301

Reduced-order thermodynamic models for

servo-pneumatic actuator chambers
J Falcao Carneiro1,2* and F Gomes de Almeida1,2
1Departamento de Engenharia Mecanica e Gestao Industrial da Faculdade de Engenharia da Universidade do Port,
Porto, Portugal
2Instituto de Engenharia Mecanica, IDMEC, Polo, Portugal
The manuscript was received on 14 September 2005 and was accepted after revision for publication on 13 March 2006.
DOI: 10.1243/09596518JSCE203

Abstract: This paper discusses thermodynamic models of air inside pneumatic actuator
chambers. In servo-pneumatics common practice, these models are simplied by neglecting
the temperature dynamics. Classical models in the literature assume the temperature inside
the pneumatic chamber either to be constant or to follow a polytropic law. Furthermore, the
mixing process of air entering the chamber and heat transfer between air and cylinder walls
is often neglected or only implicitly taken into account.
This work evaluates the impact of these simplications and order reductions in the prediction
of pressure inside the actuator chamber. Classical models are compared with several others
not only taking into account the mixing process but also explicitly including the heat transfer
between air and cylinder walls. Simulation studies show that the reduced-order models proposed in this paper can lead to a mean square error in pressure prediction of only 10 per cent
of that obtained using classical models.
Keywords: servo-pneumatic systems modelling, servo-pneumatic systems simulation

1 INTRODUCTION
In order to control a pneumatic actuator accurately, a
model of the pneumatic system has to be established.
This model includes the pressure and temperature
dynamics of the two actuator chambers and the
mechanical dynamics of the load. Therefore, even
neglecting the servo-valve and friction dynamics,
the complete model is a sixth-order model. This
is inappropriate for control purposes since it is
mathematically dicult to handle and demands a
mass or temperature observer as these variables
cannot be correctly measured during operation.
Servo-pneumatic systems are used in applications
where force or motion control is required. In both
situations the pressure inside the chambers is the
most relevant thermodynamic state variable since
the control goals directly depend on it. Therefore, the
most typical solution to reduce the order of the model
* Corresponding

author:

Faculdade

de

Engenharia

da

Universidade do Porto, Rua Dr. Roberto Frias, Porto 4200-465,

Portugal. email: [email protected]

JSCE203 IMechE 2006

is to neglect temperature dynamics and to consider

a polytropic process with an index ranging from
1 (isothermal process) to 1.4 (adiabatic reversible
process). Burrows [1] used a reversible adiabatic
approach, Zalmanzon [2], Outbib and Richard [3],
and Ning and Bone [4] an isothermal approach, and
Andersen [5] and Chitty and Lambert [6] a polytropic
approach. Furthermore, examples can be found in
the literature [710] where, although the pressure
dynamic model is deduced assuming that the temperature follows a polytropic law, a further simplication in this model is introduced by neglecting
temperature changes with respect to ambient temperature. This approach leads to a situation where
the polytropic index of pressure dynamics is tuneable
but the temperature is xed at ambient temperature.
More recently, a new approximate model of a
pneumatic cylinder thermodynamic chamber was
proposed in reference [11]; based on experimental
evidence presented in reference [12], Richer and
Hurmuzlu [11] use a polytropic-based model whose
singularity resides on the fact that it uses dierent
polytropic indexes. The charging process has an
Proc. IMechE Vol. 220 Part I: J. Systems and Control Engineering

302

J Falca o Carneiro and F Gomes de Almeida

adiabatic evolution, the discharging process an isothermal evolution, and the process due to the movement of the piston is assumed to be intermediate
between the previous two by accepting a polytropic
index equal to 1.2. Again, although the processes are
not necessarily isothermal, temperature uctuations
are neglected. The question that naturally arises is
whether these approaches, which sometimes do not
have physical meaning, provide good thermodynamic
models for pressure. Another question is which
model to choose among the existing models. Before
answering these questions an important issue is to
know whether temperature in real servo systems has
signicant changes over ambient temperature.
As observed in reference [12], when using
pneumatic cylinders for ono movements, both the
pressure and the temperature inside the cylinder
chamber experience wide variations. In that study,
experimentally measured temperatures varied from
263 K when discharging to 323 K when charging.
When using pneumatic cylinders for servo-control,
deviations of temperature from their equilibrium
values are less pronounced but are not, as usually
considered in the literature, negligible. This fact was
experimentally observed in reference [13], where the
temperature inside the discharging chamber of a
pneumatic cylinder was measured in a meter-out
velocity control set-up. In that experiment, temperature changes of approximately 30 K were measured
during a full stroke movement of the piston. Another
way of illustrating this fact is to simulate the sixthorder system. For a pneumatic cylinder of 20 mm
diameter and 100 mm stroke, which is excited
by a random white noise reference, a change of
approximately 1.5105 Pa around the equilibrium
pressure (P =5.65105 Pa) leads to temperature
0
changes of approximately 20 and 30 K around
ambient temperature (293 K). Full details of this
simulation will be given in section 4 for cylinder D,
closed-loop simulation.
This paper will focus on the thermodynamic
modelling of pneumatic cylinder chambers. As previously explained, dierent studies use dierent
reduced models but there is not, as far as the present
authors know, any work comparing them. This
paper intends to shed some light on the subject
by comparing dierent reduced-order models with
the full-order model and determining each model
performance. Whether using a reduced or a full
model, it is important to assess the inuence of the
heat transfer coecient between the air inside the
cylinder chambers and its walls. The present authors
have experimentally determined the heat transfer
coecients for three dierent industrial pneumatic
Proc. IMechE Vol. 220 Part I: J. Systems and Control Engineering

actuators. Those values were used as guidelines for

the simulation studies developed in the present work.
This paper is organized as follows. Section 2
presents the datum model of the servo-pneumatic
system used for comparison purposes. Section 3
presents the typical model reductions appearing in
the literature and proers some new approximate
reductions. These reduced-order models propose not
only dierent algebraic ways of including temperature but also dierent ways of taking into account
heat transfer through walls. In section 4 the performances of the several models presented in section 3
are compared by means of simulation studies.
Finally, the main conclusions are drawn in section 5.

2 MODEL OF A SERVO-PNEUMATIC SYSTEM

2.1 Servo-valve modelling
A pneumatic servo-valve model may be partitioned
into two parts: a dynamic part for the spool and
its actuator motion and a static part for the mass
ow stage [9]. The bandwidth of the servo-valve is
typically much higher than the bandwidth of the
pneumatic actuator. The bandwidth of the system is
therefore not limited by the servo-valve and consequently its dynamics are often neglected [9]. This
will be the approach followed in this work. Consider
a typical four-way servo-valve as schematically
presented in Fig. 1.
The air mass ows that cross each restriction 1, 2,
3, and 4, may be determined using the expression [14]
m
(x , P , P , T )
v u d u

2c
A (x )P
t v u (c1)RT

CA B A B DH
P 2/c
P (c+1)/c
d
d

P
P
u
u

1/2

P
d >0.5283 (subsonic)
P
u
P
if d 0.5283 (sonic)
P
u
if

P A (x )
0.0404 u t v
(T )1/2
u

(1)

Fig. 1 Servo-valve scheme

JSCE203 IMechE 2006

Reduced-order thermodynamic models

303

where x is the spool displacement and P , T , A (x ),

v
u u t v
and P are dened for each restriction in the ideal
d
throat of Fig. 2.
In this work, it is accepted that the areas of the
servo-valve restrictions are matched
[A (x )=A (x ); A (x )=A (x )]
1 v
4 v
2 v
3 v
and symmetric [A (x )=A (x ); A (x )=A (x )].
1
v
2 v
3 v
4
v
It is also assumed that there is no leakage of air
when the spool is at the central position and that
A (x )0 [ A (x )=0 and A (x )0 [ A (x )=0.
1 v
2 v
3 v
4 v
Finally, it is accepted that there are linear
relations between the command voltage u and
the spool displacement (x =k u) and between the
v
u
spool displacement and the area of each restriction
(A =k x , i=1, 2, 3, 4).
i
x v
From these assumptions, the relation between
command voltage and each restriction area is given
by

A =k k u,
1
u x
A =k k u,
4
u x
u0 [
A =0,
3
A =0,
2

A =0
1
A =0
4
u<0 [
A =k k u
3
u x
A =k k u
2
u x

(2)

Real servo-valves, however, have leakage of air

between the spool and sleeve that determines the
equilibrium pressure when the spool is at the central
position. With the assumptions made above, the
equilibrium pressure P is given by P =0.8077P
0
0
s
(see Appendix 2). In this work the supply pressure is
P =7105 Pa and therefore P =5.65105 Pa. The
s
0
equilibrium temperature T is the ambient temper0
ature assumed to be T =T =293 K. It is worth
amb
0
noting that, even with a fairly simple model of the
servo-valve, it suits the goals of this work since it
is focused on the thermodynamic model of the
chambers.
2.2 Mechanical modelling
Consider the pneumatic cylinder schematically
represented in Fig. 3. Applying Newtons second law
results in
Mx =P A P A F
A A
B B
f

Fig. 2 Ideal throat

JSCE203 IMechE 2006

(3)

Fig. 3 Scheme of a symmetric cylinder

where M is the external load mass plus the mass of

the moving parts of the cylinder. The frictional force
F is assumed to be entirely viscous (F =k x). Again,
f
f
f
the friction model is quite simple but suitable for
the purposes of this work. For more information on
friction modelling, see reference [15].
2.3 Thermodynamic model
Assuming that air is a perfect gas, that pressures and
temperatures are homogeneous inside the chamber,
and nally that kinetic and gravitational energies
of the uid, viscous work, and cylinder mass ow
leakages are negligible, the Reynolds transport
theorem [16] applied to mass and energy in a xed
control volume with one-dimensional inlets and
outlets gives
P dV
R
R
c1
dP

=c
+c m
Q
T c m
T
in
in
out
dt
V dt
V
V
V
(4)
T dV
RT 2
dT
=
(1c)m
(c1)

out
dt
V dt
VP
RT
(c1)

(cT T )
Q
in VP
in
PV

+m

(5)

is the heat transfer between

In these equations, Q
air inside the cylinder and its walls and T is the
in
temperature of air entering the chamber, assumed
to be ambient temperature (T =T ). This model
in
amb
is widely referenced in the literature as correctly
describing temperature and pressure evolution inside
a pneumatic chamber [7, 10, 17]. Therefore, it will be
used as the datum model in this work.

3 MODEL ORDER REDUCTION

The model given by equations (4) and (5) is not
suitable for control purposes for the reasons presented in section 1. In order to simplify this model,
the temperature is naturally the state variable to
remove since force and motion state directly depend
on pressure (see equation (3)). This reduction is
usually performed in the literature by considering
Proc. IMechE Vol. 220 Part I: J. Systems and Control Engineering

J Falca o Carneiro and F Gomes de Almeida

304

temperature to follow the polytropic law

Model M

A B

P (n1)/n
(6)
T =T
0 P
0
Another relevant issue concerns the heat transfer
through walls. It is widely accepted (see, for example,
can be
references [7], [10], and [17] to [19]) that Q
correctly determined by
=l(P, T )A (x)(T T )
Q
q
amb
where

(7)

A B

PT 1/2
(8)
l(P, T )=l
0 P T
0 0
is the heat transfer coecient [19]. However, based
on the argument that the heat transfer coecient
is dicult to determine, classical works on servopneumatics do not use equation (7). Instead, the perfect gas equation PV=mRT is directly dierentiated,
giving
dP
P dV
R
P dT
=
+ T (m
m
)+
in
out
dt
V dt
V
T dt

(9)

When using a polytropic model for temperature

evolution, equation (9) reduces to
dP
P dV
R
=n
+n T (m
m
)
in
out
dt
V dt
V

(10)

In the model represented by equation (10), n is

the polytropic index that can be adjusted from
1 (isothermal process) to 1.4 (adiabatic process).
There are several examples in the literature that use
equations (6) and (10) with a further simplication;
although to achieve equation (10) a polytropic
temperature evolution was assumed, it is common
practice to consider that temperature uctuations
over equilibrium temperature are negligible and
therefore T=T . For instance, this model was used
0
in reference [3] with n=1, in references [7] to [10]
with n being experimentally tuned, and in reference
[20] with n=1.4. In order to compare these dierent
options, models M , M , and M are dened as
1
2
3
follows.
Model M
1
T =T
0
P dV
R
dP
=
+ T (m
m
)
in
out
dt
V dt
V
Model M
T =T

0
dP
P dV
R
=n
+n T (m
m
)
in
out
dt
V dt
V
Proc. IMechE Vol. 220 Part I: J. Systems and Control Engineering

T =T

0
dP
P dV
R
=c
+c T (m
m
)
in
out
dt
V dt
V
Note that, although models M and M are parti1
3
cular cases of model M , they will appear individually
2
so that their performance can be directly compared
with the other models.
In order to enhance the quality of the previous
models, a new model was proposed in reference [11].
Based on experimental evidence presented in reference [12], the model assumes that the incoming
ow process is adiabatic, the outgoing ow process
is isothermal, and the ow process due to piston
movement lies between isothermal and adiabatic
processes. This is achieved by considering dierent
polytropic indexes in equation (10): the incoming
ow term is aected by n=1.4, the outgoing ow by
n=1, and the piston movement term by n=1.2. This
model will be called M and is dened as follows.
4
Model M
4
T =T
0
dP
P dV
R
R
=1.2
+1.4 Tm
Tm

in V
out
dt
V dt
V
The models presented so far consider that temperature uctuations over ambient temperature are
negligible. In order to study the eects of this
assumption, a model similar to M but considering
2
temperature changes inside the chamber is considered. It is called M , was used for simulation
5
purposes in reference [10] with n=1.2, and is
dened as follows.
Model M

A B

P (n1)/n
0 P
0
dP
P dV
R
=n
+n T (m
m
)
in
out
dt
V dt
V
T =T

Models M to M are the typical models used in

1
5
servo-pneumatics literature. All these use a polytropic law for temperature when replacing dT/dt in
equation (9). As a consequence, these models lose
the heat transfer process that occurs by mixing
between air entering the chamber and the air inside
it. In order to evaluate the impact of this loss, model
M was dened as being similar to model M but
6
5
with a constant temperature in the incoming ow
term.
JSCE203 IMechE 2006

Reduced-order thermodynamic models

Model M

A B

P (n1)/n
0 P
0
P dV
R
R
dP
=n
+n m
T n m
T
dt
V dt
V in in
V out
T =T

is
Model M ends the set of models where Q
6
calculated in an implicit way. As previously stated,
this approach is justied in the classical literature
by the diculty in determining the heat transfer
coecient of equation (8). However, the present
authors have developed a simple procedure to
estimate it experimentally, based on the thermal time
constant method [21], and it is therefore pertinent
to evaluate the behaviour of models explicitly
accounting for the heat transfer. Furthermore, it
would be interesting from a mathematical point of
view to simplify the heat transfer model (7). In order
to do so, note that a simplied version can be
achieved by neglecting temperature and pressure
uctuations with respect to their equilibrium values.
The heat transfer coecient can then be expressed
as l(P, T )=l(P , T )=l and the heat transfer
0 0
0
becomes
=l A (x)(T T )
Q
0 q
amb

(11)

Furthermore, considering an average heat transfer

area A9 dened as
q

K A

p
l
A
9 =A (x )= w2+ pw x +
q
q 0
0 2
2

BK

x =0
0

and a heat conductance k dened as

0
k =l A
9
0
0 q

(12)

an even more simplied heat transfer model can

be obtained by substituting equation (12) into
equation (7) to give
=k (T T )
Q
0 amb

(13)

Using equation (13) as the explicit heat transfer

model leads to model M
7
Model M

A B

P (n1)/n
0 P
0
dP
P dV
R
=c
+c T (m
m
)
in
out
dt
V dt
V

T =T

c1
k (T T )
amb
V 0

JSCE203 IMechE 2006

305

Model M does not take into account the mixing

7
process, so model M is dened as similar to model
8
M with the mixing process considered.
7
Model M
8
P (n1)/n
T =T
0 P
0
dP
P dV
R
R
=c
+c m
T c m
T
dt
V dt
V in in
V out

A B

c1
k (T T )
amb
V 0

Finally, models M and M are similar to model

9
10
M but use progressively more complex heat transfer
8
models: model M uses equation (11) and model M
9
10
uses equation (7).
Model M

A B

P (n1)/n
0 P
0
dP
P dV
R
R
=c
+c m
T c m
T
dt
V dt
V in in
V out

T =T

+
Model M

c1
l A (x)(T T )
amb
V 0 q

10

A B

P (n1)/n
0 P
0
P dV
R
R
dP
=c
+c m
T c m
T
in
in
dt
V dt
V
V out

T =T

c1
l A (x)
V 0 q

PT
(T T )
amb
P T
0 0

Note that there are some interesting relations

between models implicitly and explicitly accounting
for heat transfer through walls; if an adiabatic process is considered in M (k =0; n=1.4), this model
7 0
is equal to M with an adiabatic process (n=1.4). If
5
an adiabatic process is considered in M , M , or M
8
9
10
(k ; l =0; n=1.4), these models are equal to M with
0 0
6
an adiabatic process (n=1.4). However, if an isothermal model process is considered in M , M , M ,
7
8
9
or M (k ; l =2; n=1), these models become equal
10 0 0
to M , which is intended to model adiabatic pro3
cesses. This inconsistency is justied by the simplication process leading to M ; although the pressure
3
index of M is adiabatic, temperature changes are
3
neglected. Table 1 reviews the main features of the
reduced models.
Proc. IMechE Vol. 220 Part I: J. Systems and Control Engineering

J Falca o Carneiro and F Gomes de Almeida

306

Table 1 Features of the reduced models

Model
M
1
M
2
M
3
M
4
M
5
M
6
M
7
M
8
M
9
M
10

Explicit heat transfer

through walls

Heat transfer
by mixing

Temperature
evolution

Pressure index

m
m
m
m

m
m
m

Constant
Constant
Constant
Constant
Polytropic
Polytropic
Polytropic
Polytropic
Polytropic
Polytropic

1
n
1.4
1, 1.2, 1.4
n
n
c
c
c
c

4 PERFORMANCE COMPARISON
To compare the performance of the dierent models
when predicting pressure, several simulation studies
on two types of symmetrical cylinder were run. The
cylinders features are presented in Table 2. The tool
used to perform the simulations was MATLAB/
Simulink with a DormandPrince integrator and a
xed integration step of 1 ms.
Each cylinder with the full-order model [equations
(4), (5), and (7) for each chamber] was tested in
two types of simulation: open-loop (Fig. 4) and
closed-loop (proportional) control (Fig. 5). The openloop simulation was excited by a pseudo-random
hit sequence (PRBS) signal (implemented with a
Gaussian random number generator followed by a
sign function) and the closed-loop simulation by
a Gaussian random number generator. In order to
prevent the piston from reaching the end positions
in the open-loop simulation, the sign of the input
signal to the valve was forced to change when the
Table 2 Features of the cylinders used to test the
performances of the models
Actuator

w
(mm)

l
(mm)

D
E

20
32

100
275

V (m3)
d

A9 (m2)
q

1.571106
1.106e105

3.77103
1.54102

Fig. 4 Open-loop simulation

piston reached 96 per cent of half the stroke in each

direction. As previously stated in section 2, the bandwidth of the pneumatic servo-system is limited by
the bandwidth of the pneumatic actuator, which
typically is lower than 10 Hz. Therefore, the Gaussian
and PRBS signals have a bandwidth of approximately
10 Hz in order to excite the system fully. The most
important features of the excitation signals are
presented in Table 3.
In terms of servo-valve features, values of k k =
x u
1107 m2/V for cylinder D and k k =1106 m2/V
x u
for cylinder E were assumed. Since the maximum
input to the servo-valve was limited to 10 V, these
parameters allow a maximum ow (choked ow
at supply pressure and ambient temperature) of
approximately 100 slpm for cylinder D and 1000 slpm
for cylinder E.
The thermal conductance k [=l A (x )] of three
0
0 q 0
industrial actuators was experimentally determined
and results ranged from approximately 0.2 to 0.5 W/K.
Therefore, the simulations were made using a central
range enclosing these values plus two extreme
situations: a very adiabatic k =0.02 W/K and a
0
very isothermal k =2.5 W/K. The heat transfer
0
coecients l for the equilibrium pressure P , tem0
0
perature T , and x =0 m were determined applying
0
0
these conductances to the particular cases of cylinders
E and D. The results are shown in Table 4.
After running the full-order model simulations, the
command signals u, position x, and velocity x were
collected to make each of the reduced models run
as presented in Fig. 6.
The polytropic index n of the models in section 3
was varied from 1 to 1.4 with a step of 0.5. Therefore, a total of 912 dierent simulations (four k
0

Fig. 5 Closed-loop simulation

Proc. IMechE Vol. 220 Part I: J. Systems and Control Engineering

JSCE203 IMechE 2006

Reduced-order thermodynamic models

307

Table 3 Main features of the excitation signals used

Generator properties
(random number generator of Simulink)

Open loop
Closed loop

Cylinder
Cylinder
Cylinder
Cylinder

D
E
D
E

Mean

Variance

Initial seed

Sample time (s)

0
0
0
0

0.0003 (V2)
0.3 (V2)
0.0003 (m2)
0.0021 (m2)

666
777
666
777

0.05
0.05
0.05
0.05

1000
70

(V)
(V)
(m)
(m)

Table 4 Heat transfer coecients used in the simulation study

l (W/K m2)
0

Cylinder D
Cylinder E

k =0.02 W/K
0

k =0.1 W/K
0

k =0.5 W/K
0

k =2.5 W/K
0

5.3
1.29

26.5
6.48

132.6
32.4

663.1
162.0

Fig. 6 Simulation of reduced models

values; two cylinders; two types of simulation; six

models with nine n values and three models with
constant n values) were needed. Each combination
k cylindertype will be called an experiment E
0
j
( j=1, 2, , 16), according to the coding used in
Table 5.
An important question is how to determine the
simulation time in order to guarantee an informative
experiment. For linear systems, this problem can
be solved by determining the settling time of the
systems free response. However, for non-linear
systems, this is still an open problem and, in order
to circumvent it, the settling time t of the non-linear
s
equations describing the cylinder behaviour was
(over)estimated. This was done in simulation by providing a constant zero excitation signal to the system,
applying an external force to move the piston and
then releasing the force, which caused the cylinder
to move to an equilibrium position. Note that in the
open-loop simulation the cylinders inlets and outlets are permanently closed during the experiment
since the servo-valve is assumed to have no leakage.
These simulations were run for all the heat transfer
coecients considered in this work, for cylinders D
and E and for the open- and closed-loop simulations.
In each of these, the settling time of pressure and
temperature were determined using a 1 per cent
criterion. As an example, Fig. 7 presents the results
obtained with this simulation for cylinder D, closedJSCE203 IMechE 2006

loop simulation, and k =2.5 W/K. The initial pressure

0
and temperature of chambers A and B are P and T
0
0
and the pistons initial position is x=0. A force of
300 N (Fig. 7(a)) is applied at time 0, causing the
piston to move against an end stop positioned at
x=0.015 m (Fig. 7(b)). The force is maintained
until stationary conditions are reached. This happens
at time 1.267 s; so at this instant the force is released.
The evolution behaviours of pressure and temperature in chamber A and of pressure and temperature
in chamber B are presented in Figs 7(c), (d), (e),
and (f) respectively. The settling times were calculated using a 1 per cent criterion applied to the DP
and DT values dened in these gures. The nal
pressure and temperature of chamber A and the
nal pressure and temperature of chamber B in
this example are P =5.276105 Pa, T =292.93 K,
A
A
P =5.294105 Pa, and T =293.12 K respectively.
B
B
Table 5 presents the settling times obtained for all
the experiments.
The values underlined in Table 5 are the highest
settling times for each cylinder and experiment. The
simulation times used for performance comparison
(Table 6) were chosen to be at least ten times higher
than these values. The performance criterion was the
error between the pressure given by the complete
model (equations (4), (5), and (7)) and the pressure
given by each of the models presented in section 3.
In order to take into account pressure in both
chambers, the error vector analysed was the concatenation of the error in chamber A with the error
in chamber B.
Considering the results obtained by each model
with n leading to the lowest mean square error (MSE)
(Fig. 8), it is seen that model M has clearly worse
4
results than all the others, and will be therefore
Proc. IMechE Vol. 220 Part I: J. Systems and Control Engineering

4.019
2.979
2.975
2.571
11.498
8.161
3.740
4.784
88.676
0.43324
35.950
110.460
P
A
P
B
T
A
T
B
Cylinder E

51.661
69.632
175.052
175.192

10.268
14.007
35.070
35.108

1.995
2.881
7.070
7.087

0.431
0.666
1.467
1.479

52.290
37.427
11.338
23.363

Experiment E
12
k =0.5 W/K
0
Experiment E
4
k =0.02 W/K
0
Experiment E
15
k =2.5 W/K
0
Experiment E
3
k =0.02 W/K
0

Experiment E
7
k =0.1 W/K
0

Experiment E
11
k =0.5 W/K
0

Experiment E
8
k =0.1 W/K
0

1.285
1.069
1.063
0.936
2.783
2.197
2.032
1.743
4.457
2.605
2.603
3.221
16.902
3.505
5.846
11.338
0.108
0.124
0.249
0.259
0.247
0.513
1.079
1.107
1.522
2.222
5.205
5.305
7.844
10.762
25.821
26.280
P
A
P
B
T
A
T
B

Experiment E
13
k =2.5 W/K
0
Experiment E
9
k =0.5 W/K
0
Experiment E
5
k =0.1 W/K
0
Experiment E
1
k =0.02 W/K
0

Open loop

excluded from most of further comparisons. On the

other hand, there is no unique best model for all
experiments; so the selected performance criterion
was the average mean square error (MSE) dened as
1 NE
MSE =
MSE
Mi N
MiEj
E j

Cylinder D

Experiment E
14
k =2.5 W/K
0
Experiment E
10
k =0.5 W/K
0
Experiment E
6
k =0.1 W/K
0
Experiment E
2
k =0.02 W/K
0

Closed loop
Setting time (s)

Table 5 Pressure and temperature settling times for chambers A and B

Experiment E
16
k =2.5 W/K
0

J Falca o Carneiro and F Gomes de Almeida

308

Proc. IMechE Vol. 220 Part I: J. Systems and Control Engineering

(14)

In equation (14), M stands for the model i,

i
i=1, 2, , 10 and N is the total number of experiE
ences (N =16). Figure 9 presents the average MSE,
E
the 10 per cent and 90 per cent percentiles of the
MSE for each model on a logarithmic scale. Three
levels of error appear: the high level consisting
of models M , M , and M , the central level con1
2
3
sisting of models M and M , and the low level
5
6
consisting of models M , M , M , and M .
7
8
9
10
Analysing the six best models (Fig. 10), the best
performance are achieved by models M , M , M ,
7
8
9
and M , which are essentially indistinguishable.
10
These results suggest that the best reduced models
are M , M , M , or M . Naturally, among these
7
8
9
10
models, M would be the natural choice since it is
7
the simplest.
In terms of the expected error and dispersion of
the models, and to cope with the dierent experiments, the expected value m and standard deviation s
of the error were determined as [22]
1
m =
m
(15)
Mi N
M ,E
E Ej i j
1/2
1
(16)
s2 +s2(m
)
s =
M ,E
Mi,Ej
Mi
N
E Ej i j
Table 7 presents the overall performance results for
all reduced-order models.
These results reveal the following.

1. Model M , although intended to be a compromise

4
between the inlet and outlet processes, gives the
worst results in this comparison.
2. Taking into account temperature changes inside
the pneumatic chamber can signicantly reduce
the pressure prediction error: model M has at
5
most 40 per cent of the average MSE of models
with xed temperature (Models M , M , and M ).
1
2
3
3. Although modelling the mixing process can
slightly reduce the pressure prediction error
(model M has an average MSE of about 85 per
6
cent of model M ), a more signicant error drop
5
is obtained when taking into account heat transfer
through walls; models M , M , M , and M have
7
8
9
10
at most 63 per cent of the average MSE of the best
model not including it (model M ).
6
JSCE203 IMechE 2006

Reduced-order thermodynamic models

309

Fig. 7 Determining the minimum simulation time required to perform an informative

experiment
Table 6 Simulation times
Simulation time (s)
Open loop

Cylinder D
Cylinder E

Closed loop

k =0.02 W/K
0

k =0.1 W/K
0

k =0.5 W/K
0

k =2.5 W/K
0

k =0.02 W/K
0

k =0.1 W/K
0

k =0.5 W/K
0

k =2.5 W/K
0

300
1800

60
360

60
120

60
60

300
1200

60
600

60
120

60
60

4. There is not sucient evidence of performance

gain by considering heat transfer dependences on
area, pressure, and temperature.
However, there is a practical shortcoming in these
results; they were derived using the best n parameter
for each model and experiment which is not, for the
six best models, constant (Table 8).
From a practical standpoint, it would be useful that,
given an experimental k measure of a pneumatic
0
JSCE203 IMechE 2006

cylinder, the best model and the (constant) n parameter to use could be determined. This should
be done for dierent levels of heat transfer: an
adiabatic level corresponding to k =0.02 and
0
k =0.1 W/K, a typical level corresponding to k =0.1
0
0
and k =0.5 W/K and an isothermal level corre0
sponding to k =0.5 and k =2.5 W/K. Results from
0
0
this exercise are presented in Table 9.
Figure 11 presents the average MSE and the 90 per
cent and the 10 per cent percentiles of the MSE on
Proc. IMechE Vol. 220 Part I: J. Systems and Control Engineering

J Falca o Carneiro and F Gomes de Almeida

310

Fig. 8 Average mean square error for all models

Fig. 9 Performance comparison: log (MSE) for the nine best models
10
Table 7 Expected values, standard deviation
and average MSE for all models
Model
M
1
M
2
M
3
M
4
M
5
M
6
M
7
M
8
M
9
M
10

(Pa)
Mi
1.02104
1.76103
5.63103
4.99104
6.74102
5.45102
6.14101
1.07102
1.21102
8.04101

s (Pa)
Mi

MSE

1.99104
7.62103
9.19103
1.87104
4.89103
4.53103
3.61103
3.41103
3.35103
3.37103

4.96108
6.11107
1.15108
2.84109
2.43107
2.08107
1.30107
1.16107
1.12107
1.13107

Mi

(Pa2)

a logarithmic scale. It is interesting to note that the

three performance levels highlighted in Fig. 9 also
appear for constant n values and furthermore their
relative performances are the same. The importance
Proc. IMechE Vol. 220 Part I: J. Systems and Control Engineering

of modelling temperature changes inside the cylinder

chamber is once again revealed since model M gives
5
at most about 50 per cent of the average MSE of
models with xed temperature (models M , M ,
1
2
and M ). Furthermore, this value is reduced to 30
3
per cent for typical k values. Modelling the mixing
0
process slightly enhances the results since model M
6
has at most 96 per cent of the error of models not
considering it (model M ) and this value is reduced
5
to about 88 per cent for typical k values. The average
0
MSE of models including direct heat transfer through
walls (M , M , M , and M ) are at most approxi7
8
9
10
mately 72 per cent of the models not considering it
(model M ). Once again, there is not a signicant
6
dierence between models M , M , M , and M .
7
8
9
10
Finally, for k values belonging to the range of
0
typical industrial actuators, the model with best
results when balancing performance and complexity
JSCE203 IMechE 2006

Reduced-order thermodynamic models

311

more, it has only about 10 per cent of the error

of classical isothermal, polytropic, and adiabatic
models. The expected value of pressure prediction
error with M is 140 Pa with a standard deviation of
7
2400 Pa.

5 CONCLUSIONS

Fig. 10 Performance comparison: mean MSE for the

six best models

Table 8 Best n for the six best models

Best n
Experiment

M
5

M
6

M
7

M
8

M
9

M
10

E
1
E
2
E
3
E
4
E
5
E
6
E
7
E
8
E
9
E
10
E
11
E
12
E
13
E
14
E
15
E
16

1.35
1.40
1.40
1.40
1.35
1.40
1.40
1.40
1.30
1.35
1.40
1.40
1.15
1.25
1.35
1.35

1.30
1.35
1.40
1.40
1.30
1.35
1.35
1.40
1.25
1.35
1.35
1.40
1.15
1.20
1.35
1.35

1.35
1.40
1.35
1.40
1.25
1.40
1.35
1.40
1.15
1.35
1.25
1.35
1.05
1.15
1.10
1.20

1.40
1.40
1.40
1.40
1.40
1.40
1.40
1.40
1.15
1.30
1.35
1.30
1.05
1.15
1.15
1.15

1.40
1.40
1.40
1.40
1.40
1.40
1.40
1.40
1.15
1.35
1.35
1.35
1.05
1.15
1.15
1.20

1.40
1.40
1.40
1.40
1.40
1.40
1.40
1.40
1.15
1.35
1.35
1.35
1.05
1.15
1.15
1.20

This work has focused on the thermodynamic model

of air inside a pneumatic cylinder chamber. Although
the use of reduced-order models to describe the
pressure evolution is widespread, the choice of which
model to select is typically made in an ad hoc way.
In order to guide this choice, a comparison
between classical reduced-order models and some
new models based on the heat transfer coecient
and thermal conductance of the cylinder was performed. It was shown that the pressure prediction
of reduced-order models can be enhanced by considering, rst, the explicit heat transfer between
cylinder walls and air inside its chambers and
second, temperature changes of air inside the
cylinder.
For typical heat transfer coecients of industrial
pneumatic actuators, considering these factors may
lead to an average MSE in pressure prediction of only
10 per cent of the MSE obtained when using classical
isothermal, adiabatic, or polytropic models.

ACKNOWLEDGEMENTS
is model M with n=1.35. It is essentially undis7
tinguishable from models M , M , and M and has
8
9
10
an average MSE of about 40 per cent of model M ,
6
the next model in terms of performance. Further-

The authors would like to acknowledge Professor

Sarseld Cabral for his help on statistical topics. This
work has been partially funded by Fundacao para a
Ciencia e Tecnologia under the programme POCTI.

Table 9 Expected value, standard deviation and average MSE for all models with the best constant n
k =0.02, k =0.1 W/K
0
0
Model
M
1
M
2
M
3
M
4
M
5
M
6
M
7
M
8
M
9
M
10

k =0.1, k =0.5 W/K

0
0

k =0.5, k =2.5 W/K

0
0

m
Mi
(102
Pa)

s
Mi
(103
Pa)

MSE
(106 Pa2)

m
Mi
(102
Pa)

s
Mi
(103
Pa)

MSE
(106 Pa2)

m
Mi
(102
Pa)

s
Mi
(103
Pa)

MSE
(106 Pa2)

110
8.80
46.0
510
5.60
9.60
1.47
5.50
5.20
5.50

21.0
6.7
7.10
19.0
3.10
2.60
2.78
2.30
2.30
2.30

530
46.0
70.0
3000
9.90
7.80
7.72
5.50
5.50
5.60

110
14.0
52.0
500
11.0
9.60
2.37
1.40
0.047
0.860

20.0
7.30
7.90
19.0
3.80
3.60
2.77
2.40
2.40
2.37

520
55.0
87.0
2900
16.0
14.0
7.58
5.60
5.60
5.50

93.0
9.30
66.0
490
11.0
15.0
4.72
2.40
6.30
2.40

20.0
9.90
11.0
18
6.80
6.60
5.29
5.20
5.10
5.10

460
97.0
160
2700
47.0
45.0
27.5
26.0
26.0
26.0

JSCE203 IMechE 2006

Proc. IMechE Vol. 220 Part I: J. Systems and Control Engineering

J Falca o Carneiro and F Gomes de Almeida

312

Fig. 11 Performance of all the models except M for three levels of heat transfer with the best
4
constant n

REFERENCES
1 Burrows, C. R. Eect of position on the stability of
pneumatic servosystems. Proc. Instn Mech. Engrs,
Part C: J. Mechanical Engineering Science, 1969,
11(6).
2 Zalmazon, L. A. Components for pneumatic control
instruments, 1965 (Pergamon, Oxford).
3 Outbib, R. and Richard, E. State feedback stabilization of an electropneumatic system. Trans. ASME,
J. Dynamic Systems, Measmt, Control, 2000, 122(3),
410415.
4 Ning, S. and Bone, G. M. High steady-state accuracy
pneumatic servo positioning system with PVA/PV
control and friction compensation. In Proceedings
of the IEEE International Conference on Robotics and
automation, Washington, DC, 2002, pp. 28242829
(IEEE, New York).
5 Andersen, B. W. The analysis and design of pneumatic
systems, 1976 (John Wiley, New York).
6 Chitty, A. and Lambert, T. H. Modelling a loaded
two way pneumatic actuator. J. Measmt Control,
1976, 9(1), 1925.
7 Richard, E. De la commande lineaire et non lineaire
en position des systems electropneumatiques. PhD
Proc. IMechE Vol. 220 Part I: J. Systems and Control Engineering

10

11

12

Thesis, Institut National des Sciences Amphquees,

Lyon, 1990.
Richard, E. and Scavarda, S. Comparison between
linear and nonlinear control of an electropneumatic
servodrive. Trans. ASME, J. Dynamic Systems, Measmt,
Control, 1996, 118, 245252.
Brun, X., Belgharbi, M., Sesmat, S., Thomasset, D.,
and Scavarda, S. Control of an electropneumatic
actuator: comparison between some linear and
non-linear laws. Proc. Instn Mech. Engrs, Part I:
J. Systems and Control Engineering, 1999, 213(5),
387406.
Brun, X. Commandes lineaires et non lineaires en
electropneumatique. Methodologies et applications.
PhD Thesis, Institut National des Sciences
Amphquees, Lyon, 1999.
Richer, E. and Hurmuzlu, Y. A high performance
pneumatic force actuator system: part I nonlinear
mathematical model. Trans. ASME, J. Dynamic
Systems, Measmt, Control, 2000, 122, 416425.
Al-Ibrahim, A. M. and Otis, P. D. R. Transient air
temperature and pressure measurements during
charging and discharging processes of an actuating
pneumatic cylinder. In International Fluid Power
Exposition and Technical Conference, 1992.
JSCE203 IMechE 2006

Reduced-order thermodynamic models

13 Toshiharu, K., Tokashiki, L., and Fujita, T. Inuence

of air temperature change on equilibrium velocity
of pneumatic cylinders. Trans. ASME, J. Dynamic
Systems, Measmt, Control, 2002, 124(2), 336341.
14 McCloy, D. and Martin, H. R. Control of uid power:
analysis and design. Engineering science, 1980 (Ellis
Horwood, Chichester, West Sussex).
15 Armstrong-Helouvry, B., Dupont, P., and Canudas
de Wit, C. A survey of models, analysis tools and
compensation methods for the control of machines
with friction. Automatica, 1994, 30, 10831138.
16 White, F. Fluid mechanics, 1994 (McGraw-Hill,
New York).
17 French, I. G. and Cox, C. S. Modelling, design and
control of a modern electropneumatic actuator. IEE
Proc., 1990, 137(3), 145155.
18 Kagawa, T., Tokashiki, L., and Fujita, T. Inuence
of air temperature change on equilibrium velocity
of pneumatic cylinders. Trans. ASME, J. Dynamic
Systems, Measmt, Control, 2002, 124, 336341.
19 Det, F., Scavarda, S., and Richard, E. Simulated and
experimental study of charging and discharging of
a cylinder by using an electro-pneumatic servovalve.
In JHPS International Symposium on Fluid power,
1989, pp. 199206.
20 Pandian, S., Hayakawa, Y., Kanazawa, Y.,
Kamoyama, Y., and Kawamura, S. Practical design
of a sliding mode controller for pneumatic actuators.
Trans. ASME, J. Dynamic Systems, Measmt, Control,
1997, 119, 666674.
21 Otis, D. R. and Pourmovahead, A. An experimental
thermal time constant correlation for hydraulic
accumulators. Trans. ASME, J. Dynamic Systems,
Measmt, Control, 1990, 112, 116121.
22 Guimaraes, R. and Cabral, J. Estatstica, 1999
(McGraw-Hill, Lisboa).

M
M
i
MSE
MSE
n
P
P ,P
A B
P
s
P ,P
u d
P
0
Q
R
slpm
T
T
amb
T
in
T
u
T
0
V
V
d
x, x, x
x
v
x
0

APPENDIX 1
Notation
A ,A
A B
A

A9
q
A
t
A ,A ,A ,A
1 2 3 4
E
j
F
f
k
f
k ,k
x u
k
0
l
m

areas of chambers A and B

respectively (m2)
heat transfer area (m2)

c
l
l

average heat transfer area (m2)

throat area (m2)
servo-valve restriction areas (m2)
experiment j
frictional force (N)
friction coecient (N s/m)
servo-valve parameters (mm)
(mm /V)
thermal conductance at
equilibrium conditions (W/K)
actuator stroke (mm)
mass ow entering or leaving the
cylinder chamber (kg/s)

JSCE203 IMechE 2006

313

mass of the moving parts of the

actuator plus load mass (kg)
model i
mean square error in the pressure
prediction (Pa2)
average mean square error in the
pressure prediction (Pa2)
polytropic index
absolute pressure inside the
actuator chamber (Pa)
absolute pressures of chambers A
and B respectively (Pa)
absolute supply pressure (Pa)
absolute upstream and
downstream pressures respectively
(Pa)
absolute equilibrium pressure (Pa)
heat transfer (W)
perfect gas air constant (J/kg K)
ISO standard litres per minute
(T=293.15 K; P=101 325 Pa)
temperature inside the actuator
chamber (K)
ambient temperature (K)
temperature of the air entering the
actuator (K)
upstream temperature (K)
equilibrium temperature (K)
volume of the actuator chamber
(m3)
dead volume (m3)
piston displacement (m), velocity
(m/s) and acceleration (m/s2)
spool displacement (mm)
central position of the piston
(mm)
ratio of specic heats for air
heat transfer coecient (W/K m2)
heat transfer coecient at
equilibrium conditions (W/K m2)
actuator diameter (mm)

APPENDIX 2
Equilibrium pressure
Consider the half-bridge model of a servo-valve
represented in Fig. 12. m
and m
represent the leak1
2
ages of restriction 1 and restriction 2 (see Fig. 1) and
the spool is at the central position. At equilibrium
T =T , A =A , and m
=m
. In the typical situation
s
1
2
1
2
where P 3.6P , there are three possible situations:
s
atm
Proc. IMechE Vol. 220 Part I: J. Systems and Control Engineering

314

J Falca o Carneiro and F Gomes de Almeida

Equalizing m
and m
in the rst situation gives
1
2

2c
A P
1 s (c1)RT
=

CA B A B DH
P 2/c
P (c+1)/c

P
P
s
s

1/2 PA
2c
2 1/(c1)
2
c+1
(c1)R
T 1/2

1/2

(17)

Fig. 12 Half-bridge model of a servo-valve

P=P , P [0.5283P , P ], P=P , P [1/0.5283P ,

1 1
s s
2 2
atm
0.5283P ], and P=P , P [P , 1/0.5283P ]. In
s
3
3
atm
atm
the rst situation, m
is subsonic and m
is sonic.
1
2
In the second situation, m
and m
are sonic. In the
1
2
third situation, m
is sonic and m
is subsonic.
1
2

Proc. IMechE Vol. 220 Part I: J. Systems and Control Engineering

The solution for equation (17) when the uid is

air, which is assumed to be a perfect gas, gives
P=0.8077P . Note that the same exercise when
s
applied to the second and third situations would
result in false propositions.

JSCE203 IMechE 2006