Escolar Documentos
Profissional Documentos
Cultura Documentos
, Frank Willems
,
, Maarten Steinbuch
u = [u
EGR
, u
V TG
]
w = [
e
, W
f
]
z = [, O
2
%]
(2)
From model t and validation, it is seen that this
model can accurately describe the dynamic responses
of the air and exhaust ows in the engine.
Figure 2 illustrates the selection of the controlled
outputs. Emission legislation limits are converted to
targets for NO
x
and PM emissions in the dierent en-
gine operating points. This is realized indirectly, by
EGR
VTG
NO
x
PM
Engine speed
Fuel
O
2
%
Engine
Model
Emission
Model
Fig. 2: Scheme illustrating the selection of the con-
trol outputs
1 1.5 2 2.5 3 3.5 4
11
12
13
14
15
16
17
18
19
20
21
O
2
%
VTG
EGR
Fig. 3: The equilibrium values of and O
2
% result-
ing from sweeping the VTG and EGR valve at 1500
rpm and 100 mg/inj. Steps in the EGR valve are
2.5% and in the VTG angle 5%.
specifying the corresponding air-fuel ratio and EGR
ow rates, including transients. Also, related quan-
tities such as intake and exhaust pressure, fresh air
ow, or burned gas fraction are applied [1, 2]. How-
ever, in this study air-fuel ratio and oxygen concen-
tration O
2
% in the intake manifold are applied. This
selection is mainly driven by the strong correlation
between these quantities and the emissions to be con-
trolled. The formation of PM is strongly linked with
. Especially low values of lead to excessive PM
formation. The formation of NO
x
is strongly linked
to the O
2
%. A lower O
2
% lowers the combustion
temperature and also less oxygen is present near the
ame front. Both eects are benecial for lowered
NO
x
formation.
3. SYSTEM ANALYSIS
In this section we investigate the properties of
the diesel engine airpath that are of interest for the
design of the control system using the mean value
engine model. More precisely: eects like dynamic
coupling, the feasible range of the outputs and un-
certainty are studied.
3.1 Static analysis
Figure 3 shows the output values resulting from
a sweep of the input values. This graph shows that
the range of and O
2
% that can be obtained is lim-
ited. Opening the EGR valve will result in a lower
O
2
% and lower . Opening the VTG vanes results in
an increased O
2
%. The eect on depends on the
conditions. The range of obtainable and O
2
% is
clearly limited by the selected hardware.
166
AVEC 10
0 2 4 6 8 10
0.15
0.1
0.05
0
0.05
O
2
%
EGR step
0 2 4 6 8 10
0.05
0
0.05
0.1
VTG step
0 2 4 6 8 10
0.03
0.02
0.01
0
0.01
Time
0 2 4 6 8 10
0.04
0.03
0.02
0.01
0
0.01
Time
Fig. 4: EGR (left) and VTG (right) step responses.
In black: nominal; colored: disturbed.
Setpoint selection The goal of the controller will
be to make sure that desired values for and O
2
%
are obtained in practise. It remains to choose these
setpoint values.
Given the range of feasible points in Figure 3, the
best point has to be chosen. The point that is cho-
sen determines the tradeo that is made between the
dierent emissions and engine eciency. Generally
a high O
2
% gives a high eciency, low PM, but high
NO
x
. A reasonable tradeo setpoint value that we
will use in this study is O
2
% = 16%.
The setpoint for is the result of a tradeo between
PM emissions, BMEP and pumping losses. In this
paper we will use = 2 as a setpoint.
3.2 Dynamic analysis
Important for the design of a feedback controller
will be how the system responds to a change in the
control inputs. The black line in Figure 4 shows
the response of the outputs to a unit step change
in the inputs starting from nominal conditions. This
response clearly shows the coupling between all in-
puts and outputs that is present in the system. This
means that any single-input single-output SISO con-
trol strategy will not be able to get optimal perfor-
mance from this system. A multi-input multi output
(MIMO) control approach is required to obtain the
best performance.
The time scales needed for reaching a value close to
the equilibrium are between one and three seconds
for all responses. The eect of VTG on , is the
slowest. Furthermore, non-minimum phase behavior
is observed in the VTG- response. This complicates
the design of a high performance feedback controller.
3.3 Eect of uncertainties
Model uncertainties and changing ambient con-
ditions will have an eect on both the dynamic re-
sponse of the system and the resulting equilibrium
values for and O
2
%. We want to maintain the
setpoint values of and O
2
%. This is the reason
Engine
Controller
specications
Engine
model
Legislation &
Road usage
Control
model
Optimization
criterion
Controller
synthesis
Controller
Fig. 5: Diagram showing the main steps followed in
control design process.
we want to apply a feedback control that adapts the
VTG and EGR to counteract these disturbances and
uncertainties. The uncertain dynamic response will
cause robustness problems when standard feedback
strategies are used.
We implicitly study the sensitivity to changing
conditions by choosing an operating point that de-
viates from nominal. Due to the nonlinearity of the
system, this generates a dierent response. This way
we can study the sensitivity to changing conditions
without having to change the model.
For the design of the control model, in section 4.2,
81 dierent operating points are included. A selec-
tion of 5 responses that deviate 5% from nominal
in engine speed, fueling, EGR valve position and/or
VTG vane angle is shown in Figure 4. The dier-
ences between the step responses are of the same
order of magnitude as the individual step responses
themselves. This makes it very important to take
robustness into account for control design.
4. CONTROL DESIGN
The purpose of the controller is to adapt the VTG
and EGR valve in such a way that the desired intake
conditions for both and O
2
% are maintained, re-
gardless of disturbances or uncertainties in the be-
havior of the engine. To achieve this, feedback con-
trol is indispensable. For the feedback controller, we
assume and O
2
% are both directly measurable.
As seen in the previous section, we are faced with a
system that exhibits strong coupling between inputs
and outputs. Also, nonlinearity plays an important
role, which results in a large uncertainty to the local
dynamic response. Robust performance in the face
of this uncertain response is a main requirement for
the controller. To be able to take this into account in
a systematic way, we propose to use the -synthesis
framework [13] for the design of the controller.
Using the -synthesis framework presents possibili-
ties for a robust control design which optimizes the
worst case performance for an uncertain system. In
Figure 5, the followed steps for this design are de-
picted schematically. Following this scheme results
in a framework that can be fully automated. The
engine design and modeling have been completed
already. Also, the emission legislation, road usage
and control specications are not the subject of this
study. This leaves three topics for this section: mak-
ing a control model, formulating the optimization
criterion and nally controller synthesis.
167
AVEC 10
Engine
Model: P EGR, VTG
, O2%
We
Wu
Controller
+
ze
zu
Reference
, O2%
Fig. 6: A schematic representation of the placement
of the performance weights.
10
4
10
3
10
2
10
1
10
0
10
1
10
4
10
3
10
2
10
1
10
0
10
1
10
2
10
3
M
a
g
n
it
u
d
e
Frequency in Hz
Fig. 7: In blue, green and red the weights on respec-
tively O
2
%, and both control inputs.
4.1 Optimization criterion
The optimization criterion will determine how the
tradeo between robustness, maintaining the desired
outputs and use of the system inputs is made. Ro-
bustness of the controller will be taken into account
by including model uncertainty in the control model.
This will be covered in section 4.2.
The -framework makes it possible to design fre-
quency dependent weights on the signals in the sys-
tems. Using DK-iteration we then design a controller
that gives the best weighted performance. We will
use weights on the setpoint error, W
e
, and control in-
put, W
u
. In Figure 6 this is schematically depicted.
A higher weight on a signal means that more eort
will be used to make sure that the performance tar-
get for that signal is reached. Four individual targets
can be identied. The rst two are to maintain the
desired and O
2
% during real world driving. The
second two targets are to limit the use of the two
input signals. Especially, highly uctuating input
signals should be penalized, because they would de-
crease the lifetime of the actuators.
To make a meaningful weight on the tracking
performance of and O
2
%, it is important to rst
consider the conditions of operation. Two dierent
sources of disturbances are expected. The rst source
is slowly varying disturbances caused by e.g. chang-
ing ambient conditions, heat up of the engine or age-
ing. As these disturbances are typically much slower
than the dynamics of our system, integral control will
be a suitable solution. The second source of distur-
bance is the changing torque request by the driver.
This disturbance is measured and hence can be coun-
teracted using feedforward techniques. However, due
to the uncertain response of the system, this feedfor-
ward will be imperfect and an error still remains.
The error will likely have about the same frequencie
distribution as the original disturbance. The Euro-
pean Transient Cycle (ETC) is a typical example of
real world driving. It represents urban, suburban,
and highway driving. Moreover, it is used as part of
the European emission legislation tests. By studying
the frequency contents of the ETC, we have a repre-
sentative and objective way to quantify the expected
disturbances due to the driver input. The ETC fre-
quency distribution can be described fairly accurate
by a rst order transfer function. This rst order
estimate of the ETC will be combined with an inte-
grator that penalized the low frequency disturbances.
The integrator zero is placed a factor 3 lower in fre-
quency than the pole of the ETC part. True integral
weights would cause numerical problems when used
during synthesis. Therefore, to facilitate control syn-
thesis using the Matlab dksyn algorithm, a pole is
added to the integral weight at a frequency a factor
10.000 before the integrator zero. Now, to make the
nal weight W
e
on tracking error of and O
2
%, we
add a scaling gain to make the tradeo in allowed
setpoint errors between these two outputs. For this,
we look at the expected errors without feedback con-
trol and we end up choosing the gain on three times
higher than on O
2
%.
The weight W
u
on the input signal is employed
to prevent large control actions and highly uctu-
ating control actions. This weight helps to prevent
saturation of the input signal, though no guarantees
can be given. It is important to keep in mind that
control actions that are within the required perfor-
mance region, i.e. up to the dropo of W
e
are needed
for accurately keeping the setpoint. Penalizing these
frequencies will decrease the performance. We use
a rst order weight on both system inputs that in-
creases after the dropo of W
e
. Dierent scalings are
not needed this time, because the inputs are both
normalized. In Figure 7 the frequency response of
the weights is shown.
4.2 Control model
The applied mean value engine model is not suit-
able for controller synthesis. To make controller syn-
thesis practical, we are much better o with a linear
time invariant (LTI) model. For each operating point
there is an LTI model that is equivalent to the gen-
eral simulation model. These LTI models are only ac-
curate in a small neighborhood around the nominal
point. Disturbances will cause the system to leave
the nominal point and the response of the system to
a change in the input parameters will change as well
as shown in Figure 4. To take this into account, we
will not describe the system by a normal LTI model,
but by an uncertain LTI model. By identifying the
behavior of the system in a set of operating points
located around the nominal point, we investigate the
changes in the behavior of the system. An uncertain
LTI model will then be made that represents the pos-
168
AVEC 10
sible responses of the system. We will use an additive
uncertainty model of the form:
P
(z) = P
nom
(z) +T
1,1
w
1,1
(z)
1,2
w
1,2
(z)
2,1
w
2,1
(z)
2,2
w
2,2
(z)
(3)
With P
nom
the nominal system behavior, T a trans-
formation matrix and
i,j
an uncertain complex value
such that |
i,j
| < 1 and w
i,j
a transfer function to de-
scribing the maximum uncertainty. It is worth not-
ing here that P
nom
will not be equal to the system
behavior at nominal conditions.
To make the uncertain system model P
, a series
of steps is needed. The rst step is to generate a
set of possible system behaviors. For this a nominal
point of operation needs to be chosen. We will use
1500 rpm, 100 mg/inj, VTG 70%, EGR 60%, re-
sulting in = 2 and O
2
% = 16% (see section 3.1.1).
Next a range of operating points is dened. For all
four parameters of the operating point (engine speed,
fueling, VTG and EGR) we will use low, high and
nominal values. The dierence between low, high
and nominal is 10%. Combining the three possible
values of the four parameters gives 3
4
= 81 dierent
variations of the system.
To identify the local behavior at the 81 operat-
ing points the engine model is used. Small steps
are applied at the control inputs. Next a method
called approximate realization [14] is applied to iden-
tify an LTI system that gives the same dynamic re-
sponse. This method will yield very accurate ap-
proximations, because we use simulation data free of
measurement errors or noise. The dierence between
the LTI approximation and the simulation data is
negligibly small. A convex combination of these 81
LTI systems is expected to represent all possible be-
haviors of the system.
The uncertain model should accurately describe
the identied set of of behaviors. For this we need
to choose a P
nom
, four w
i,j
and a matrix T.
The nominal model P
nom
that is calculated is de-
signed to make the required uncertainty as small as
possible. A general optimization algorithm is ap-
plied to calculate on a logarithmically distributed set
of frequencies which nominal response will give the
smallest required uncertainty radius. Next, output
error optimization is used to calculate a low order
t on the calculated optimal response. For each of
the four elements of the response, a fourth order t
is made. Combining the models for all individual
input-output elements into a single LTI system gives
the nominal system model.
A transformation matrix T is calculated to take
into account some of the structure that is present
in the uncertainty. An optimization algorithm is
applied to calculate a transformation that puts as
much of the low frequent uncertainty in the diagonal
blocks. This should reduce the conservatism that
is introduced by using the elementwise uncertainty
model.
Using the just made P
nom
, the largest dierence
10
2
10
1
10
0
10
1
10
4
10
3
10
2
10
1
M
a
g
n
i
t
u
d
e
EGR input
10
2
10
1
10
0
10
1
10
4
10
3
10
2
10
1
VTG input
10
2
10
1
10
0
10
1
10
4
10
3
10
2
10
1
Frequency in Hz
M
a
g
n
i
t
u
d
e
10
2
10
1
10
0
10
1
10
4
10
3
10
2
10
1
Frequency in Hz
Fig. 8: The magnitude frequency response of T
1
P
.
Nominal model (line) and the uncertainty range
(dashed).
between the nominal LTI model and the set of iden-
tied system behaviors is calculated at the same log-
arithmic set of frequencies for each of the four ele-
ments of the transfer function. This results in an un-
certainty radius at each of the frequencies. When the
t of the nominal system is close to the optimal data,
this uncertainty radius will be close to the smallest
uncertainty radius that was calculated before. Now
an overbound on the uncertainty radii is made by
calculating four SISO transfer functions w
i,j
. Only
the magnitude of these functions are of importance,
because complex uncertainties are used. Hence the
phase of the uncertainty is completely undetermined.
The combination of the nominal model with the
uncertainty model gives the nal uncertain system
model P
. In Figure 8 T
1
P
0 20 40 60 80
1480
1500
1520
1540
1560
E
n
g
i
n
e
s
p
e
e
d
i
n
r
p
m
Time in seconds
0 20 40 60 80
98
100
102
104
106
F
u
e
l
i
n
m
g
/
i
n
g
Time in seconds
Fig. 9: In blue, a simulation without feedback control
of an engine undergoing disturbances in engine speed
and fueling. The nominal operating point is 1500
rpm and 100 mg/inj. In green, the same simulation,
but now with the feedback controller switched on.
we simulate its performance using step disturbances.
These disturbances are particularly challenging for
the controller and show the transient performance of
the feedback controlled system. In Figure 9 the re-
sults are shown. The feedback controlled simulation,
depicted in green is able to keep the setpoint error
small by adapting the input signal. The time needed
for the feedback controlled system to adapt to the
disturbance is about two seconds. This is indeed fast
compared to the real life disturbances, which makes
the controller suitable for accurate control in real life
conditions.
6. CONCLUSION & FUTURE RESEARCH
A method for controller design for EGR-VTG
control of a heavy duty diesel engine is presented.
The design is fully model-based and suitable to be
automated. It features optimal performance and ro-
bustness against model uncertainties.
Future research is needed to make the controller
suitable for the complete engine operating range. A
real life test on a heavy-duty diesel engine can show
the practical benets. The ultimate goal of my re-
search is to formulate a control strategy for obtaining
optimal fuel consumption and low emissions from a
diesel engine.
REFERENCES
[1] M. van Nieuwstadt et al., Egr-vgt control
schemes: experimental comparison for a high-
speed diesel engine, IEEE Control Systems
Magazine, 2000.
[2] A. Stefanopoulou, I. Kolmanovsky, and
J. Freudenberg, Control of variable geometry
turbocharged diesel engines for reduced emis-
sions, IEEE Transactions on Control Systems
Technology, vol. 8, no. 4, pp. 733745, 2000.
[3] M. Amman et al., Model-based Control of the
VGT and EGR in a Turbocharged Common-
Rail Diesel Engine: Theory and Passenger
Car Implementation, SAE paper 2003-01-0357,
Jan. 2003.
[4] H. Ferreau et al., Predictive control of a real-
world diesel engine using an extended online ac-
tive set strategy, Annual Reviews in Control,
vol. 31, 2007.
[5] G. Stewart and F. Borrelli, A model predictive
control framework for industrial turbodiesel en-
gine control, in Proc. of IEEE Conference on
Decision and Control, 2008.
[6] Y. Wang, I. Haskara, and O. Yaniv, Model-
based quantitative feedback control of egr rate
and boost pressure for turbocharged diesel en-
gines, in Proc. of American Control Confer-
ence, 2008.
[7] J. Wahlstrom, L. Eriksson, and L. Nielsen,
Egr-vgt control and tuning for pumping work
minimization and emission control, IEEE
Transactions on Control Systems Technology,
Accepted for publication.
[8] M. Jankovic, M. Jankovic, and I. Kolmanovsky,
Constructive lyapunov control design for tur-
boharged diesel engines, IEEE Transactions
on Control Systems Technology, vol. 8, no. 2,
pp. 288299, 2000.
[9] E. Aleri, A. Amstutz, and L. Guzzella, Gain-
scheduled model-based feedback control of the
air/fuel ratio in diesel engines, Control Engi-
neering Practice, vol. 17, 2009.
[10] X. Wei, Gain scheduled H