Control Engineering Practice 18 (2010) 873–881

Contents lists available at ScienceDirect

Control Engineering Practice

journal homepage: www.elsevier.com/locate/conengprac

Adaptive internal model control with application to fueling control

Daniel Rupp , Lino Guzzella
Department of Mechanical and Process Engineering, ETH Zurich, ML K, Sonneggstrasse 3, 8092 Zurich, Switzerland

a r t i c l e in fo abstract

Article history: This paper presents an adaptive internal model controller for stable but not necessarily minimum-
Received 22 December 2008 phase SISO plants and its application to the air/fuel ratio control system of a spark-ignited engine. The
Accepted 17 March 2010 internal model of the controller is formulated in an output-error structure that can be adapted by using
Available online 8 April 2010
standard adaptive laws. The method is applied to an air/fuel ratio control system with a reduced-order
Keywords: internal model and unknown sensor dynamics. Experiments on an engine test bench demonstrate the
Adaptive control capability of the adaptive controller to recover the performance and robustness properties of the
Internal model control control system in the case of an aged oxygen sensor.
Output-error model & 2010 Elsevier Ltd. All rights reserved.
Air/fuel ratio
Oxygen sensor
Sensor ageing

1. Introduction automated design strategy based on H1 control. The authors of

The most common pollution abatement system for SI port-
injection engines is the three-way catalytic converter. It derives
its name from its ability to simultaneously reduce NOx and
oxidize CO and HC. State-of-the-art systems are capable of
removing more than 98% of the pollutants. However, this can
only be achieved by operating the engine within very narrow air/
fuel-ratio (AFR) limits, which requires a very precise control
system. Aside from the feedforward controller, which compen-
sates for excursions of the AFR during engine transients, the
feedback controller uses the measurement data from a wide-
range oxygen sensor (also called lambda sensor) to compensate
for all kinds of disturbances.
For several reasons, the dynamic behavior of an AFR sensor
may change considerably during its lifetime. To some extent, a
robust controller can mitigate this change. However, the trade-off
between robustness and performance may be in conflict with the
increasing demands on the effectiveness of the control system.
This paper presents an adaptive control concept which is capable
of sustaining a high level of performance even in the presence of
substantial changes in the sensor dynamics.
AFR control has been addressed by many researchers in the
field. In Jones, Ault, Franklin, and Powell (1995) and Stroh,
Franchek, and Kerns (2001) recursive identification is used to tune
a model-based AFR controller online. The authors of Roduner,
Onder, and Geering (1997) and Alfieri, Amstutz, and Guzzella
(2009) focussed on the design of the feedback controller and its
 Corresponding author. Tel.: + 41 44 632 2453; fax: + 41 44 632 1139.
E-mail address: daniel.rupp@alumni.ethz.ch (D. Rupp).
Guzzella, Simons, and Geering (1997) employed feedback linear-
ization for the control of the air/fuel ratio. Adaptive strategies
based on Kalman filtering techniques have been described in
Turin and Geering (1995) and recently in Muske, Jones, and
Franceschi (2008). Due to the inherent time delay in engine
systems Smith predictors, or more generally, internal model
controllers (IMC) (Balenovic, Backx, & Hoebihk, 2001; Inagaki,
Ohata, & Inoue, 1990) and their adaptive counterparts (Yildiz,
Annaswamy, Yanakiev, & Kolmanovsky, 2008) have been gaining
interest with the context of AFR control. The requirement for the
adaptation of the oxygen sensor dynamics has evolved only
recently in the automotive industry. In Rupp and Guzzella (2009)
an IMC controller is proposed that is tuned iteratively to cope
with changes in the dynamics of oxygen sensors.
This paper deals with the adaptation of the internal model of
the IMC that regulates the AFR. The methodology that was
developed for the adaptive AFR control is introduced in a general
framework, such that the same concept can be applied to
practically any other internal model control systems with stable
but not necessarily minimum-phase plants. The idea of an
adaptive internal model has already been pursued earlier. The
authors of Datta and Ochoa (1996) suggest the use of a series–
parallel identification model based on the equation error method,
which implies that the output of the internal model is calculated
based on the input and the measured output of the plant. The
method presented in this paper proposes to use the output error
method, where the output of the internal model is calculated
based on the input to the plant and prior model outputs. As
pointed out by the authors of Datta and Ochoa (1996), in contrast
to the equation error method, with the output error method a
guarantee for global stability often cannot be given. However, the

0967-0661/$ - see front matter & 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.conengprac.2010.03.011
874 D. Rupp, L. Guzzella / Control Engineering Practice 18 (2010) 873–881
local stability results obtained are shown here to be sufficient in
the case of AFR control, which is a representative application of
internal model controllers. It is also shown that by foregoing
global stability the robustness of the adaptation regarding
external disturbances and the control performance can be
increased. These are important features when adaptive control
methods are used for processes such as AFR control where only a
reduced-order model is available.
In contrast to classical adaptive control systems, for the
method described in this paper, typically a substantial amount
of prior information about the model is available, i.e. the initial
internal model is already a good approximation of the dominant
plant dynamics. In classical model reference adaptive controllers,
the adaptation signals are used to compensate for disturbances in
order to match a reference model. With the method presented
here, these signals are used solely to improve the internal model
of the IMC, while the resulting adapted but static IMC handles the
rejection of disturbances. The new method can thus be seen as a
stable way to increase controller performance by online closed-
loop system identification. This implies that the adaptation needs
to be turned on only when the model deviates from the plant.
2. Air/fuel ratio control of SI engines
Lambda sensors are an essential part of modern AFR control
systems. Based on the measurement data provided by the oxygen
sensor, the controller in the engine control unit (ECU) regulates
the correct amount of fuel to be injected. Due to ageing and harsh
operating conditions, a wide-range oxygen sensor can undergo
substantial changes in its dynamics, which have a detrimental
effect on the performance of the control system.
2.1. The air/fuel ratio dynamics of an SI engine
Fig. 1 shows the fuel path of an SI engine. Its dynamics
comprise several delays, the wall-wetting dynamics, the gas
mixing dynamics, and the sensor dynamics. The input uf to the
plant is a multiplicative factor for the nominal injection duration,
and the output ls is the signal produced by the lambda sensor,
which is assumed to measure the lambda value lmC at the main
confluence point of the exhaust pipes. A detailed analysis of this
system can be found in Onder, Roduner, and Geering (1997).
However, for control purposes, the complete fuel path from the
fuel injector to the main confluence point can be approximated
fairly well by a first-order low-pass element with a time constant
t and a transport delay d (Guzzella & Onder, 2004)
DlmC 1
Pe ðsÞ ¼ ¼  esd , ð1Þ
Duf t  sþ1
where s is the Laplace variable and D stands for a small deviation
from the nominal value. For small deviations of the nominal
injected fuel mass the linear model Pe(s) also covers the wall-
wetting dynamics. This can be seen in Guzzella and Onder (2004,
Fig. 1. Main dynamic subsystems in the fuel dynamics of a spark-ignited engine.
p. 190) where the injected fuel mass is varied from  10% to + 10%
of the nominal fuel mass without any compensation of the wall-
wetting dynamics. The two parameters of this model are
dependent on the operating point of the engine, as determined
by the engine speed and the mass of air in the cylinders. The
values of the two parameters are typically available from engine-
specific maps that are identified off-line. A typical range for both
parameters is from 30 ms for high speed and high air mass up to
0.5 s for low speed and low air mass. The relative errors of the
maps that are used in production-type cars are less than 20% for
all operating points except very low air mass and speed. The
wide-range lambda sensor can also be modeled as a first-order
low-pass element with the time constant ts :
Dls 1
Ps ðsÞ ¼ ¼ : ð2Þ
DlmC ts  s þ 1
The time constant of a new state-of-the-art wide-range lambda
sensor is about 50 ms, but during its lifetime this time constant ts
can increase to 1 s.
2.2. Internal model controller
The main goal of the feedback controller is to compensate for
the disturbances in the AFR as a result of potentially rapidly
varying torque commands issued by the driver. For several
reasons, such as easy tunability and parametrization as well as
robustness, an IMC as illustrated in Fig. 2 is used. The performance
and robustness of a control system governed by a model-based
controller obviously is strongly connected to the quality of the
model available. To illustrate this, let us assume that the relative
uncertainty of the model PðsÞ ^ with respect to the plant P(s)
satisfies the inequality
 
PðjoÞPðj
^ oÞ

  o W2 ðoÞ, 8o
 ^ oÞ
Pðj 
for all frequencies o, where W2 ðoÞ is the frequency-dependent
uncertainty bound. The IMC control system is robustly stable if
the inequality
1 1
jQ ðjoÞj o  , 8o ð3Þ
^ oÞj W2 ðoÞ
jPðj
holds for all frequencies o (Morari & Zafiriou, 1989). In the ideal
case, when no disturbances or model uncertainties are present,
the internal controller Q(s) acts as a pure feedforward controller.
It is thus typically designed to invert the plant, at least in the
control-relevant frequency range. However, (3) shows that the
model uncertainty W2 ðoÞ acts as a border line in the tradeoff of
robustness versus performance. Specifically, given a fixed internal
controller Q(s), the robustness can be increased or recovered by
improving the precision of the model.
Fig. 2. Internal model control system with correction signal hT  nðu, yÞ
^ for the
internal model, which will be defined in Section 3.1 below.
 D. Rupp, L. Guzzella / Control Engineering Practice 18 (2010) 873–881 875

3. Adaptation of the internal model—the nominal case and

hT ^ ^ ^ T
y ¼ ½an1 a n1 ,an2 a n2 , . . . ,a0 a 0 
In this section an adaptive law is derived for the nominal case
in which no disturbances are acting on the system and in which such that the model transfer function resulting from (9) matches
the order of the plant is fully known. the plant P(s):
^ þ hT  bn
NðsÞ NðsÞ
u
3.1. Plant matching adaptation parameters ¼ : ð10Þ
^ þ hT  an
DðsÞ DðsÞ
y

For the stable SISO plant

NðsÞ 3.2. Error model
PðsÞ ¼ ð4Þ
DðsÞ
with The adaptation parameter vector h can be written as the sum
of the plant-matching adaptation parameter vector h and the
m m1
NðsÞ ¼ bm  s þ bm1  s þ    þ b1  s þ b0 adaptation parameter error vector /:
and h ¼ h þ /: ð11Þ
n n1
DðsÞ ¼ s þ an1  s þ    þ a1  s þa0 , With (11), Eq. (9) can be written as
with known orders n and m and known estimates a^ i and b^ i of the ^ þ ðhT þ /T Þ  an Þ  fyg
ðDðsÞ ^ þðhT þ /T Þ  bn Þ  fug
^ ¼ ðNðsÞ ð12Þ
y y u u
coefficients ai and bi, respectively, an IMC is designed based on the
model with /u ¼ hu hu
and /y ¼ hy hy .
By using the plant-matching
^ condition (10), Eq. (12) can be rewritten as
^ ¼ NðsÞ ,
PðsÞ ð5Þ
^ NðsÞ 1
DðsÞ y^ ¼  fug þ  f/Tu  bn  fug/Ty  an  fygg:
^ ð13Þ
DðsÞ DðsÞ
with
The second term on the right-hand side of (13) is now expanded
^
NðsÞ ¼ b^ m  sm þ b^ m1  sm1 þ    þ b^ 1  s þ b^ 0 as follows:
and NðsÞ ^
DðsÞ 1
y^ ¼  fug þ   f/Tu  bn  fug/Ty  an  fygg:
^ ð14Þ
^ DðsÞ ^
DðsÞ DðsÞ
DðsÞ ¼ sn þ a^ n1  sn1 þ    þ a^ 1  s þ a^ 0 : ð6Þ
To account for the discrepancy between the plant and the model, With (7), Eq. (14) can be reformulated as
the signal hT  nðu, yÞ,^ consisting of the adaptation parameter ^
NðsÞ DðsÞ
^ is introduced at the
vector h and the auxiliary signal vector nðu, yÞ, y^ ¼  fug þ  f/T  ng:
^ DðsÞ DðsÞ
output of the internal model PðsÞ, as illustrated in Fig. 2. In the
following the arguments of the signal n are omitted for space Thus, the matching error e~ can now be written as
reasons. The auxiliary signal vector n A Rm þ 1 þ n is defined as ^
( ) DðsÞ
bn  fug e~ ¼ yy^ ¼   f/T  ng: ð15Þ
1 DðsÞ
n¼  ^ , ð7Þ
^
DðsÞ an  fyg

with 3.3. Adaptive law

m m1 T
bn ¼ ½s ,s , . . . ,s,1
The question treated in this section is how to adapt the
and adaptation parameter vector h such that the matching error e~ goes
to zero asymptotically. To this end, the following adaptive law is
an ¼ ½sn1 ,sn2 , . . . ,s,1T ,
chosen
where the term A  fxg implies that the signal vector x is filtered
h_ ¼ C  n  e,
~ ð16Þ
with the linear filter A. The signal y^ can be written as
^ where C 4 0 is the user-defined positive definite adaptation gain
NðsÞ
y^ ¼  fug þ hT  n: ð8Þ matrix. The resulting adaptive control system is depicted in Fig. 3.
^
DðsÞ In Appendix A it is shown that if the signal r is bounded and if the
The adaptation parameter vector can be written as h ¼ ½hTu , hTy T ,
where the vectors hu A Rm þ 1 and hy A Rn are assigned to the input
^ respectively. Thus, the signal y^ is
signal u and the output signal y,
obtained as follows:
^
NðsÞ bn an
y^ ¼  fug þ hTu   fughTy  ^
 fyg,
^
DðsÞ ^
DðsÞ ^
DðsÞ
which can be rewritten as
^ þ hT  an Þ  fyg
ðDðsÞ ^ þ hT  bn Þ  fug:
^ ¼ ðNðsÞ ð9Þ
y u

The (unknown) matching adaptation parameter set h ¼ ½hT T
u , hy 
is defined as

hT ^ ^ ^ T
u ¼ ½bm b m ,bm1 b m1 , . . . ,b0 b 0  Fig. 3. Structure of the adaptive IMC.
876 D. Rupp, L. Guzzella / Control Engineering Practice 18 (2010) 873–881

transfer function
^
DðsÞ
Pe~ ðsÞ ¼ ð17Þ
DðsÞ
is strictly positive real (SPR), the adaptive system shown in Fig. 3
is globally stable and the error e~ converges to zero as t-1.
Furthermore, if n is a persistently exciting signal, then
lim J/J ¼ 0,
t-1

which implies that the coefficients of the internal model converge

to the ‘‘true’’ coefficients of the plant.

3.4. Analysis of the SPR condition

The transfer function Pe~ ðsÞ of the error dynamics as defined in Fig. 4. Maximal allowable relative errors of o0 and z for different values of z, such
(17) is the crucial part of the stability analysis of the adaptive that (23) remains SPR.
control system. Global stability can be guaranteed if Pe~ ðsÞ is SPR.
Since Pe~ ðsÞ contains the unknown denominator polynomial of the is obtained. This equation is useful to analyze the SPR robustness
plant PðsÞ, its strictly positive realness cannot be determined prior of a given setup of the control system.
to the adaptive control design. However, since the parameters of
the internal model PðsÞ ^ of the initial control design are typically Example. Given is the second-order plant
derived using an off-line system identification procedure they can NðsÞ
PðsÞ ¼
be assumed to represent the dynamics of the plant quite well. s2 þ 2zo0  s þ o20
From this assumption it follows that the transfer function Pe~ ðsÞ is
and its associated model
close to unity and thus SPR.
Based on the ideas presented in Karimi, Kunze, and Longchamp ^
NðsÞ
^ ¼
PðsÞ ,
(2007), a method was developed to determine a region for the s2 þ 2z^ o^ 0  s þ o^ 20
coefficients of the polynomial D(s) given the polynomial DðsÞ ^ such
that the transfer function Pe~ ðsÞ is SPR. To this end, consider the where z is the damping ratio, o0 is the natural frequency, and z^
transfer function and o ^ 0 are their corresponding estimates. The order of the
^
numerator polynomial NðsÞ is equal to the order of the numerator
^
DðsÞ ^
DðsÞ
Pe~ ðsÞ ¼ ¼ , ð18Þ polynomial N(s). With (17) the error model of this setup can be
DðsÞ ^
DðsÞ þ DðsÞ written as
where s2 þ 2z^ o ^ 20
^ 0  sþo
Pe~ ðsÞ ¼ : ð23Þ
DðsÞ ¼ dn1  a^ n1  sn1 þ    þ d1  a^ 1  s þ d0  a^ 0 ð19Þ s2 þ 2zo0  s þ o20

is an unknown polynomial, with the coefficients a^ 0 , a^ 1 , . . . , a^ n1 For this error model, Eq. (21) can be solved analytically. The areas
from (6) and the unknown tuning factors d0 ,d1 , . . . ,dn1 . By below the graphs in Fig. 4 represent the allowable relative error of
assumption, the transfer function in (18) has relative degree 0 and o0 and z, for different values of z, such that (23) remains SPR.
is analytic in Re½s Z0, where Re[s] stands for the real part of s, Clearly, the maximal allowable error is small for weakly damped
thus Pe~ ðsÞ is SPR if and only if (Ioannou & Tao, 1987) systems, whereas for sufficiently damped systems a large relative
" # error can be accepted. It is also worth mentioning that the graphs
^ oÞ
Dðj shown in Fig. 4 are independent of the (nominal) value of o0 .
Re 4 0, 8 o A ð1,1Þ:
^ oÞ þ DðjoÞ
Dðj
4. Adaptation under real conditions
In Kim and Meadows (1971) it is shown that if and only if a
transfer function is SPR then its inverse is SPR as well. Thus, in
For obvious reasons, the assumptions that the plant order is
order to guarantee strict positive realness the inequality
fully known and that no disturbance is acting on the system are
" #
DðjoÞ not valid when the adaptive controller is applied to a real system.
1 þRe 4 0, 8 o A ð1,1Þ ð20Þ The use of a low-order model results in the loss of the strictly
^ oÞ
Dðj
positive realness of the error dynamics Pe~ ðsÞ, which in turn leads
has to hold. With (19), expression (20) can be written as to the loss of the global stability of the adaptive control system.
" # However, the local stability results that are presented in this
a ^ ðjoÞT
1 þRe N  d 4 0, 8 o A ð1,1Þ, ð21Þ section prove to be sufficient in the case of AFR control, which can
^ oÞ
Dðj be seen as a representative application of internal model
with d ¼ ½dn1 , . . . d0 T and aN^ ðsÞ ¼ ½a^ n1  sn1 , . . . , a^ 1  s, a^ 0 T . In- controllers.
equality (21) can now be discretized at the frequencies oi of To deal with the difficulties evolving from neglected system
interest. As a result a linear matrix inequality (LMI) dynamics and disturbances, the following modified version of the
2 3 adaptive law (16) is used:
a ^ ðjo1 ÞT Z
6 N 7 g tk n  e~
6 Dðj
^ o1 Þ 7 htk ¼ htk1 þ dt, ð24Þ
6 7 T tk1 1 þ nT  n
6 ^ 7
IN þ Re6 7  d 40 ð22Þ
6 7 where htk is the adaptation parameter vector h at the time tk, g is
6 a ^ ðjoN ÞT 7
4 N 5 the adaptation gain, and T¼tk tk  1 is the adaptation interval,
^ oN Þ
Dðj
respectively. The modification incorporates the normalization and
 D. Rupp, L. Guzzella / Control Engineering Practice 18 (2010) 873–881 877

the averaging of the update step, which is often referred to as For reasons of symmetry, Eq. (30) applied to condition (27) then
hybrid adaptation (Elliott, 1982). yields
X
1
T
Re½Pe~ ðjoi Þ  Re½ni  n i  4 0: ð31Þ
4.1. Stability analysis i ¼ 1

This condition is always satisfied if Pe~ ðsÞ is SPR and if the matrix
With (11), (15), and (24) the adaptation parameter error vector T
ni n i is positive definite due to persistent excitation. However,
/ at time tk can be written as
expression (31) allows the formulation of a less conservative
Z
g tk n  Pe~  f/Ttk1  ng condition for stability. The SPR condition is relaxed to a positivity
/tk ¼ /tk1  dt: condition of the energy of the signal n weighted by the real parts
T tk1 1þ nT  n
of the transfer function Pe~ . For the specific case of the adaptation
Since /Ttk1 is constant, it can be taken out of the integral of the internal model, the condition (31) implies that the main
Z energy of the signals u and y^ should be in the frequency range
g tk n  Pe~  fnT g
/tk ¼ /tk1  dt  /tk1 , where the structure of the internal model P^ is flexible enough to
T tk1 1 þ nT  n
closely match the plant P(s). Given the quasi-stationary hybrid
which can now be rewritten as adaptation and a small adaptation gain g, the frequency content of
" Z # the adaptation signal hT  n is mainly governed by the signal n. The
g tk n  Pe~  fnT g reference signal r and the plant output y thus are the two main
/tk ¼ I dt  /tk1 : ð25Þ
T tk1 1 þ nT  n ‘‘sources’’ of frequencies in the control system. The frequency
content of the reference signal r is thus to be chosen by the user in
To analyze the stability properties of (25), the following lemma is such a way that the sum of its weighted energy enters positively
used: in (31) to account for the negative entries due to disturbances and
noise in the non-SPR frequency region. In the case of air/fuel ratio
Lemma 1. Consider the sequence of the vector x:
control, this condition is easy to meet since the air/fuel ratio can
xk ¼ x0 for k ¼ 0 be excited quite strongly without too much impact on the tailpipe
emissions. However, this method is suitable also to be applied to
xk ¼ ½Igk  F k   xk1 for k ¼ 1,2, . . . , ð26Þ any other practical problems of the same type.
where F k and gk are the matrix F and the adaptation gain g at step k,
respectively. If the matrix gk  F k is bounded and if the matrix 5. Application to air/fuel ratio control
Rk ¼ F k þ F Tk satisfies the inequality
Rk 4 0 8 k, ð27Þ The internal model of the IMC for the AFR control can be split
into two transfer functions, as shown in Fig. 5. The transfer
then (g 4 0 such that for all gk A ð0, g Þ xk -0 asymptotically as
 
function P^ e covers the dynamics and the delay time of the fuel
k-1.
path, which are assumed to be known, whereas P^ s represents the
An outline of the proof of this lemma is given in Appendix B. In first-order dynamics of the healthy lambda sensor with the time
the following, this lemma is used to weaken the SPR condition constant t^ s,0 :
under the assumption that the adaptation gain g is small. For the 1=t^ s,0
P^ s ¼ :
further analysis, the averaging analysis introduced by Anderson s þ 1=t^ s,0
et al. (1986) is adopted. Assuming that the signal vector nðtÞ is T-
According to Section 3.1 and by using the fact that the static gain
periodic, i.e.
of the adapted model is equal to one at all times, the signal x can
X
1
2pi be defined as
nðtÞ ¼ ni ejoi t with oi ¼ , ð28Þ
i ¼ 1
T 1=t^ s,0
x¼  fy^ e yg:
^ ð32Þ
s þ 1=t^ s,0
and where ni is the amplitude of the i-th harmonic. The signal
Pe~  fnT g can be written as By inserting (32) into (8), the signal y^ can be derived as
X
1 1=t^ s,0  ð1 þ yÞ
Pe~  fnT g ¼ Pe~ ðjoi Þ  fni ejoi t g: y^ ¼  fy^ e g,
s þ 1=t^ s,0  ð1 þ yÞ
i ¼ 1
where y is the adaptation parameter. Furthermore, the matching
For the evaluation of the positivity condition (27) the matrix
error can be written as
Z
1 T s þ 1=t^ s,0
n  P e~  fnT g e~ ¼   ff  ðy^ e yÞg,
^
T 0 s þ1=ts
is considered without the normalizing scalar 1 þ nT  n. Thus, with
(28) it can be written as
8 !T 9
Z < X =
1 T X 1
joi t
1
jo i t
ni e  Pe~ ðjoi Þ  ni e : ð29Þ
T 0 : ;
i ¼ 1 i ¼ 1

Considering the periodicity of the signals and the fact that for the
complex conjugate n i of ni it holds that x i ¼ ni , (29) can be
written as
X
1
T
Pe~ ðjoi Þ  ni  n i : ð30Þ
i ¼ 1 Fig. 5. AFR control system with an IMC and excitation at the input of the plant.
878 D. Rupp, L. Guzzella / Control Engineering Practice 18 (2010) 873–881

of being the output of the internal model y. ^ The system was

sampled at ts ¼10 ms. The excitation signal ue was chosen as a
square signal with an amplitude of 0.05 and a periodic time of 2 s.
The engine was operated at two different operating points: the
first is at a speed of n¼2250 rpm, with the air mass in each
cylinder of mc ¼0.22 g, which corresponds to a relative load of
about 40% (OP1), while the second is at a speed of n ¼1250 rpm,
with the air mass in each cylinder of mc ¼0.12 g, which
corresponds to a relative load of about 20% (OP2). The resulting
time constant and delay time of the fuel path are t ¼ 0:07 s and
d ¼ 0:08 s at OP1 and t ¼ 0:25 s and d ¼ 0:2 s at OP2, respectively.
Fig. 7 depicts the estimated sensor time constants as a result of
the adaptation of the internal model at OP1 for different initial
values ts,0 . The adaptation gains were chosen as g ¼ 50 for the
initial values ts,0 ¼ 0:1 s, g ¼ 100 for ts,0 ¼ 0:4 s, g ¼ 200 for
Fig. 6. Comparison of the first-order sensor model (grey) with the measurement ts,0 ¼ 0:7 s, and g ¼ 350 for ts,0 ¼ 1 s, respectively. Fig. 8 shows
data (dotted) at an engine speed of n¼ 2000 rpm and a relative load of 35%. For the the result of the adaptation at OP2. The adaptation gains were
excitation of the system the injected fuel mass was varied step-wise. chosen as g ¼ 100 for the initial value ts,0 ¼ 0:1 s, g ¼ 200 for
ts,0 ¼ 0:4 s, g ¼ 350 for ts,0 ¼ 0:7 s, and g ¼ 500 for ts,0 ¼ 1 s,

where f ¼ yy , with the matching adaptation parameter respectively. At both operating points the output error method
y ¼ 1=ts 1=t^ s . In this simple case the error dynamics presented in this paper results in a more accurate adaptation of
the internal sensor model. The adaptation at OP2 shown in Fig. 8
s þ 1=t^ s,0
Pe~ ¼ results in a larger variance of the converged values of the
s þ1=ts
estimated sensor time constant. The reasons for this variance
are SPR for every positive time constant t^ s,0 . being larger were discussed in the previous section.
Certainly, the quality of the engine model Pe(s) has an influence
on the result of the adaptation. Typically, for operating points
with low air mass flows and low engine speeds, the portion of
unmodeled dynamics and disturbances in the control-relevant
frequency range increases. This leads to an increased uncertainty
of the adapted parameter as the measurement results in the
following section show.
The first-order model (2) yields a very good approximation of
the dynamics of the sensor. Fig. 6 depicts the comparison of the
output of the model and the measurement result of an artificially
slowed-down lambda sensor when the injected fuel mass is
varied. For the identification of the model parameter a fast
lambda sensor was mounted right next to the slow sensor. By
assuming that the dynamics of the fast sensor can be neglected,
its signal can be used as the input signal for the identification.
During the adaptation, the control system is excited by the signal
ue at the input of the plant, as depicted in Fig. 6, rather than by the
reference signal r because that signal is more compatible with
Fig. 7. Estimated sensor time constant during adaptation of the internal model for
existing diagnosis functions in the production-type electronic
the output error (black) and the equation error method (grey), respectively, at OP1.
engine control unit. The dashed line indicates the time constant of the sensor resulting from open-loop
identification.

6. Experiments

The measurements were performed on an engine test bench at

the Institute for Dynamic Systems and Control at ETH Zurich,
Switzerland. The engine was a production type SI, 5-cylinder, 2.5l
device with port fuel injection. In order to test the algorithm for
different time constants, instead of using the slow lambda sensor,
the AFR in the exhaust pipe was measured with a fast lambda
sensor and then filtered with a model of a slow sensor before the
signal was fed into the controller. The effective time constant that
results from the adaptation is dependent on the quality of the
given model parameters t and d of the fuel path. Therefore, in all
experiments the time constant of the sensor was determined by
open-loop identification to get a reference value for the adaptive
algorithm. The output error method presented in this paper was
compared with the equation error method presented in Datta and
Fig. 8. Estimated sensor time constant during adaptation of the internal model for
Ochoa (1996). The difference between the output error method the output error (black) and the equation error method (grey), respectively, at OP2.
and the equation error method is solely that in the latter the The dashed line indicates the time constant of the sensor resulting from open-loop
auxiliary signal vector n is a function the measurement y instead identification.
 D. Rupp, L. Guzzella / Control Engineering Practice 18 (2010) 873–881
Fig. 9 shows the error signal e~ defined as the difference
between the measured lambda and the output of the internal
model during adaptation at OP2, with the initial value ts,0 ¼ 0:1 s.
The amplitudes of the error obtained using the equation error
method are smaller than the ones of the output error method
during the entire experiment. This is due to the fact that in the
equation error method the disturbed lambda signal is directly
used to calculate the model output and thus a part of the
disturbances cannot be seen in the error signal e. ~ The IMC
concept, on the other hand, is based on the idea that the signal e~ is
an estimate of the disturbances, which implies that the output
error method is more suitable to be used as an internal model of
an IMC.
At both operating points, the adaptation algorithms were also
tested under the scenario that the sensor time constant is
ts ¼ 0:1 s and that the algorithms are initiated with t^ s,0 ¼ 0:5 s.
The adaptation gain was chosen as g ¼ 500 for OP1 and g ¼ 300 for
OP2, respectively. The reduced sensor time constant decreases the
attenuation of disturbances that pass the sensor and thus affect
the adaptation. To illustrate this Fig. 10 shows the measured
lambda signal and the output of the internal model during the
first and the last 10 s of the adaptation at OP2. The comparison of
the output and equation errors is shown in Fig. 11. The equation
error method results in a considerable bias at OP2 due to the large
disturbances at this operating point. The effect of small static
disturbances of 2% that were introduced at the input of the plant
was also investigated. The method presented in this paper results
Fig. 9. Adaptation error e~ of the output error (black) and the equation error
method (grey), respectively, at OP2 with the initial value ts,0 ¼ 0:1 s.
Fig. 10. Measured lambda (black) and output of the internal model (grey) during
the first and the last 10 s of the adaptation at OP2 with the output error method as
shown in Fig. 11. The dashed line indicates the excitation signal ue at the input of
the plant.
879
Fig. 11. Estimated sensor time constant during adaptation of the internal model
using the output error (black) and the equation error method (grey), respectively,
at both operating points. The dashed line shows the result of the adaptation with a
static disturbance of 2% at the input of the plant. The straight dashed line indicates
the time constant of the sensor resulting from open-loop identification.
Table 1
Normalized sums of the squared control error signal during the (warm) start phase
of the FTP-75 cycle with a sensor time constant of 0.5 s in three different scenarios
(second to fourth columns) and with a production-type sensor (last column).
t^ s ¼ 0:05 s t^ s ¼ 0:5 s t^ s ¼ 0:5 s and inv. Production type
P 1 0.75 0.63 0.58
Dl2
in a performance that is similar to that obtained without a static
disturbance. The estimate of the equation error method, however,
did not converge due to the strong correlation of the static
disturbance appearing in both the output y^ of the internal model
and the measured output y. While robustness measures such as
an error dead zone can prevent this behavior, they cannot
improve the precision of the adaptation.
Table 1 clearly shows the increase in performance that can be
achieved with an adapted controller. It shows the sum of the
squared control error signal during the (warm) start phase of the
FTP-75 cycles for a sensor time constant of 0.5 s in three different
controller scenarios: no adaptation of the internal model
(t^ s ¼ 0:05 s), adapted internal model (t^ s ¼ 0:5 s), and adapted
internal model inversion of the sensor dynamics in the controller
Q(s). For first, the second, and the fourth columns the internal
controller Q(s) is chosen as
t  sþ1
Q ðsÞ ¼  ,
d  s þ1
where t is the time constant and d is the delay time of the engine
model Pe(s), while for the third column
ðt  s þ 1Þðt^ s  s þ 1Þ
Q ðsÞ ¼  ,
ðd  s þ 1Þ2
where t^ s is the adapted time constant of the sensor. The last entry
in Table 1 shows the result achieved with a production-type
sensor (Bosch LSU 4.2). In order to obtain an emission-relevant
result in all four scenarios, the calculation of the square sum of the
lambda deviations is based on the signal of a production-type
sensor rather than on the signal of the slow sensor that is fed into
the controller. Thus, under stationary conditions, the internal
model can be adapted efficiently with the method introduced
in this section. However, in a passenger car, the adaptation has to
be performed under highly transient conditions such as they
occur in normal driving situations. During engine transients, the
disturbances that act on the AFR control loop due to the imprecise
880 D. Rupp, L. Guzzella / Control Engineering Practice 18 (2010) 873–881

operating points and for various values of the sensor time

constant as well as the initial value of the model time constant.
The adaptive IMC was able to improve the performance of the
control system even in the presence of large disturbances and
uncertainties in the model describing the dynamics in front of the
sensor.

Appendix A. Global stability of the adaptive system in Fig. 3

The dynamics of the adaptation error e~ in (17) can be written

in state-space form as

x_ ¼ Axþ b½/T  n,

Fig. 12. Adaptation of time constant during the FTP-75 driving cycle. Three e~ ¼ cx þ d½/T  n ð33Þ
different initial values are shown in the case of a sensor time constant of ts ¼ 0:5 s.
with the system matrices fA,b,c,dg. Since Pe~ is SPR, it follows from
Lemma 4 in Meyer (1965) that there exist the symmetric positive-
feedforward compensation of the AFR have a detrimental effect on
definite matrices P and Q , a vector q, and a scalar e 4 0 such that
the performance of the adaptive algorithm. It is thus important to
ensure that the adaptation is active only when the engine is AT P þPA ¼ qqT eQ
operating near stationary conditions. To that end, a heuristic
pffiffiffi
decision logic has been developed that rejects the parameter Pbc ¼ dq: ð34Þ
update of the hybrid adaptive law in (24) whenever an
unacceptable engine transient has occurred since the last Note that the standard Kalman–Yacubovich Lemma cannot be
update step. The main ‘‘sources’’ of potential disturbances of the applied since the error dynamics transfer function Pe~ is not
feedback control systems are the movements of the throttle. Thus, strictly proper, and due to the potential zero-pole cancelations the
in order to obtain a measure for the movements of the throttle, pair ðA,bÞ cannot be guaranteed to be fully controllable. Now
the integration of the sum of squares of the deviations from the consider the Lyapunov function candidate
throttle position a at the beginning of the integration in (24) is
Vðx, /Þ ¼ xT Pxþ /T C1 /
calculated
Z with Vðx, /Þ 4 0, 8x, / a 0. The derivative of Vðx, /Þ with respect to
1 tk
Ja ¼ ðaðtk1 ÞaðtÞÞ2 dt: the time t can be written as
T tk1
T _ T C1 / þ /T C1 /
V_ ðx, /Þ ¼ x_ Px þ xT P x_ þ / _: ð35Þ
The update of the adaptation is then rejected if the measure Ja is
larger than a certain predefined threshold value. Furthermore, for With the adaptive law (16) and the error dynamics (33), and after
the reasons discussed in Section 5, the update of the adaptation is some straightforward algebraic manipulations, (35) can be
also rejected if the engine is idling. Fig. 12 depicts the adaptation written as
during the hot-start phase of an FTP-75 driving cycle for various
initial values. At certain points of the adaptation, the value of the V_ ðx, /Þ ¼ xT ðAT P þ PAÞx þ2xT Pb½/T  n2e½
~ /T  n,
estimated time constant t^ s remains constant for some time due to and thus with (34)
the decision logic. pffiffiffiffiffiffi
V_ ðx, /Þ ¼ xT ½qqT þ eQ x2d½/T  n2 þ 2 2dxT q½/T  n,
which can be written as
7. Conclusions
V_ ðx, /Þ ¼ exT QxvT v, ð36Þ
The problem of sensor ageing in AFR control systems is pffiffiffiffiffiffi T
with v ¼ ½x q 2d½/  n and thus V_ ðx, /Þ o 0, 8 x, / a 0, which
T
addressed by presenting a method to adapt the internal model of
proves the global stability of the adaptive system. The stability
an IMC during closed-loop operation. Inspired by the IMC
implies that e~ and / are globally bounded. Thus, since r is
principle, an output error model structure was used as the basis
assumed to be bounded and the transfer function Q(s) is stable
of the adaptive algorithm. The analysis of the theoretical stability ^ ^ ^
and proper, and the transfer functions PðsÞ, an =DðsÞ and bn =DðsÞ
properties can be reduced to the analysis of the dynamics Pe~ of the
from (7) are stable and strictly proper, it follows that the signals u
adaptation error e. ~ This has been investigated in a general
and y^ and thus n and n_ are also bounded. From the boundedness
framework and then studied in detail on a general second-order
of fng and / it follows that x_ in (33) is also bounded. Thus the
example to demonstrate the applicability of the method to other
second derivative of Vðx, /Þ with respect to time
industrial processes. The strictly positive realness of Pe~ is a
T T
condition for global stability of the adaptive system. However, V€ ðx, /Þ ¼ ex_ QxexT Q x2
_ v_ v,
assuming that a reasonable initial estimate of the internal model pffiffiffiffiffiffi pffiffiffiffiffiffi
^ with v_ ¼ ½x_ T q þ 2de~ nT CT n 2d½/T n_  is bounded and thus
PðsÞ of the IMC is known, the SPR requirement can be replaced by _
V ðx, /Þ is uniformly continuous. Thus according to Barbalat’s
more practical conditions on the frequency content of the
lemma it holds that
reference signal r. The adaptive controller has been applied to
an AFR control system to adapt the time constant of an aged lim V_ ðx, /Þ ¼ 0
t-1
oxygen sensor. The disturbances due to the cyclic operation of the
engine considerably depend on the operating point of the engine and thus from (36) it follows that
and the actual time constant of the oxygen sensor. The experi-
lim x ¼ 0
ments on the test bench have thus been performed at several t-1
 D. Rupp, L. Guzzella / Control Engineering Practice 18 (2010) 873–881 881

and exists and yields an upper bound gk 40 of gk 8k. Since F k is

bounded, there exists a time-invariant upper bound V0 ðxÞ with
lim ½/T n ¼ 0 ð37Þ
t-1 V0(0)¼ 0 such that Vk rV0 8k. Furthermore, since limJxk J-1 Vk
ðxk Þ ¼ 1, it follows that xk -0 asymptotically as k-1.
and thus with (33)
~ ¼ 0:
lim eðtÞ
t-1 References
Now, introducing / as the converged value of / such that
Alfieri, E., Amstutz, A., & Guzzella, L. (2009). Gain-scheduled model-based feedback
T control of the air/fuel ratio in diesel engines. Control Engineering Practice, 17,
/ n¼0 ð38Þ 1417–1425.
Anderson, B. D.O., Bitmead, B. R., Johnson, C. R., Jr., Kokotovic, P., Kosut, R. L., &
and by multiplying (38) by the vector n and integrating over the
Mareels, I. M.Y., et al. (1986). Stability of adaptive systems passivity and
period T it follows that averaging analysis. Cambridge, MA: The MIT Press.
Z t þ T  Balenovic, M., Backx, A., & Hoebihk, J. (2001). On a model-based control of a three-
n  nT dt  / ¼ 0: way catalytic converter. In SAE World Congress No. 2001-01-0937.
t Datta, A., & Ochoa, J. (1996). Adaptive internal model control: Design and stability
analysis. Automatica, 32, 261–266.
Thus, perfect model matching, i.e. / ¼ 0, can only be achieved if Elliott, H. (1982). Hybrid adaptive control of continuous time systems. IEEE
(T 4 0 such that the matrix Transactions on Automatic Control, 27(2), 419–426.
Guzzella, L., & Onder, C. (2004). Introduction to modeling and control of internal
Z tþT combustion engine systems. Berlin, Heidelberg: Springer-Verlag.
n  nT dt Guzzella, L., Simons, M., & Geering, H. P. (1997). Feedback linearizing air/fuel-ratio
t controller. Control Engineering Practice, 5, 1101–1105.
Inagaki, H., Ohata, A., & Inoue, T. (1990). An adaptive fuel injection control with
is positive definite, which can be seen as a condition for the internal model in automotive engines. IEEE Proceedings, 78–83.
persistent excitation of the signal n. Ioannou, P., & Tao, G. (1987). Frequency domain conditions for strictly positive real
This completes the proof. functions. IEEE Transactions on Automatic Control, 1, 53–54.
Jones, V. K., Ault, B. A., Franklin, G. F., & Powell, D. (1995). Identification and air–
fuel ratio control of a park ignition engine. IEEE Transactions on Control Systems
Technology, 3(1), 14–21.
Appendix B. Outline of the proof of Lemma 1 Karimi, A., Kunze, M., & Longchamp, R. (2007). Robust controller design by linear
programming with application to a double-axis positioning system. Control
Engineering Practice, 15, 197–208.
Consider the Lyapunov function candidate Kim, W. H., & Meadows, H. E. (1971). Modern network analysis. New York: Wiley.
Meyer, K. R. (1965). On the existence of Lyapunov functions for the problem of
1 T
Vk ¼ x x Z0: Lur’e. SIAM Journal on Control, 3, 373–383.
2 k k Morari, M., & Zafiriou, E. (1989). Robust process control. Englewood Cliffs, NJ:
Prentice-Hall.
With (26)
Muske, K. R., Jones, J. C.P., & Franceschi, E. M. (2008). Adaptive analytical model-
1 T 1 based control for SI engine air fuel ratio. IEEE Transactions on Control Systems
DVk ¼ Vk þ 1 Vk ¼ x x  xT x ð39Þ Technology, 16, 763–768.
2 kþ1 kþ1 2 k k Onder, C. H., Roduner, C. A., & Geering, H. P. (1997). Model identification for the A/F
can be written as path of an SI engine. SAE SP-1736, electronic engine controls (pp. 65–73).
Roduner, C. A., Onder, C. H., & Geering, H. P. (1997). Automated design of an air/
1 T 1 g fuel controller for an SI engine considering the three-way catalytic converter in
DVk ¼ x ½IgF k T ½IgF k xk  xTk xk ¼ xTk ½gF Tk F k ðF k þF Tk Þxk : the h1 approach. In 5th IEEE Mediterranean conference on control and systems.
2 k 2 2
Rupp, D., & Guzzella, L. (2009). Iterative tuning of internal model controllers with
Since F k þ F Tk is positive definite, it follows that F k and F Tk F k are application to air/fuel ratio control. IEEE Transactions on Control Systems
Technology, 18, 177–184.
positive definite as well, thus the inverse of ðF Tk F k Þ exists and thus,
Stroh, D. J., Franchek, M. A., & Kerns, J. M. (2001). Fueling control of spark ignition
a solution of the inequality engines. Vehicle System Dynamics, 36, 329–358.
Turin, R. C., & Geering, H. P. (1995). Model-reference adaptive A/F-ratio control in
gk I oðF k þF Tk ÞðF Tk F k Þ1 an SI engine based on Kalman filtering techniques. In Proceedings of the 1995
American control conference (pp. 4082–4090).
in the sense of Yildiz, Y., Annaswamy, A., Yanakiev, D., & Kolmanovsky, I. (2008). Adaptive air fuel
ratio control for internal combustion engines. In Proceedings of the American
gk IðF k þF Tk ÞðF Tk F k Þ1 o0 control conference (pp. 2058–2063).