International Journal of Control

Vol. 79, No. 5, May 2006, 449464

Automotive engine hybrid modelling and control
for reduction of hydrocarbon emissions
P. R. SANKETI*, J. C. ZAVALA and J. K. HEDRICK
Vehicle Dynamics Lab, Department of Mechanical Engineering,
University of California, Berkeley, USA-94720
(Received 29 June 2005; in nal form 24 November 2005)
Automotive engine models vary in their complexity depending on the intended application.
Pre-prototype performance prediction models can be very complex in order to make accurate
predictions. Controller design models need to be as simple as possible since model-based
controllers must operate in real time. This paper develops hybrid models for engine control
that incorporate time and events in their formulation. The resulting hybrid controllers have
the capability of switching between two alternative control modes. The rst mode is designed
to reduce the raw hydrocarbon (HC) emissions while the second mode tries to increase the
temperature of the catalytic converter as rapidly as possible during the initial transient or
cold start period. Reachability, as a tool for system analysis, is used to verify the properties
of the closed loop system.
1. Introduction
The synthesis of automotive engine controllers generally
employs simplied models, in particular mean value
models, which represent the engine behaviour well for
most conditions. There are, however, phases of the
engine operation for which more accurate description
is required. It is then that hybrid schemes can oer
better modelling capabilities. In particular, these can
be used eectively for the warm-up period of the
engine, which is the focus of this paper.
As much as 80% of the hydrocarbon (HC) emissions
in a typical engine drive cycle come from the initial
12 minutes of operation, commonly termed as the
coldstart period. There are three main factors why
a signicant portion of the hydrocarbon emissions
occur during the coldstart: the rst is that the cold
engine walls weaken the ame propagation due to the
heat transfer from the gas to the walls; the second is
that the catalytic converter is not active at temperatures
below 350

C; and the third is that the oxygen sensor

does not reach its operating temperature during
coldstart. With emission standards getting stricter
every year, coldstart emissions reduction is an area
that needs immediate attention.
Developing a correct system model is a crucial part of
a control synthesis cycle. The model must consistently
represent the behaviour of the plant in the desired
regions of operation. This is dicult to achieve for the
case of a combustion engine. The challenge is to nd
a level of representation that is detailed enough as to
contain the important dynamics of the system, but at
the same time is simple enough as to be useful for
control synthesis. A common approach in automotive
engine modelling is the use of mean value models
representing continuous ow of air and fuel into the
engine and also continuous production of torque and
pollutants, as presented in Aquino (1981). Very complex
models exist that can accurately predict the performance
during the coldstart period but are too complex to run in
real time. An intermediate level of complexity and detail
of description is achieved by hybrid models. The cycle to
cycle variation of the engine variables make the hybrid
paradigm a natural approach to engine modelling. In
Balluchi et al. (2000a), a hybrid model for calculating *Corresponding author. Email: [email protected]
International Journal of Control
ISSN 00207179 print/ISSN 13665820 online 2006 Taylor & Francis
http://www.tandf.co.uk/journals
DOI: 10.1080/10556780600605079
active torque generation is introduced that captures
the car manufacturers requirements related to force
requests in the longitudinal motion of the car. The
hybrid plant model is relaxed to a continuous time
model and the corresponding continuous time optimal
control problem is solved by applying classical methods.
In Baotic et al. (2003), hybrid system theory is used to
obtain a state-feedback optimal control law for an
electronic throttle. After modelling the electronic
throttle as a piecewise ane (PWA) system, an optimal
control law is derived for such a hybrid system via
dynamic programming. A similar approach is applied
to a multi-object adaptive cruise control problem in
Mobus et al. (2003). In Giorgetti et al. (2005), hybrid
tools are used to model and optimally control direct
injection stratied charge engines.
Alternative technologies have been proposed to
improve the coldstart emissions performance of the
engine. One such example is a catalyst that reacts to
the environment to achieve faster light-o, as shown
in Tanaka et al. (2001). In most of the cases where
physical changes to the catalyst have been proposed,
improvements have been shown in the performance of
the control of emissions. However, extra cost is added
as new devices and materials are incorporated into the
system. Simplied control oriented thermal models of
the catalytic converter and the engine are developed
in Shaw and Hedrick (2002), Shaw et al. (2002). Such
models are particularly useful for controller design.
Controllers with multiple control inputs have been
developed, although exhaust gas temperature, ignition
timing and air-fuel ratio (AFR) continue to be used
the most. Lean-limit control to reduce the HC emissions
is used in Lee et al. (2001). A practical approach is
followed in Chan and Hoang (1999): maintaining high
idle speed with a high value of ignition retard (HVIR)
with excess air factor; both together give high engine
exhaust temperature (T
exh
). In Sun and Sivashankar
(1998) the trade-o between catalyst light-o and raw
engine-out HC, and the eect of dierent operating
constraints on the catalyst light-o is discussed.
Isolated engine and catalyst models are used in Shaw
and Hedrick (2003) to determine optimum engine-out
parameter proles that could reduce the overall tailpipe
HC emissions. It was assumed in this paper that AFR
was available for measurement, or that an observer
could be used. One such observer was developed in
Tunestal and Hedrick (2001), using in-cylinder
measurements to predict the engine AFR through a
heat-release analysis. In most of the attempts to make
a model-based controller, the main focus has been on
faster catalyst light-o achieved by increasing T
exh
using ignition retard.
In this paper, the concepts of hybrid systems are
applied to the area of engine modelling and control
for coldstart analysis. A hybrid model for engine
coldstart purposes is developed. Mean value model-
based non-linear controllers for catalyst temperature
and raw HC emissions are developed. Further, a
hybrid controller is designed which uses the two
non-linear controllers as its modes to reduce the HC
emissions. The catalyst subsystem and the closed loop
system properties are veried using reachability
analysis with the aid of the Levelset toolbox devel-
oped by Professor Ian Mitchell, Assistant Professor,
Department of Computer Science, University of
British Columbia in Vancouver, Canada. The paper is
organized as follows. In x 2, a hybrid model of the
engine for coldstart is described. In x 3, a derivation of
the mean value and hybrid controllers is oered and
the trajectories of the closed loop system using both
mean value and hybrid models are displayed. Finally,
in x 4, reachable sets of the catalyst subsystem and
the closed-loop system are calculated. The reachability
analysis is done using the mean value engine model
with both the mean value and the hybrid controllers.
2. Engine hybrid model
2.1 Mean value engine model
A mean value model of the engine is described in this
section, which will form the basis of the hybrid model,
described in x 2.2. The mean value model uses the
main elements such as the air ow, the fuel ow rate
and the rotational inertia of the engine from the model
proposed in Cho and Hedrick (1989). The rst state of
the model is the mass air ow m
a
through the intake
manifold. Its dynamics can be dened as the dierence
between the mass ow rate into and out of the manifold
_ m
a
_ m
ai
_ m
ao
. 1
The mass ow rate running into the manifold is calcu-
lated as the maximum ow rate multiplied by throttle
angle and pressure ratio scaling factors
_ m
ai
_ m
ai, max
TCo PRIP
m
, P
a
. 2
The mass ow rate out of the manifold is a function of
engine displacement, V
e
; intake manifold volume, V
m
;
intake manifold pressure, P
m
; mass of air in the mani-
fold, m
a
; engine speed, o
e
and volumetric eciency, j
v
;
and is given by
_ m
ao

V
e
j
v
P
m
, o
e
m
a
o
e
4 V
m
. 3
450 P. R. Sanketi et al.
The volumetric eciency can be calculated using a
2-D look-up table with inputs o
e
and P
m
or a simpli-
ed function with similar behaviour. The fueling
dynamics consist of the fuel ow rate into the intake
port _ m
fo
as its state. As the fuel is injected into the
intake ports, part of the fuel vaporizes and part of
it deposits on the intake manifold as a liquid.
The fuel deposited on the intake manifold wall aects
the in-cylinder air-fuel ratio as it becomes part of the
air stream. A model of these dynamics, as proposed
in Aquino (1981) and used in Souder and Hedrick
(2004), is given by
_ m
fo
_ m
fv
_ m
ff
4
t
f
m
ff
_ m
ff
1 c _ m
fc
5
_ m
fv
c _ m
fc
, 6
where, c represents the portion of the fuel that enters
the cylinder directly as vapour, (1 c) is the portion
of fuel that is deposited on the manifold walls, _ m
fc
is
the commanded fuel mass ow rate, _ m
fv
is the mass
ow rate of fuel entering the cylinder directly as
vapour, _ m
ff
is the mass ow rate of fuel entering the
cylinder from the fuel puddle on the manifold walls.
The time constant t
f
is modeled as a constant here for
simplicity. Equations (4), (5) and (6) can be combined
into a single equation for use in the AFR controller
design
m
fo

1
t
f
_ m
fo
c m
fc

1
t
f
_ m
fc
. 7
The dynamics of the state due to the rotational inertia
of the engine can be given by
J
e
_ o
e
T
i
T
a
, 8
where J
e
is rotational inertia of the engine, T
i
is the
indicated torque and T
a
is the accessory load. T
i
is
modelled as a function of air mass per cylinder,
air-fuel ratio inuence, spark timing inuence and
crankshaft speed
T
i

c
T
_ m
ao
AFIz SPI
o
e
, 9
where AFIz is the normalized air-fuel ratio inuence
on torque production and SPI is the normalized
spark advance/retard inuence on torque production.
c
T
is a constant which is roughly the same for engines
with the same compression ratio.
The dynamics of the exhaust gas temperature T
exh
,
based on the analysis provided in Shaw and Hedrick
(2002), can be given by
_
T
exh

1
t
e
T
exh
ST AFI , 10
where, ST 7.5 600, Spark timing in deg after
top dead centre and AFI AFR inuence factor.
The catalyst model describes the thermal interaction
between the catalyst mass, the exhaust gas and the
chemical reactions taking place in the catalyst.
The dynamics are described by equation (11) through
(15) as proposed in Shaw et al. (2002).
mC
p
T
cat
dt

dQ
dt

gen

dQ
dt

in

dQ
dt

out
11
dQ
dt

gen
j
conv
H
gen
12
dQ
dt

in
h
in
A
in
T
exh
T
cat
13
dQ
dt

out
h
out
A
out
T
cat
T
atm
14
j
conv
1 exp a
1
z z
0
z

m
1
a
2
T
cat
T
0
T

m
2

,
15
where m is the catalyst mass, C
p
is the catalyst specic
heat, T
cat
is the catalyst temperature, j
conv
is the
conversion eciency of the catalyst, H
gen
is the heat
generated by the conversion of the pollutants inside
the catalytic converter, h
in
and h
out
are respectively
the inner and outer eective heat transfer coecients
of the catalyst.
2.2 Engine hybrid model
A hybrid model was developed from the mean value
model described in the previous section. Hybridness
was introduced in the subsystems that oered higher
modelling accuracy by being in a hybrid form. The
strokes of the engine dene a dierent set of dynamics
inside each cylinder and aect the production of
torque, pollutants and heat. The stroke of the engine
also determines the amount of exhaust gas and hydro-
carbons produced at a given time. The variation of
these variables from stroke to stroke will be considered
here as the main element of hybridness. The crankshaft
angle () is the decision variable for the transition from
one discrete state to another. Let us review the changes
to the proposed model due to the introduction of
hybridness. A diagram of the hybrid engine model is
Reduction of hydrocarbon emissions 451
shown in gure 1. The model consists of the following
subsystems: intake manifold, intake port, thermal
behaviour of the engine, torque production and catalyst.
The crankshaft angle (t) is used to index the moments
at which the exhaust mass corresponding to the combus-
tion cycle of each cylinder is available at the exhaust
port. In this manner, the production of hydrocarbons
is delayed by a combustion cycle. The intake manifold
integrates the dierence given in equation (1). The
intake port subsystem uses equation (3) to determine
_ m
ao
, however, the mass ow of air is separated for
each cylinder. This, mixed with the fuel, results in the
exhaust gas after being subjected to the combustion
cycle delay. The AFR and _ m
fo
are also calculated
inside the intake port. In the intake port, the angle
(t) forms the index of _ m
ao
. The exhaust temperature
and the raw emissions are calculated in the thermal
block according to equation (10) and (21). The torque
model contains a nite state machine for each cylinder,
consisting of four states: Intake I, Compression C,
Expansion E, and Exhaust X. The transition from one
state to the other is based on the position of the cylinder.
The top dead centre (TDC) is the starting position for
the intake stroke and is indexed with t 0. At the
end of the intake stroke (t ), there is a transition
to the compression mode. The expansion mode C
starts at t 0 and the exhaust mode X at t .
The torque produced by each stroke is generated by
continuous functions f
1
through f
4
, corresponding
respectively to the four modes. Each of the torque func-
tions f
i
have the following inputs: the air mass ow rate
_ m
ao
, the air-fuel ratio AFR, the spark timing delay
and the engine speed o
e
. The torque function f
i
has a
negative prole for the intake, compression and exhaust
strokes, whereas it is positive for the expansion stroke.
The mean value of f
i
along the full cycle is positive.
The torque model accounts for the delay due to the
combustion cycle.
2.3 Comparison of models
In our experience, there are two main aspects that make
the mean value engine model dierent from the hybrid
engine model.
.
The mean value model gives a good approximation
of the behaviour of the engine during warmed-up
operation and it is simple. On the other hand, it
is not suitable to capture all the events that are of
interest for the study of the coldstart period.
. The hybrid model is more complex and more dicult
to validate, however it can be useful to describe
instantaneous events and short period transients,
which are important during coldstart.
The hybrid model presented here is dierent fromthe one
developed in Balluchi et al. (2000a) mainly in the
following aspects. The rst dierence is that we assume
that there is no load during the rst few seconds of
operation of the engine, hence our model does not
include the drivetrain dynamics. The second dierence
is found in the torque generation. In our model, the
torque generated by each cylinder is a continuous func-
tion with dierent proles in dierent strokes. During
the compression, exhaust and intake strokes, the torque
is negative due to the force that is applied by the other
pistons to compress, expel or suck the cylinder charge.
Figure 1. Engine hybrid model.
452 P. R. Sanketi et al.
The positive torque is produced during the expansion
phase. In the model described in Balluchi et al. (2000a),
the torque is produced by a zero-order hold during
the expansion phase. In this manner, the torque is
characterized by constant levels for each stroke.
However, in Balluchi et al. (2001), a memoryless generic
function is used for the torque generation of the model,
allowing for non-constant proles during the four
strokes. Both of the models presented in Balluchi et al.
(2000a, 2001) use the stroke state, the piston
position, the mass and fuel loaded into the cylinder and
the spark timing as parameters of the torque
function. The third dierence is the number of states
in the FSM for the cylinders. In our model there are
four states, whereas in the model presented in Balluchi
et al. (1998) there is one state for each stroke, plus two
more corresponding to the expansion stroke under
advance and retard of spark timing.
3. Hybrid control
This section deals with the design of a hybrid controller
for minimizing the tailpipe HC emissions during the
coldstart period using the mean value engine model.
Initially, the problem is described in terms of the
inputs and outputs of the system. The control laws
designed are then explained. Further, the results are
presented and compared with those of mean value
controllers. The performance of the hybrid controller
is assessed by testing it with the hybrid engine model.
Finally, some notes on the stability of the proposed
controller are added.
3.1 Problem description
The objective is to minimize the cumulative tailpipe
hydrocarbons. The tailpipe HCs depend on the raw
HCs produced by the engine and the eciency of the
catalyst. The catalyst is said to have achieved light-o
when its eciency is above 50%. There is a trade-o
between how fast the catalyst light-o can be achieved
and how much the raw HCs can be reduced. Thus,
just reducing the raw emissions does not necessarily
mean minimizing the tailpipe emissions. The control
inputs available to optimize the performance index are
throttle angle o, commanded fuel injection rate _ m
fc
and the spark timing . It is assumed that the engine
speed sensor, linear air-fuel (UEGO) sensor, exhaust
gas temperature and catalyst temperature sensors and
HC analyser (raw and tailpipe emissions) are available
for measurement throughout the coldstart period.
However, in practice, the AFR sensors are not active
until around 10s, and the HC analyser delay is too big
to be used in real-time control. In Tunestal and
Hedrick (2001), an observer for the AFR using the
in-cylinder pressure measurements is designed. The
in-cylinder pressure is directly measured using a piezo-
electric pressure sensor. The raw HCs can be predicted
knowing the AFR and estimating the amount of fuel
injected. The amount of fuel injected can be predicted
using the injection duration. Alternatively, a black-
box model for the raw HCs can be developed by
analysing the constant delay in the FID HC analyser
measurements.
3.2 Hybrid controller
3.2.1 Motivation. Mean value controllers have been
developed in literature to control the engine exhaust
gas temperature T
exh
(Shaw and Hedrick 2003) and
the AFR (Souder and Hedrick 2004). In the context
of coldstart emissions control, T
exh
is important to get
the catalyst warmed up faster, whereas keeping the
AFR lean reduces the raw emissions itself. In this
paper, two high level dynamic surface controllers
(Song et al. 2002) for catalyst temperature and raw
HCs were developed using T
exh
and the AFR as
synthetic inputs respectively. As mentioned before,
there is a trade-o between the two. The two dynamic
surface controllers were designed without incorporating
the coupling between them, i.e. each controller tries to
achieve its own objective. Hence, a mean value control-
ler which consists of these two controllers running in
parallel with static gains may not exploit the trade-o.
By designing a switching controller consisting of two
modes where one of the two controllers is preferred in
each, the coupling between them is made use of and
either of the two objectives is not highly penalized.
In one mode, fast catalyst light-o is favoured and
in another reducing the raw HC emissions is favoured.
Achieving similar optimality with controllers having
constant gains would require designing a multi-input
multi-output (MIMO) dynamic surface control,
which is not easy. Also, the switching algorithm is
more robust with respect to varying driving conditions.
Next we discuss the components of the controller.
3.2.2 Modes of hybrid controller. The overall
hierarchical structure of the controller is shown in
gure 2. The topmost level controller is the hybrid
controller, followed by catalyst temperature (T
cat
) and
engine out raw emissions (
_
HC
out
) dynamic surface
controllers, T
exh
and AFR controllers, spark timing
and commanded fuel injection rate controllers in that
order. As seen in gure 3, the high-level hybrid
controller consists of two modes, namely T
cat
control
dominant mode and
_
HC
out
control dominant mode.
In the
_
HC
out
control dominant mode, the gain z
1b
on
Reduction of hydrocarbon emissions 453
the T
cat
controller (described further in x 3.2.2)
represented by surface S
1
is low and the gain z
3b
on
the
_
HC
out
controller (described further in x 3.2.3)
represented by surface S
3
is high. In the T
cat
control dominant mode, the gain z
1a
on T
cat
controller
is high whereas the gain z
3a
on
_
HC
out
controller is
low. The dynamic surface controller for T
cat
uses T
exh
as the control variable whereas the dynamic surface
controller for
_
HC
out
uses AFR as the control variable.
The low level T
exh
and AFR controllers use spark
timing () and fuel injection rate ( _ m
fc
) as the inputs
respectively. These are designed as single-inputsingle-
output (SISO) controllers.
The high-level control strategy is as follows. Initially,
the controller is in T
cat
control dominant mode where
the gain on T
cat
control is high. Hence, there is an
eort to increase the catalyst temperature helping fast
catalyst light-o, however at the expense of raw HC
emissions. When the raw HCs reach a certain upper
bound L
1
, the hybrid controller switches to the
_
HC
out
control dominant mode, where the gain on the T
cat
control is low. In this mode, there is no special eort
to increase the catalyst temperature and the focus is
on getting the raw emissions rate as low as possible.
When the raw emissions reach a certain lower bound
L
2
, the hybrid controller switches back to T
cat
control
dominant mode. The switching continues till the catalyst
light-o is achieved. The constituents of the hybrid
controller are described in detail next.
3.2.3 Controller derivation T
cat
control. Catalyst
temperature is strongly dependent on the engine exhaust
temperature, which is strongly dependent on the ignition
timing. Using dynamic surface control, we control T
cat
treating T
exh
as a synthetic input. We dene a sliding
surface equal to the dierence between the actual and
desired value of the catalyst temperature.
S
1
T
cat
T
cat, d
_
S
1

_
T
cat

_
T
cat, d
.
16
Substituting for the dynamics of the catalyst
temperature from equation (11), we get,
_
S
1

_
Q
gen

_
Q
in

_
Q
out
mC
p

_
T
cat, d
.
Denoting the catalyst internal surface area and heat
transfer coecient as A
in
and h
in
respectively, we have
_
Q
in
h
in
A
in
T
exh
T
cat
.
Similarly,
_
Q
out
h
out
A
out
T
cat
T
atm
,
where T
atm
is the ambient temperature.
Treating T
exh
as the input, we design the control law
to obtain
_
S
1
z
1
S
1
,
where z
1
is a positive gain. This leads to the following
equation
"
T
exh


_
T
cat, d
z
1
S
1
mC
p

_
Q
gen

_
Q
out
h
in
A
in
T
cat
, 17
Figure 2. Hierarchy of controllers.
Figure 3. Modes of hybrid controller.
454 P. R. Sanketi et al.
where
"
T
exh
is the synthetic input. To track the desired
value of the synthetic input, we need to nd its deriva-
tive, which can lead to too many terms called the
explosion of terms problem (Swaroop et al. 2000).
Also, the term
"
T
exh
may include uncertainties which
can lead to problems in dierentiation. Hence, the
desired value of T
exh
to be tracked is found by passing
the synthetic input through a low-pass lter so that
explosion of terms problem and taking unknown
derivatives is avoided
t
T
_
T
exh,d
T
exh,d

"
T
exh
. 18
Then, we dene a sliding surface based on the dierence
between the actual and the desired exhaust gas
temperature. This part of the T
exh
tracking controller
is based on Shaw and Hedrick (2003).
S
2
T
exh
T
exh,d
_
S
2

_
T
exh

_
T
exh,d
.
19
Knowing that T
exh
depends on the spark timing ,
we design the control law to obtain
_
S
2
z
2
S
2
where z
2
is a positive gain.
Using the plant dynamics given by equation (10),
we get the control law as,

1
7.5
t
e
AFI
T
exh
t
e

_
T
exh,d
z
2
S
2

600

. 20
3.2.4 Controller derivation Engine exhaust HC
control. The engine-out HC emissions denoted by
_
HC
out
are strongly dependent on the AFR. Therefore,
AFR is treated as a synthetic input to control HC
out
.
The inversion of the following expression in terms
of AFR is used to devise the controller
_
HC
out
_ m
f
r
c
1
r
c
exp a

EVO

0

m

, 21
where, r
c
is the compression ratio,
EVO
is the exhaust
valve opening angle and

0
k
1
k
2
k
3
AFR 14.7
2
k
4
k
2
, k
3
, k
4
, a and m being the model parameters.
Dene a sliding surface as the dierence between the
engine out HC emissions rate and the desired rate.
S
3

_
HC
out

_
HC
out, d
22
Dierentiating,
_
S
3


HC
out
0 23
Since the calculation of

HC
out
is complex, it will be
dicult to invert that equation in terms of AFR.
Hence, we pass
_
HC
out
through a rst order lter to
obtain
_
HC
f, out
as follows:
t
p

HC
f, out

_
HC
f, out

_
HC
out
)

HC
f, out

1
t
p

_
HC
out

_
HC
f, out
. 24
Substituting equation (24) in equation (23), and using
AFR as a synthetic input, we design the controller to get
_
S
3

1
t
p

_
HC
out

_
HC
f, out
z
3
S
3
z
3
being a positive gain. After some algebra, we get
the synthetic input as
"
AFR X
1,m
1

EVO

0
k
4

1
k
3

1,2
14.7, 25
where
X
1
a
log
_ m
f
r
c
1

_
HC
f, out
t
p
z
3
S
3
r
c
" #
.
Again, this is passed through a lter to get the desired
AFR
d
) and
_
AFR
d
. As mentioned before, explosion of
terms and taking unknown derivatives is avoided using
this method
t
A
_
AFR
d
AFR
d

"
AFR. 26
To track the desired AFR, we dene a sliding surface
as follows:
S
4
_ m
fo
,
,
_ m
ao
AFR
d
)
_
S
4
m
fo
_ ,,
27
Reduction of hydrocarbon emissions 455
where, _ m
ao
is the manifold out air ow rate. This part
of the AFR tracking controller is based on Souder and
Hedrick (2004). Using the fuelling dynamics given by
equation (7),
_
S
4

1
t
f
_ m
fo
_ m
fc
_ ,.
The commanded fuel ow is used as the input to obtain
_
S
4
z
4
S
4
,
where z
4
is a positive gain. After further simplication,
we get the following control law
_ m
fc
_ m
fo
t
f
m
ao
AFR
d

_ m
ao
_
AFR
d
AFR
d
2
z
4
S
4
" #
. 28
3.3 Results and analysis
3.3.1 Controller performance. The controllers designed
were applied to the engine and the catalyst models
and the performances were simulated at idle
condition. First, a mean value controller that uses the
catalyst in the feedback loop in real-time was simulated.
If the catalyst is not in the feedback loop, such as the
controller described in Shaw and Hedrick (2003), then
the inputs to the catalyst have to be calculated oine,
and hence, the catalyst properties such as temperature
and eciency cannot be controlled directly. Also,
having the catalyst in the closed loop is useful in
accounting for ageing of the catalyst. The mean value
controller comprises a T
cat
controller and a
_
HC
out
controller described in xx 3.2.2 and 3.2.3 respectively.
The controllers work in parallel, but their gains are
tuned constant. Figure 4 shows the performance of the
controller (mean value controller implies T
cat
controller
and
_
HC
out
controller running in parallel with tuned
static gains), where the cumulative HC emissions are
seen to be 7.5 g. As mentioned previously, this controller
with constant gains does not really take advantage of
the trade-o between the two controllers. Next, the
hybrid controller (described in x 3.2) in which the gains
on the two lower level controllers can be changed in
real-time was simulated.
Figure 5 shows the performance of the hybrid control-
ler. Even though the catalyst light-o is not as fast,
the cumulative tailpipe HC emissions are about 3.5 g
as compared to around 7.5 g in the previous controller.
Figure 6 indicates the hybrid nature of the controller.
The oscillations in the HC ppm graph correspond to
the switching of the hybrid controller between its two
modes. By doing so, it exploits the trade-o between
the raw emissions and the catalyst light-o. Hence, the
hybrid controller performs better than the mean value
controller. On zooming into gure 6, it is found that
switching takes place approximately every two engine
cycles (refer to gure 7). It is feasible to control the
AFR in practice in that interval since the delays
associated with the measurements of AFR and HC are
at most one combustion cycle. In case the HC
measurements were not available, HC measurements
can be observed through the estimation of AFR (using
in-cylinder pressure measurements) and the amount of
fuel injected.
Finally, the hybrid controller was assessed with
the hybrid model described in x 2.2. Its performance
0 10 20 30 40 50 60 70
0
500
Catalyst Temperature
T
e
m
p

(

C
)
0 10 20 30 40 50 60 70
0
0.5
1
E
t
a
h
c
0 10 20 30 40 50 60 70
0
5
10
Cumulative Tailpipe HC
H
C
t
p

C
u
m

(
g
)
Time (s)
Figure 4. Mean value controller: Catalyst eciency and
cumulative tailpipe HC.
0 10 20 30 40 50 60 70
0
500
Catalyst Temperature
T
e
m
p

(

C
)
0 10 20 30 40 50 60 70
0
0.5
1
E
t
a
h
c
0 10 20 30 40 50 60 70
0
2
4
Cumulative Tailpipe HC
H
C
t
p

C
u
m

(
g
)
Time (s)
Figure 5. Hybrid controller: Catalyst eciency and
cumulative tailpipe HC.
456 P. R. Sanketi et al.
is shown in gures 8 and 9. The performance of
the hybrid control is very similar to that of the mean
value engine model. We believe that the hybrid model
can be advantageous during the rst few cycles of
operation of the engine, when the amount of
hydrocarbon emissions are the highest. During
those cycles, it can be useful to determine how many
hydrocarbons are emitted due to each cycle and how
the fuel dynamics in the manifold inuences those
emissions. The design of the controller can be then
adapted to consider those factors.
The switching time of the controller was found to be
related to the values of the gains of the two modes of
the controller. As the gains are changed, the dynamics
of the closed-loop system is aected and the controlled
variables react at dierent velocities. However, the
performance of the controller is not aected
signicantly by small variation in the switching times.
Thus, the change of the gains in the modes of the
controller allows for adjustment of the switching
times if required. Regarding the synchronization
of the controller with the engine cycle, our model
does not time the fuel injection events or spark
events, hence, synchronization is not required.
Synchronization becomes crucial during the controller
implementation, when the fuelling dynamics have to
be considered for fuel injection timing.
0 1 2 3 4 5
2000
4000
6000
8000
10000
12000
Raw and Tailpipe HC Comparison (Detail)
H
C

p
p
m

R
a
t
e
Time (s)
TP
Exh
Figure 7. Hybrid controller: Tailpipe and raw HC
comparison (Detail).
0 10 20 30 40 50 60 70
0
500
Catalyst Temperature
T
e
m
p

(

C
)
0 10 20 30 40 50 60 70
0
0.5
1
E
t
a
h
c
0 10 20 30 40 50 60 70
0
2
4
Cumulative Tailpipe HC
H
C
t
p

C
u
m

(
g
)
Time (s)
Figure 8. Hybrid controller with hybrid model: Catalyst
eciency and cumulative tailpipe HC.
0 5 10 15 20 25 30 35 40
2000
4000
6000
8000
10000
12000
Raw and Tailpipe HC Comparison
H
C

p
p
m

R
a
t
e
Time (s)
TP
Exh
Figure 9. Hybrid controller with hybrid model: Tailpipe and
raw HC comparison.
0 5 10 15 20 25 30 35 40
2000
4000
6000
8000
10000
12000
Raw and Tailpipe HC Comparison
H
C

p
p
m

R
a
t
e
Time (s)
TP
Exh

Figure 6. Hybrid controller: Tailpipe and raw HC

comparison.
Reduction of hydrocarbon emissions 457
3.3.2 Notes on the stability of the sliding surfaces. To
analyse the convergence properties of the controller,
it is important to look at the evolution of the sliding
surface S
i
0 as a function of time.
Figures 10 and 11 show respectively the evolution
of the sliding surfaces when the mean value controller
and the hybrid controller are used with the mean value
engine model. All the surfaces reach the boundary
layer near zero without any instabilities. The presence
of oscillations in T
exh
is due to the hybrid nature of
the controller. Figure 12 shows the evolution of surfaces
when the hybrid controller is used with hybrid engine
model. Though there are inherent delays in the hybrid
model, the surfaces still converge well. To analyse the
eect of sensor delays and transport delays on the
controller performance, the mean value and hybrid
models were subjected to such delays of the order of a
combustion cycle (besides the delays already present in
the formulated models). Figures 13 and 14 demonstrate
the eect of delays on the evolution of the surfaces for
mean value and hybrid controllers. As seen from these
gures, instabilities usher in for a short period and the
surfaces stabilize after some time. When the delay is
increased to more than a combustion cycle, for both
the mean value and the hybrid controllers, the closed
loop system becomes unstable.
4. Reachability analysis
4.1 Motivation
Even though ecient tools are available for simulation
of the engine plant, the question of stability and evolu-
tion of the states under a hybrid modelling scheme has
not been answered completely. One reason is that
model simulations cannot cover all possible trajectories
inside a set. An approach to the problem is to apply
the concepts of reachable sets, which would give an indi-
cation of how the states can evolve with time given the
bounds on inputs and disturbances. Excellent examples
of application of reachability analysis in hybrid control-
ler synthesis can be found in Balluchi et al. (2000b) and
Tomlin et al. (1998). In particular, our objective is to
verify the properties of the controllers that we have
designed. First a backwards reachable set of the open-
loop catalyst subsystem is calculated which establishes
the required capability of the controller. Further, to
verify the stability and safety of the controller, both
the forward and backward reachable sets of the system
under closed-loop control are calculated.
0 10 20 30 40 50 60 70
20
0
20
S
1
,

r
a
d
/
s
S1, speed surface
0 10 20 30 40 50 60 70
2
0
2
10
18
S
2

(
k
g
)
S2, manifold air mass
0 10 20 30 40 50 60 70
500
0
500
S
3

(
K
)
S3, exhaust temperature surface
0 10 20 30 40 50 60 70
50
0
50
S
4

(

)
S4, Air to fuel ratio surface
0 10 20 30 40 50 60 70
400
200
0
Time (s)
S
5

(
K
)
S5, Catalytic Converter Temperature
Figure 10. Sliding surfaces: Mean value controller with
mean value model.
0 10 20 30 40 50 60 70
20
0
20
S
1
,

r
a
d
/
s
S1, speed surface
0 10 20 30 40 50 60 70
1
0
1
10
18
S
2

(
k
g
)
S2, manifold air mass
0 10 20 30 40 50 60 70
500
0
500
S
3

(
K
)

S3, exhaust temperature surface
0 10 20 30 40 50 60 70
5
0
5
S
4

(

)
S4, Air to fuel ratio surface
0 10 20 30 40 50 60 70
500
0
500
Time (s)
S
5

(
K
)
S5, Catalytic Converter Temperature
Figure 11. Sliding surfaces: Hybrid controller with mean
value model.
458 P. R. Sanketi et al.
4.2 Introduction to reachability
As engineering systems have become more complex,
methods for verifying the correct behaviour of such
systems have been developed. Model checking, in
particular, is an important issue in this regard. It is
a verication method in which the state space of the
design is explored in order to determine whether the
system can enter into an unsafe or incorrect state.
Many model checking algorithms attempt to compute
a reachable set, which can be of two types. A forwards
reachable set is the set of states that can be reached by
system trajectories which start in a given set of initial
states. A backwards reachable set is the set of states
that can give rise to trajectories which subsequently
pass through some given set of target states. The
target of a backwards reachable set can dened as
safe or unsafe. The rst type is used to determine
the states from which the system will eventually reach
a desired safe set. The latter case is to determine
from which states, the system will eventually reach an
undesired unsafe set.
Consider the backwards reachable set from a target
set T of a continuous system with dynamics
_ x fx, a, b, where x is the state of the system, a is
an input seeking to keep the system from entering T,
and b is an input seeking to drive the system into T.
0 10 20 30 40 50 60 70
0 10 20 30 40 50 60 70
0 10 20 30 40 50 60 70
0 10 20 30 40 50 60 70
0 10 20 30 40 50 60 70
10
0
10
S
1
,

r
a
d
/
s
S1, speed surface
2
0
2
10
18
10
3

S
2
,

[
k
g
]
S2, manifold air mass
1000
0
1000
S
3
,

[
K
]

S3, exhaust temperature surface
1
0
1
S
4
,
[

]

S4, Air to fuel ratio surface
500
0
500
Time [s]
S
5
,

[
K
]
S5, Catalytic Converter Temperature
Figure 12. Sliding surfaces: Hybrid controller with hybrid
model.
0 10 20 30 40 50 60 70
10
0
10
S
1
,

r
a
d
/
s
S1, speed surface
0 10 20 30 40 50 60 70
1
0
1
10
18
S
2

(
k
g
)
S2, manifold air mass
0 10 20 30 40 50 60 70
1000
0
1000
S
3

(
K
)

S3, exhaust temperature surface
0 10 20 30 40 50 60 70
1
0
1
10
3
S
4

(

)
S4, Air to fuel ratio surface
0 10 20 30 40 50 60 70
500
0
500
Time (s)
S
5

(
K
)
S5, Catalytic Converter Temperature
Figure 13. Sliding surfaces with delays: Mean value
controller with mean value model.
0 10 20 30 40 50 60 70
10
0
10
S
1
,

r
a
d
/
s
S1, speed surface
0 10 20 30 40 50 60 70
1
0
1
10
18
S
2

(
k
g
)
S2, manifold air mass
0 10 20 30 40 50 60 70
1000
0
1000
S
3

(
K
)

S3, exhaust temperature surface
0 10 20 30 40 50 60 70
2
0
2
10
3
S
4

(

)
S4, Air to fuel ratio surface
0 10 20 30 40 50 60 70
500
0
500
Time (s)
S
5

(
K
)
S5, Catalytic Converter Temperature
Figure 14. Sliding surfaces with delays: Hybrid controller
with hybrid model.
Reduction of hydrocarbon emissions 459
Considering T as an unsafe set, a should be considered
a controller keeping the system safe, and b consists of
disturbances or model uncertainties which are assumed
to try to make the system unsafe. Computation of the
backwards reachable set is normally encoded as a
terminal value Hamilton-Jacobi (HJ) PDE, the same
as an initial value PDE, but with the time running
backwards (Mitchell 2005). The reach set is represented
by an implicit surface function x, t. The terminal
value of the implicit surface function is same as the
target set, so x, 0 should be a representation of T.
Evolution of the backwards reachable set is
accomplished by solving the following equation
backwards in time.
D
t
x, t min0, Hx, D
x
x, t 0, 29
where
Hx, p max
a
min
b
p
T
fx, a, b 30
H(x, p) being the Hamiltonian. (In case of T being a
safe set, the input a will minimize the Hamiltonian.)
The solution x, t is an implicit surface representation
of the nite time backwards reachable set. For our
analysis, a Level Set Method Toolbox (Mitchell 2005)
developed by Prof. Ian Mitchell was used to calculate
the backwards reachable set. The toolbox is designed
for solving initial value PDEs. For autonomous systems,
converting the terminal value PDEs to the initial
value PDE form used in the toolbox simply requires
multiplying f by 1 (Mitchell 2005).
4.3 Backwards reachability of catalyst subsystem
As stated before, catalyst light-o is one of the essential
factors in coldstart control. With an aim to nd out the
time required to achieve the light-o of the catalyst
starting from a particular initial condition, we analyze
the reachability of the catalyst subsystem. The back-
wards reachable set at any instant t gives the range of
T
cat
in which the system should start so as to reach
the light-o temperature within time t. Obviously, the
reachable set at time t 0 s is a very small interval
around the catalyst light-o temperature, which is
approximately 350 to 400

C. Note that the target set

here is the set that the system wants to reach.
The reach set can only grow as the time increases. The
reach set, hence, gives an idea of the input to be given
to the system so as to be able to achieve the light-o
within a given time period, which in turn reects on
the required capability of the controller.
4.3.1 Results and analysis. Referring to the dynamics
of the catalyst temperature given by equation (11), the
AFR and the T
exh
can be considered as the inputs to
the catalyst subsystem. The AFR is assumed to be
constant around stoichiometric value and the only
active input is T
exh
.
The state equation of T
cat
is ane in the input, and
thus calculating the Hamiltonian is not very dicult.
The maximum input that can be given to the catalyst
is limited to a certain constant value. The results
obtained for an AFR of 14.8 and a maximum input of
400

C are shown in gure 15. The y-axis represents

the level set as a function of catalyst temperature
which is represented on the x-axis (for higher dimen-
sions, just the surface x 0 is shown). The gure
shows how the level set grows with time. The reachable
set that we are interested in is the interval in which the
level set is negative. For example, the reach set at
t 50 s is the interval between 310 K (37

C) and 675 K
(402

C), i.e. if the catalyst is at 37

C to start with,
then the light-o can be achieved within 50 s if the
T
exh
can be maintained at least at 400

C. From control-
lers standpoint, if the system starts at 37

C, we know
that an input of T
exh
400

C is required to achieve
the light-o in 50 s. If you want to achieve a faster
light-o, then a higher level of input is needed.
Figure 16 indicates that to achieve a light-o in 30 s
starting from the same initial point 37

C, an input
level of 700

C is required. Hence, this analysis provides

useful inputs for control design. It can be seen that the
reach set moves in the right direction and it does not
have any abrupt changes in direction.
300 350 400 450 500 550 600 650 700
0
50
100
150
200
250
300
350
T
cat
(K)
L
e
v
e
l

s
e
t

f
u
n
c
t
i
o
n
Catalyst Reachability Analysis
t=0 s
t=10 s
t=20 s
t=30 s
t=40 s
t=50 s
Figure 15. Backward reachability analysis of catalyst:
T
exh, max
400

C.
460 P. R. Sanketi et al.
The analysis was repeated for various values of AFR.
The results indicated that for either rich or very lean
values of AFR, more time is required for the catalyst
light-o.
It is interesting to note that in practice, once the
engine warms-up and T
exh
reaches its maximum
value, it would take hours for the engine to cool
down. This means that the input can be applied in
only one direction. However, it is ne for our analysis,
since it makes sense to apply the maximum possible
input to reach the target set from a given initial condi-
tion in minimum possible time. In practice, spark
timing can be used as an input to eectively achieve
the required T
exh
.
4.4 Forward reachability of closed-loop system
In order to analyse the stability of the controller
designed, the forward reachable set of the closed-loop
system was calculated. We wanted to show that the
states remain bounded under the closed-loop control.
Initially, the forward reachable set for the whole
system comprising of 5 states was attempted.
However, the reach set calculation was very slow, even
on a coarse grid. Hence, a subset of 3 states was
analysed. The simplications made were that the
engine speed and the amount of air inside the intake
manifold were xed, which is reasonable at idle.
The values of the constant states were chosen such
that the trajectories of the original closed loop system
matched those of this simplied version.
Figure 17 shows the forward reach set at t 50 s of
the closed loop system starting from a set which is the
set of all possible initial conditions (the _ m
fo
axis is
scaled by 10
5
for better visualization. The variables
corresponding to the dierent axes are X _ m
fo

10
5
kg/s, Y T
exh
K, Z T
cat
K). It was observed
that the reach set does not move as much as expected
(found through various ODE simulations) in the direc-
tion of T
cat
. Another reach set was calculated with the
initial conditions being the set corresponding to the
values expected at around 30 s. Also this reach set was
found not to move much in the T
cat
direction.
On further research, it was found that the result was
similar unless the initial set was chosen corresponding
to the time when the dynamics of the system slow
down. The reason for this behaviour, as pointed out
by Prof. Ian Mitchell, is that the system consists of
dynamics that have dierent time scales. In particular,
the fuel dynamics are much faster than the catalyst
temperature dynamics. Unfortunately, the level set
code deals poorly with such systems. They require
very small timesteps, and the reach set might quickly
converge to an incorrect xpoint. Hence, the forward
reach set of the closed-loop system with the hybrid
controller was calculated starting from 50 s when the
system dynamics are expected to settle relatively soon.
The reach set was found to grow, although at a slow
rate and then it became invariant in all directions.
The reach set is shown in gures 18 and 19. Hence,
all the states remain bounded under the closed-loop
control.
250 300 350 400 450 500 550 600 650
0
50
100
150
200
250
300
350
Catalyst Reachability Analysis
T
cat
(K)
L
e
v
e
l

s
e
t

f
u
n
c
t
i
o
n
t=0 s
t=10 s
t=20 s
t=30 s
Figure 16. Backward reachability analysis of catalyst:
T
exh, max
7008C.
Figure 17. Forward reachable set for closed-loop system.
Reduction of hydrocarbon emissions 461
4.5 Backward reachability of closed-loop system
The forward reachability analysis presented in the
previous section helped us verify the stability of the
closed-loop system. It would be useful to also verify
the possibility of the system reaching an unsafe set.
This can be done by calculating the backwards
reachable set of the closed-loop system starting from
an undesired state. If the backwards reach set does not
intersect the set of possible initial states within the
time of interest, then the system can be termed safe
with respect to the given unsafe state. In
particular, we are interested in nding out if the catalyst
temperature can rise to very high values, thus damaging
the catalyst.
The same simplied system as that in the previous
section was used for analysis. However, only the mean
value controller was used since the methods for calculat-
ing the backwards reachable set for a hybrid system are
still under development. The target or unsafe set for
the backwards reachability analysis in this case was
a small ball around the center given by catalyst tempera-
ture of 1023 K (750

C) and the other states close to their

steady state values. Figure 20 shows the backwards
reachable set. The specied target set can be reached
from values of T
exh
and _ m
fo
outside the boundaries of
the set shown, but we are interested in determining if
the reach set ever intersects the set of initial conditions.
It can be seen that at t 60 s, the set reaches 644 K
(371

C) in the T
cat
direction. Since we are analysing
the coldstart problem, it is valid to assume that the
catalyst temperature is close to the atmospheric
temperature when the engine starts. Hence, any initial
condition of the engine does not belong to the back-
wards reachable set we obtained. In other words, the
backwards reachable set does not intersect the set of
all possible initial conditions. Thus, we can conclude
that the closed-loop system is safe with respect to
the specied unsafe set.
5. Conclusions
Utilizing the structure of a mean value model as the
base, hybridness was introduced in selected engine
subsystems. The resulting hybrid model accounted for
the dierent strokes during the coldstart operation of
Figure 18. Forward reachable set for closed-loop system.
Figure 19. Forward reachable set for closed-loop system:
Another view.
Figure 20. Backward reachable set for closed-loop system.
462 P. R. Sanketi et al.
the engine. Even though this model describes reasonably
well the engine torque generation, it can still be extended
to include other event based eects, such as valve timing,
spark timing, etc. The complexity of the interaction of
the engine subsystems is a limiting factor in the inclusion
of all the hybrid elements in the model.
Mean value model based exhaust gas temperature
and air-fuel ratio controllers were used to synthesize
a dynamic surface controller where the eects of the
catalyst were included in the feedback loop for real-
time control. In this way, catalyst temperature control
combined with raw HC emissions control was achieved.
As an approach to optimizing the trade-o between
minimizing the raw HC emissions and decreasing the
catalyst light-o time, a hybrid controller was derived.
The simulations show a decrease in the production of
cumulative HC. Hence, the application of the hybrid
automata for improvement in coldstart control is
demonstrated.
In order to investigate how the states of the system
evolve with time, the LevelSet Toolbox developed by
Prof. Ian Mitchell was used for reachability analysis
of the system. The rst objective was to calculate the
backwards reachable set for the catalyst subsystem.
The reach set obtained matches the expected behaviour
of the catalyst and gives a description of the required
capability of the catalyst temperature controller.
Further, forward reachability analysis was performed
for the closed-loop system in order to verify the stability
of the proposed hybrid controller. The resulting reach
set was not as expected for certain initial conditions
due to the multiple time scales in the system dynamics.
The fuelling dynamics are much faster than the catalyst
dynamics. The reach set starting from close to steady
operation was found to be invariant after some time,
thus indicating the stability of the hybrid controller.
Furthermore, to verify the safety property of the
controller, a backwards reachable set of the closed-
loop system was calculated starting with an unsafe set
where the catalyst temperature is very high. However,
only the mean value controller was used since the
methods for calculating the backwards reachable set
for a hybrid system are still under development.
The backwards reach set was found not to intersect
with the set of possible initial conditions of the system,
thus concluding the safety of the controller with respect
to the specied unsafe set.
In summary, hybrid modelling and control provides
us with useful tools to address the coldstart problem.
Acknowledgements
Authors acknowledge the nancial support provided
by the National Science Foundation (under by NSF
Cooperative Agreement No. CCR-0225610), Toyota
Motor Corporation and CONACYT (Consejo
Nacional de Ciencia y Technologa de Mexico).
The authors would like to thank Prof. Shankar Sastry,
Prof. Ian Mitchell, Dr. Jonathan Sprinkle and Prof.
Claire Tomlin for their help. We would like to acknowl-
edge in particular the eort and time put in by Prof.
Mitchell in assisting us with the reachability analysis.
We would also like to thank the reviewers of this
paper for their useful suggestions and comments.
References
C.F. Aquino, Transient a/f control characteristics of the 5 liter central
fuel injection engine, SAE 810494, 1981.
A. Balluchi, L. Benvenuti, M.D. Di Benedetto, C. Pinello and
A. Sangiovanni-Vincentelli, Automotive engine control and
hybrid systems: Challenges and opportunities, in Proceedings of
the IEEE, 88, pp. 888912, 2000a.
A. Balluchi, L. Benvenuti, M.D. Di Benedetto, G.M. Miconi, U. Pozzi,
T. Villa, H. Wong-Toi and A.L. Sangiovanni-Vincentelli, Maximal
safe set computation for idle speed control of an automotive
engine, 1790, pp. 3244, 2000b.
A. Balluchi, L. Benvenuti, M. Di Benedetto, A. Ferrari, C. Pinello and
A. Sangiovanni-Vincentelli, Hybrid systems and the design of
embedded controllers for automotive engine management, in
Proceedings of the 37th IEEE Conference on Decision and Control,
TP01-7, pp. 26562661, 1998.
A. Balluchi, M.D. Di Benedetto, C. Pinello and A.L. Sangiovanni-
Vincentelli, Mixed models of computation in the design of
automotive engine control, in Proceedings 40th IEEE Conference
on Decision and Control, Orlando, Florida, pp. 26562661, 2001.
M. Baotic, M. Vasak, M. Morari and N. Peric, Hybrid systems
theory based optimal control of an electronic throttle,
in Proceedings of the American Control Conference, Denver,
Colorado, USA, Vol. 6, pp. 52095214, 2003.
S.H. Chan and D.L. Hoang, Modeling of catalytic conversion of
co/ch in gasoline exhaust at engine coldstart, SAE Technical
Paper 1999-01-0452, 1999.
D. Cho and J.K. Hedrick, Automotive powertrain modelling for con-
trol, Journal of Dynamic Systems, Measurements and Control, 111,
pp. 568576, 1989.
N. Giorgetti, A. Bemporad, I. Kolmanovsky and D. Hrovat, Explicit
hybrid optimal control of direct injection stratified charge engines,
Proceedings of the IEEE International Symposium on Industrial
Electronics, Dubrovnik, Croatia, June, pp. 247252, 2005.
A.T. Lee, M. Wilcutts, P. Tunestal and J.K. Hedrick, A method of
lean air-fuel ratio control using combustion pressure measurement,
Society of Automotive Engineers of Japan, Inc. and Elsevier Science,
Vol. 22, No. 4, pp. 389393, 2001.
I. Mitchell, A Toolbox of Level Set Methods. Version 1.1 beta,
Vancouver, Canada: University of Bristish Columbia, 2005
R. Mobus, M. Baotic and M. Morari, Multi-object adaptive
cruise control, Hybrid Systems: Computation and Control, 2623,
pp. 359374, 2003.
B. Shaw and J.K. Hedrick, Coldstart engine combustion modelling
to control hydrocarbon emissions, in Proceedings of 15th
Triennial World Congress of the International Federation of
Automatic Control, Barcelona, Spain, 2126 July, 2002.
B. Shaw and J.K. Hedrick, Closed-loop engine coldstart control to
reduce hydrocarbon emissions, in Proceedings of the American
Control Conference, San Diego, CA, USA, pp. 13921397, 2003.
B. Shaw, G.D. Fischer and J.K. Hedrick, A simplified coldstart
catalyst thermal model to reduce hydrocarbon emissions, in
Proceedings of 15th Triennial World Congress of the International
Reduction of hydrocarbon emissions 463
Federation of Automatic Control, Barcelona, Spain, 2126 July,
2002.
B. Song, A. Howell and K. Hedrick, Robust stabilization and
ultimate boundedness of dynamic surface control systems via
convex optimization, International Journal of Control, 75,
pp. 870881, 2002.
J. Souder and J.K. Hedrick, Adaptive sliding mode control of air-fuel
ratio in internal combustion engines, International Journal of
Robust and Nonlinear Control, 14, pp. 525541, 2004.
J. Sun and N. Sivashankar, Issues in cold start emission control
for automotive IC engines, in Proceedings of the American
Control Conference, Philadelphia, PA, 2126 June, Vol. 3,
pp. 13721376, 1998.
D. Swaroop, J.K. Hedrick, P.P. Yip and J.C. Gerdes, Dynamic
surface control for a class of nonlinear systems, IEEE
Transactions on Automatic Control, 45, pp. 18931899, 2000.
H. Tanaka, M. Uenishi and I. Tan, An intelligent
catalyst, SAE Technical Paper 2001-01-1301, 2001.
C. Tomlin, J. Lygeros and S. Sastry, Synthesizing controllers for
nonlinear hybrid systems, in First International Workshop,
HSCC98, Hybrid Systems: Computation and Control, Lecture
Notes in Computer Science, T. Henzinger and S. Sastry, Eds, 1386,
1998, pp. 360373.
P. Tunestal and J.K. Hedrick, Cylinder air/fuel ratio estimation
using net heat release data, 3rd IFAC Workshop on Advances
in Automotive Control, Karlsruhe, Germany, March 2001.
464 P. R. Sanketi et al.