2009 American Control Conference

Hyatt Regency Riverfront, St. Louis, MO, USA

June 10-12, 2009

FrC10.1

Fine and simplified dynamic modelling of complex hydraulic systems

Wilber Acuna-Bravo, Enrico Canuto, Stefano Malan, Davide Colombo, Marco Forestello and Riccardo Morselli
Abstract This paper deals with the dynamic modelling of a
complex electro-hydraulic system. Modelling is based on
physical laws and the system knowledge. The main idea is to
obtain a simple and reliable model that can be used for
controller synthesis and implementation by using the
architecture of Embedded Model Control. Simplifications are
made by taking as basis the ideas of singular perturbation. A
subsequent identification procedure is made in order to acquire
some important parameters, required for carrying out a
simulation.

I.

INTRODUCTION

Model building is a constant issue in control engineering

since it is needed for many different tasks, such as analysis,
control design, training, etc. The accuracy level relies on the
type of application it is intended for.
The paper is concerned with fine and simplified modelling
of hydraulic systems in view of control design and testing.
The adopted approach follows the framework of the
Embedded Model Control given in [1]. In this approach the
control law is designed and implemented around a discretetime simplified dynamics of the plant to be controlled,
enhanced with the uncertainty dynamics to be rejected to
guarantee performance. The problem of model simplification
is well known in the control literature, where the main goal
is to look for efficient algorithms capable of reducing the
dynamics within a specified bound of the modelling error
between the complex and the simplified model under the
same class of commands. Different techniques have been
proposed: singular perturbations, singular values, etc., all of
them addressing linear dynamics.
Here a rather different approach is followed from two
different standpoints. The so called true model, here called
fine model, which for complex hydraulic systems is
usually assembled from standard component models and
commercial packages like AMESIM, is studied for each
component to enhance the dynamics within a frequency
bandwidth, which is sufficiently larger than the target
W. Acuna-Bravo, E. Canuto and S. Malan are with Politecnico di Torino,
Dipartimento di Automatica e Informatica, Corso Duca degli Abruzzi 24,
10129 Torino, Italy (e-mail: stefano.malan@polito.it)
D. Colombo and M. Forestello are with Centro Ricerche FIAT S.C.p.A.,
Vehicle Systems, Active Systems, Strada Torino 50, 10043 Orbassano
(Torino), Italy (e-mail: davide.colombo@crf.it)
R. Morselli is with CNH Italia S.p.A.., Innovation Competence Center, V.
delle Nazioni 55, 41100 Modena, Italy (e-mail: riccardo.morselli@cnh.com)

978-1-4244-4524-0/09/$25.00 2009 AACC

control bandwidth. Under this approach a simplified model

version is obtained, which is referred as neglected
dynamics in [1] also called unstructured uncertainties in
control literature. This is not an easy task, since component
data from manufacturers do not care for dynamics but very
often only for static performance, especially when the
system includes hydraulic compensation devices designed
and arranged to ensure them. On the other hand, such models
do not need to be very accurate, say to know exact time or
frequency responses, especially when their dynamics are
expected to lie beyond the target control bandwidth. Usually,
as it will be shown, a triple of dynamics have to be
accounted for, namely, mass dynamics of the moving parts
of valves, distributors, pumps, usually of the second order,
pressure dynamics (continuity equation) of the lumped
hydraulic capacities of each device; hydraulic links, of
accumulators, usually first order dynamics, and finally
electromechanical dynamics of driving solenoids when
appropriate. The pressure dynamics is usually interconnected
to mass dynamics through the flow permitted by the moving
parts, where pressure enters in a nonlinear way if flow
becomes turbulent. Solenoid dynamics is usually the fastest
one, being regulated by appropriate electronics. The result is
a set of nonlinear state equations with their working
conditions and limits, where state variables are position,
velocity, pressure, current; commands are solenoid driving
voltage; measurements are pressures, flows, currents;
disturbances are leakages, flows, discharge pressures,
frictions, etc. An equilibrium point is then looked for, not for
dynamic linearization but to fix the intermediate point of the
different variables, for instance zero-hydraulic positions, and
their working limits.
Simplification method looks very similar to singular
perturbation method [2][3] and may be facilitated to certain
extent by hydraulic systems to be designed as the
interconnection of well defined lumped dynamics. As a
baseline when cut-off frequencies (or eigenvalues) of a set of
equations are outside a pre-specified bandwidth, equations
are reduced to be static by zeroing the derivative term on the
LHS of the equations. From a hydro-mechanical standpoint
it corresponds to assume negligible fluid volume variation
p iVi / i [m3/s] due to pressure rate Vi / i , where Vi is the

5480

Ka

Pilot cylinder

Load

Pilot cylinder

Pilot valve

Pilot solenoid

fp

Solenoid

xc
Pl

xa

Cradle

Swash plate

Kb

Hydraulic actuator
Mechanical feedback

Pi

Bias spring

Fig. 2.
Fig. 1.

Spool
Vp

ld

Axial piston

variable pump

Kc

Feedback spring

Relief valves

Essential hydraulic circuit.

capacitance volume, 1 / i the fluid compressibility; to

assume small inertia forces mk xk of the moving parts under
the balance of hydraulic, visco-elastic and solenoid forces or
to assume a small current error rate d ( I ref I ) / dt . This
method assumes weak interconnection between different
dynamics, which may not occur and therefore must be
carefully examined by frequency or modal analysis of the
overall equations or part of them.
The paper is organized as follows, section 2 is devoted to
system description and problem statement; section 3 boards
the development of a fine model for the system under study.
Sections 4 and 5 deal with the simplification of the complex
obtained model, as well as the transfer function analysis;
section 6 considers an identification process of some
parameters. Finally some conclusions are presented.
II.
PLANT DESCRIPTION
A.
The hydraulic circuit
The plant under study is a subset of a more complex
hydraulic circuit supplying loads to off-highway vehicles.
The essential elements of the network shown in Fig. 1 are
1) an axial-piston variable pump driven by an electrohydraulic actuator,
2) a set of relief valves to limit pressure in the circuit,
3) the hydraulic load, composed by several parallel loads,
part of them to be continuously supplied (priority loads)
and part of them on demand,
4) a load-sensing control system, not represented in Fig. 1,
fed by load measurements and in charge of regulating
the pump flow according to load demand.
The paper will restrict to the sole pump and the electrohydraulic actuator, while the load is simplified to be a
passive pressure drop device as in Fig. 1 (port B).

Basic axial pump diagram.

B.
The pump
The pump is an axial-piston variable type (see Fig. 2), the
flow being varied by tilting a swash plate against a bias
spring with stiffness K b . The plate tilting and the subsequent
flow regulation may be obtained in two ways:
1) active regulation: the plate tilt is measured by a suitable
sensor and fed to a control loop commanding the
electro-hydraulic actuators [4][5];
2) mechanical feedback: the hydraulic actuator stroke xa
(the pilot cylinder in Fig. 2) is servoed to the spool
position xc of a proportional valve by means of an
elastic link (the feedback spring in Fig. 2) having
stiffness K a .
The second (passive) solution is more compact and robust as
the sensor becomes useless and only one pilot cylinder may
be employed. Careful design ensures the pump flow Qp to
be proportional to the current is of the pilot solenoid, under
steady state conditions, less some hysteresis due to friction.
The goal of the paper is to derive the state equations of the
pump and the actuator (fine model) and then to simplify
them (design LTI model) within the bandwidth required by
the load-sensing control.
A complete derivation of the open-loop pump dynamics,
starting from the dynamics of each pump piston, can be
found in the literature [6][7][8]. As shall be proved, pump
dynamics is not the core of the target model, due to
mechanical feedback which replaces the pump closed-loop
and therefore simplifies current to flow dynamics.
III.
FINE MODEL
A.
Mechanical equivalence
The mechanical chain from solenoid to plate can be
represented as in Fig. 3 by masses and springs. Hydraulics
enters through the pressure forces.

5481

Modelling assumptions are:

1) electrodynamics of the solenoid is neglected, under
assumption of a suitable current regulator, with
bandwidth better than 1 kHz,
2) each body is considered as a second order mass-spring
system with position and rate as state variables;
3) rigidly connected elements as plate and actuator body
will be reduced to a single body;
4) equations are given in a variational way with respect to
a well specific equilibrium condition corresponding to
zero spool position, namely xc = 0 , and the central plate
position, defined by the zero actuator stroke, i.e. xa = 0 .
note the resulting equations are not small variation
equations, but large variations, holding in the whole
range of the state variables.
Kc , Fc0
K c , Fc 0

Kb , Fb 0
(disc)
Cradle

xcA

Sd Pl

Actuator
(piston)

s Is

Spool
Ka , Fa0

Sa Pa

xc,max
Rb

Fig. 3.

X a , va

Ra

X c , vc

Rc

Mechanical equivalent

At the end a state equation system of the sixth order is

obtained, driven by the solenoid current is and the pump
discharge pressure Pl , in the following referred to as line
pressure.
B.
Spool equation
According to sign conventions in Fig. 3 and the above
guidelines, the second order spool equation is given by
mc 
xc ( t ) = Ac ( xc , xc ) ( K c + K a ) xc K a xa + s is
(1)
xc ( 0 ) = xc 0 , xc ,max xc xc ,max , I max is I max
where xc is the spool displacement, mc the spool mass,
K c << K a the stiffness of the bias spring defining the spool
rest position xc ,max . Finally, Ac accounts for friction and
fluid forces, whereas s is the solenoid electro-mechanical
constant. The symmetric current range defined by I max is the
flow regulating range, corresponding to the actual current
range
(2)
0 < I min I s = I min + is + I max I min + 2 I max .
The current region below I min is used to move the spool in
the regulation range, an issue not treated here.
C.
Actuator equations
As the pilot cylinder body is assumed to be rigidly
connected to the swash plate through a spherical joint, only
the hydraulic equation is considered here. Denote the current
hydraulic volume with Va and the active area with S a ;
neglecting leakages, the continuity equation is written in
terms of the total actuator internal pressure Pa as follows

Pa ( t ) = Qa ( t ) + Sa xa ( t ) , Pa 0,
Va

(3)

where is the compressibility coefficient and Qa is the

input/output flow, which is assumed as positive when going
out [5]. Equation (3) must be completed with the flow
equations
Qa ( t ) = Qa ,in ( t ) = a xc Pl Pa , xc < 0, Pl Pa 0
Qa ( t ) = 0, xc = 0
Qa ( t ) = Qa ,out ( t ) = a xc Pa , xc > 0, Pa 0

(4)

a = ha 2 / ,
where is the fluid density and ha is the equivalent height
of the orifice, assumed constant at any xc . Note the flow Qa
is assumed to be proportional to the spool displacement xc
which may be approximate at the extremes of the spool
stroke. Note further the flow is zero for xc = 0 .
D.
Plate dynamics
As was already said, a simplified equation of the plate
dynamics is reported, which is however coherent with the
simplified equations reported in the literature [8]. By
assuming actuator and plate rigidly connected, a single
equation may be written, and directly in the actuator stroke
xa , which is proportional to plate tilt d through the
kinematic link
(5)
xa = d l d ,
with ld being the actuator arm. Then, the force balance on
the actuator and plate ensemble reads
ma 
xa ( t ) = Aa ( xa , xa ) K b xa K a ( xa + xc ) pa S a pl S d
(6)
xa ,max xa xa ,max
where Aa accounts for friction and fluid forces, ma is the
equivalent mass of the ensemble and S d << S a is the total
active area of the line pressure in the pump pistons,
generating a force over the swash plate. Equation (6) being a
variation equation in xa , has been written in terms of the
pressure variations
pa = Pa Pa , pl = Pl Pl
(7),
where the fixed pressures Pa , Pl are defined by the bias
spring preload.
E.
Pump output dynamics
The swash plate tilting allows to express the discharge fluid
volume as proportional to d and then to xa . The rotation
of the tilted pistons allows expressing flow as proportional to
1
shaft angular frequency f p = p ( 2 ) in Hertz units, p
being the shaft angular rate. At the end, the pump flow
equation can be written as follows,
1 x (t )
Qp ( t ) = Vp f p + a
(8)
= Q p + q p ( t )
2 x
a ,max

i.e. as the sum of the mean flow Qp and its variation q p ( t ) .

5482

pl
xc 0
Qa

is

xc

Ka

pa

1
Sa

a P ( xc )

x a

xa

Vp f p
xa ,max

qp

ql

Vl

xa

u u

p l

pl

pu

Fig. 4.

Block scheme of the simplified model.

xc =

Load dynamics
F.
Load dynamics may be very complex; here it is assumed a
single fluid volume Vl at the line pressure Pl supplied by
the pump flow, and discharged by a variable flow depending
on Pl . Actually, the output flow must be complemented with
the actuator input flow Qa ,in only existing when xc 0. The
corresponding equation holds

1/
Pl =
Q p + Qa uu ( Pl Pu ) u , xc < 0
Vl

1/
Pl =
Q p uu ( Pl Pu ) u , xc 0,
Vl

(9)

where the output flow may be driven by the load pressure

Pu Pl and/or the aperture degree 0 u 1 , and
1 u 2 .
IV.

SIMPLIFIED MODEL

The methodology followed here takes advantage of the so

called singular perturbation method [2][3][9], that averages
state variables, inputs and outputs over a given time period
T . To this end, small parameters, denoted as = T , premultiplying the n -th derivative (for example, n = 1 for the
case of continuity equations and n = 2 in case of massspring equations), make them so small to be approximated to
zero, except along small time intervals comparable with ,
playing the role of a time constant. Moreover,
c = mc / K c 5 ms holds for the spool equation,
a = Va / ( Qa / Pa ) 0.1 ms for the actuator hydraulics
and b = ma / Kb 5 ms for the plate equation. Roughly
speaking, if the load-sensing control averages over times
longer than T = 50 ms , the above dynamics can be
neglected, with an error that can be evaluated by the
perturbation method.
A.
Spool dynamics
Singular perturbation applied to equation (1) and
K c << K a provides

s is
Ka

xa ,

(10)

showing the spool position playing the role of the control

error between the current command and the actuator
displacement; since xc = 0 at steady state condition, from
equation (10) xa is proportional to is and pump flow
variation q p in equation (8) can be rewritten as
i
(11)
q p (t ) = Vp f p s ,
I max
which is the standard form of the linear pump characteristics.
B.
Actuator and plate equation
Singular perturbation applied to equations (3) and (6)
provides
xa = Qa / Sa
(12)
K x + s is + pl Sd
K x + s is ,
pa = b a
 b a
Sa
Sa
where second equation relates the actuator pressure to
command and actuator displacement.
Load equation
C.
Assuming for simplicity u = 1 in (9), and q p from (8), the
load equation holds

p l = ( bp xa + Sa x a uu ( pl pu ) ) , xc < 0
Vl
(13)

p l = ( bp xa uu ( pl pu ) ) , xc 0
Vl
where bp = V p f p xa ,max and Pl defines the equilibrium for
a constant Pu .
The block diagram for this simplified version of model is
depicted in Fig. 4.
V.

TRANSFER FUNCTION ANALYSIS

As a next step, a transfer function synthesis is done. As can

be seen from equations (10)-(13), the simplified model
remains non-linear for variations of Pa and Pl , more
specifically the signal Qa is a non-linear function of the
difference between the line and actuator pressure, whom in
turn depends on the variation of spool position xc . Once

5483

again an analysis for each condition of xc (positive or

negative) must be done.
The linearization procedure is based on fixating the
variations of signals with respect to their nominal values.
Pa or
Pl Pa as a fixed
i.e., taking the functions
parameter.
A.
Positive error xc > 0
This situation corresponds to an increase of pump output
flow, it means, there is no feedback flow to the pilot valve.
From (10)-(13) and replacing the pilot valve flow (4), the
linear system is given by
P

a a
0
a a s
0
xa
Sa
xa S a K a
is
p =
p +

u u pu
l bp
l
0
u u
Vl

Vl
Vl
pl 0 1 xa
q = b 0 p
l p
l
(14)
where ql is the line flow (see Fig. 4).
Negative error xc < 0
B.
This situation corresponds to a drop in the output pump
flow. Additionally there is a flow request from the pilot
valve, in order to reach a minimum pressure level to reduce
the plate inclination. By following same guidelines as for
xc > 0 , the linear state space is given by

P
0
a

xa
Sa
xa +
p =

u u pl
l p
b p a P
Vl
Vl

P
s
a
S
K

a a
+
a P s
VK
l
a

is

u u pu
Vl
0

0
pl
q =
l b p a P

(15)

1 xa
+ P s
0 pl a
Ka

0
is
,
0 pu

uu

sVl + u u
0

Pa ,
xc > 0
(18)
f p ( Pl , Pa ) =
Pl Pa , xc < 0.
It is worth noting that transfer function given by (16) is valid
for the non linear system as well; more exactly this
approximation is valid in the case of small variations of
terms pa and pl .
Function (17) reveals the existence of a zero when there is a
flow drop in the pump. A further analysis to this fact reflects
a time constant of the zero around z 10 ms , which can be
considered negligible with respect to predominant dynamics
within the system.
In order to complete the modelling process, some parameters
must be estimated. To this aim, an identification process is
effectuated, regarding, for a first instance, only the path from
is to ql , that is to say the element 2,1 of matrix (16). The
proposed model obeys to a first order delayed system, i.e.,
G0 std
(19)
Gid ( s ) = G21 ( s ) e std =
e ,
1+ s
with td , the time delay and time constant respectively and
G0 corresponds to the DC gain. Note that the time delay td
(not present in model (16)) is added just to better fit the
acquired data and it is due not only to the neglected
dynamics but mainly to signal transmission delays and
sensor delays of the experimental setting. The estimation
procedure is based on a non linear least square optimisation
process.
As a first step in the identification, it is necessary to acquire
the seed for the optimisation process. This initial information
is obtained from the data of tests made over the plant. This
identification process is effectuated only for the growing
flow zone, i.e., xc > 0 .
The cost function to be minimized is given by
N

J = ( ql , m ql ,i ) ,
2

(20)

i =1

where P = Pl Pa .
The matrix transfer function of systems (14) and (15) is
G (s) =

f p ( Pl , Pa ) f z ( s, xc )

sS a + a f p ( Pl , Pa ) ( sVl + u u )
=

f p ( Pl , Pa ) f z ( s, xc )

sS a + a f p ( Pl , Pa )

where s stands for the Laplace variable, = a s K a and

the functions
xc > 0
bp ,
(17)
f z ( s, xc ) =
bp + sS a , xc < 0

(16)

where terms ql , m , ql ,i are the measured and estimated flow

respectively. The estimated flow is generated in the search
process by evaluating the estimated parameters into (19). It
should be clarified that the data set is finite and discrete,
however, it is being evaluated in the continuous time transfer
function (19) in order to generate the required sequence ql ,i .
The used optimisation algorithm is called Nelder-Mead
simplex search method and it is codified in Matlab under the
fminsearch() function name. From the identification
process, applied to nine different data sets of acquired
signals, parameters value and uncertainty are obtained:

5484

 = 100 2 ms,

G0 = 100 0.7,
VI.

td = 55 5.5 ms.

SIMULATION RESULTS

0.08
0.06
Estimation error: ql,m-ql,i

In order to show the effectiveness of the obtained models,

some simulations were carried on and compared with the
acquired data.
Fig. 5 shows simulation results of the optimization process
of Section V. , for one data set, and of the simplified model
described in the block scheme of Fig. 4 of Section IV. Dash
dotted and solid lines correspond to measured current input
and measured flow output respectively. Dashed line
represents the simulated flow output obtained at the end of
the optimization procedure, while dotted line represents the
simulated flow output obtained from block scheme of Fig. 4.
Note that all these signals are normalized, for the sake of
confidentiality, and that the measured signals are filtered in
order to eliminate the dither contribution.

of them with a local compensation, in the procedure shown

here, for the sake of simplicity this load system has been
assumed to be a pressure drop. The called fine model has
been obtained from the physical system behaviour, together
with the analysis of preloaded equilibriums. Once this fine
model was presented, a simplified version was obtained,
turning into a local linearized system model, that anyway
shows a time behaviour strictly fitting actual acquired data.
At the end of the paper a simple identification process was
presented, with the purpose of acquire some important
parameters required to complete the system modelling.
Future work will be devoted to the analysis and modelling of
the load system, as well as the corresponding simulation of
the obtained model.

1
0.8
0.6
0.4

0.02
0
-0.02
-0.04
-0.06

is
0.2

0.04

-0.08

ql,m

2
time [s]

ql,i
0

ql,s
0

Fig. 5.

0.5

1.5

2
time [s]

2.5

3.5

Fig. 6.

Estimation error

REFERENCES

Pump responses, measured and identified.

[1]

Canuto, E. Embedded Model Control: Outline of the theory. ISA

Trans. Vol 46, no.3, 2007, pp.363-377.

Moreover, Fig. 6 presents the final estimation error

ee ( t ) = ql , m ( t ) ql ,i ( t ) obtained.
As can be seen, the distance between the responses is small
enough as to be accepted the parameters. The variations
observed in the plot, are mainly due to quantization effects
within the measured signal.

[2]

Kokotovic P., Khalil H.K., OReilly J. Singular perturbation methods

in control: analysis and design, Academic Press, 1986.

[3]

Kevorkian, J. and J.D. Cole, Perturbation Methods in Applied

Mathematics. Springer-Verlag, 1981.

[4]

Akers A, Lin S.J. Control of an axial piston pump using a single-stage

electro-hydraulic servovalve, in Proc. American Control Conference,
1982, pp. 1865-1870

[5]

Kugi, A. Non-linear Control Based on Physical Models. In: Lecture

Notes in Control and Information Series 260. 2001, Springer.

[6]

Zhang, X., J. Cho, S.S. Nair, and J. Manring Reduced Order

Modeling of the Dynamics of an Indexing Valve Plate Pump. In:
Proc. of the American Control Conference. 2000, pp.3595-3599.

[7]

Zhang, X., J. Cho, S.S. Nair, and J. Manring Damping on the Swash
Plate of an Axial-Piston Pump. In: Proc. of the American Control
Conference. 2000, pp.3590-3594.

[8]

Zhang, X., J. Cho, S.S. Nair, and J. Manring New swash plate
damping model for hydraulic axial-piston pump. Journal of Dynamic
Systems, Measurement, and Control. Vol. 123, 2001, pp.463-469.

[9]

OMalley, R.E. Introduction to Singular perturbations. Academic

Press, 1974.

VII.

CONCLUSIONS

In this note, a modelling process of a complex hydraulic

system has been shown. The system under study is highly
complex and non-linear, however as a first step just the basic
feeding component has been analyzed, with a subsequent
simplification of the obtained model.
A key issue in this modelling process relies in the load
attached to the pump, since this load may affect in a large
extent the general system behaviour. Even though in the real
system the load is composed of a set of mechanic-hydraulic
elements such as non-return valves or load distributors, most

5485