AbstractDue to their high power-to-weight ratio and low cost,

pneumatic actuators are attractive for robotics and automation

applications; however, achieving fast and accurate control of their
position have been known as a complex control problem. A
methodology for obtaining high position accuracy with a linear
pneumatic actuator is presented. During experimentation with a
number of PID classical control approaches over many operations of
the pneumatic system, the need for frequent manual re-tuning of the
controller could not be eliminated. The reason for this problem is
thermal and energy losses inside the cylinder body due to the
complex friction forces developed by the piston displacements.
Although PD controllers performed very well over short periods, it
was necessary in our research project to introduce some form of
automatic gain-scheduling to achieve good long-term performance.
We chose a fuzzy logic system to do this, which proved to be an
easily designed and robust approach. Since the PD approach showed
very good behaviour in terms of position accuracy and settling time,
it was incorporated into a modified form of the 1
st
order Tagaki-
Sugeno fuzzy method to build an overall controller. This fuzzy gain-
scheduler uses an input variable which automatically changes the PD
gain values of the controller according to the frequency of repeated
system operations. Performance of the new controller was
significantly improved and the need for manual re-tuning was
eliminated without a decrease in performance. The performance of
the controller operating with the above method is going to be tested
through a high-speed web network (GRID) for research purposes.

KeywordsFuzzy logic, gain scheduling, leaky integrator,
pneumatic actuator.

I. INTRODUCTION
XPERIMENTATION with a number of conventional (or
"classical") three-term controllers shows that, as repeated




Manuscript received September 15, 2005.
M. G. Papoutsidakis is a Research Associate at the Intelligent Systems
Laboratory, Dept of Automation, Technological Institute of Piraeus, Athens,
Greece. ( TEI Piraeus, dept. Automation, P.Ralli & Thivon 250, Athens
12244, Greece. phone; 0030-210-5381427, fax; 0030-210-5451128, e-mail:
[email protected]).
Dr. George E. Chamilothoris is a Professor at the Dept of Automation of
the Technological Institute of Piraeus and the Head of the Intelligent Systems
Laboratory, Dept of Automation, Technological Institute of Piraeus (e-mail:
[email protected]).
F.Dailami, N.Larsen and Dr.A.Pipe are lecturers and researchers in the
Intelligent Autonomous Systems Laboratory, University of the West of
England, Bristol, Frenchay Campus, Coldharbour Lane, BS16 1QY, UK.



operations accumulate, the characteristics of the pneumatic
actuator change requiring frequent re-tuning of the controller
parameters (PID gains). Furthermore, three-term controllers
are found to perform poorly in recovering the closed-loop
system after the application of load or other external
disturbances. The key reason for these problems lies in the
non-linear exchange of energy inside the cylinder relating, in
particular, to the complex friction forces that develop on the
piston-wall interface. Therefore, another methodology should
be followed in order to improve the performance of such
system. There are two well-known fuzzy logic approaches
used in control systems. The first is the Mamdani method that
has been used to create PI controllers but only a few PID, and
mostly in trivial control problems requiring low accuracy and
minimal performance [2], [5]. Since there are also no
guarantees of stability and robustness, in the case of a
complex nonlinear system, such as the pneumatic actuator rig,
it is not appropriate to adopt this approach. The second fuzzy
logic method, implemented in this project, is the Tagaki-
Sugeno approach. It is well known that Takagi-Sugeno fuzzy
models can provide an effective representation of complex
nonlinear systems in terms of fuzzy sets and fuzzy reasoning
and, in the recent past, researchers made this clear in their
research projects [9], [7], [8], [2]. Using this approach, we
have incorporated a number of PD controllers, one into each
of the consequent parts of the fuzzy rules. We then hand-over
between controllers, based on rule activations, thus achieving
a smooth gain-schedule. Gain scheduling is a popular
nonlinear control design approach, which has been widely and
successfully applied in fields ranging from aerospace to
process control. Whilst much of the classical gain scheduling
theory originates from the 1960s, there has recently been a
considerable increase in interest in gain scheduling [4], [6],
[1]. Although a wide variety of control methods are often
described as gain scheduling approaches, they normally
share a divide and conquer type of design procedure
whereby the nonlinear control design task is decomposed into
a number of linear sub-problems. This divide and conquer
approach is the source of much of the popularity of gain
scheduling methods since it enables well established linear
design methods to be applied to nonlinear problems. This is
the most important advantage of this method and is the reason
it was chosen for implementation in this research project.
According to earlier researches like [12], [13], [14] the
performance of such a control method could be tested over the
well-known European high-speed network (GEANT). The
experimental rig consists of a pneumatic cylinder with 80mm
Accurate Control of a Pneumatic System using
an Innovative Fuzzy Gain-Scheduling Pattern
M. G. Papoutsidakis, G. Chamilothoris, F. Dailami, N. Larsen, and A Pipe

E
World Academy of Science, Engineering and Technology 8 2005
189



of stroke, 25mm bore and a proportional pneumatic servo
valve, model FESTO, MPYE-5-. An LVDT is used as a
position sensor and the controller is a Siemens/Infineon
80C164CI, programmed in the C language.

II. THE GAIN-SCHEDULER BASED ON FUZZY LOGIC AND THE
CONTROLLER DESIGN
One of the most important parts of the gain-scheduler is the
variable input to the fuzzy system and the way this variable is
calculated in real time during system operations. By theorising
that temperature is, in some way, the main factor in changing
the friction forces experienced by the piston and that this is
the main factor in determining new PD controller gain values,
we managed to build an empirical model of the features and
the influence of temperature on the system. This model was
based on a leaky-integrator, which can be used to model
thermal effects [3]. An electrical circuit theory interpretation
of the leaky-integrator is shown in figure 1. It is actually
producing an independent variable, which from now on will
be called alpha in the rest of this paper. The integrator is
charged by a current pulse every time a pneumatic control
action is undertaken and is discharged by the simulated
resistor. The "alpha generator" is tuned for just one
operating regime, for instance every five minutes over a six-
hour period, then tested for its generalisation abilities in
different regimes, for example every fifteen minutes over
another six-hour period, without further tuning. Of course, this
structural modelling of what is believed to be the nature of the
underlying effect is still based in empirical evidence, rather
than mathematical/physical analysis of the cause of the effect.
However, a structural model is much more powerful than
simply fitting the equation of a curve to the observed effect
from conducting experiments for just one operating regime;
this model, it is hoped, will generalise across operating
regimes adequately. The alpha variable will be used finally as
the input that modifies the gain schedule via the fuzzy logic
gain-scheduler. The value of alpha, which is the output of the
leaky-integrator, is generated automatically every second from
an appropriate C code program. This program also includes
the fuzzy logic algorithm and will be finally applied to the
controller of the pneumatic system.
The Vc voltage across the capacitor represents the value of
alpha, the integrator part is the capacitor itself and the leaky
part is the resistance. A discrete time version of this circuit,
with one-second sampling time, is the alpha generator.
The Tagaki-Sugeno fuzzy logic approach was implemented
in the system as stated earlier, thus implementing a fuzzy-gain
scheduling method. Earlier research works in this fuzzy gain-
scheduling field were accomplished with success in nonlinear
systems [10], [11]. In fact, the fuzzy gain scheduling structure
implemented here represents a variant from the work
referenced above, and the mainstream of Tagaki-Sugeno fuzzy
gain scheduling work. A thought experiment, which stays
within the application domain of this investigation, will be
employed to illustrate this variance. Let us assume, for the
moment, that the plant dynamics are such that different gain
schedules are required at different operating positions and
piston velocities of the pneumatic actuator although, actually,
this was not found to be the case in work to date. It would be
useful, under these hypothetical circumstances, to have fuzzy
rules that operated in the inputs domains of the position error
and velocity of position error so as to be able to schedule
changing Kp and Kd gains at different operational points. This
is the form of most other work that uses the Tagaki-Sugeno
system for problems similar to ours. Such work is definitely
not excluded from further work carrying on from this project,
but that is not what is being done here. In this work, the
position and velocity of position errors are inputs to the
system (as for other work), but the weighted Kp and Kd gains
that they are multiplied by are changed as a result of changes
in the alpha value, and independent variable. Although this is
not a major difference, the authors are unaware of any other
work of this nature for control problems of this type. The
fuzzy controller that achieved the satisfactory results
presented below used just four rules, and was therefore
straightforward to tune. Each rules antecedent identified an
equilateral triangular membership function in the alpha value
input space and described a sum of products for position error
and its derivative; each rule consequent defining a different
pair of Kp and Kd gain values.

III. BENEFITS VALIDATION OF THE PATTERN VIA
EXPERIMENTAL PROCESS
A very large number of experiments were designed in order
to evaluate the benefits of such a technique in the pneumatic
actuator position control problem presented here. The idea
was to test the behaviour of the system over long time periods
on different days. The subject of this investigation was to
record how the optimal values for the two gain values (Kp,
Kd) change over time so that the system operates with the best
position accuracy. A large number of tests with the pneumatic
actuator took place initially using the PD controller alone. In
each different test the controller was manually retuned in
order to record the values of the Kp and Kd gains that
provided the best system response. Three different
experiments were designed to cover the range of the gain
changes. The first experiment includes 52 different tests over
six hours of total experimentation. That is, every five minutes
one test was done for the first two hours, then one test was
done every thirty minutes for the next two hours and finally
during the last two-hour interval one test again every five
minutes was done. The second experiment of 21 tests had the
same overall duration of six hours. It was again split into three
continuous intervals of two hours each, like the previous
experiment, but the first and the last parts were split into
fifteen minute (rather than five minute) time intervals. Finally
the third experiment of 243 tests had the same structure, with
three major intervals again of two hours each, but this time
during the first and the last two-hour intervals, one test every
minute took place. These experiments took place on different
days to allow the system to settle between tests, and they
were designed to provide us with an understanding of how the
gain values must change over time after many operations. The
results from the first experimentation of five minutes repeated
tests were then used to tune the alpha generator. It was also
used to create the fuzzy systems input membership functions
and rules, including the corresponding gain values for the rule
World Academy of Science, Engineering and Technology 8 2005
190



consequent parts. All the above process was repeated again
with the fuzzy logic controller implemented in the system,
because the comparison of the gain values between the two
different control methods was necessary. So, another three
experiments were performed during different days with 316
tests in total, but keeping the same time intervals as before.
This generated another three tables of fuzzy logic gain values.
In figure 1, a representative set of values for the proportional
Kp gain value for both classical and fuzzy methods is plotted.
Kd values varied in a similar way, and are not shown for
clarity. The plot shows experimental data from the fifteen
minute tests, and is illustrative of the general trends. It is
noticeable that both curves rise between the first two hours
interval, then they settle down between the second two hours
interval and rise again in the last two hours of the
experimentation. The fact that within the second two-hour
interval both profiles reduce to lower values than they were at
the end of the first interval is because the experimentation
frequency changes. During the second two hour period the
gains are recorded every thirty minutes and not every fifteen
like in the first period. The system seems to recover partly
back to the initial condition when the experimentation
frequency decreases. In the third two-hour period, both curves
rise again.
Kp Classical & Kp Fuzzy Logic estimation
1
1.2
1.4
1.6
1.8
2
2.2
0 60 120 180 240 300 360
Time (mins)
K
p

G
a
i
n
Kp Classical Kp Fuzzy

Fig. 1 The combinational plot of classical and fuzzy proportional
gain values

IV. THE PERFORMANCE IMPROVEMENT DUE TO ALPHA
GENERATOR IMPLEMENTATION
The reason for implementing all the above in the
experimental rig is to create a robust controller of improved
performance relative to a fixed control gain system. The new
controller was applied to the pneumatic rig and tested for
different values of air pressure and the piston position
accuracy was checked out over the whole piston stroke. We
consider that reducing the steady state error of a pneumatic
positioning system like this is the most important aim. We
also consider that the settling time, which is the time taken by
the piston to reach the 95% of its final displacement at any
point of the stroke, as the second critical factor in order to
evaluate the performance of our experimental rig. In section 3
we described the experimental process that took place to
compare the classical control method with the frequent re-
tuning with the automatically tuned fuzzy control method. The
improvement of the system when the alpha generator
technique and the gain scheduler were applied to the system
was obvious. In the following figures we provide all the
details of that excellent performance of the system in contrast
with the earlier behaviour with only the classical controller
applied to it. For simplicity we illustrate two of the three
different experiments, those with the 5 and 15 minutes
operations respectively, over 6 hours operations in total. In
figure 2 the recorded steady state error for both processes of 5
and 15 minutes testing, with the classical and the fuzzy
approach, is shown.
S-S Error VS 6h Readings (5'+15'mins exp)
-0.5
0
0.5
1
1.5
2
2.5
3
0 1 2 3 4 5 6
Time (h)
S
-
S

E
r
r
o
r

(
m
m
)
15'_fuzzy 5'_fuzzy 15'_classical 5'_classical

Fig. 2 The combinational plot of the system S-S error with the
classical and the fuzzy methods

Although the steady state error is initially the same in all
cases since the initial tuning is common, as time passes the
system starts to demonstrate an unpredictable performance
when it actuates without frequent re-tuning (green and red
line). On the contrary when the system actuates with the
automatic gain scheduler and the fuzzy controller it shows an
excellent performance and the steady state error is never
bigger than 0.3mm throughout the six hours of
experimentation. In addition to this, we also record the settling
time of the system for exactly the same experimentation, and
alike with the position error provided before, the
combinational plot of the settling time of the system is as
shown in Fig. 3.
Settling Time VS 6h Readings (5'+15'mins exp)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0 1 2 3 4 5 6
Time (h)
T
i
m
e

(
s
e
c
)
15'_retune 5'_retune 15'_NOretune 5'_NOretune

Fig. 3 The combinational plot of the system settling time with the
classical and the fuzzy methods
World Academy of Science, Engineering and Technology 8 2005
191



It is, once again, confirmed that when our pneumatic
positioning system actuates with the fuzzy controller the
overall performance is improved and the settling time remains
close to 0.1 seconds.

V. CONCLUSIONS AND FURTHER RESEARCH STUDY
It is well-known that pneumatic positioning systems are
very complex and highly non-linear systems due to the
compressibility of air and the multiple physical phenomena
that exist within a pneumatic cylinder. In fact, all classical
control methods that were implemented faced difficulties in
achieving the goal of position control over long term
operation and needed frequent retuning in order to operate
properly. Although it was possible to retune the controller
each time manually, this clearly would not be suitable for an
implemented system. The empirically derived fuzzy control
approach adopted here was quite straightforward to design;
yet yielded acceptable performance over long term operation
and benefits from the analytical basis of the manually tuned
PD-control approach. With these control methods applied to
the system, the basic conclusion is that this form of fuzzy
controller is a robust solution that allows effective control of a
plant with time-variant dynamics. During this work it has
become clear that further research could improve the
performance of the system. However, perhaps even more
worthy of attention is the control of a pneumatic system with
multiple actuators operating together. The development of a
combinational actuator circuit, which will control all degrees
of freedom of a robot arm, is the final and most challenging
further research subject that, it is hoped, will be investigated
in the near future. In addition to this, like stated earlier, this
project constitutes the first attempt to build a fully controllable
remote system. This control system will be shortly tested over
a real time high-speed network (GRID), established between
two well-known Universities in Budapest-Hungary and
Athens-Greece.
ACKNOWLEDGEMENT
The first author acknowledges financial support by the
Greek State Scholarships Foundation. Work presented in the
paper was partly funded under the Hungary-Greece Joint
Research and Technology project by the Operational
Programme Competitiveness of the Community Support
Framework for Greece.
REFERENCES
[1] Gerard Scorletti, Lauren El Chaoui, Improved Linear Matrix Inequality
Conditions for Gain Scheduling and Related Control Problems,
International Journal of Robust and Nonlinear Control, issue 8, pp.845-
877, 1998.
[2] Hao Ying, Yongsheng Ding, Shaokan Li, Shihuang Shao, Comparison
of Necessary Conditions for Typical Tagaki-Sugeno and Mamdani
Fuzzy Systems as Universal Approximators, IEEE Transactions on
Systems, Man and Cybernetics, Vol 29, No 5, September 1999.
[3] Jianfeng Feng, Effects of Correlated and Synchronised Inputs to Leaky
Integrator Neuronal Model, Computational Neoroscience Laboratory,
The Babraham Institute, Cambridge, UK
[4] King, R., A. Stathaki, Fuzzy Gain Shceduling Control of Nonlinear
Processes, Department of Electrical and Computer Engineering,
University of Patras, Greece.
[5] Lee, C.-C., "Fuzzy logic in control systems: fuzzy logic controller-parts
1 and 2", IEEE Transactions on Systems, Man, and Cybernetics, Vol. 20,
No. 2, pp 404-435, 1990.
[6] Leith, D., W.E. Leithhead, Survey of Gain-Scheduling Analysis &
Design Department of Electronic & Electrical, University of
Strathclyde, Glasgow, Scotland.
[7] La Garca-Prez, Jos M. Caas, Mara C.Garca-Alegre, Pablo Yez,
Domingo Guinea, Fuzzy control of an electropneumatic actuator, IEEE
Trans. System, Man and Cybernetics, Jan.1996.
[8] Ming-Chang Shih and Niarn-Liarng Luor, Self-Tuning Neural Fuzzy
Control the Position of a Pneumatic Cylinder Under Vertical Load
,IEEE Trans. on IE, Vol. 39, No. 6, pp. 472-489, 1992.
[9] Moreno Llagostera, Control of a Pneumatic Servosystem using Fuzzy
Logic, Proceedings of 1
st
FPNI-PhD Symp. Hamburg 2000, pp 189-
201.
[10] Pauli Viljamaa, Fuzzy gain Scheduling and Tuning of Multivariable
Fuzzy Control-Methods of Fuzzy Computing in Control Systems, PhD
Thesis, Tampere University of Technology, 2002.
[11] Zhen-Yu Zhao, Masayoshi Tomizuka, Fuzzy Gain Shceduling of PID
Controllers, IEEE Transactions on Systems, Man and Cybernetics,
vol.23 No5, 1993.
[12] M. J. Mineter, C. H. Jarvis and S. Dowers, From Stand-alone Programs
Towards GRID-aware Services and Components: a Case Study in
Modelling & Software, Environmental Modelling & Software, vol 18,
issue 4, April 2003, Pages 379-391.
[13] Brian Tierney, William Johnson, Jason Lee and Mary Thompson, A
Data intensive Distributed Computing Architecture for GRID
Applications, Future Generation Computer Systems, vol 16, Issue 5,
March 2000, p473-481.
[14] Agranat D., Engineering Web Technologies for Embedded
Applications, Internet Computing, vol 3, No 3, June 1998.
















World Academy of Science, Engineering and Technology 8 2005
192