and I terms but it has been widely reported, design. The Simulink environment provides a drag–and–drop
that even after careful instruction, plant operators still have interface for building mathematical models of systems in the
difficulty installing and operating such systems. It is obvious conventional block diagram form.
that a feedback system is of little value if the tuning
parameters are poorly selected. For systems with long time The toolbox which has been developed contains numerous
constants (of the order of several minutes) the semi–empirical MATLAB functions and Simulink blocks to assist many of
methods favoured by some users can be both time consuming the tasks described above. There are many types of process
and tedious. input signals, process controller structures and interfaces for
typical laboratory process training equipment. The standard
Consequently, any help offered to the process operator to ease Simulink block libraries may also be included in any
this tedious task and at the same time produce consistent activities being performed.
closed–loop behaviour on the process is clearly very
desirable. In many process industries (e.g. the water industry) The typical activities of data collection, data pre–processing,
it is often found that the system dynamics display appreciable process modelling, controller design, implementation and
time delay. The use of PI(D) control in these situations is not evaluation can be performed very rapidly. The simple drag–
recommended. Instead, some form of dead–time and–drop interface allows for easy adjustment of the various
compensation is necessary. Whilst improved performance is parameters, the display of process variables, the storage of
achievable it is accomplished at the cost of a more complex data, etc.
 Control 2004, University of Bath, UK, September 2004 ID-095

3 The Test System. test rig or an industrial process. Signal conditioning units
have been built to allow the sampling of voltages or currents.
The test system used to demonstrate the application of the Typical scales which are used include -10 to +10V, 0 to 10V,
toolbox was the process trainer as shown in Figure 1. This 4 to 20mA.
apparatus is known informally as the ‘hairdryer’ for obvious
reasons. The system consists of a fixed speed blower which To read or measure signals from a sensor (process output) via
propels a current of air through the plastic tube. The air flow the ADC we use the function:
rate may be modified by the use of a damper. The air is
warmed using a simple heating device consisting of a mesh of » y = changet(chan_no);
wires just inside the entrance to the tube. The heating effect is
controlled by the voltage supplied to the mesh. The air To write or send signals to an actuator (process input) via the
temperature may be measured at one of three locations along DAC we use the function:
the tube with a thermistor. Hence the test equipment forms a
simple single–input single–output system. » chanput(chan_no,sig_lev);

There are some other functions available which assist the

collection of data from processes in real–time. These are very
flexible but require some simple programming knowledge. It
is generally much easier to use the Simulink drag–and–drop
environment and this is the route which is described here.

The Simulink model is shown in Figure 2. There are several

points to note about the blocks used here. Firstly, the real–
time block is important as this forces the simulation to run in
real–time. Hence the data collection is triggered at the correct
frequency. In this case the sample time was chosen to be
100ms. The sample rate which can be achieved for a given
experiment is dependent on many factors: the windows
operating system, the Simulink software and model
complexity and the PC processor speed. On a typical lab PC
the system can achieve 100ms sample times for the standard
Figure 1: Process trainer test system used for the data data collection and controller implementation tasks. When
collection, modelling, controller design and controller working with processes which require faster sampling than
evaluation studies. this, then some kind of commercial system is needed.
Examples being LabVIEW, xPC target for windows or
dSPACE.
4 The Computer Hardware/Software
A simple step test is to be performed so a step block is used as
The software toolbox runs on a PC which already has
the input (parameters specified are the step time, the initial
MATLAB and Simulink installed. The PC is also fitted with
value and the final value). There are other signals that may be
an Amplicon data acquisition card. Driver software has been
used such as a Pseudo Random Binary Sequence, a sine wave,
written for several types of card so that the software can be
an impulse, etc., but the step is probably the simplest. The
used on any of the computers in our laboratories. This allows
PT326 block is used to specify the DAC and ADC channels to
the reading and writing of analogue signals to and from the
be utilised for communication to the process. There is no
test equipment.
actual model of the process within this block. The process
inputs and outputs are collected and stored in variables u and
y respectively. Both are also sent to the scope block which
5 The Controller Design Procedure reveals a real–time plot as the data collection is progressing.
The main stages involved in the design of a feedback
Note that a variable, tv, is also recorded. This variable stores
controller are described in the subsections below. Each
information about the actual time a sample was collected .
subsection will show how the software is used and the typical
Hence, it can be compared to the variable t to ensure that the
results achieved at each stage.
simulation is running in real–time and that there are no
problems in achieving the sample rate specified on the PC
being used.
5.1 Data Collection
When using the software toolbox, data may be received from The user defines a fixed sample time and then the data
and sent to the outside world via the interface card in the PC. collection experiment is controlled at this fixed sampling rate.
These features may be used to collect data from a laboratory The user also defines starting and stopping times.
 Control 2004, University of Bath, UK, September 2004 ID-095

5.3 Model Estimation

Once the data has been suitably pre–processed a process
model may be estimated from the data. In this initial example
the model structure chosen is that of a first order plus dead–
time (FOPDT) equation. This model structure is relatively
simple and is used by many controller design techniques as
described comprehensively by O’Dwyer [2].

The model structure is shown in transfer function form in

Equation 1. It is very simple and requires only three
parameters (K, T and D) to be specified.

Ke − Ds
Figure 2: Example of a simple Simulink model used for the G (s ) = (1)
collection of data from a process in real–time. 1 + sT

The results are presented graphically on the screen as the test The function required to calculate the three parameters is
is being performed. A typical result is shown in Figure 3. This shown below:
is a data collected from the test rig using a sample time of
» [K,T,D] = fopdt(t,y,u);
0.1s. The collected data was plotted using the function:

» plotresp(t,y,u) The software uses the method of Nishikawa et al [3]. In this

particular case the software calculates the parameters to be as
shown in Table 1. Other methods of model parameter
estimation based upon optimisation techniques may also be
used. As can be seen the process dead–time is slightly smaller
than the process time constant.

Parameter Description Value

K Process Gain 1.13
T Process Time Constant 0.43
D Process Time Delay 0.36

Table 1: Parameters of the FOPDT Model estimated using the

data collected during the process step test.
Graphical output is also generated automatically by the
software. The plots in Figure 4 show the input step as well as
Figure 3: Data collection from an open–loop process under the actual collected sampled data (the dots) plus the
the influence of a step change in the process input. continuous line which represents the output from the FOPDT
model to the same step input. This kind of information is very
useful for the user because it provides knowledge about the
5.2 Data Pre–Processing accuracy of the characterisation.
As can be seen from Figure 3, the step response covers only a
small time period of the graph. Also there is some transient
response information at the start of the plot relating to the 5.4 Controller Design
process reaching its initial test level. This should be removed The actual design of the controller is relatively straight
before any analysis is done on the data. This is achieved by forward once the FOPDT model parameters have been
using the MATLAB function: estimated. Another function may be used to design a simple
feedback PI or PID controller for the above process. The
» [t,y,u] = aacrop(t,y,u); function inputs are the model gain, K, time constant, T, and
delay, D, (all evaluated above) to calculate the controller
The user simply uses the mouse to select the desired start and parameters. The actual function call is:
stop times for the data. In this case the data between 7 and 15s
was selected. » [Kc,Ti,Td] = despid(K,T,D,’ZN’,2)
 Control 2004, University of Bath, UK, September 2004 ID-095

Figure 4: Results of the model estimation using data collected Figure 6: Controller performance for the PI design.
from the process.
The additional inputs define which controller design method
to use (in this case Ziegler–Nichols has been used) and which
controller structure to use (1 = P, 2 = PI, and 3 = PID). The
controller is defined by the controller gain, Kc, the controller
integral action time, Ti, and the controller derivative time, Td.
The function call shown performs a Ziegler–Nichols
controller design, but several other design methods are
available.

The results for a PI controller designed using the method of

Daniel & Cox [1] yields the following controller parameters.

Parameter Description Value

Kc Controller Gain 0.52
Ti Controller Integral Time 0.43 Figure 7: Controller performance for the pPI design
Td Controller Derivative Time 0.00
5.5 Controller Evaluation
Table 2: Parameters of the FOPDT Model estimated from the
step test data collected. The controller may then be evaluated on either the model (in
simulation) or on the actual process itself (in real–time). Here,
the controller will be evaluated in real–time on the process.
The Simulink model is set up as shown in Figure 5 and, as
can be seen, the set point input is a sequence of multi–level
steps. The process set–point (r), the controller output (u) and
the process variable (y) are all logged for the duration of the
trail. The controller is evaluated over a total time period of 1
minute with the sample time set to 100ms. The response of
the closed–loop control system can be seen in Figure 6.

The advantages of performing a controller design in this way

are numerous. The design process is flexible, simple to
implement and the software provides easy access to any of the
signals around the loop for analysis purposes.

Figure 5: Simulink model used to implement and then

evaluate the PI controller on the test process in real–time.
 Control 2004, University of Bath, UK, September 2004 ID-095

5.6 Advanced Control Acknowledgements

As a final exercise, an advanced controller was designed and The authors would like to thank the staff and students (both
implemented using the software. This only requires the past and present) of the Control Systems Centre at the
replacement of the PID block with the new controller block. University of Sunderland for their contributions to this work.
In this case a pPI (predictive PI) controller was designed. The
same trial as was done with the PI controller was then References
performed using this more advanced controller.
[1] Cox, C.S., Daniel, P.R. & Lowdon, A. (1997)
As can be seen from Figure 7, the pPI controller provides Quicktune, A reliable automatic strategy for determining
much tighter control than the PI controller. The responses to PI and pPI controller parameter using a FOPDT model.
the set–point changes are less oscillatory and there is less Control Eng Practice, 5(10), pp 1463–1472.
overshoot. This is as expected and the difference would be
more marked if the process time delay was larger in [2] O’Dwyer, A. (2000) A Summary of PI and PID
comparison to the time constant. Controller Tuning Rules for Processes with Time Delay.
IFAC Conf – PID2000 Universitat Politècnica de
Catalunya, Terrassa, Spain, April 5–7, 2000.
6. Conclusions
[3] Nishikawa, Y., Sannomiya, N., Ohta, T. & Tanaka, H.
The MATLAB/Simulink Toolbox software described here (1984) A method for auto–tuning of PID control
provides an ideal environment for research and development parameters. Automatica, 20, pp 321–332.
studies. All the stages of controller design such as data
collection, process modelling and controller testing may be
performed using the software.

It is important to emphasise that the design carried out above

hardly exploits the wide range of design features available
with the Toolbox. Model development for continuous–time
situations also includes SOPDT (second order plus dead time)
and SOZPDT (second order with zero plus dead time)
characterisations. In addition three different forms of time
delay compensation are available and may be immediately
implemented using individual Simulink blocks. Advanced
discrete–time modelling is also facilitated. The discrete–time
models derived are used for model predictive control designs.
Two methodologies are available. These are GPC
(generalised predictive control) and PIP (proportional–
integral–plus).

The software described here has the advantage over other

loop tuning software primarily in terms of its flexibility. New
modelling algorithms and controller structures may be easily
incorporated into the library. Users may develop their own
tailor–made routines to suit the particular application being
studied. The software can also be extended by taking
advantage of the wide range of other MATLAB and Simulink
libraries already available. This provides the opportunity of
using statistical analysis and performance evaluation, neural
network modelling, fuzzy logic control, and many other
ready–made solutions.

The main problem faced with using this software is the final
implementation of more complex control strategies within
off–the–shelf controllers or PLCs. This is possible in some
devices that have specific configuration software. It may also
be possible to implement coded versions of the control
strategies in PLCs using the IEC1131 standard.