Professional Documents
Culture Documents
在装有共轨喷射系统的内燃机中,空气/燃料混合物的精确计量严格依赖压力调节,特别是压缩天然气(Cng)喷射系统的精确计量是很难达到的,因为
Abstract 气体的压缩性使得燃料输送过程更加复杂,由于控制器设计需要喷射系统的模型,本文提出了一种基于物理的状态空间模型。一种新颖的CNG喷射系
统,模型参数只依赖于定义良好的几何数据和燃料特性,通过对不同工况下的仿真和实验结果进行比较,验证了该模型的有效性
In internal combustion engines equipped with the common rail injection system the accurate metering of the air/fuel mixture strictly
depends on the pressure regulation. Accuracy in metering is difficult to be achieved especially for compressed natural gas (CNG)
injection systems, as the gas compressibility makes the fuel delivery process more complex.
Since the controller design requires a model of the injection system, this paper presents a physics-based state-space model of an
innovative CNG injection system. The model parameters only depend on well defined geometrical data and fuel properties. Comparing
simulation and experimental results in different operating conditions validates the model.
Further, the proposed model is used for designing a generalized predictive controller for the injection pressure regulation, which is
implemented in few steps. Experimental results show the effectiveness of the proposed approach.
r 2008 Elsevier Ltd. All rights reserved.
Keywords: Injection pressure control; Common rail; Modelling injection systems; Compressed natural gas; Polluting emissions reduction; Predictive
control
0967-0661/$ - see front matter r 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.conengprac.2008.01.008
ARTICLE IN PRESS
P. Lino et al. / Control Engineering Practice 16 (2008) 1216–1230 1217
Recently, to overcome these difficulties, the FIAT Re- the simplicity is a focal point, this paper is specifically
search Center, Valenzano branch, Italy, experimented an aimed at defining a simple enough control-oriented model,
innovative common rail injection system for CNG engines which at the same time ensures reliable data used to design
(Amorese et al., 2004; Maione et al., 2004), in which an a predictive control law. All assumptions are clearly
Electronic Control Unit (ECU) controls the opening/ specified at each stage of the modelling procedure. Final
closing rate of the electro-valve feeding the rail to achieve results are validated by comparison with experimental
an effective control of the pressure in the rail and, hence, an data.
accurate metering of the fuel by means of the electronically This paper is organized as follows. Section 2 describes
driven injectors. In addition, this controller can modify the the new CNG common rail injection system. Section 3
working point of the pressure in the rail, for adapting it to develops the state-space model of the injection system used
the changes of load, to the request of acceleration, etc. in Section 4 to derive the predictive control law for the
This paper describes a model-based, predictive control of injection pressure regulation. In Section 5, both experi-
the pressure in common rails of injection systems for GNG mental and simulation results are presented to validate the
engines. Obviously this technique is closely dependent on proposed model and to evaluate the controller perfor-
the assumed process model, which is the main block of the mances. Section 6 gives the conclusions.
predictive control. In other words, only a model generating
accurate output predictions fits the purpose (Qina & 2. The CNG injection system
Badgwell, 2003; Rossiter, 2003). Unfortunately, there are
relatively few papers which describe modelling and control The main elements of the CNG injection system of Fig. 1
of injection systems for CNG engines. On the contrary, are a fuel tank storing high pressure gas, a pressure
similar problems are frequently considered for diesel reducer, a common rail and six electro-injectors. The fuel
engines, so that many methodologies can be transferred coming from the tank supplies the pressure reducer before
to CNG injection systems. In particular, a well known reaching the common rail and feeding the electronically
modelling method (Heywood, 1988; Streeter, Wylie, & controlled injectors. By supplying gas to the intake
Bedford, 1998), also used in this paper, splits the system manifolds, injectors lead to the proper air/fuel mixture.
into its main volumes and describes the continuous flow of Note that the large volume of the common rail helps in
the fluid through these volumes. However, if the fuel damping the oscillations due to the operation of both
compressibility and the pipeline flexibility are taken into pressure controller and injectors. Namely, combining the
account, the description is often in terms of partial electronic control of rail pressure with optimum design of
differential equations (PDE) (Catania, Dongiovanni, Mit- the rail volume reduces the pressure oscillations inside the
tica, Negri, & Spessa, 1996; Cantore, Mattarelli, & Boretti, rail and leads to a more accurate fuel metering. More
1999; Ficarella, Laforgia, & Landriscina, 1999; Hountalas precisely, since the pressure ratio between the intake
& Kouremenos, 1998; Kouremenos, Hountalas, & Kour- manifold and the rail pressure is below the so-called
emenos, 1999). These detailed models are useful to ‘‘critical value’’, the flow rate depends only on the rail
diagnose faults and to explore different system configura- pressure (see Section 3 for details). Hence, the injected fuel
tions and alternative geometrical and functional designs, amount can be metered acting on rail pressure and
etc. However, for control problems, they are too complex injection timings, which are precisely driven by the ECU.
and contain too large amount of parameters to be useful. The output of a piezo-resistive pressure sensor inside the
Fortunately, a recognized and important property of rail is processed to drive the injectors and then to close the
feedback control is that one can often achieve satisfactory loop with the pressure reducer consisting of two chambers.
levels of performance by means of relatively simple models, The inflow section of the main chamber is variable with the
provided that they capture the relevant features of the axial displacement of a spherical shutter over a conical
controller design. In the proposed study, for designing a seat. The pressure of the second (control) chamber is
model-based predictive controller, there is no need to regulated by a solenoid valve. A piston placed between
describe all the physics and internal behaviour of the these chambers varies the shutter opening, which controls
injection process to get a model providing reliable the fuel flow rate. When the solenoid valve is opened, the
prediction. Namely, only the more significant character- fluid increases the pressure on the upper surface of the
istics can be taken into account leaving most secondary piston, which is pushed down together with the shutter.
effects un-described mathematically. In this way, without Consequently, more fuel enters the main chamber and
losing essential validity, the injection system can be forces the pressure to increase. When the solenoid valve is
represented by ordinary differential equations (ODE) closed, the pressure on the upper side of the piston
(Aespy, Engja, & Skarboe, 1996; Desantes, Arregle, & decreases. Since the direction of the resulting force is
Rodriguez, 1999; Fuseya, Nishimura, Sato, & Tanaka, upward, the piston moves up and allows a preloaded spring
1999; Gauthier, Sename, Dugard, & Meissonnier, 2005; to close the shutter of the main chamber.
Morselli, Corti, & Rizzoni, 2002), by transfer function The solenoid valve, which admits the fluid in the control
(Kuraoka, Ohka, & Ohba, 1989), or state-space description chamber, encompasses an electromagnet and a moving
(Lino, Maione, & Rizzo, 2007; Maione et al., 2004). Since anchor integral with a spherical shutter. With the
ARTICLE IN PRESS
1218 P. Lino et al. / Control Engineering Practice 16 (2008) 1216–1230
pressure
solenoid
reference
valve
ECU
p1 pressure
sensor
tank control
chamber
pressure
bypass p2
sensor
orifice
common rail
piston
main rod
chamber
shutter
injectors
pressure reducer
p0 pressure
sensor
temperature, cd 1 is an experimental coefficient to account where possible changes due to mechanical part motions
for non-uniform mass flow rates (Mulemane, Han, Lu, (for example in the main and control chambers) are
Yoon, & Lai, 2004). neglected, without sensibly affecting the model accuracy.
The second case, when pK =pH oð2=ðg þ 1ÞÞg=ðg1Þ (‘‘criti- Hence the perfect gas law (Streeter et al., 1998; Zucrow &
cal value’’), is the usual situation considered in this paper. Hoffman, 1976), p ¼ rRT ¼ mRT=V , leads to
Then the expression of the mass flow rate becomes (Zucrow RT K
& Hoffman, 1976) p_ K ¼ _ HK m
ðm _ KL Þ. (4)
VK
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ðgþ1Þ=ðg1Þ The integration of (4) yields the pressure in the generic
g 2
m_ HK ¼ cd2 pH AHK , (2) control volume. Eqs. (1)–(2) are used to model the common
RT H g þ 1
rail injection system presented in Figs. 1 and 3.
which is independent of pK (‘‘critical flow condition’’).
Now, if m_ HK and m _ KL are the mass flow rates in and out 3.2. Injection system modelling
of the container K of volume V K , and rK is the gas density,
it holds (see Fig. 2): Let p0 , p1 , p2 , p3 be the pressures in the tank, in the
control chamber circuit, in the main circuit and in the
dðrK V K Þ dr intake manifold, respectively. To derive the system state
_ HK m
m _ KL ¼ ¼ VK K , (3)
dt dt equations, pressures p1 and p2 are chosen as state variables.
Then p_ 1 and p_ 2 are expressed in terms of the differences
between the mass flow rate flowing in and that flowing out
LEGEND of their corresponding volumes, according Eq. (4). Eqs. (1)
Flow direction and (2) are used to express the mass flow rates in terms of
Mechanical motion the pressures. Note that the flow sections are affected by
p1 direction
m· 12 Pressure forces
the mechanical parts motion, which is influenced by the
p1 direction pressures in the control volumes.
m· 01
From Fig. 3, m _ 02 , m
_ 12 , m
_ 23 are the flows affecting the
p0
main circuit. The flow between the tank and the main
_ 02 , depends on the instantaneous flow section A02 ,
circuit, m
S1 p1
p1 thus on the shutter and the piston axial displacement hs .
m· 12 The shutter and piston dynamics can be described by
applying the Newton’s second law of motion to the forces
S2 p2 p
m· 23 2
acting upon each of them. Let hs ¼ 0 when the shutter is
closed, and hs 40 when the shutter and the piston move
hs down. To calculate the forces acting on the shutter, by
m· 02
S0 assumption, the pressure gradient p0 p2 is fully applied
p0 on the inlet minimal section (Figs. 3 and 4(a)). The force F 0
p3 due to p0 , acting on the lower spherical cap portion of the
shutter surface S 0 , can be computed as follows. Since the
Fig. 3. Block scheme of the injection system showing the fuel flows and direction of pressure forces is normal to the shutter surface,
the hydraulic forces involved. the overall pressure force in the motion direction is
main chamber p2
ds 5
s 4.5
exact
4
Inlet section [mm2]
hs 3.5
3
dsM 2.5
approximated
90-s dsm 2
ls
d0 / 2 1.5
1
d
0.5
dS 0
p0 0 0.1 0.2 0.3 0.4 0.5
Shutter displacement [mm]
Fig. 4. Computation of the pressure reducer inlet section, A02 , and of the pressure forces acting on the shutter: (a) valve geometry; (b) comparison of
accurate and approximated computation.
ARTICLE IN PRESS
1220 P. Lino et al. / Control Engineering Practice 16 (2008) 1216–1230
obtained by integrating the elementary contribution and the lower base radius is given by
depending on the angle y:
dM m
s ¼ d s þ 2l s cos bs ¼ d 0 cos bs þ hs sin 2bs . (12)
dF 0 ¼ p0 dS cos y. (5)
Finally, it is straightforward to calculate the inlet section:
Since the elementary surface is
p m p
d0 1 A01 ¼ ðd þ d M s Þl s ¼ ð2d 0 cos bs þ hs sin 2bs Þhs sin bs
dS ¼ pd 0 sin y dy ¼ pd 20 sin y dy, (6) 2 s 2
2 2 p
’ d 0 hs sin 2bs . (13)
the substitution of (6) in (5) gives 2
The last expression on the right side of (13) follows by
dF 0 ¼ 12p0 pd 20 sin y cos y dy. (7) neglecting the h2s -term: a simplification justified by the
Hence, the integration within the angular interval ½90 small value of hs (Fig. 4(b)).
bs ; 180 yields Since the flow m _ 12 between the control and the main
circuits is approximately stationary, it can be determined
Z 180
1 by the following equation (Streeter et al., 1998):
F0 ¼ p0 pd 20 sin y cos y dy
90bs 2 rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
p2
1 m_ 12 ¼ cd cL A12 ðp p2 Þ, (14)
¼ p0 pd 20 cos2 bs . (8) RT 1
4
The pressure p2 applies on the surface S2 including the where cL takes into account the effect of kinetic energy
piston bottom surface and the upper spherical cap surface losses in the bypass orifice minimal section A12 and in the
of the shutter, so that the resulting force F 2 is the sum of pipe cranks. Eq. (14) assumes that no reversal flows
two opposite contributions. By assuming the rod base occur.
surfaces negligible, the former contribution is p2 S1 , In dependence of engine speed and load, the ECU sets
because the bottom and the top surfaces of the piston are the injectors opening time intervals. The whole injection
approximately equal. As for F 0 , the latter contribution is cycle takes place in a 720 interval, with a delay between
obtained by integrating the elementary force p2 dS cos y each injection command that depends on the number of
within the angular interval ½0; 90 bs . Therefore, the injectors. Since this model neglects the injectors opening
overall force due to p2 is and closing transients, only two conditions can occur:
injectors closed, or opened with a flow section A23 . This
Z 90bs
1 simplification does not introduce a considerable error, as
F 2 ¼ p2 S 1 þ p2 pd 20 sin y cos y dy the transients expire within 1 ms, while reduces the system
0 2
order and computational effort, because it is not necessary
1
¼ p2 S 1 pd 20 cos2 bs . (9) to calculate the shutter axial displacement and speed. Since
4
critical flow condition always holds, the injection mass flow
Finally, the force F 1 ¼ p1 S1 due to p1 applies on the piston m_ 23 must be calculated by (2) with AHK ¼ ET inj A23 , and
upper surface S 1 . where ET inj (injectors energizing time) is the injectors
Summing up, if the viscous friction term, the piston and timing, normalized with respect to the injection cycle
the shutter inertias are neglected in comparison with the duration. More in detail, if n is the number of injectors,
large pressure forces, the force balance gives ET inj is a square signal of amplitude f1; 2; . . . ; ng depending
F 0 þ F 1 þ F 2 ks hs F s0 F c ¼ 0, (10) on the number of injectors simultaneously opened, and on
the period variable with the engine speed.
where ks is the spring constant and F s0 is the spring The pressure dynamics within the control chamber
preload, i.e. the force applied by the spring when the circuit depends on the incoming flow from the tank m _ 01 ,
shutter is closed. Finally, the coulomb friction F c is and the outgoing flow towards the main circuit m _ 12 .
assumed to be constant and evaluated experimentally. The displacement of the shutter driven by the solenoid
Hence, hs is obtained by solving (10), allowing to compute valve regulates the flux m _ 01 . Since its inertia is
the main circuit minimal inlet section A01 by means of negligible, the inlet section of the valve can be considered
simple geometrical considerations. completely opened or completely closed, depending
The flow section A01 , perpendicular to the flow direction, on the actual driving current (energized/not-energized
is the lateral surface of a truncated cone (see Fig. 4(a)). If bs circuit). With this assumption, the value of the inlet
is the slope of the conical seat, d 0 the shutter diameter, the section can be calculated by using (13) (being the shutter
cone slant height l s and the upper base diameter d m s are, and seat geometries the same as in Fig. 4(a), with the
respectively, appropriate sizes) and by putting hs ¼ ET sv hmax , where
(
l s ¼ hs sin bs ; ET sv (solenoid valve energizing time) is a square signal,
(11) assuming values in f0; 1g, representing the valve energizing
dm s ¼ d 0 cos bs
conditions.
ARTICLE IN PRESS
P. Lino et al. / Control Engineering Practice 16 (2008) 1216–1230 1221
give the relationship between the output yðp2 Þ and the matrix form y^ ¼ G~u þ r, where y^ ¼ ½yðt ^ þ 1Þ; . . . ;
inputs du1 and du2 , respectively. Considering du1 as the ^ þ NÞT , u~ ¼ ½DuðtÞ; . . . ; Duðt þ N 1ÞT , r ¼ ½rðt þ 1Þ;
yðt
control signal, it is possible to derive a family of . . . ; rðt þ NÞT and G is a lower triangular N N matrix,
AutoRegressive eXogenous (ARX) models suitable for whose elements are Gðl; mÞ ¼ gl;lm for lXm, and Gðl; mÞ ¼
the controller design from the first element of the transfer- 0 for lom (Rossiter, 2003).
function matrix. With some notational abuse, dropping the If w ¼ ½wðt þ 1Þ; wðt þ 2Þ; . . . ; wðt þ NÞT is a sequence of
symbol d and using uðtÞ instead of u1 ðtÞ, and regarding t as future reference-values, a cost function taking into account
the discrete-time variable, yields the future errors, i.e. the difference between future output
and reference values, and the control action variations can
ð1 a1 q1 ÞyðtÞ ¼ ðb0 q1 b1 q2 ÞuðtÞ, (20)
be introduced:
where q1 is the shift operator.
J ¼ EfðG~u þ r wÞT ðG~u þ r wÞ þ l~uT u~ g, (26)
4.2. Controller design where l is a sequence of weights on future control actions.
The minimization of the cost function J with respect of u~
Model predictive control techniques predict the output gives the optimal control law for the prediction horizon N:
from a process model and then impress a control action
able to drive the system to a reference trajectory (Rossiter, u~ ¼ ðGT G þ lIÞ1 GT ðw rÞ. (27)
2003). The control action is applied to systems represented
Since the first element of u~ is DuðtÞ, the current control
by an AutoRegressive Integrated eXogenous (ARIX)
action is
model:
Aðq1 ÞyðtÞ ¼ Bðq1 Þuðt 1Þ þ xðtÞ=D, (21) uðtÞ ¼ uðt 1Þ þ gT ðw rÞ, (28)
where xðtÞ is a zero mean white noise, Aðq1 Þ and Bðq1 Þ where gT is the first row of ðGT G þ lIÞ1 GT ; at each step
are polynomials in q1 of degrees na and nb, respectively, the first computed control action is applied and then the
and D is the discrete derivative operator ð1 q1 Þ. The j- optimization process is repeated after updating all vectors.
step division of 1 by DAðq1 Þ gives the following By placing infinite weights on control changes after N U
Diophantine equation: steps, where N U is the control horizon, all next control
actions are taken to be equal to uðt þ N U Þ. A higher value
1 ¼ E j ðq1 ÞAðq1 ÞD þ qj F j ðq1 Þ, (22) for N U results in a more active control action, while a
where E j ðq1 Þ and qj F j ðq1 Þ are, respectively, the quotient lower value reduces the computational effort, as u~ reduces
and the remainder polynomials of ðj 1Þth and ðj þ naÞth to a N U 1 vector. Obviously a good choice puts the
order. By putting together the previous equations it can be prediction horizon equal to the system rise time. The above
calculated the system output in j steps: concepts are applied to design a GPC for the rail pressure
considering the family of ARX models (20), because it
yðt þ jÞ ¼ F j ðq1 ÞyðtÞ þ E j ðq1 ÞDBðq1 Þuðt þ j 1Þ belongs to the class represented by (21). It can be shown
þ E j ðq1 Þxðt þ jÞ. (23) (Rossiter, 2003) that the resulting control law becomes
Since the last term in (23) only refers to future noise DuðtÞ ¼ k1 wðtÞ þ ðk2 þ k3 q1 ÞyðtÞ þ k4 Duðt 1Þ, (29)
samples, it cannot be known in advance, so putting G j ¼
E j ðq1 ÞBðq1 Þ gives the following equation of the j-step where ½k1 ; k2 ; k3 ; k4 depends on N and N U (Table 2).
optimal predictor: A simple scheduling strategy based on a sensitivity
analysis of linear model coefficients has been implemented
^ þ jjtÞ ¼ G j ðq1 ÞDuðt þ j 1Þ þ F j ðq1 ÞyðtÞ,
yðt (24) in simulation, consisting in what follows. A fixed value of
1 1
where G j ðq Þ and F j ðq Þ are polynomials of degrees ðj þ the prediction error is chosen to select the linear model
nb 1Þ and na, respectively. Let rðt þ jÞ be the component used to compute the controller parameters. Simulation and
of yðt þ jÞ depending on known values at time t: experimental tests point out that the tank pressure is the
main variable affecting controller performances, allowing
rðt þ 1Þ ¼ ½G 1 ðq1 Þ g10 DuðtÞ þ F 1 ðq1 ÞyðtÞ, to set different main families of sub-models. The pressure
drop between control chamber and common rail is
rðt þ 2Þ ¼ ½G 2 ðq1 Þ q1 g21 g20 qDuðtÞ þ F 2 ðq1 ÞyðtÞ, approximately constant apart the operating condition,
thus only x2 is considered for scheduling. Finally, u2
..
. determines the final linear model for control, also taking in
mind that Að2;2Þ (and thus u2 ) had a weaker influence on the
rðt þ jÞ ¼ ½G j ðq1 Þ qðj1Þ gj;j1 gj0 model coefficients. The complete set of GPC parameters
can be computed off-line and stored in a look-up table as a
qj1 DuðtÞ þ F j ðq1 ÞyðtÞ, (25)
function of the chosen operating conditions. Clearly,
1 1 ðnbþj1Þ
where G j ðq Þ ¼ gj0 þ gj0 q þ þ gj;nbþj1 q . increasing the number of operating conditions for the
Then it is possible to express (24), for j ¼ 1; . . . ; N, in the scheduling strategy would require a larger amount of
ARTICLE IN PRESS
P. Lino et al. / Control Engineering Practice 16 (2008) 1216–1230 1223
memory to store controller parameters, though not performances. Due to safety reasons, air replaced CNG
increasing the computational effort. as test fluid during the experiments, without substantially
Since the GPC law gives the change with respect of the affecting the fluid-dynamic behaviour. Table 3 compares
previous control action, it is necessary to integrate (29) to the values of model parameters for the two cases. Fig. 5
apply the whole input. The control action is bounded in the shows the experimental setup consisting of three main
range [0%, 100%], thus an anti wind-up system to avoid subsystems, namely the common rail injection system, a PC
undesired oscillations in the control loop has been systems equipped with a 250 MHz dSPACEs board
introduced. (ds1104) and a PC system equipped with a National
Instruments acquisition board. Moreover, a programma-
ble MF3 development master box took the role of ECU.
5. Simulation and experimental results The test rig replaced the fuel tank with a compressor,
providing air at a constant input pressure p0 , and included
5.1. Experimental set-up an injection system with prototype injectors, designed to
operate within a 4–20 bar pressure interval, and sending air
This section describes the validation experiments of the to a discharging manifold. Two different configurations
proposed model and the evaluation of the controller were considered during tests, using n ¼ 4 and 6 injectors,
respectively. The former refers to open loop operations,
Table 2
carried out to collect data for model validation, the latter
GPC notation
refers to closed loop operations, to evaluate the controller
Symbol Description performances.
The system operating conditions were determined by the
AðÞ, BðÞ, EðÞ, F ðÞ, GðÞ Polynomials of ðq1 Þ
engine driving shaft speed u (engine speed in the following),
g Subsidiary column vector for computation of u
E Expected value the intake manifold air flow, which varies with engine load
G Matrix composed of elements of GðÞ and power request, the input air pressure p0 and the
J Cost function instantaneous rail pressure p2 . Since the test rig did not
k GPC gains include the driving shaft and the intake manifold, the
N, N U GPC prediction and control horizons
engine speed and load signals were artificially generated by
q1 Discrete time shift operator
r ¼ ½rðÞ Column vector of known components of y the National Instrument PC system using the labVIEWs
u~ ¼ ½DuðÞ Column vector of future control actions software. To inject the proper fuel amount corresponding
w ¼ ½wðÞ Column vector of future reference values to the actual working conditions, the ECU imposed the
y^ ¼ ½yðÞ
^ Column vector of predicted outputs injection timings, in terms of opening time interval tj (as a
D Discrete derivative operator function of engine load and rail pressure) and frequency
l Sequence of weights on future control actions
xðÞ Zero mean white noise n u=120 (as a function of engine speed), by drawing the
appropriate values from a look-up table. Eventually, tj
Table 3
Nonlinear model parameters
could be set independently from the working conditions using the 16 bit A/D-D/A converters that are integrated in
during open loop operations. the dSPACE board. During open loop operations, the
The rail pressure reference, which was established by the control signal duty cycle was artificially generated bypass-
ECU depending on the engine speed and load, was sent to ing the controller.
the dSPACE PC system to drive the solenoid valve by
means of a 24 V power driver, both in open loop and closed 5.2. Model validation
loop operations. Alternatively, the dSPACE Control Desk
software was capable of artificially generating the rail The solution of (15) in the MATLAB/Simulink environ-
pressure reference bypassing the ECU. The 24 V power ment enabled the comparison between simulation and
driver properly supplied the solenoid valve by performing a experiments. The data necessary for the model validation
peak-hold modulation through a 30 kHz TTL signal were obtained in open loop operations with the deactivated
(Fig. 5). The resulting driving current, which was applied pressure controller by properly generating the driving
for a tph time interval to the valve electromagnetic circuit, signal of the solenoid valve. With a constant 40 bar input
consisted of a 1.6 ms peak phase followed by a variable pressure p0 , three cases covering typical operating condi-
duration hold phase, of 6 and 3.5 A amplitudes, respec- tions were analysed: (a) with constant injection timings and
tively. The valve control period tc was defined as the time varying duty cycle, (b) with both duty cycle and injection
interval between two leading edges of the current signal, set timings held constant, (c) with both duty cycle and
to 50 and 100 ms during the closed loop and open loop injection timings varying.
experimental tests, respectively. Finally, the hold phase Firstly, with constant injectors driving command, the
duration was regulated by varying the tph =tc ratio, namely system response to step variations of the duty cycle of the
the control signal duty cycle. valve driving signal was determined. Fig. 6 compares
During closed loop operations, the control algorithms simulation results with the performed experiments. The
implemented in the MATLAB/Simulinks environment operating conditions were a 2400 rpm engine speed and
imposed the signal duty cycle. The dSPACE code generator 8 ms injectors exciting time interval, when two pairs of
compiled the Simulink program and then the real-time opposite duty cycle step variations were applied (3–9%,
executable code was downloaded to the board memory. 9–3%, 3–12%, 12–3%), at 1.5, 28, 56.8, 82.8 s time
Then the board processor received the feedback from the instants, respectively. With a positive step variation, the
rail pressure sensor and applied the appropriate control pressure increased in the control chamber (see Fig. 6(a)),
action to the solenoid valve. The signals were processed making the regulator inlet section to stay open longer.
ARTICLE IN PRESS
P. Lino et al. / Control Engineering Practice 16 (2008) 1216–1230 1225
12 12
6 6
4 4
simulated
simulated
2 2
0 20 40 60 80 100 120 0 20 40 60 80 100 120
Time [s] Time [s]
x1 (control chamber pressure) [bar]
13.6 12.7
simulated
simulated
13.2 12.3
13 12.1
Fig. 6. Simulation and experimental results with duty cycle step variations and constant injectors driving signal: (a) control chamber pressure; (b) rail
pressure; (c) detail on control chamber pressure; (d) detail on rail pressure.
As a consequence of the larger mean air inflow coming pressure disturbances propagating towards the rail was not
from the compressor, the rail pressure raised (see Fig. 6(b)). taken into account. In addition, the distributed losses along
Conversely, with a negative step reducing the duty cycle, the pipes connecting common rail and pressure reducer
both the control chamber and the rail pressures fell to were neglected. However, a detailed representation of these
lower levels. As shown in Fig. 6(c), the pressure in the phenomena would require a PDE description without
control chamber had an oscillating behaviour within the adding essential information useful for control.
control period: the pressure increased (decreased) when the In a second test, duty cycle was held constant (9%),
solenoid valve was open (closed). Shortening the control while engine speed and injection timings varied. Figs. 7(a)
period could smooth away these pressure variations. and (b) depict model output in terms of control chamber
The waves amplitude difference between experiment and rail pressures. Simulation started from a steady state
and simulation in Fig. 6(c) is mainly due to model condition corresponding to a u ¼ 2200 rpm engine speed
simplifications. In fact, by introducing both the second and tj ¼ 3 ms injectors exciting time interval. At time 4.5 s,
order dynamics of solenoid valve shutter and piston, a 4000 rpm speed step was applied, and the injection time
better results can be obtained. Obviously, introducing interval was raised to 11 ms. In these conditions, the
these high order dynamics increases the model complexity, current duty cycle was no longer able to maintain the initial
resulting in a higher order controller. On the other rail pressure, because of the increased amount of injected
hand, the oscillating behaviour should not affect the air. Besides, the air flow between main circuit and control
prediction, which is computed on the base of mean value chamber made the control chamber pressure to decrease.
inputs. At time 21 s a complete cut-off was applied, i.e. the
Moreover, higher frequency, superimposed rail pressure injectors were kept closed in the whole injection cycle. So
oscillations due to injections within a cycle occurred (see control chamber and rail pressures raised because the air
Fig. 6(d)). In this case, simulation results exhibited wider was no more sent to the discharging manifold. Finally, at
oscillations than experiments, synchronously with the time 29 s, u and tj were taken back to initial values. Results
solenoid valve operations. Modelling assumptions, indeed, show that the model approximates the real system
considered the reducer main chamber and the common rail behaviour both during the steady state condition and
as a whole control volume. Consequentially, the decou- during the cut-off transient, even though it had a smoother
pling effect of the main chamber volume with respect to dynamics, probably due both to unmodelled nonlinearities
ARTICLE IN PRESS
1226 P. Lino et al. / Control Engineering Practice 16 (2008) 1216–1230
14 14
6 6
4 4
simulated simulated
2 2
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time [s] Time [s]
Fig. 7. Simulation and experimental results with a constant duty cycle and when applying step variations of engine speed and injection time interval: (a)
control chamber pressure; (b) rail pressure.
12 12
x1 (control chamber pressure)
simulated
simulated
10 x2 (rail pressure) [bar] 10
8 8
[bar]
6 6
4 4
2 experimental 2
experimental
7 14
υ (engine speed) [rpm x 1000]
6 12
u (duty cycle) [%]
5 10
4 8
3 6
2 4
1 2
0 0
0 20 40 60 80 100 120 140 160 0 20 40 60 80 100 120 140 160
Time [s] Time [s]
Fig. 8. Simulation and experimental results when varying duty cycle and engine speed, with a constant tj : (a) control chamber pressure; (b) rail pressure;
(c) engine speed; (d) control signal duty cycle.
(as the pressure discontinuity at time 6 s testifies) and 5.3. Controller performances
imperfect parameters tuning.
Finally, the system behaviour for a constant load (resulting The tuning of the controllers was referred to models
in a 3 ms opening time interval), while varying engine speed linearized at the starting equilibrium point, according to
and solenoid valve driving signal was evaluated. The engine design steps of Section 4. However, the values of the
speed was composed of ramp profiles (Fig. 8(c)), while the controller parameters was not scheduled during experi-
duty cycle changed abruptly within the interval ½1%; 13% ments. Namely, the set relevant to initial working condition
(Fig. 8(d)). Figs. 8(a) and (b) show the accordance of the was only considered, as preliminary results on the test
resulting dynamics with the expected behaviour. A maximum bench were satisfactory. Two different experiments allowed
error of 10% confirms the model validity. to evaluate the controller performances for different engine
ARTICLE IN PRESS
P. Lino et al. / Control Engineering Practice 16 (2008) 1216–1230 1227
24 30
12 15
8 10
4 pressure reference
5
0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]
2 50
υ (engine speed) [rpm x 1000]
1.8 40
1.4 20
1.2 10
1 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]
Fig. 9. System response for speed and load ramp variations, in presence of a GPC with N ¼ 5 (0.25 s) and N U ¼ 1 (0.05 s): (a) rail pressure; (b) injectors’
exciting time interval tj ; (c) engine speed; (d) control signal duty cycle.
load and speed, letting the ECU to generate the appro- Analogous working conditions were reproduced in a test
priate injection timings and rail pressure reference, and where a classical PI controller tuned by a trial and error
imposing the rail pressure reference with constant speed approach and including a simple anti-wind-up system
and load manually. replaced the GPC (Fig. 10). In this case, the imposed
The first experiment was performed increasing and engine speed profile was more favourable than in the
decreasing ramp profiles of the engine speed and load, previous one (being the speed bounded within the
and a GPC with a N ¼ 5 (0.25 s) prediction horizon and a 950–1600 rpm interval, see Fig. 10(c)), so it produced a
N U ¼ 1 (0.05 s) control horizon. The rail pressure reference smaller pressure disturbance due to air injection. Fig. 10(a),
was provided by the ECU in dependence of the current clearly shows good tracking performances of the PI
working conditions set by the National Instrument PC controller during the rising transient, and at the same time
system and had a sort of ramp profile as well. The input air a considerable undershoot of the rail pressure before to
pressure coming from the compressor was always 30 bar. reach the steady state condition, with a large settling time.
Fig. 9 shows the engine speed accelerating from 1000 to In the second experiment the controller performances
1800 rpm and then decelerating to 1000 rpm, within a 20 s were evaluated by applying a step reference variation
time interval (Fig. 9(c)). With a ramp slop not exceeding a taking engine load and speed constant, as well as a
certain value, the control action guaranteed a good constant input pressure from the compressor. As stated
reference tracking (Fig. 9(a), during time intervals [0, 14] earlier, in this experiment, the reference pressure was
and [22, 40]). Starting from time 14 s, owing to the manually generated, and it was no more set by the ECU
request of a quick pressure reduction the control depending on the engine speed and load. Figs. 11 and 12
action closed the valve completely (Fig. 9(d)) by imposing depict the system behaviour in presence of the GPC and the
a duty cycle equal to zero. Thanks to injections, the rail classical PI controllers, respectively. Fig. 11(a) refers to a
pressure (Fig. 9(a)) decreased to the final 5 bar reference step reference variation from 5 to 20 bar, while keeping a
value, with a time constant depending on the system constant 1460 rpm engine speed (Fig. 11(c)) and a 30 bar
geometry. Due to the saturation of the actuation input pressure. The ECU reduced the injection time
variable the maximum error amplitude could not further interval from the initial 19–4 ms (Fig. 11(b)), so that the
decrease. Fig. 9(b) shows the injectors’ exciting time during same air amount was injected during a cycle when the rail
the experiment. pressure raised. Fig. 11(d) shows that the control action
ARTICLE IN PRESS
1228 P. Lino et al. / Control Engineering Practice 16 (2008) 1216–1230
24 30
12 15
8 10
4 pressure reference 5
0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]
1.8 50
υ (engine speed) [rpm x 1000]
1.6 40
1.2 20
1 10
0.8 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time [s] Time [s]
Fig. 10. System response for speed and load ramp variations, in presence of a standard PI: (a) rail pressure; (b) injectors’ exciting time interval tj ; (c)
engine speed; (d) control signal duty cycle.
24 50
tj (injection time interval) [ms]
20
x2 (rail pressure) [bar]
40
16
30
pressure reference
12
20
8
rail pressure
4 10
0
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
Time [s] Time [s]
2 50
υ (engine speed) [rpm x 1000]
1.8 40
u (duty cycle) [%]
1.6 30
1.4 20
1.2 10
1 0
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
Time [s] Time [s]
Fig. 11. System step response for a 1460 rpm engine speed, in presence of a GPC with N ¼ 5 (0.25 s) and N U ¼ 1 (0.05 s): (a) rail pressure; (b) injectors’
exciting time interval tj ; (c) engine speed; (d) control signal duty cycle.
ARTICLE IN PRESS
P. Lino et al. / Control Engineering Practice 16 (2008) 1216–1230 1229
24 50
0
0 3 6 9 12 15 18 0 3 6 9 12 15 18
Time [s] Time [s]
2 50
υ (engine speed) [rpm x 1000]
1.8 40
1.4 20
1.2 10
1 0
0 3 6 9 12 15 18 0 3 6 9 12 15 18
Time [s] Time [s]
Fig. 12. System step response for a 1330 rpm engine speed, in presence of a standard PI: (a) rail pressure; (b) injectors’ exciting time interval tj ; (c) engine
speed; (d) control signal duty cycle.
during the rising transient was the maximum possible. parameters appear unsuited to enable good performances
Moreover, the step response presented a negligible over- in every working condition.
shoot, and the limited amplitude pressure oscillations close
to set-point were basically due to the injection process. 6. Conclusion
Clearly, increasing prediction horizon could result in a
sluggish response, while further experiments may show that The work presented in this paper investigated the
increasing the control horizon does not result in a better feasibility of the electronic control of the injection pressure
rail pressure behaviour. in a new common rail injection system for CNG engines.
As for the PI controlled system, the experimental test Better performances were obtained by taking advantage of
was performed by applying a 7–15 bar step variation some well-established diesel technologies. The main im-
(Fig. 12(a)), considering a constant 1330 rpm engine speed provements consist in a better fuel metering and in the
(Fig. 12(c)) and a 100 bar input pressure. Further, the capability of on-line adaption of the injection pressure to
injection time interval was reduced from 40 to 15 ms by the different operating conditions. Even though a direct
ECU to keep the injected fuel amount constant. Fig. 12(d) assessment has not been performed in the present work,
shows that owing to the higher input pressure a reduced other authors proved that performances and pollutant
control effort was necessary to maintain the pressure at set emissions reduction can benefit from a precise control of
point. Even though conclusive remarks cannot be drawn the injection pressure.
due to different operational conditions, it seems that the To develop a control law for the injection pressure
GPC enables a better closed loop dynamics than the PI. In regulation, a model of the injection system in state-space
this case, the rise time was considerably longer than the form was presented. The parameters of the model were
GPC case, and the system response showed a 17% directly known or easy determined by geometrical data. The
overshoot. However, it must to be underlined that the comparison of simulation results with experimental data
higher input pressure could have negatively influenced the showed that the system dynamics was modelled with a good
system overshoot, as more air could enter the control accuracy and for a wide range of operating conditions. The
chamber for the same control action, while the higher load, proposed model was used to design a generalized predictive
which resulted in a longer tj (Fig. 12(b)), could have controller for the injection pressure control. Experimental
increased the system rise time. In the end, the PI results confirmed the effectiveness of the approach.
ARTICLE IN PRESS
1230 P. Lino et al. / Control Engineering Practice 16 (2008) 1216–1230
In conclusion, the model simplicity makes it suitable for Hountalas, D. T., & Kouremenos, A. D. (1998). Development of a fast
on-line use or for inclusion in a complete engine model for and simple simulation model for the fuel injection system of diesel
engines. Advances in Engineering Software, 1(29), 13–28.
emissions control. Moreover, the proposed GPC law seems
International Gas Union (2005). Global opportunities for natural gas as a
promising with respect to the classical PI, also in view of transportation fuel for today and tomorrow. Report on study group 5.3
the development of nonlinear predictive control strategies. ‘‘Natural gas vehicles (NGV)’’.
Future developments would also consider the MIMO Kouremenos, D. A., Hountalas, D. T., & Kouremenos, A. D. (1999).
control of injection pressure and injectors timings under Development and validation of a detailed fuel injection system
a common control law. simulation model for diesel engines. SAE Paper 1999-01-0527.
Kuraoka, H., Ohka, N., & Ohba, M. (1989). Application of H-infinity
optimal design to automotive fuel control. In Proceedings of American
References control conference, Pittsburg, Pennsylvania.
Kvenvolden, K. A. (1999). Natural gas hydrate occurrence and issues.
Aespy, V., Engja, H., & Skarboe, L. V. (1996). Fuel injection system Annals of the New York Academy of Sciences, 715, 232–246.
design, analysis and testing using bond graph as an efficient modeling Lino, P., Maione, B., & Rizzo, A. (2007). Nonlinear modelling and control
tool. SAE Paper 962061. of a common rail injection system for diesel engines. Applied
Amorese, C., De Matthaeis, S., De Michele, O., & Satriano, A. (2004). Mathematical Modelling, 9(31), 1770–1784.
The gaseous fuel option: LPG and CNG. In Procedings of the Maione, B., Lino, P., De Matthaeis, S., Amorese, C., Manodoro, D., &
international conference on vehicles alternative fuel system & environ- Ricco, R. (2004). Modeling and control of a compressed natural gas
mental protection, Dublin, Ireland. injection system. WSEAS Transactions on Systems, 3(5), 2164–2169.
Cantore, G., Mattarelli, E., & Boretti, A. (1999). Experimental and McCormick, R. L., Graboski, M. S., Alleman, T., Herring, A. M., &
theoretical analysis of a diesel fuel injection system. SAE Paper 1999- Nelson, P. (1999). In-use emissions from natural gas fueled heavy-duty
01-0199. vehicles. SAE Paper 1999-01-1507.
Catania, A. E., Dongiovanni, C., Mittica, A., Negri, C., & Spessa, E. Morselli, R., Corti, E., & Rizzoni, G. (2002). Energy based model of a
(1996). Study of automotive diesel injection-system dynamics under common rail injector. In Proceedings of the IEEE conference on control
control. SAE Paper 962020. applications, Glasgow, Scotland.
Desantes, J. M., Arregle, J., & Rodriguez, P. J. (1999). Computational Mulemane, A., Han, J. S., Lu, P. H., Yoon, S. J., & Lai, M. C. (2004).
model for simulation of diesel injection systems. SAE Paper 1999-01- Modeling dynamic behavior of diesel fuel injection systems. SAE
0915. Paper 2004-01-0536.
Dyntar, D., Onder, C., & Guzzella, L. (2002). Modeling and control of Pan, C. P., Li, M. C., & Hussain, S. F. (1998). Fuel pressure control for
CNG engines. SAE Paper 2002-01-1295. gaseous fuel injection systems. SAE Paper 981397.
Ficarella, A., Laforgia, D., & Landriscina, V. (1999). Evaluation of Qina, S. J., & Badgwell, T. A. (2003). A survey of industrial model
instability phenomena in a common rail injection system for high speed predictive control technology. Control Engineering Practice, 7(11),
diesel engines. SAE Paper 1999-01-0192. 733–764.
Fuseya, T., Nishimura, T., Sato, Y., & Tanaka, H. (1999). Analysis on Rossiter, J. A. (2003). Model-based predictive control: A practical
common-rail injector using bond graph simulation program. SAE approach. New York, USA: CRC Press.
Paper 1999-01-2936. Streeter, V., Wylie, K., & Bedford, E. (1998). Fluid mechanics (9th ed.).
Gauthier, C., Sename, O., Dugard, L., & Meissonnier, G. (2005). New York, USA: McGraw-Hill.
Modelling of a diesel engine common rail injection system. In Weaver, C. S. (1989). Natural gas vehicles—a review of the state of the art.
Proceedings of the 16th IFAC world congress, Prague, Czech Republic. SAE Paper 892133.
Heywood, J. (1988). Internal combustion engine fundamentals. New York, Zucrow, M., & Hoffman, J. (1976). Gas dynamics. New York, USA:
USA: McGraw-Hill. Wiley.