5th IFAC Conference on

5th IFAC Conference on Available

Engine
5th IFACand Powertrain
Conference onControl, Simulation and online at www.sciencedirect.com
Modeling
Engine
5th IFACand Powertrain
Conference
Changchun, China, onControl,20-22,
September Simulation
2018 and Modeling
Engine and Powertrain
Changchun, Control,20-22,
China, September Simulation
2018 and Modeling
Engine and Powertrain
Changchun, Control,20-22,
China, September Simulation
2018 and Modeling
Changchun, China, September 20-22, 2018 ScienceDirect
IFAC PapersOnLine 51-31 (2018) 700–705

Recursive
Recursive Engine
Engine In-Cylinder
In-Cylinder Pressure
Pressure Estimation
Estimation Using
Using Kalman
Kalman Filter
Filter and
and
Recursive Engine In-Cylinder Pressure
Structural
Recursive Engine In-Cylinder Estimation
Vibration
Pressure Signal
Estimation Using Kalman Filter and
Structural Vibration SignalUsing Kalman Filter and
Structural
Structural Vibration
Vibration Signal
Signal
Runzhe Han*, Christian Bohn*, Georg Bauer*
Runzhe Han*, Christian Bohn*, Georg Bauer*
*Institute of Electrical Runzhe
InformationHan*, ChristianClausthal
Technology, Bohn*, Georg Bauer*
University of
*Institute of Electrical Runzhe
InformationHan*, ChristianClausthal
Technology, Bohn*, Georg Bauer*
University of Technology,
Technology, Germany,
Germany,
(e-mail:
*Institute
(e-mail: [email protected],
of [email protected], [email protected],
Electrical Information Technology, Clausthal University
[email protected], [email protected]).
of Technology, Germany,
[email protected]).
*Institute of Electrical Information Technology, Clausthal University of Technology, Germany,
(e-mail: [email protected], [email protected], [email protected]).
(e-mail: [email protected], [email protected], [email protected]).
Abstract: The
Abstract: The cylinder
cylinder pressure
pressure signal
signal isis aa very
very useful
useful indicator
indicator for for modern
modern high-performance
high-performance internal-
internal-
combustion
Abstract:
combustion (IC)
The
(IC) engines.
cylinder
engines. Unfortunately,
pressure direct
signal is direct
Unfortunately, measurements
a very measurements of
useful indicatoroffor the
the cylinder
modern
cylinder pressure are
high-performance
pressure are impractical
internal-
impractical
The cylinder
Abstract:installing
because pressure
cylinder signal isis adifficult
transducers very useful
and indicator
conditions forinmodern high-performance
IC engine cylinders internal-
combustion (IC) engines.
because installing cylinderUnfortunately,
transducers is direct measurements
difficult and conditions of the in cylinder
IC enginepressure areare
cylinders
adverse.
impractical
are adverse.
combustion
Therefore, (IC) engines.
differentcylinder Unfortunately,
methodstransducers direct
have been investigated measurements
to reconstructof the
the cylinder
cylinder pressure
pressure are
from impractical
externally
because
Therefore,installing
different methods have been is difficult
investigated and
to conditions
reconstruct in
the IC engine
cylinder cylinders
pressure fromare adverse.
externally
because
measured installing such
cylinder
as thetransducers is difficult and and
conditions in IC engine cylinders are adverse.
Therefore,signals,
measured different
signals, methods
such
engine
as the have
engine
structure
been
structure
vibration
investigated the crank
to reconstruct
vibration and the crank theangular
cylinderspeed.
angular
In this
pressure
speed. from
In this
paper, some
externally
paper, some
Therefore,
of different
existingsignals,
methods methods have
areasreviewed been investigated
firstly. Thenvibrationto reconstruct
a noveland recursive the cylinder
cylinder pressure
pressure from
estimation externally
method
measured
of existing methods suchare the engine
reviewed structure
firstly. Then a novel the
recursivecrank angular
cylinder speed.
pressure In this paper,
estimation some
method
measured
based signals,
on using thesuch as thefilter
Kalman enginewithstructure
theThen vibration
engine and the
structure crank angular
vibration speed.
inputInisthis
signalpressure
as paper, some
proposed. Two
of existing
based methods
on using are reviewed
the Kalman firstly.
filter with theThen
enginea novel recursive
structure vibrationcylinder as inputestimation
signalpressure is proposed. method
Two
of existing
combustion methods
metrics, are reviewed
pressure peak firstly.
(P ) and a
peak novel
locationrecursive
(P ), cylinder
were used for estimation
evaluating the method
proposed
based on using
combustion the Kalman
metrics, pressure filter
peak with
(P maxthe engine structure vibration
) and peak location (P loc
), were signal
used as
for input is
evaluatingproposed.
the Two
proposed
based
method.on using
The the Kalman
estimation filter
results werewithmaxthe engine structure vibration
max
compared with the
loc
loc signal as input is proposed. Two
combustion
method.
combustionThemetrics, pressure
estimation
metrics, resultspeak
pressure were(P
peak max) and peak
(Pcompared
max) and peak the experimental
withlocation (Ploc), were
experimental
location (Ploc), were
data
data collected
used
collected
used
from
from aa four-cylinder
for evaluating
for evaluating
the proposed
four-cylinder
the proposed
diesel
method.
diesel engine
engine and
and showed
The estimation
showed the
results
the effectiveness
were compared
effectiveness of
of the
the proposed
with method.
the experimental
proposed method. data collected from a four-cylinder
method. The estimation results were compared with the experimental data collected from a four-cylinder
diesel
© 2018,engine
Keywords:IFAC and
Kalman showed
filters,the
(International effectiveness
Federation
control of thesystem
of Automatic
applications, proposed
Control) method.
Hosting byinternal
identification, Elseviercombustion
Ltd. All rights reserved.
engines.
diesel engine
Keywords: and
Kalman showed
filters,the effectiveness
control of thesystem
applications, proposed method.
identification, internal combustion engines.
Keywords: Kalman filters, control applications, system identification, internal combustion engines.
Keywords: Kalman filters, control applications, system identification, internal combustion engines.

1. INTRODUCTION AND MOTIVATION function (RBF) networks (Du et al., 2001).

1. INTRODUCTION AND MOTIVATION function (RBF) networks (Du et al., 2001).
The 1.
measurement INTRODUCTION
and analysis AND
of the MOTIVATION
cylinder pressure signal function
1.2. (RBF) networks (Du et al., 2001).
1. INTRODUCTION
The measurement and analysis of AND the MOTIVATION
cylinder pressure signal function 1.2. Crank
Crank (RBF)Angular Speed-Based
networks
Angular (Du et al.,Reconstruction
Speed-Based 2001).
Reconstruction
can measurement
The
can provide crucial
provide crucial information
and analysis
information for
of thefor engine
cylinder
engine combustion
pressure signal The
combustion 1.2. fluctuating
Crank Angular Speed-Based
waveform of engine Reconstruction
speed versus crank angle
The measurement
control, harmful and analysis
emissions of theand
control cylinder
fault pressure signal
diagnosis in IC also1.2. Crank Angular
The fluctuating Speed-Based
waveform of engine Reconstruction
speed versus crank angle
can provide
control, harmful crucial
emissionsinformation
control for fault
and engine combustion
diagnosis in IC The contains
fluctuating information
waveform of about
engine the
speedcylinder-by-cylinder
versus crank angle
can provide crucial information for engine combustion also contains
The fluctuating information
waveform about the
of engine2006).speedcylinder-by-cylinder
versus
engines. harmful
control,
engines.
control,
Unfortunately,
emissions
Unfortunately,
harmful emissions
forcontrol
for
direct and
direct
controlinand
measurements
fault diagnosis
measurements
fault diagnosis
the inhigh
the IC combustion
high
in IC alsoalso contains pressure
information(Johnsson,
about the How crank
cylinder-by-cylinder
angle
the speed
temperature
engines. and the
Unfortunately, high for pressure
direct cylinders
measurements causethe cylinder
high combustion
contains pressure
information(Johnsson,
about 2006).
the How the speed
cylinder-by-cylinder
temperature
engines. and the high for
Unfortunately,
pressure transducers to have
pressure
adirect
in measurements
limited
cylinders causethe
lifetime,cause
whichcylinder high fluctuation
cylinder
makes
combustionvaries
fluctuation
combustion varies
with the
pressure
with the
pressure
cylinder pressure
(Johnsson, 2006). changes
cylinder pressure
(Johnsson, 2006).
How thehas
changes
How thehas
been
speed
been
speed
temperature
pressure and the high
transducers to pressure
have a in cylinders
limited lifetime, which makes explored
fluctuation by many
varies researchers.
with the cylinder Moro et
pressure al. (2002)
changes used
has the
been
temperature
the method and the
expensive.high pressure
Additionally,in cylinders
sometimes cause cylinder
it ismakesalso FRFexplored
fluctuation by many
varies researchers.
with the cylinder Moro et
pressure al. (2002)
changes used
has the
been
pressure
the method transducers
expensive. to have a limited
Additionally, lifetime,
sometimes whichit ismakesalso FRF between
explored by many the cylinder
researchers. pressure
Moro andand
et al.the crank
(2002) angular
used the
pressure
difficult transducers
to find a to have
suitable a limited
place to lifetime,
mount which
cylinder pressure between
explored by the cylinder
many researchers. pressure
Moro et the(2002)
al. crank used
angular
the
the method
difficult to findexpensive.
a suitable Additionally,
place to mount sometimes
cylinderit it pressure
is also speed converted
FRF between by
the aa time
cylinder domain
pressure modelandand theapplied an
crank angularFRF
the method
transducers. expensive.
As aaa suitable Additionally,
consequence, sometimes is also
the FRFspeed converted
between by time
the cylinder domain model
pressure pressure and applied
and the crank an FRF
angular
difficult to find place tothemountreconstruction
cylinder pressureof the mapping to
transducers.
difficult
cylinder
As
topressure
find consequence,
a suitable
from place tothe
indirect
reconstruction
mount cylinderhas
measurements
of
pressure
been
speed converted
mapping
speed to improve
converted
by a time
improve
by a
the
time
cylinder
domain
the domain model
cylindermodel and
reconstruction
and applied
pressure an FRF
reconstruction
applied an FRF
transducers.
cylinder pressure As a fromconsequence, the
indirect the reconstruction
measurements hasof ofbeen
the accuracy
mapping under to improvetime-varying workingpressure
the cylinder points. reconstruction
Connolly and
transducers.
investigated As
by manya consequence,
researchers since reconstruction
the mid 1980s. the
Three accuracy
mapping under
to improvetime-varying
the working
cylinder points.
pressure Connolly
reconstructionand
cylinder
investigated pressure
by many from indirectsince
researchers measurements
the mid has Three
1980s. been Yagle
accuracy (1994)
under modeled
time-varying the cylinder
working pressure
points. via the
Connolly crank
and
cylinder pressure from indirect measurements has been Yagle (1994) modeled the cylinder pressure via the crank
main approaches
investigated
main approachesby many have
have
been investigated:
researchers
been since the mid
investigated:
engine
1980s.
engine Three accuracy
structure
structure angular
Yagle
under from
speed
(1994)
time-varying
modeled a the
working
statistical
cylinder point points.
of
pressure
Connolly
view.
via Shiao
the
and
and
crank
investigated by many researcherscrank since the mid 1980s. Three Yagleangular(1994)speed modeled
from a statistical
the cylinder point of view. via Shiao and
vibration-based
main
vibration-based
main approaches
reconstruction,
approachesreconstruction,
have been investigated:
have been
angular
crank angular
investigated:
engine
engine
speed-based
structure Moskwa
speed-based
structure angular speed (1995)from and aAl-Durra
statisticalet al. pressure
point (2011)
of view.have the
used
Shiao
crank
the
and
reconstruction, and
vibration-based combination of
reconstruction, engine
crank structure
angular Moskwa
angular
vibration extended
speed-based (1995)
speed and
from aAl-Durra
statistical et al.
point (2011)
of have
view. used
Shiao the
and
reconstruction,
vibration-based and combination
reconstruction, of engine
crank structure
angular vibration
speed-based Moskwa
extended sliding
(1995)
sliding observer
and Al-Durra
observer and
and the
theet Kalman
al.
Kalman(2011) filter
have
filter respectively
used the
respectively
and crank angular
reconstruction, andspeed-based
combinationreconstruction,
of engine structurerespectively. Moskwa
vibration extended (1995) and Al-Durra et al. (2011) have used the
and crank angular
reconstruction, andspeed-based
combinationreconstruction,
of engine structure vibration based
respectively. based on
on an
sliding
an
extended sliding
engine
engine energy
observer
energy andmodel
observer and the model whereas
the Kalman
whereas
Kalman
artificial
filter
artificial neural
respectively
neural
filter respectively
and crank angular speed-based reconstruction, respectively. networks (ANNs)
engineapproaches have been
1.1. Engine
and crank
1.1. Engine
Structure
angular Vibration-Based
speed-based
Structure
Reconstruction
reconstruction,
Vibration-Based respectively.
Reconstruction
based on an
networks
based on (ANNs)
an engine
energy model
approaches
energy model been investigated
whereas
havewhereas artificial neural
investigated
artificial
by
by
neural
1.1. Engine Structure Vibration-Based Reconstruction Jacob
networks et al. (1999),
(ANNs) Gu et
approaches al. (1999),
have Potenza
been et al.
investigated(2007),by
The
1.1. fast pressure change in a cylinderReconstruction
during combustion Jacob et al. (1999), Gu et al. (1999), Potenza et al. (2007),
The Engine
fast Structure
pressure change in a cylinder during combustion networks
Vibration-Based and
Jacob Bennett (ANNs) approaches have been investigated by
et al. (2014).
(1999), Gu et al. (1999), Potenza et al. (2007),
leads to engine structure vibrations. The potential of these and Bennett (2014).
The
The
fast
leadsfast pressurestructure
to engine
pressure
change vibrations.
change
in a cylinder
in a cylinder
during
Theduring
potential of these Jacob
combustion
combustion
et al. (1999), Gu et al. (1999), Potenza et al. (2007),
and Bennett (2014). Engine Structure Vibration and Crank
vibrations
leads to
to engine
vibrations to recover the
structure
recover the cylinder
vibrations.
cylinder pressure
The has
pressure been
potential
has been explored
of these 1.3.
explored and
1.3. Combination
Bennett
Combination (2014).ofof Engine Structure Vibration and Crank
leads
by to engine
Polonowksi structure
et al. vibrations. The potential of these
vibrations
by Polonowksito recover
et the(2007),
al. cylinderArnone
(2007), pressureet
Arnone has
et al.
al. (2009),
been explored
(2009), and
and 1.3. Angular
Combination
Angular Speed-Based
of Engine
Speed-Based Reconstruction
Structure Vibration and Crank
Reconstruction
vibrations to recover
Taglialatela et al. the cylinder
(2009). The pressure
results has been explored
showed that the 1.3. Combination of Engine Structure Vibration and Crank
by Polonowksi
Taglialatela et et al.
al. (2007),
(2009). The Arnone
results al. (2009),
etshowed that and
the It has Angular
been Speed-Based
shown that Reconstruction
both the vibration and the crank
by Polonowksi
vibration signal et al. (2007),information
Arnone et related
al. (2009), to and It hasAngular
been shown Speed-Based
that both Reconstruction
the vibration and the crank
Taglialatela
vibration et al. contains
signal (2009). The
contains results showed
information related that
to
the angular
the
the It has speed
been contain
shown information
that both theabout the cylinder
vibration and the pressure
crank
Taglialatela
combustionsignal et al. (2009).
process,contains
and can easilyThe results
be measured,showed that
normally the
on butangular
It has speed
beeninshown contain information
both theabout
thatfrequency the cylinder
vibration and the pressure
crank
vibration
combustion process, and can information
easily be measured,related to the
normally on mainly
angular speed different
contain information regions
about the (Johnsson,
cylinder 2006).
pressure
vibration
the cylindersignalhead orcontains
the information related vibration
to the angularbut mainly speed in contain
differentinformation
frequency about regions the(Johnsson,
cylinder 2006).
pressure
combustion process,
the cylinder process,
head or theandengine
engine
block.
can easily be However,
block. measured,the
However, normally
the vibration on In butJohnsson
mainly in(2006), a complex
different frequency RBF network
regions was proposed
(Johnsson, 2006).
combustion
signal also head contains and can easily be measured,
other non-combustion normally
information on In
but Johnsson
mainly in(2006), a
different complex
frequency RBF network
regions was
(Johnsson, proposed
2006).
the cylinder
signal also head or the engine
contains block. However, the
other non-combustion vibration for
information the reconstruction
In Johnsson (2006), aof the
complex cylinder
RBF pressure
network using both
was proposed the
the cylinder
because the cylinder or the engine
pressure block. However,
is non-combustion
not the unique source the vibration for the
In Johnsson
causing vibration reconstruction
(2006), of the
a complex cylinder pressure using
RBF network was proposed both the
signal also
because also contains
the cylinder other
pressure is non-combustion
not the unique source information and the
for the reconstruction
causing vibration crank angular
ofangular speed.
the cylinder pressure using both the
signal
engine contains other information and the crank speed.
for the reconstruction of the cylinder pressure using both the
engine structural
because
because
the cylinder
structural
theresponse
cylinder
vibrations
pressure (Bennett,
vibrations
pressure (Bennett,
is(FRF)
2014).
is not the unique Based
source
2014). source
not the between
unique
on
on the
causing
Based causing the More
vibration and the crank angular speed.
frequency
engine structural
frequency response function
vibrations (Bennett,
function(Bennett, 2014). Based
(FRF) between the on the More cylinder
cylinder
the cylinder vibration and the
cylinder
pressure
pressure
reconstruction
crank angular speed. methods
reconstruction
can be found
methods can be found
engine
pressure structural
and the vibrations
vibration signals, various2014).
methods Based have on the
been in
MoreJia et al.
cylinder (2016).
pressure Many of the
reconstruction proposed
methods methodologies
can be found
frequency response
pressure andresponse function
the vibration (FRF)
signals, between
various methods the cylinder been Morein Jia cylinder
et al. (2016).
pressureMany of the proposed
reconstruction methodsmethodologies
canstill
be found
frequency asfunction (FRF) between thehavecylinder haveJia shown promising results. However,
proposedthere
investigated,
pressure and
investigated,
such
the
suchvibration
as
inverse
signals,
inverse
filtering
various
filtering
applying
methods
applying
cepstrum
have
cepstrumbeen in
have
in Jia
et al. (2016).
shown
et al. promising
(2016).
Many
Many
of the
results.
of theHowever,
proposed there still exist
methodologies
methodologies exist
pressure
techniques and(Azzoni
the vibration
and signals, various
Paniero,1990; methods
Ren,applying have been some
1999; El-Ghamry have showndrawbacks. For
promising example,
results. for
However, inverse
there FRF-based
still exist
investigated,
techniques (Azzoni such as
and inverse filtering
Paniero,1990; Ren,applying cepstrum
1999; El-Ghamry some
have showndrawbacks. For
promising example,
results. for
However, inverse FRF-based
there still exist
investigated, orsuch asdomain
inverse filtering cepstrum and methods (with
et al., 2005)(Azzoni
techniques
et al., 2005) or
timeand
time
smoothing
Paniero,1990;
domain smoothing
techniques
Ren, (Gao
1999; El-Ghamry
techniques (Gao and methods
some (with or
some drawbacks.
drawbacks. or without
For example,
without
For
cepstrum
cepstrum
example,
fortechniques),
fortechniques),
inverse
we
we may
inverse FRF-based
FRF-basedmay
techniques
Randall, (Azzoni
1999), but and
also Paniero,1990;
nonlinear Ren, 1999;
methods El-Ghamry
like radial basis encounter
methods (with spectrum leakage,
or without ill-conditioned
cepstrum techniques), inversion,
we may or
et al., 2005) or time domain smoothing techniques (Gao and encounter spectrum leakage, ill-conditioned inversion, or
Randall,
et 1999),
al., 2005) but also
or time domain nonlinear
smoothing methods like radial
techniques (Gaobasisand methods
encounter
(with or without cepstrum techniques), we may
spectrum leakage, ill-conditioned inversion, or
Randall, 1999), but also nonlinear methods oflike radial basis
Randall, ©
2405-8963 2018, IFAC
1999), but also(International
nonlinearFederation
methods like Automatic Control) Hosting
radial basis encounter by Elsevier
spectrum Ltd. All rights reserved.
leakage, ill-conditioned inversion, or
Peer review©under
Copyright 2018 responsibility
IFAC of International Federation of Automatic
744Control.
Copyright © 2018 IFAC
10.1016/j.ifacol.2018.10.161 744
Copyright © 2018 IFAC 744
Copyright © 2018 IFAC 744
 IFAC E-CoSM 2018
Changchun, China, September 20-22, 2018 Runzhe Han et al. / IFAC PapersOnLine 51-31 (2018) 700–705 701

FRF variations problems. Also, inverse FRF-based methods working points to show whether this is the case for the engine
can only be used for systems with an input number less than used in this study. It can be seen in Fig. 2 that the coherence
or equal to the output number. ANNs normally need large estimates show constant high values up to a certain cutoff
amounts of acquired data to train the network. For engine frequency nearly independent on the loads tested. Thus, it can
energy model-based methods, the engine energy model has to be concluded that the relationship between the cylinder
be known and needs to be calibrated first, which can be pressure and the vibration can approximately be seen as
expensive and time-consuming in practice when used for linear in this frequency range. For higher speeds the cutoff
different types of engines. Also, a majority of the researches frequency is more or less constant at about 1 kHz and only
have been performed on one- or two-cylinder engines which reduces significantly for low speeds (e.g., 1200 rpm). For this
are relatively small in size with low power output. reason, the collected signals for this study were low-pass
filtered with a unified cutoff frequency of 1 kHz.
The aim of this study therefore was to overcome the above
drawbacks by using a novel approach tested with a four-
cylinder diesel engine. The basic idea was to identify a linear
model with four input signals (cylinder pressure) and one
output signal (engine structure vibration) first, and then to
regard the other three cylinder pressures as the delays of the
pressure of cylinder No. 1 in every engine cycle. This leads to
a single-input single-output (SISO) system which can be used
with the vibration signal to estimate the pressure of cylinder
No. 1 based on using the disturbance estimation method
proposed by Bohn et al. (2004). In this approach, the reason
for using the delays is to guarantee the observability for the
state estimation. To the best of the authors’ knowledge, such Figure 2. Coherence between cylinder pressure and vibration.
a cylinder pressure reconstruction algorithm has not been
reported previously in the literature. In the following, based on the signals analysis in Fig. 2, the
modeling between the cylinder pressure and the vibration is
The remainder of this paper is organized as follows. In Sec. introduced first, and then a novel cylinder pressure estimation
2, the engine test bench used for evaluation is introduced. method is derived using delay blocks.
Sec. 3 illustrates the main estimation algorithm for the
cylinder pressure followed by Sec. 4, where experimental 3.1. Model Identification
identification and estimation results are presented. Finally, In this study, the model G with four input signals (cylinder
some discussions and conclusions are given in Sec. 5. pressure) and one output signal (vibration) is assumed to be
linear time-invariant. Both subspace identification method
2. TEST BENCH DESCRIPTION
and prediction-error method (PEM) (Ljung, 1999) were used
Fig. 1 describes a four-cylinder diesel engine with its for the model G identification. The model G in discrete time
measurement system used as the test bench for evaluation. is represented by the innovations form
The cylinder pressure, the vibration acceleration (used as the
vibration signal), and the crank angular speed signals can be xt (k  1)  At xt (k )  Bt ut (k )  K t et (k )
, (1)
collected synchronously using the measurement system (the yt (k )  C t xt (k )  et (k )
sampling frequency was chosen to be 20 kHz). The cylinder
where xt (k )  p is the state vector, ut (k )  4 is the input
pressure transducers used in this study are piezoelectric
signal vector, and yt (k )  is the output signal of the model
transducers which are integrated in glow plugs. The cylinder
(1) (the model G). et (k ) is a scalar innovations sequence,
pressure pegging was done using polytropic model-based
of which the autocovariance is  2 .
least-squares algorithm (Lee et al., 2008). This experiment
setup was used for measurements within typical ranges of 3.2. Augmented Model Derivation
angular speed (1200 – 3000 rpm) and load (60 – 180 Nm). The cylinder pressure signal in discrete-angle domain can be
approximately expressed as a Fourier series with amplitudes
ai and phases  i (Bohn et al., 2005), that is
n 1 1
Pm ( (k ))  a0   i 1 ai sin( i (k )  i ) , (2)
2
where  (k ) is the engine crank angle, Pm ( (k ))  denotes
the cylinder No. m pressure signal. n was chosen to be 40
according to the spectrum analysis of the cylinder pressure.
Figure 1. Experimental setup.
Based on (2), the cylinder pressure signal in discrete-time
3. CYLINDER PRESSURE ESTIMATION ALGORITHM domain can be modeled as the output of an autonomous state
space with the state transition equation (Bohn et al., 2004)
Earlier in this paper, it was noted that the vibration
measurement contains information about the cylinder xPm (k  1)  APm ( f (k )) xPm (k ) (3)
pressure. A coherence analysis was made under nine different
and the output equation

745
IFAC E-CoSM 2018
Changchun, China, September 20-22, 2018 Runzhe Han et al. / IFAC PapersOnLine 51-31 (2018) 700–705
702

Pm (k )  CPm xPm (k ) . (4) and yd (t )  is the output of the model (6). Ad ( f (t )) and
Cd ( f (t )) are time-varying matrices which depend on f (t ) ,
The matrices APm ( f (k )) and C Pm are given as Bd is a constant matrix.
1 0 0  The discretized model of the system (6) can be expressed as
0 A ( f (k )) 
APm ( f (k ))  
Pm , 1  xd (k  1)  Gd ( f (k )) xd (k )  H d ( f (k ))ud (k )
  , (7)
0 yd (k )  Cd ( f (k )) xd (k )
 
0 0 APm , n-1 ( f (k ))  where Gd ( f (k )) and Hd ( f (k )) are time-varying matrices
and which depend on f (k ) .
Additionally, if we use the conceptual discrete time-varying
CPm  1 CPm , 1 CPm , n -1 
. delay block directly instead of using the above procedure (i.e.,
The individual block entries in these block matrices can be from (5) to (7)), its dimension can be very large which is not
represented as suitable for practical applications. The specific discretization
process from (5) to (7) can be referred to Han et al. (2018).
 cos(i  2 f (k )Ts ) sin(i  2 f (k )Ts ) 
APm , i ( f (k ))    By combing the discretized delay systems with the model G
  sin(i  2 f (k )Ts ) cos(i  2 f (k )Ts )  according to Fig. 3, a SISO model can be obtained, and
and substituting P1 (k ) and V (k ) (the vibration signal) into the
transformed SISO model, we can obtain
C Pm , i  1 0 ,
xs (k  1)  As ( f (k )) xs (k )  Bs ( f (k )) P1 (k )  K s et (k )
where Ts is the sampling time, and f (k ) is the instantaneous , (8)
V (k )  Cs xs (k )  et (k )
engine cycle frequency which can be computed using the
instantaneous crank angular speed. where xs (k )  r
is the state vector of the model (8).
The pressure signals of the other cylinders can be seen as the The matrices As ( f (k )) , Bs ( f (k )) , C s , K s , and xs (k ) are
delay versions of the pressure signal of cylinder No. 1 since given as
the phase difference is 180 degrees between every two
consecutive engine strokes and the pressure signal shapes   0 0 0  
corresponding to different cylinders are nearly the same in  C ( f ( k )) 0 0  
A B  d
 
every engine cycle. As a result, a SISO system as shown in 
t t

 0 C ( f ( k ))
d
0  
Fig. 3 can be obtained by combining three same delay    ,
systems with the model G.
A ( f ( k )) =
s
  0 0 C ( f ( k )) 
d

  G ( f ( k )) 0 0  
 
d

H
0  d
( f ( k ))C ( f ( k ))
d
G ( f ( k ))
d
0

  0  
H ( f ( k ))C ( f ( k )) G ( f ( k )) 
d d d

Bs ( f (k ))  1 0 0 0 Bt
T
T
H d ( f (k )) 0 0
T
 ,
Cs  Ct 0 0 0 ,
T
K s   K t T 0 0 0
 ,
and
Figure 3. Discrete-time SISO system.
T
The conceptual continuous time-varying transfer function of xs (k )   xt T (k ) xd(1) T (k ) xd( 2) T (k ) xd(3) T (k ) 
 ,
the delay system in Fig. 3 can be represented as
where xd( i ) (k ) denotes the state of the i-th delay system.
H (s)  eTd (t ) s , (5)
By augmenting the SISO model (8) with the state of the
where Td (t ) is the time-varying dead time which corresponds cylinder pressure signal model (represented by (3) and (4)),
to a crankshaft rotation angle of 180 degrees, i.e., we can obtain an augmented model as follows
Td (t )  1 (4 f (t )) , where f (t ) denotes the continuous-time
instantaneous engine cycle frequency. xa (k  1)  Aa ( f (k )) xa (k )  K a et (k )
, (9)
V (k )  Ca xa (k )  et (k )
By using Padé approximation (Vajta, 2000), we can obtain a
rational function approximation of (5), and then transform the where xa (k )  w
is the state vector of the model(9).
approximation into a linear time-varying state-space model
The matrices Aa ( f (k )) , C a , K a , and xa (k ) are given as
xd (t )  Ad ( f (t )) xd (t )  Bd ud (t )
, (6)  As ( f (k )) Bs ( f (k ))C P1 
yd (t )  Cd ( f (t )) xd (t ) Aa ( f (k ))   ,
 0 AP1 ( f (k )) 
where xd (t )  q
is the state vector, ud (t )  is the input,

746
 IFAC E-CoSM 2018
Changchun, China, September 20-22, 2018 Runzhe Han et al. / IFAC PapersOnLine 51-31 (2018) 700–705 703

Ca  Cs 0 , collected under different engine working points. Two

datasets, A and B, are given for model identification and
T
K a   K s T 0
 ,
validation, as shown in Table 1. The numbers 1, 2, and 3
represent the working points 2100 rpm at 60 N·m, 120 N·m,
and and 180 N·m, respectively while 4, 5, and 6 represent the
working points 3000 rpm at 60 N·m, 120 N·m, and 180 N·m,
 xs (k )  respectively. The numbers 7, 8, and 9 represent the working
xa ( k )   . (10)
 xP1 (k )  points 1200 rpm at 60 N·m, 120 N·m, and 180 N·m,
respectively. The symbol ‘⋯’ denotes the transient process
Since the model (9) is uniformly completely stochastic between two steady-state working points.
observable, the Kalman filter for the augmented model (9) can
be guaranteed to be uniformly asymptotic stable. Additionally, Table 1(a). Model identification dataset A
the state estimation for the system (9) can work. 1 ⋯ 2 ⋯ 3 ⋯ 4 ⋯ 5 ⋯ 6
3.3. Kalman Filter Implementation Table 1(b). Model validation dataset B
With the model (9) and the vibration signal V (k ) , the
Kalman filter can be used for the estimation of xa (k ) , and 7 ⋯ 8 ⋯ 9
then it can be used for the reconstruction of the cylinder We applied subspace methods with different options to the
pressure according to (4). Considering the aspect of the dataset A, as well as the PEM. Different model order
numerical computation, the Potter’s square-root Kalman filter candidates were chosen according to the gap in the singular
(SRKF) (Simon, 2006) was implemented. However, due to value spectrum, for the prediction horizons we used the
the use limit (e.g., no process noise) of the Potter’s SRKF a defaults.
forgetting factor should be involved to tune the uncertainty
error covariance to avoid the ‘sleep’ of the filter. The cylinder The best results (fitness) were achieved for a model order of
pressure estimation algorithm is illustrated as follows. 7 that shown in Table 2. The different models correspond to
subspace estimates with canonical variate analysis (CVA)
Algorithm 1: Cylinder pressure estimation algorithm and multivariable output-error state space (MOESP)
Step 1. Assume W (k  1)  I w and the initial state weightings, with focus = ‘prediction’ or focus = ‘simulation’
estimation value xˆ a (k  1)  0 , where k  1 , W (k ) is the (CVA(s) and MOE(s) for simulation focus) (Ljung, 1999).


square root of the a posterior estimate error covariance P (k )
, I w is a w  w identity matrix; As for the role of the PEM in this paper, the parameter
Step 2. The prediction xˆ a  (k )  Aa ( f (k )) xˆ a (k 1) ; estimates are initialized using the subspace approach and then
Step 3. W  (k )  Aa ( f (k ))W (k )  , where W  (k ) is the the parameter values are refined using the prediction error
square root of the a prior estimate error covariance P  (k ) ,  minimization approach (with CVA weighting and focus =
 is the forgetting factor and 0    1 ; ‘simulation’).
Step 4. F (k )  W  (k )Ca T ; The measure used in Table 2 is
Step 5.  (k )  1 (F T (k )F (k )   2 ) ;
Step 6. K (k )   (k )W  (k )F (k ) , where K (k ) is the  y  yˆ 
 fitness  100  1  ,
Kalman filter gain;  y mean( y) 
Step 7. xˆ a (k )  xˆ a  (k )  K (k )(V (k )  Ca xˆ a  (k )) ;  
Step 8.  (k )  1 (1    (k )) ; where  denotes the Euclidean norm, y is the validation
Step 9. W (k )  W  (k )   (k ) K (k )F T (k ) ; data output and ŷ is the response of the model.
Step 10. According to (10) and the estimation xˆ a (k ) in
Table 2. Model identification results for 7-th order models
Step 7, the estimation xˆ P1 (k ) can also be obtained, and then
based on using (4) the cylinder No. 1 pressure estimation CVA MOESP CVA(s) MOE(s) PEM
 Pˆ1 (k )  C P1 xˆ P1 (k ) ; Dataset A 61.42 51.71 68.96 69.05 68.98
Step 11. By using the delay system (7), the pressure Dataset B 51.20 52.10 65.19 66.00 64.85
signals of other cylinders can be simultaneously computed By comparing the identification results in Table 2, the
using the estimation Pˆ1 (k ) ; ‘simulation’ focus is better than the ‘prediction’ focus, and
Step 12. Return Step 2 with k  k  1 . the identified model with 7-th order and MOE(s) was finally
To be more, in Algorithm 1, F (k ) ,  (k ) and  (k ) are the chosen for the estimation.
intermediate variables. 4.2. SRKF Estimation Results
4. EXPERIMENTAL RESULTS Based on the identified model and the measurements from the
test bench, the pressure signal for cylinder No. 1 was
In this part, the identification results for the model G are estimated under nine different engine working points (100
shown first, and then the estimation results for the pressure of engine cycles chosen for each working point). The estimation
cylinder No. 1 are given. The pressure estimates of other results for a typical cycle are shown in Figs. 4, 5, and 6. It
cylinders are not given since the limited space. can be seen that the estimation works quite well in the region
4.1. System Identification Results around the pressure peak but not in the engine intake and
exhaust strokes (the fluctuations at the bottom).
The input-output data for the identification of G were

747
IFAC E-CoSM 2018
Changchun, China, September 20-22, 2018 Runzhe Han et al. / IFAC PapersOnLine 51-31 (2018) 700–705
704

Table 3. Estimation results evaluation based on the identified model

1200 rpm, 1200 rpm, 1200 rpm, 2100 rpm, 2100 rpm, 2100 rpm, 3000 rpm, 3000 rpm, 3000 rpm,
60 N·m 120 N·m 180 N·m 60 N·m 120 N·m 180 N·m 60 N·m 120 N·m 180 N·m
Mean (Pmax error) –17.74% 12.00% –4.45% –3.56% –5.02% –5.42% 9.05% 3.09% 6.03%
SD (Pmax error) 0.83% 1.06% 2.81% 0.22% 0.76% 0.64% 0.22% 0.35% 0.40%
Mean (Ploc error) 2.24 0.35 1.00 1.91 2.30 1.71 2.11 1.34 1.91
SD (Ploc error) 0.17 0.15 0.57 4.08 0.26 0.32 0.46 0.47 0.44

To evaluate the effectiveness of the proposed method, the

relative error of pressure peak Pmax and the absolute error of
peak location Ploc have been investigated given by
Pmax  Pmax, est
Pmax error = 100% ,
Pmax
and
Ploc error  Ploc  Ploc, est ,
where Pmax, est and Ploc, est are corresponding estimated values.
Table 3 shows the mean and standard deviation (SD) of these Figure 6. Estimation results at 3000 rpm.
errors for all the working points considered. It can be seen
that for Pmax estimation, the mean of the Pmax error varies
from –17.74% to 12.00%, but most of the absolute values of
the mean of Pmax error are less than 6%. Also most of the
absolute values of the SD of Pmax error are less than 1.10%,
which indicates the estimation stability.
For Ploc estimation, the absolute values of the mean of Ploc
error do not exceed 2.30° for all working points, and the
corresponding SD values are very small except the test point
2100 rpm and 60 N·m, which indicates the estimation
framework is very effective for Ploc estimation. Since most
other researches have obtained Pmax errors within the range of
Figure 7. Estimation results at 120 N·m using an ideal model.
5 – 10% and Ploc errors within 5 (Bennett, 2014), it can
indicate the potential of the proposed method here. Table 4. Estimation results evaluation using an ideal model
1200 rpm, 2100 rpm, 3000 rpm,
120 N·m 120 N·m 120 N·m
Mean (Pmax error) 5.10% 4.46% 0.93%
SD (Pmax error) 0.64% 0.37% 0.39%
Mean (Ploc error) 1.05 1.86 0.64
SD (Ploc error) 0.17 0.61 0.44

Moreover, the limit of the proposed method can be tested

under an ideal condition by assuming the identified model to
be an ideal linear time-invariant model between the four input
cylinder pressure signals and one artificially calculated output
vibration signal. In Fig. 7 and Table 4, the results for this
Figure 4. Estimation results at 1200 rpm. ideal case are shown under the working points 1200 rpm,
2100 rpm, and 3000 rpm at 120 N·m, respectively. It is
obvious that with a perfect model, the estimation results are
improved generally. It can also be seen that the fluctuations
in the estimation signal are much lower than before, so that
they are mainly caused by the identification error of the
model G.
5. DISCUSSIONS AND CONCLUSIONS
5.1. Discussions
The estimation results are consistent with the coherence
analysis in Fig. 2, that is to say both the estimation and the
Figure 5. Estimation results at 2100 rpm.
coherence results are better for high speeds (e.g., 3000 rpm). It

748
 IFAC E-CoSM 2018
Changchun, China, September 20-22, 2018 Runzhe Han et al. / IFAC PapersOnLine 51-31 (2018) 700–705 705

should be noted that the model used for all working points is 12(8), 1029-1039.
just a linear time-invariant model which is not suitable for low Bohn, C., Magnor, O., and Schultalbers, M. (2005). State observer based
analysis of crankshaft speed measurements with application to misfire
speeds (e.g., 1200 rpm).
detection. Proceedings of IEEE International Conference on Control
As seen in Table 4, if the model is accurate, the estimation and Automation. Budapest, June. 239-244.
results can be improved greatly especially for a low speed. In Connolly, F.T. and Yagle, A.E. (1994). Modeling and identification of the
combustion pressure process in internal combustion engines.
summary, two main parts can affect the estimation results in
Mechanical Systems and Signal Processing, 8(1), 1-19.
this study: Du, H., Zhang, L., and Shi, X. (2001). Reconstructing cylinder pressure from
1) Modeling: The identification accuracy of the model G has vibration signals based on radial basis function networks. Proceedings
of the Institution of Mechanical Engineers, Part D: Journal of
a big effect on the estimation according to the results Automobile Engineering, 215(D6), 761-767.
comparison between Table 3 and Table 4. Several factors can El-Ghamry, M., Steel, J.A., Reuben, R.L., and Fog, T.L. (2005). Indirect
determine the model quality: i) inaccurate measurements (e.g., measurement of cylinder pressure from diesel engines using acoustic
cylinder pressure transducers can be affected by the high emission. Mechanical Systems and Signal Processing, 19(4), 751-765.
temperature in cylinders), ii) inaccurate input or output Gao, Y. and Randall, R.B. (1999). Reconstruction of diesel engine cylinder
number (e.g., the cylinder pressure is not the only one source pressure using a time domain smoothing technique. Mechanical
Systems and Signal Processing, 13(5), 709-722.
causing engine structure vibrations), and iii) inaccurate model
Gu, F., Jacob, P.J., and Ball, A.D. (1999). Non-parametric models in the
structure assumption (e.g., the linear assumption on G may monitoring of engine performance and condition. Part 2: non-intrusive
lead to some errors if nonlinearity is severe). estimation of diesel engine cylinder pressure and its use in fault
detection. Proceedings of the Institution of Mechanical Engineers, Part
2) Conceptual time-varying delay block: The use of the
D: Journal of Automobile Engineering, 213(D1), 135-143.
conceptual time-varying delay block is based on the Han, R., Bohn, C., and Bauer, G. (2018). A Novel Inverse Filtering Method
assumption that the shapes of the pressure signals of the four for Systems with Multiple Input Signals. Accepted to appear in
cylinders are the same in every engine cycle which is not Proceedings of the 18th international conference on control,
strictly true because of some factors, such as cylinder-to- automation and systems. PyeongChang, October.
cylinder injection and air-EGR dispersions, so this assumption Jacob, P.J., Gu, F., and Ball, A.D. (1999). Non-parametric models in the
can lead to some errors in estimation. Furthermore, the Padé monitoring of engine performance and condition. Part 1: modeling of
non-linear engine processes. Proceedings of the Institution of
approximation of the conceptual continuous time-varying Mechanical Engineers, Part D: Journal of Automobile Engineering,
delay block can also bring some errors (Han et al., 2018). 213(D1), 73-81.
5.2. Conclusions Jia, L., Jeffrey, D.N., and Jason, R.B. (2016). Review of sensing
methodologies for estimation of combustion metrics. Journal of
In this paper, the cylinder pressure estimation results show Combustion, 2016, Article ID 8593523.
the effectiveness of the proposed method over different Johnsson, R. (2006). Cylinder pressure reconstruction based on complex
radial basis function networks from vibration and speed signals.
engine speeds and loads. This method can also be used for
Mechanical Systems and Signal Processing, 20(8), 1923-1940.
the cylinder pressure estimation for IC engines with more Lee, K., Yoon, M., and Sunwoo, M. (2008). A study on pegging methods for
cylinders. Based on the estimation results and discussions, noisy cylinder pressure signal. Control Engineering Practice, 16(8),
further studies for the current framework improvement can 922-929.
include nonlinear system identification for the model G, the Ljung, L. (1999). System identification theory for the user. Prentice Hall,
effects from the conceptual time-varying delay block on Upper Saddle River, New Jersey.
estimation, and the optimal mounting location of the Moro, D., Cavina, N., and Ponti, F. (2002). In-cylinder pressure
reconstruction based on instantaneous engine speed signal. Journal of
vibration accelerometer for estimation. Additionally, as
Engineering for Gas Turbines and Power, 124, 220-225.
mentioned earlier in this paper, the crank angular speed also Polonowksi, C.J., Mathur, V.K., Naber, J.D., and Blough, J.R. (2007).
contains information about the cylinder pressure, so the Accelerometer based sensing of combustion in a high speed HPCR
identification of the model G can involve the crank angular diesel engine. Proceedings of SAE World Congress & Exhibition.
speed as another output in the future. Detorit, April. SAE Technical Paper 2007-01-0972.
Potenza, R., Dunne, J.F., Vulli, S., and Richardson, D. (2007). Multicylinder
REFERENCES engine pressure reconstruction using NARX neural networks and crank
kinematics. International Journal of Engine Research, 8(6), 499-518.
Al-Durra, A., Canova, M., and Yurkovich, S. (2011). A model-based
Ren, Y. (1999). Detection of knocking combustion in diesel engines by
methodology for real-time estimation of diesel engine cylinder
inverse filtering of structural vibration signals. University of New
pressure. Journal of Dynamic Systems, Measurement, and Control,
South Wales, Sydney, New South Wales.
133(3), 031005-031005-9.
Shiao, Y. and Moskwa, J.J. (1995). Cylinder pressure and combustion heat
Arnone, L., Manelli, S., Chiatti, G., and Chiavola, O. (2009). In-cylinder
release estimation for SI engine diagnostics using nonlinear sliding
pressure analysis through accelerometer signal processing for diesel
observers. IEEE Transactions on Control Systems Technology, 3(1),
engine combustion optimization. Proceedings of SAE 2009 Noise and
70-78.
Vibration Conference and Exhibition. Ann Arbor, May. SAE Technical
Simon, D. (2006). Optimal state estimation: Kalman, H infinity, and
Paper 2009-01-2079.
nonlinear approaches. John Wiley & Sons, Hoboken, New Jersey.
Azzoni, P. and Paniero, A. (1990). Cylinder pressure reconstruction by
Taglialatela, F., Cesario, N., Porto, M., Merola, S.S., Sementa, P., and
cepstrum analysis. Proceedings of the 22nd International Symposium
Vaglieco, B.M. (2009). Use of accelerometers for spark advance
on Automotive Technology and Automation. Florence, May. 629-637.
control of SI engine. SAE International Journal of Engines, 2(1), 971-
Bennett, C. (2014). Reconstruction of gasoline engine in-cylinder pressures
981.
using recurrent neural networks. University of Sussex, Brighton.
Vajta, M. (2000). Some remarks on Padé-approximations. Proceedings of the
Bohn, C., Cortabarria, A., Härtel, V., and Kowalczyk, K. (2004). Active
3rd TEMPUS-INTCOM Symposium. Veszprém, September. 1-6.
control of engine-induced vibrations in automotive vehicles using
disturbance observer gain scheduling. Control Engineering Practice,

749