IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 55, NO.

11, NOVEMBER 2008 3891

Estimation of Gasoline-Engine Parameters
Using Higher Order Sliding Mode
Qarab Raza Butt and Aamer Iqbal Bhatti, Member, IEEE
AbstractAutomotive-engine control and fault diagnostics
largely depend upon the accuracy of the nonlinear models used.
The structure of these nonlinear models is generally agreed
upon. However, the model parameters are mostly difcult to
obtain. This paper presents the development of second-order
sliding-mode technique with real twisting algorithm for estimation
of more than one parameter from a single dynamical equation
of the nonlinear model. The system under study is a mean
value engine model of a naturally breathing gasoline engine. The
parameters estimated are throttle bodys discharge coefcient,
load torque, and indicated torque as a function of inlet manifold
pressure. The estimated variables are used to compensate for the
unmodeled dynamics, modeling inaccuracies, and approximations
which arise from the assumptions made for the development of
mathematical model of a real-world system. The resulting model
is a better description of the actual engine dynamics and gives
good agreement to real engine data. The data are acquired from a
production model vehicle equipped with an electronic control unit
compliant to OBD-II standard. The observer designed is simple
enough for implementation, and estimated parameters can also be
used for engine-controller design and fault-diagnosis work.
Index TermsDischarge coefcient, indicated torque, load
torque, mean-value modeling, parameter estimation, sliding-mode
observer (SMO).
NOMENCLATURE
T
HE FOLLOWING variable names, acronyms, and no-
tations have been used in this paper. All units are in
SI (metric) system until and otherwise mentioned.
Variables and Constants
Throttle opening angle (in degrees).
A
E
Effective throttle area (in square meters).
C
1
Speed-density constant.
C
d
Throttle discharge coefcient.
C
d
Correction in throttle discharge coefcient.
J
e
Inertia of engine.
Engine speed (in radians per second).
Fuel split parameter.
Ratio of specic heat capacities.
Fraction of fuel injected before IVC.
m
ai
Air ow rate across throttle (in kilograms per second).
m
fi
Fuel ow rate at injector (in kilograms per second).
Manuscript received January 15, 2008; revised August 18, 2008. Current
version published October 31, 2008. This work was supported by the Higher
Education Commission (HEC) of Pakistan.
Q. R. Butt is with the Center for Advanced Studies in Engineering, Islamabad
G-5/1, Pakistan (e-mail: [email protected]).
A. I. Bhatti is with Mohammed Ali Jinnah University, Islamabad 46000,
Pakistan (e-mail: [email protected]).
Digital Object Identier 10.1109/TIE.2008.2005930
m
fo
Fuel ow rate into cylinder (in kilograms per second).
m
ff
Fuel-lm mass ow (in kilograms per second).
m
fv
Fuel-vapor mass ow (in kilograms per second).
m
ff2
Fuel ow injected after IVC (in kilograms per second).
m
ff3
Fuel ow injected before IVC (in kilograms per
second).
m
fsl
Fuel ow lagged by wall wetting (in kilograms per
second).
N Engine speed (in revolutions per minute).
P
a
Ambient air pressure (in kilopascals).
P
m
Intake manifold pressure (in kilopascals).
R Universal gas constant (in kilojoules per kilogram
Kelvin).
T
i
Indicated torque (in Newton meters).
T
b
Brake torque (in Newton meters).
T
p
Pumping torque (in Newton meters).
T
f
Frictional torque (in Newton meters).

f
Slow fuel time constant (in seconds).
T
m
Intake manifold temperature (in Kelvin).
T
a
Ambient air temperature (in Kelvin).

v
Volumetric efciency of engine.
V
d
Engine displacement (in cubic meters).
V
m
Intake manifold volume (in cubic meters).
C
r
Compression ratio.
C
v
Specic heat capacity at constant volume (in kilojoules
per kilogram Kelvin).
H
k
Percentage of heat converted into brake work (25%).
Acronyms
ANN Articial neural network.
BDC Bottom dead center.
EKF Extended Kalman lter.
EPA Environmental protection agency.
IVC Intake valve closing.
OBD-II Onboard diagnostics revision II.
PW Pulsewidth (in degrees).
SI Spark ignition.
SOI Start of injection.
TDC Top dead center.
I. INTRODUCTION
Most modern cars are equipped with onboard diagnostic
programs in their electronic control units (ECUs) to control
and monitor the engine operations. Due to the strict legisla-
tive regulations on environmentally harmful engine-exhaust
emissions set by EPAs and a growing end-user demand of
0278-0046/$25.00 2008 IEEE
3892 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 55, NO. 11, NOVEMBER 2008
engine-fuel economy, the airfuel ratio (AFR) of in-cylinder
mixture is required to be as closer to its stoichiometric propor-
tions as possible. The requirement of AFR improvement has
raised a need for investigation into improvement of mean-value
engine models (MVEMs) and their parameter estimation.
Some of the critical parameters governing the AFR control
and overall engine performance are discharge coefcient of
throttle body, indicated torque, and load torque. In the fol-
lowing lines, some recent work on these parameters is briey
summarized.
A. Throttle-Discharge Coefcient
The air mass ow across the throttle body of an automotive
engine is represented by the gas equation of isentropically
compressible air mass ow through an orice. Since a throt-
tle/buttery valve does not act like an orice, its representa-
tive mathematical model has modeling inaccuracies. Moreover,
while modeling a throttle valve, the air ow is assumed to be
unidirectional. Turbulent effects and backows are neglected.
It is also assumed that air path is smooth, ow is laminar,
and there are no abrupt bends and notches. In the real world,
these conditions hardly exist; therefore, the modeled equation is
unable to correctly predict the air mass ow across the throttle
valve. To accommodate these assumptions and inaccuracies,
a coefcient of discharge is introduced in gas equations. It is
represented by C
d
throughout this paper. This correction factor
is dened as a ratio of actual to the ideal ow rates of air mass
and can be written as
C
d
=
m
actual
m
ideal
(1)
where m
actual
is the actual mass ow rate determined experi-
mentally and m
ideal
is the ideal mass ow rate given by the ow
equation. When the engine is running, any one of the ow rates
in (1) would not be equal to zero; hence, C
d
would be well
dened within the operating range of engines. In steady-state
conditions (constant input (throttle opening angle) and constant
engine load), discharge coefcient is nearly a constant number.
But at varying input and engine load, it is neither a constant
parameter nor a variation linear [2], [6], [7], [22][24], [29].
The coefcient of discharge has been taken as a constant
parameter by most researchers in MVEMs available in pub-
lic literature (see, for example, [9], [13], [15], [18][20]). A
few researchers have taken discharge coefcient as a varying
parameter. It has been modeled as pressure-ratio-dependent
parameter and merged into the pressure-ratio function [22]. It
has been modeled as a parameter depending on upstream and
downstream pressures around throttle valve, area of throttle
valve, and temperature of air [23]. Its dependence on operating
point has been indicated in [29], and a fourth-order polynomial
in throttle angle has been suggested for mass-ow correction.
It has also been modeled as a function of throttle area and
ratio of pressures [24, pp. 327]. In [7], discharge coefcient
has been modeled as a function of engine speed and throttle-
plate opening angle. In [6], the coefcient of discharge has been
modeled as a function of pressure ratio and throttle opening
area. In [2], the throttle mass ow has been modeled as a
function of throttle opening angle and engine angular speed.
Discharge coefcient of throttle body has not been expressed
uniquely and/or generically by references cited earlier; there-
fore, it still needs further investigation. In this paper, a rst-
order sliding-mode observer (FOSMO) is used to estimate the
discharge coefcient as an engine signal.
B. Indicated Torque
Indicated-torque estimation is important for engine-control
and fault-diagnostic schemes. One approach of computing
indicated engine torque is based on engine geometry and
in-cylinder combustion pressure measurements with pressure
transducers inside combustion chamber, a variable which is not
easily measurable and requires expensive sensors [26]. This
is not a cost-effective solution for mass-production vehicles
on long-term basis. Moreover, an additional in-cylinder sensor
may raise certain mounting and maintenance problems and also
engine assembly and packaging issues. Therefore, estimation of
indicated torque from sensors available on all modern vehicles
(like velocity and manifold pressure sensors, etc.) is cheaper
and requires no additional hardware.
Different techniques for estimation of indicated torque are
available in public literature. In [26], nonlinear estimator
based on three-term backpropagation neural-network technique
(ANN) is used to estimate indicated torque. In [8], [16], and
[29], SMOs have been presented to estimate indicated torque.
In [30], an engine-speed-based estimation of indicated torque is
presented. Engine speed is estimated with an estimator based on
nite-automaton theory. In [12], indicated torque is estimated
by frequency-domain analysis of engine speed.
In this paper, indicated torque is estimated as a function of
manifold pressure using higher order sliding mode (HOSM).
A coefcient a
1
is estimated to represent indicated torque as a
function of inlet manifold pressure (P
m
).
C. Load Torque
In an engine laboratory facility or a test bed, the load torque
is created by controllable brakes and dynamometers and, hence,
measurable through brake or dynamometer data. In a produc-
tion vehicle, load torque is created at random by driver com-
mands and road conditions and, hence, not directly measurable.
Therefore, it is estimated or computed by different techniques
using sensor data of variables accessible for measurements. For
example, in [12], load torque is estimated by frequency-domain
analysis of engine speed. In this paper, load torque (L
T
) is
estimated using HOSM.
D. Sliding-Mode Technique in Parameter Estimation
The sliding-mode techniques have been used by many re-
searchers for estimation of immeasurable and/or uncertain
parameters and inaccessible system states for linear-and
nonlinear-systemmodels in a variety of applications like robots,
automotive processes and applications, ac and dc motors, etc.
(see, for example, [5], [12], and [21]). An SMO has a simple
design and implementation structure and is robust against mod-
eling errors and perturbations due to parametric variations. EKF
is another observer scheme for nonlinear systems, but it requires
BUTT AND BHATTI: ESTIMATION OF GASOLINE-ENGINE PARAMETERS USING HIGHER ORDER SLIDING MODE 3893
an online linearization at each operating point about which a
system is being observed; thus, increasing the computational
load signicantly. The main advantages of SMOs are simplicity
of design, convergence in nite time, and robustness against
varying and uncertain parameters.
This paper presents SMOs for online estimation of discharge
coefcient, indicated torque, and load torque and is organized
as follows.
In Section II, formulation of gasoline-engine model is de-
scribed in brief. Section III contains description of observer
scheme developed for estimation. Section IV presents re-
sults of simulations and model tuning with estimated parame-
ters. Concluding remarks are made in Section V. Constants
and coefcients represented by lengthy expressions are given
in the Appendix. References are given at the end of this
paper.
II. MODEL DESCRIPTION
A three-state nonlinear engine model (MVEM) based on
engine-physics principle and PV diagram is presented in this
section. It is assumed that the engine is a four-stroke four-
cylinder gasoline engine. It is also assumed that cylinders are
paired in two so that pistons of two cylinders move simultane-
ously around TDC and BDC, but only one cylinder is red at
a time, thus, all the four principle processes, namely, suction,
compression, power generation, and exhaust can be taken as
continuous over time in the engine. The uctuations during
power generation due to gradual decrease in pressure during
gas-expansion process in power strokes are averaged by mean
effective pressure (MEP) which is computed using Otto cycle.
The choked-ow conditions across the buttery valve and
uctuations of manifold pressures due to periodic phenomena
have been neglected. It is also assumed that the temperature
of the inlet manifold remains unchanged for small intervals
of time; therefore, temperature dependence in the manifold
is ignored. Equation representing the rotational dynamics has
been developed using Otto cycle for modeling of combustion
process, computation of maximum cycle pressure, temperature,
MEP, and indicated torque (T
i
). Frictional-torque equation is
modeled using a little variation of an empirical relationship
from [19] and [20], given as T
f
= (1/2)V
d
{97000 + 15N +
5N
2
10
3
}. Pumping torque is taken as T
p
= Tb/. The equa-
tion of rotational dynamics is derived from =
_
(T
b
/J
e
)dt
using equation of brake torque as T
b
= T
i
T
p
T
f
L
T
and
given in a simplied form. Thus, a nonlinear generic model
can be represented as a set of dynamical equations given as
follows:
m
fo
= m
ff2
+ m
ff3
+ m
fsl
(2)

P
m
= C
1

v
P
m
+
RT
m
V
m
m
ai
(3)
=a
1
P
m
a
2
a
3

2
L
T
(4)
where the coefcients and parameters in these equations are
given in the Nomenclature and Appendix.
III. OBSERVER DESIGN
In [3], a second-order SMO (SOSMO) based on Real Twist-
ing Algorithm is used to estimate more than one parameter in
one equation of an LTI system. The same concept is extended
for estimation of more than one parameter in one equation
of a nonlinear system model (i.e., MVEM). Real Twisting
Algorithm has a simpler structure among twisting algorithms
and does not require time derivative of a sliding function.
A FOSMO is used to estimate one parameter in manifold
dynamics equation, and a SOSMO is used to estimate two
parameters: indicated torque coefcient (a
1
) and load torque
(L
T
) in rotational dynamics equation. Model equations are
written in the following form for this purpose:

P
m
= C
1

v
P
m
+ A
k
(C
d
C
d
)f(P
m
)u() (5)
=a
1
P
m
a
2
a
3

2
L
T
. (6)
Differentiation of (6) and putting (5) and (6) into the resulting
equation yields the following:
= a
1
C
1

v
P
m
+ a
1
A
k
(C
d
C
d
)f(P
m
)u()
a
1
a
2
P
m
+ a
2
2
+ 3a
2
a
3

2
+ a
2
L
T
2a
1
a
3
P
m

+ 2a
2
3

3
+ 2a
3
L
T
. (7)
Dening a third-order system description by new state vari-
ables as given as follows:
q
1
=P
m
q
2
=
q
3
= .
The dynamic system given by (5)(7) can be written as
q
1
= C
1

v
q
1
q
2
+ A
k
(C
d
C
d
)f(q
1
)u() (8)
q
2
=a
1
q
1
a
2
q
2
a
3
q
2
2
L
T
= q
3
(9)
q
3
= a
1
C
1

v
q
1
q
2
+ a
1
A
k
(C
d
C
d
)f(q
1
)u()
a
1
a
2
q
1
+ a
2
2
q
2
+ 3a
2
a
3
q
2
2
+ a
2
L
T
2a
1
a
3
q
1
q
2
+ 2a
2
3
q
3
2
+ 2a
3
L
T
q
2
. (10)
For system dened earlier, a SMO is dened by the
following:

1
= C
1

1
q
2
+A
k
(C
d
)f(q
1
)u()+(
1
q
1
) (11)

2
=
3
(12)

3
=(
2
q
2
). (13)
The difference between measured and estimated variables
is the sliding surface in an SMO. Here, the FOSM-based
discontinuous function is (
1
q
1
) = k
s
sign(s
1
); k
s
is the
gain of FOSMO, s
1
=
1
q
1
is sliding surface for FOSM,
3894 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 55, NO. 11, NOVEMBER 2008
and (
2
q
2
) is a second-order sliding-mode function deter-
mined by real twisting algorithm such that
(
2
q
2
) =
_

M
sign(s
2
), if s
2
s > 0

m
sign(s
2
), if s
2
s 0
(14)
where
s =
_
0, k = 0
s
2
kn
s
2
(k1)n
, k 1
(15)
and n is sampling period assumed to be the same as that of
engine signals, k is the sample number, s
2
=
2
q
2
is the
sliding surface for SOSMO, and
M
,
m
satisfy the condition:

M
>
m
.
The observation error dynamics e
i
=
i
q
i
are given by
e
1
=(e
1
) C
1

v
(
1
q
1
)q
2
+ A
k
C
d
f(q
1
)u() (16)
e
2
=
3
q
3
(17)
e
3
=(e
2
) {a
1
C
1

v
q
1
q
2
+ a
1
A
k
C
d
f(q
1
)u()
a
1
A
k
C
d
f(q
1
)u()}
a
1
a
2
q
1
+ a
2
2
q
2
+ 3a
2
a
3
q
2
2
+ a
2
L
T
2a
1
a
3
q
1
q
2
+ 2a
2
3
q
3
2
+ 2a
3
L
T
q
2
. (18)
A. Stability Analysis
Analysis of error dynamics of FOSMO is investigated using
Lyapunov conditions. Referring to [17, Art. 3.4.2, Th. 3.2
and 3.3] for local and global stability in Lyapunovs direct
method, let V (e
1
) = (1/2)e
2
1
be the candidate Lyapunov func-
tion for error dynamics.
According to these theorems, if V (e
1
) has continuous
rst-order derivatives such that we have the following three
conditions, then the equilibrium point at the origin is globally
asymptotically stable.
1) V (e
1
) is positive denite.
2)

V (e
1
) is negative denite.
3) V (e
1
) as e
1
.
It implies that derivative of error function V (e
1
) must
be negative denite for convergence. The conditions for
convergence are analyzed as

V (e
1
)=e
1
e
1

V(e
1
)=e
1
{C
1

v
q
2
e
1
+A
k
C
d
u()f(q
1
)k
s
sgn(e
1
)} .
Assuming that b is a positive number such that
[A
k
C
d
u()f(q
1
)]
min
b [A
k
C
d
u()f(q
1
)]
max
where A
k
C
d
u()f(q
1
) assumes its maximum value at wide-
open-throttle conditions, i.e., 90

, and engine is running at

no load (L
T
= 0) and minimum value, when engine is idling
( at its minimum value which is 8.8

in experimental data and

may vary around this value from one vehicle to the other), and
rearranging the earlier equation, the following can be written:

V C
1

v
q
2
e
2
1
+ e
1
(b k
s
sgn(e
1
)) .
From the earlier equation, it can be seen that

V is negative
denite in one condition only, which is when b k
s
. In other
words, the global asymptotic stability of error dynamics is
guaranteed when gain of SMO is greater than or equal to b;
otherwise, error dynamics may or may not be stable. The
value of SMO gain chosen for simulations is obtained by using
engine experimental data in earlier equations (for stability and
convergence analysis of HOSM, the reader is referred to [1],
[27], and [28]).
A rst-order sliding motion would occur at e
1
= 0 and a
second-order sliding motion at e
2
= e
2
= 0 (see [1] and [28,
Ch. 3 and 4]). It should be noted that e
1
and e
2
are independent
sliding surfaces, and using appropriate gains, convergence of
both is ensured.
The following three algebraic relationships are obtained from
(16)(18) as a result:
= A
k
C
d
f(q
1
)u() (19)

3
=q
3
= a
1
q
1
a
2
q
2
a
3
q
2
2
L
T
(20)
= a
1
C
1

v
q
1
q
2
+ a
1
A
k
(C
d
C
d
)f(q
1
)u()
a
1
a
2
q
1
+ a
2
2
q
2
+ 3a
2
a
3
q
2
2
+ a
2
L
T
2a
1
a
3
q
1
q
2
+ 2a
2
3
q
3
2
+ 2a
3
L
T
q
2
. (21)
From (19), the discharge-coefcient correction is computed
with the help of the following:
C
d
=

A
k
f(q
1
)u()
. (22)
From (20) and (21), we dene
3
and as follows:

3
=
3
+ a
2
q
2
+ a
3
q
2
2
(23)
=
_
a
2
2
q
2
+ 3a
2
a
3
q
2
2
+ 2a
2
3
q
3
2
_
(24)
and get the following:

3
=a
1
q
1
L
T
(25)
=(G
1
+ G
2
)a
1
+ G
3
L
T
(26)
where G
1
, G
2
, and G
3
have the following:
G
1
=A
k
(C
d
C
d
)f(q
1
)u() (27)
G
2
= (C
1

v
q
1
q
2
+ a
2
q
1
+ 2a
3
q
1
q
2
) (28)
G
3
=(a
2
+ 2a
3
q
2
). (29)
From (22), the correction in discharge coefcient C
d
is
computed, and corrected value is found in relation to the as-
sumed nominal value of C
d
while developing the model, which
is unity. The modeling errors in throttle mass-ow equation are
compensated in this way (see gures showing the comparison
of engine data and tuned model for manifold pressure in later
sections.
The system of equations (25) and (26) for higher order
sliding mode can be written as
_

_
=
_
q
1
1
(G
1
+ G
2
) G
3
_ _
a
1
L
T
_
. (30)
BUTT AND BHATTI: ESTIMATION OF GASOLINE-ENGINE PARAMETERS USING HIGHER ORDER SLIDING MODE 3895
Fig. 1. Block diagram representing the structure of observer, model, compensated model, inputs and outputs of observer and both models, and their
interconnections. The interconnection of Sliding Mode Observer and Tuned Engine Model shows the online feedback of estimated parameters for model
tuning. The self-feedback of SMO comes from implementation of higher order sliding function using real twisting algorithm from (14) and (15).
It can be shown that the 2 2 matrix in (30) is nonsingular
for the operating range of the engine. The system of equations
represented by (30) is nonlinear. However, while simulating a
nonlinear system on a digital computer, the system of equations
is solved at each sampling point/time step. Thus, at each
sampling instance, the 2 2 matrix in (30) assumes some
constant values, and the system of equations can be taken as
a linear system at that instance. Therefore, instantaneous values
of parameters can be found by using the relationship as follows:
_
a
1
L
T
_
=
_
q
1
1
(G
1
+ G
2
) G
3
_
1
_

_
. (31)
The sequences of instantaneous values obtained for both
parameters (a
1
and L
T
) using (31) would represent the sampled
versions of their continuous-time nonlinear representations. As
we make the sampling interval smaller, the resulting discrete
sequence of each parameter comes closer to its continuous-time
nonlinear representations. The sample time has been selected
as 0.1, 0.01, 0.001, and 0.00001 s in four set of simulation and
trial experiments. Results presented in this paper come from
the experiment with sampling time of 0.01 s. The parameter
a
1
is used to nd out indicated torque, and parameter L
T
represents load torque. The next section conrms the validity
of the parameter-estimation scheme through actual trials on a
production vehicle.
IV. TRIALS AND RESULTS
The estimation method formulated in previous section has
been implemented and tested for computation of parameters.
Computer simulations of engine model formulated in Section II
has been implemented and parameterized with values which
correspond to a 1.3-L production vehicle engine. The vehicle is
equipped with an ECU which is compliant to OBD-II. The data
are logged through OBD-II data scanner and logging software.
The observer is parameterized with signals obtained to
make estimates of parameters a
1
, L
T
, and C
d
. The estimated
values of these unknown parameters are used in physical/
rst-principal-based MVEM to compensate for the modeling
inaccuracies and tune the model to real engine signals (Fig. 1).
In this experiment, engine data are obtained while vehicle is
in neutral gear and engine input (throttle angle) is varied (with
accelerator pedal). The contributions in engine load are given
as follows:
1) engine rotating/reciprocating masses, e.g., pistons, con-
necting rods, crankshaft, ywheel and power-train com-
ponents engaged, etc.;
2) air-conditioning system switched on at t = 250 s;
3) other loads are windows motorized glasses, wiper mo-
tors, and other miscellaneous minor loads which may
occur during a vehicle operation.
The throttle opening angle of vehicle while idling is equiv-
alent to 8.8

. Throttle opening is kept constant at steady state

for a while for six distinct intervals at four values which
are visible from the plot of throttle opening angle shown in
Fig. 2. The engine data used as driving data for the SMO;
throttle angle variations, manifold pressure, and angular speed
in this experiment are shown in Fig. 2. The uctuations in
manifold pressure which start at about t = 300 s are attributed
to switching the air conditioner on and off.
The sliding-surface functions of both rst and second orders
are shown in Figs. 3 and 4. The engine speed is observed
through HOSM, and the corresponding sliding function is
shown in Fig. 3. The observer reaches sliding phase in 25 s
and recovers with every transient in a similar time interval. The
maximum error in angular speed is around 20%. This error
can be further reduced by increasing the HOSM gains but at
the cost of increased chattering in the injection signal. The
SMO gains are set at 150 and 25, respectively. The gains are
tuned to be barely high enough to make sliding occur. The
sliding function of the FOSMO consists of observation error in
manifold pressure shown in Fig. 4.
The observer reaches sliding phase in less than 15 s while the
maximum error is around 10% despite very heavy transients.
The FOSMO gain is chosen to be 25 to drive the observer into
sliding phase.
Once the observers are successfully running in the sliding
phase (despite transients, they recover their sliding states), the
3896 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 55, NO. 11, NOVEMBER 2008
Fig. 2. Engine input and output signals, throttle angle, manifold pressure, and
angular speed.
Fig. 3. Engine and observer angular speeds with estimation error.
Fig. 4. Engine and observer manifold pressures with estimation error.
Fig. 5. Indicated torque parameter computed by second-order sliding-mode
scheme.
Fig. 6. Estimated engine indicated torque, load torque, and discharge
coefcient.
computed injection signals are used for the required parameter
estimation. The indicated torque is computed by rst estimating
coefcient a
1
and then multiplying it with manifold pressure.
The switching of the injection signal causes values of a
1
to
become negative. The negative values are discarded as they
do not have any physical justication. The estimation of a
1
is
shown in Fig. 5. The average value of the estimation is a good
t against its theoretical value of 2.6.
Fig. 6 shows indicated torque, load torque, and discharge
coefcient. It is hard to determine whether these parameter
estimates reach their true values or not. However, a good acid
test would be to inject these estimates back into the original
nonlinear model and look at the delity of the tuned model
responses to the actual engine data.
BUTT AND BHATTI: ESTIMATION OF GASOLINE-ENGINE PARAMETERS USING HIGHER ORDER SLIDING MODE 3897
Fig. 7. Untuned model, tuned model, and engine angular speeds.
Fig. 8. Untuned (model with theoretical values) and tuned model (model with
computed estimates) and engine data plot of manifold pressure.
This test will be reported later. At the moment, it will
sufce to say that load torque and indicated torque are in good
agreement with each other and low throttle values.
The real test of these parameter estimates is conducted
by tuning the engine nonlinear model with the computed
estimates. Untuned (model with theoretical values) and tuned
model (model with computed estimates) outputs along with
experimental data are compared as shown in Figs. 7 and 8.
Signicant improvement can be seen in manifold pressure
with the use of the estimated parameters. On the average, the
modeling error of 3035 kPa is reduced to 57 kPa, which
is a considerable improvement. Similar is the case for engine
speed. For angular-rotation case, the modeling error has been
reduced by magnitude. The errors can be explained on the
grounds of structural uncertainties like neglecting of air bypass
valve and inaccuracies in other parameters like volumetric
efciency and friction torque. Another important factor is the
nonlinear coupling terms between the manifold-pressure and
the rotational dynamics equations. It is observed that the two
equations do better prediction on their own.
V. CONCLUSION
A parameter-estimation scheme for a nonlinear automotive-
engine model based on SOSMO has been presented, and its
validity has been demonstrated on the real engine data taken
from a production vehicle. The approach is generic and can
be used for any engine make, thus providing a very effective
tool for simultaneous estimation of critical engine parameters
with easy-to-tune estimation algorithm. A novel feature of the
scheme is that two engine parameters are estimated from a
single nonlinear equation of rotational dynamics. Apart from
the model structure, no other a priori information is assumed.
This approach can be effectively employed for robust control
and condition monitoring of automotive systems.
APPENDIX
The coefcients and parameters in (2)(7) are given by
expressions as follows:
m
ff2
= m
fi
(1 ) (32)
m
ff3
= m
fi
(33)
d
dt
( m
fsl
) =
m
fi
(1 ) m
fsl

f
(34)
=
_
1, if
PW

IV C

SOI

IV C

SOI

PW
, if
PW
>
IV C

SOI
(35)
m
ai
=u()A
E
P
a
C
d

2
( 1)RT
a
f(P
m
) (36)
where
f(p
m
)=

_
_
_
P
m
P
a
_2

_
P
m
P
a
_
+1

_
(37)
u()=(1 cos ) (38)
a
0
=

Je( + 1)
(39)
a
1
=
a
0
V
d
4

C
2
r
_
C
1
r
1
_
(H
k
Q)
_
1
1
C
1
r
_
( 1)(C
r
1)C
v
T
m
(AFR)

10
3
(40)
a
2
=
a
o
V
d
4
(41)
a
3
=
a
o
V
d
.(0.05)
18 10
4
(42)
A
k
=
RT
m
V
m
P
a
A
E

2
( 1)RT
a
. (43)
ACKNOWLEDGMENT
The authors would like to thank the reviewers for sparing
their precious time and giving comments and suggestions for
the improvement of this paper.
REFERENCES
[1] A. Levant, Principles of 2-sliding mode design, Automatica, vol. 43,
no. 4, pp. 576586, Apr. 2007.
[2] R. Scattolini, A. Miotti, G. Lorini, P. Bolzern, P. Colaneri, and
N. Schiavoni, Modeling, simulation and control of an automotive gaso-
line engine, in Proc. IEEE Int. Conf. Control Appl., Munich, Germany,
Oct. 46, 2006, pp. 27482753.
3898 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 55, NO. 11, NOVEMBER 2008
[3] T. Floquet, J. A. Twiddle, and S. K. Spurgeon, Parameter estimation via
second order sliding modes with application to thermal modelling in a
high speed rotating machine, in Proc. IEEE Int. Conf. Ind. Technol.,
Mumbai, India, Dec. 1517, 2006, pp. 26352639.
[4] H. Shraim, B. Ananou, L. Fridman, H. Noura, and M. Ouladsine, Sliding
mode observers for the estimation of vehicle parameters, forces and states
of the center of gravity, in Proc. 45th IEEE Conf. Decision Control,
San Diego, CA, Dec. 1315, 2006, pp. 16351640.
[5] Z. Xu and M. F. Rahman, An adaptive sliding stator ux observer for a
direct-torque-controlled IPM synchronous motor drive, IEEE Trans. Ind.
Electron., vol. 54, no. 5, pp. 23982406, Oct. 2007.
[6] P. Anderson, Air charge estimation in turbocharged spark ignition
engines, Ph.D. dissertation, Dept. Elect. Eng., Linkping Univ.,
Linkping, Sweden, 2005.
[7] J. A. F. Vinsonneau, D. N. Shields, P. J. King, and K. J. Burnham,
Improved SI engine modelling techniques with application to fault detec-
tion, in Proc. IEEE Int. Conf. Control Appl., Glasgow, U.K., Sep. 1820,
2002, pp. 719724.
[8] I. Haskara and L. Mianzo, Real-time cylinder pressure and indicated
torque estimation via second order sliding modes, in Proc. Amer. Control
Conf., Arlington, VA, Jun. 2527, 2001, pp. 33243328.
[9] E. Hendricks and J. B. Luther, Model and observer based control
of internal combustion engines, in Proc. MECA, Salerno, Italy, 2001,
pp. 920.
[10] A. Chevalier, M. Muller, and E. Hendricks, On the validity of mean value
engine models during transient operation, presented at the SAE World
Congr., Detroit, MI, 2000, SAE Tech. Paper 2000-01-1261.
[11] J. K. Pieper and R. Mehrotra, Air/fuel ratio control using sliding mode
methods, in Proc. Amer. Control Conf., San Diego, CA, Jun. 1999,
pp. 10271031.
[12] F. Bengtsson, Estimation of indicated and load torque from engine speed
variations, M.S. thesis, Dept. Elect. Eng., Linkping Univ., Linkping,
Sweden, 2006. Vehicular Systems.
[13] E. Hendricks, Engine modelling for control applications: A critical
survey, Meccanica, vol. 32, no. 5, pp. 387396, Oct. 1997. (DTU).
[14] Y. Shiao and J. J. Moskwa, Model-based cylinder-by-cylinder airfuel
ratio control for SI engines using sliding observers, in Proc. IEEE Int.
Conf. Control Appl., Dearborn, MI, Sep. 1518, 1996, pp. 347354.
[15] R. W. Weeks and J. J. Moskwa, Automotive engine modeling for real-
time control using MatLab/Simulink, presented at the International
Congr. & Expo., Detroit, MI, 1995, SAE Tech. Paper 950417.
[16] G. Rizzoni, S. Drakunov, and Y.-Y. Wang, On-line estimation of indi-
cated torque in ic engines via sliding mode observers, in Proc. Amer.
Control Conf., Seattle, WA, Jun. 1995, pp. 21232127.
[17] J.-J. E. Slotine and L. Weiping, Applied Nonlinear Control. Englewood
Cliffs, NJ: Prentice-Hall, 1991.
[18] E. Hendricks and S. C. Sorenson, Mean value modeling of spark ignition
engines, presented at the SAE Technical Paper Series, Detroit, MI, 1990,
Paper 900616.
[19] P. Andersson, L. Eriksson, and L. Nielsen, Modeling and architecture
examples of model based engine control, in Proc. 2nd Conf. Comput.
Sci. Syst. Eng., Linkping, Sweden, 1999, pp. 19.
[20] J. Heywood, Internal Combustion Engine Fundamentals. New York:
McGraw-Hill, 1988.
[21] A. Stotsky and I. Kolmanovsky, Simple unknown input estimation
techniques for automotive applications, in Proc. Amer. Control Conf.,
Arlington, VA, 2001, pp. 33123317.
[22] E. Hendricks, A. Chevalier, M. Jensen, and S. C. Sorensen, Modeling
of the intake manifold lling dynamics, presented at the International
Congr. & Expo., Detroit, MI, 1996, SAE Tech. Paper 960037.
[23] I. Arsie, F. De Franceschi, C. Pianese, and G. Rizzo, ODECSA Com-
puter Code for the Optimal Design of S.I Engine Control Strategy, 1996.
[24] G. P. Blair, Design and Simulation of Four Stroke Engines. Warrendale,
PA: SAE, 1999.
[25] D. Khiar, J. Lauber, T. Floquet, and T. M. Guerra, An observer design
for the instantaneous torque estimation of an IC engine, presented at the
IEEE-VTS Vehicle Power and Propulsion (VPP) Conf., Illinois Inst. of
Technol., Chicago, IL, Sep. 79, 2005.
[26] Y. H. Zweiri and L. D. Seneviratne, Diesel engine indicated torque esti-
mation based on articial neural networks, in Proc. IEEE/ACS AICCSA,
Amman, Jordan, 2007, pp. 791798.
[27] A. Levant, Sliding order and sliding accuracy in sliding mode control,
Int. J. Control, vol. 58, no. 6, pp. 12471263, 1993.
[28] L. Fridman and A. Levant, Higher Order Sliding Modes in Sliding Mode
Control in Engineering. New York: Marcel Dekker, 2002.
[29] M. Scherer, C. Arndt, and O. Loffeld, Inuence of manifold pressure
pulsations to mean value models in air fuel ratio control, in Proc. 5th
IEEE Mediterranean Conf. Control Syst., Paphos, Cyprus, Jul. 2123,
1997.
[30] M. Hellstrom, Engine speed based estimation of the indicated engine
torque, M.S. thesis, Dept. Elect. Eng., Linkping Univ., Linkping,
Sweden, 2005.
Qarab Raza Butt received the B.S. degree in me-
chanical engineering from the University College
of Engineering, Taxila, Pakistan, in 1989, the Post-
graduate Diploma in computer system software and
hardware from the Computer Center, Islamabad,
Pakistan, in 1990, and the M.S. degree in control
engineering from the Center for Advanced Studies
in Engineering, Islamabad, in 2004, where he is
currently working toward the Ph.D. degree.
Since 1990, he has been working in industry as
an Installation, Fabrication, and Design Engineer.
He is the rst author and coauthor of six conference papers. His research
interests include mathematical modeling of dynamic systems for control and
fault diagnostics.
Aamer Iqbal Bhatti (M05) received the B.S. de-
gree in electrical engineering from the University
of Engineering and Technology, Lahore, Lahore,
Pakistan, in 1993, the M.S. degree in control sys-
tems from Imperial College of Science, Technol-
ogy and Medicine, London, U.K., in 1994, and the
Ph.D. degree in control engineering with a research
on idle-speed control of the Ford Mondeo Engine
and also the application of nonlinear control to
high-temperature furnaces from the University of
Leicester, Leicester, U.K., in 1998.
He continued his stay at Leicester University while doing postdoctoral
research on fault diagnostics and control of high-powered diesel engines funded
by Caterpillar. In 1999, he returned to Pakistan and started working for ERDC,
a consultancy rm providing services in the eld of aerospace controls where
he worked on nonlinear simulations of air vehicles, system identication, con-
troller design for aerospace vehicles, and data-acquisition experiments design.
In 2001, he was with Communications Enabling Technologies, Islamabad,
Pakistan, where he worked on enhancements of line echo cancellers used in
VoIP. Later on, he cofounded the Center for Advanced Studies in Engineering,
an engineering education institution, and CARE, an R&D company. At CARE,
he led a team that designed a radar signal processor and an ELINT system.
Since 2007, he has been with Mohammed Ali Jinnah University, Islamabad,
where he is a Professor of DSP and control systems with the Department
of Electronic Engineering. He is the rst author or coauthor of more than
35 refereed international papers, including four journal publications. His re-
search interests include sliding-mode applications and radar signal processing.