Anda di halaman 1dari 13

636 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO.

3, MAY 2013

Air-to-Fuel Ratio Switching Frequency Control for


Gasoline Engines
Jason Meyer, Stephen Yurkovich, Fellow, IEEE, and Shawn Midlam-Mohler

Abstract—Modern gasoline internal combustion engines use a even the best pre-catalyst AFR controllers cannot eliminate
variety of technologies to enhance the efficiency of fresh air induc- all deviations from stoichiometry. While trying to regulate
tion. These technologies, which include variable valve timing and to stoichiometry, most pre-catalyst AFR controllers produce
variable intake geometry systems, also make it more difficult to
predict the mass of fresh air that is trapped during the induction uncontrolled limit-cycle type behavior (e.g., chattering). Oscil-
stroke of the engine because they not only affect the residual gas lating the AFR around stoichiometry extends the conversion
fraction of the trapped air charge, but also the wave dynamics of efficiency range of a TWC [1], [2]. However, as shown in [3],
the system. As the number of controllable actuators increases, this the amplitude and frequency of these oscillations strongly affect
estimation problem becomes even more difficult. As these technolo- the conversion efficiency; therefore, uncontrolled oscillations
gies continue to develop, the importance of robustness in air-to-
fuel ratio control continues to grow. This paper presents an air-to- due to limit cycle behavior should be avoided.
fuel ratio control algorithm based on a switching frequency regu- Two types of oxygen sensors, binary exhaust gas oxygen
lator that has favorable robust stability properties in the presence (EGO) sensors and universal exhaust gas oxygen (UEGO)
of both input and model errors. Instead of modeling the air path sensors, can be used to provide a feedback measurement of
system with a simplified model, this control architecture considers the composition of the combustion mixture. A UEGO sensor
the air estimate as a control input. As a result, air estimation er-
rors behave like input errors, not modeling errors. By using the indicates the AFR of the combustion mixture, whereas a binary
rich-to-lean and lean-to-rich air-to-fuel ratio switching frequencies EGO sensor only indicates whether the mixture is rich or lean.
of the pre-catalyst exhaust gas oxygen sensor as the primary feed- For a narrow range of AFR values near stoichiometry, however,
back signal, the control laws are completely independent of the pa- a binary EGO sensor does provide magnitude information.
rameters of the plant model. The performance of this controller is Within this range, a binary EGO has a much higher sensitivity
demonstrated both with a robust stability analysis and through a
vehicle-based experimental validation. to AFR than a counterpart UEGO sensor. As long as an accurate
AFR estimator can be designed, an EGO-based control scheme
Index Terms—Air-to-fuel ratio (AFR), delay systems, emissions,
can reject disturbances as well as a UEGO based control
engine control, robustness, stability.
scheme. Given that the desired AFR target is stoichiometry,
the higher sensitivity provided by a binary EGO sensor enables
I. INTRODUCTION more precise AFR control during steady-state operation. For
this reason, many of the gasoline powered automobiles in
production today still use low cost binary EGO sensors.
S INCE the advent of the three-way catalyst (TWC), most
air-to-fuel ratio (AFR) control architectures consist of a
primary feedback loop based on a pre-catalyst AFR measure-
Most advanced control formulations such as controllers
[4], Lyapunov-based controllers [5], model-predictive con-
trollers [6] and linear parameter varying controllers [7], [8]
ment and a secondary feedback loop based on a post-catalyst
require UEGO sensors, and are developed using overly simpli-
AFR measurement. Including this study, the emphasis of most
fied plant models. The Lyapunov-based control law developed
AFR control research is pre-catalyst AFR control because
in [5] for example is based on a simplified filling and emptying
most of the control authority resides within this control loop.
model along with basic curve fitting approximations. Similarly,
It has been well documented that a TWC has the highest
[4] uses a simple discrete state-space model and [7] uses a first
steady-state conversion efficiencies when the incoming exhaust
order lag model. Most of these models do not account for the
gases are produced from stoichiometric combustion. If the
time varying transport delays of the exhaust system. This type
pre-catalyst AFR could be regulated perfectly to stoichiom-
of unmodelled dynamics leads to uncontrolled oscillations. Not
etry, then tail-pipe emissions would be minimized. However,
only are the performances of these controllers limited by these
potentially significant modeling errors, but stability results
Manuscript received November 10, 2011; accepted January 24, 2012. Manu-
proven for an ideal system may become invalid.
script received in final form February 15, 2012. Date of publication March 21,
2012; date of current version nulldate. Recommended by Associate Editor U. Sliding mode controllers possess some inherent robustness
Christen. characteristics, and several formulations such as [9]–[11] are de-
J. Meyer and S. Midlam-Mohler are with the Department of Mechanical
signed based on binary EGO sensor feedback. However, these
Engineering, Center for Automotive Research, The Ohio State University,
Columbus, OH 43212 USA (e-mail: meyer.506@osu.edu). formulations tend to rely on empirical models. Even with a
S. Yurkovich is with the Department of Systems Engineering, University of very accurate empirical engine model, sliding mode controllers
Texas at Dallas, Richardson, TX 75080 USA.
struggle to precisely regulate the AFR during steady state. Chat-
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. tering is an inherent problem common to all sliding mode con-
Digital Object Identifier 10.1109/TCST.2012.2188631 trollers. For systems which are characterized by time varying

1063-6536/$31.00 © 2012 IEEE


MEYER et al.: AIR-TO-FUEL RATIO SWITCHING FREQUENCY CONTROL FOR GASOLINE ENGINES 637

delays such as internal combustion engines, chattering is even note that this type of control is only effective when the plant
more difficult to overcome. As a result, sliding mode based AFR delay is predicted exactly. In direct contrast to the control ob-
controllers are able to reject disturbances well but have diffi- jective, phasing errors brought on by delay misestimation cause
culty precisely regulating the EGO switching frequency once uncontrolled limit cycle behavior. As shown in [17], the plant
stoichiometry is reached. delay varies markedly and is very difficult to predict, especially
Because current control formulations are not able to precisely during transients.
regulate the AFR trajectory to stoichiometry, the average AFR By recasting the problem into a EGO switching period con-
may be slightly rich or slightly lean even when the controller is trol problem, it is possible to track a desired AFR trajectory in
meeting its control objective. This result is further compounded terms of amplitude, frequency and offset. In such a scheme, the
by sensor errors and sensor dynamics. The composition of measured AFR is not directly compared to a predicted AFR,
the exhaust stream can cause oxygen sensor deception due to so the phasing and delay estimation issues are eliminated. The
varying diffusion rates of the gas constituents, thus indicating switching period based AFR controller proposed in this paper
an AFR or switching point different from the actual AFR makes use of the fact that the average AFR (or equivalently
incident on the sensor [12]–[14]. Although these errors can be the AFR offset error) can be estimated by observing the rate by
mitigated with a secondary AFR control loop, they still have a which a pre-catalyst EGO sensor switches from rich to lean and
negative impact on the tailpipe-out emissions. lean to rich. This discrete time controller operates in the engine
In this paper, a unified EGO-based AFR control architecture event domain. This domain is a subset of the crank angle do-
built on a switching frequency regulation design is presented. main where the number of events per engine cycle is equal to the
This control design is based on the controller in [15] which number of cylinders. For a four stroke engine with cylin-
was shown to outperform similar designs such as the phase-lock ders, the number of crank angle degrees per event is .
loop controller developed in [16]. Instead of trying to directly Consider the control diagram of an AFR switching period con-
regulate the AFR to a constant value of stoichiometry, an AFR troller depicted in Fig. 1. Beginning from the top center of this
switching period control structure accurately tracks a periodic diagram and moving clockwise, the control system consists of
waveform that has a desired amplitude, frequency, and offset. a plant model, an EGO sensor switching detector, a coordinated
By actively controlling the AFR fluctuations around stoichiom- set of closed loop compensators and finally a set of feed-forward
etry, which are left uncontrolled in regulation controllers, the contributions.
stored oxygen mass within the TWC can be directly maintained The plant model upon which this controller is developed in-
at its optimal level, thus maximizing the catalyst conversion ef- cludes both the physical engine and a fuel dynamics compen-
ficiency. This switching frequency based AFR controller detects sator. Define the fuel/air equivalence ratio (EQR) of the com-
disturbances by comparing the duration of time between output bustion mixture as
switches to some desired value. Even though magnitude infor-
mation is not directly provided by a binary EGO sensor, the (1)
average AFR can be inferred from the switching period. Be-
cause the feedback signal is actually the switching period, not where is the trapped fuel mass, is the trapped air
the switching frequency, this controller will henceforth be re- mass and is the stoichiometric AFR. The input to the fuel
ferred to as an AFR switching period controller. dynamics compensator is the desired EQR corrected to over-
One of the key benefits of this control architecture is its ro- come air and fuel estimation errors. Using this input, the fuel
bustness to disturbances and modeling errors. Instead of mod- dynamics compensator calculates a desired fuel mass according
eling the air system with a low order control-oriented model, to
the control system considers the trapped air mass estimate as
an input. Since production vehicles already include algorithms (2)
to accurately estimate the trapped air mass, many of the typ-
ical modeling errors that affect other control formulations can where is the desired trapped fuel mass, is the esti-
be eliminated without any additional calibration effort. As an mated trapped air mass and is the time index in engine events.
added benefit, air estimation errors affect the system as input Following this, the fuel dynamics compensator adjusts the in-
disturbances, not as modeling errors which are more difficult jected fuel mass to ensure that the trapped fuel mass is indeed the
to compensate. With this control architecture, the time varying desired fuel mass. A fuel dynamics compensator can be gener-
plant delay is accounted for without the need for an estimator, ated by inverting one of the well known wall wetting dynamics
rendering its performance invariant to modeling errors. models such as [18] or [19].
With the inclusion of the fuel dynamics compensator, is
II. SWITCHING PERIOD CONTROLLER treated as the input to the plant model. The output of the plant
If the plant delay is known perfectly and a UEGO sensor is is a binary signal produced by the pre-catalyst EGO sensor
used, then the AFR control problem posed in this paper could be which indicates the relative richness/leanness of the exhaust
formulated into a tracking control problem. In this ideal case, the stream. The switching detector monitors the pre-catalyst EGO
measured AFR could be compared to a predicted AFR to gen- measurement signal to produce an output which repre-
erate an error signal. Many well known techniques such as the sents the number of events since the last EGO switch and an
internal model principle could be used to develop a controller output which identifies the occurrence of an EGO switch.
that asymptotically drives this error to zero. It is important to The output takes a null value of zero when an EGO
638 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 3, MAY 2013

Fig. 1. AFR switching period control diagram.

switch is not detected and indicates the achieved switching pe- switching occurrence. The late switching compensator con-
riod when an EGO switch is detected according to tinues to make corrections until the output switches. The error
signal fed to the late switching compensator is governed by
if EGO switches
(3)
otherwise if
(5)
otherwise
where is the event number of the previous output switch.
Each output of the switching detector is compared to the desired where is the size of the deadzone region.
switching period to produce a feedback error signal. The desired Conversely, the early switching compensator makes a one-time
switching period is one half times the period of the dither signal. integral correction whenever the output switches earlier than
Although any periodic signal can be tracked by switching period the desired switching period. The error signal fed to the early
control, this paper focuses on sinusoidal signals. When the pe- switching compensator satisfies
riodic dither waveform is a sinusoid, it can be described by if
otherwise.
(6)
(4) With these definitions, the switching period control laws can
now be defined.
where is the reference dither waveform in units of EQR, By developing the AFR switching period controller in the
is the dither amplitude in units of EQR and is the dither EQR domain, the complex air and fuel dynamics can be pulled
period in units of engine events. outside of the control model. Obviously, the accuracies to which
The dither period and amplitude are two of the most impor- the trapped fuel and air mass can be estimated strongly influ-
tant design parameters. Their numerical values may help to pro- ence the performance of any AFR controller. With this structure,
vide some context. The experimental validation of this control however, the error in predicting these quantities are viewed as
architecture was performed on a 2006 Saturn Ion powered by a additive input disturbances. The corrected EQR signal fed to the
2.4 liter, four-cylinder gasoline engine with variable intake and fuel dynamics compensator can be described by
exhaust valve timing. For this engine, the dither period and am-
plitude were chosen to be 25 events and 0.0115, respectively. (7)
The rationale used to select these values is based on the stability where is the desired EQR offset, is the contribution from
properties of the control system and is discussed in Section III. the switching period controller and is a lumped disturbance
Closed loop corrections are made when the EGO sensor variable which represents the combined effect of air estimation
output switches either slower than desired via the late switching errors, fuel estimation errors, sensor errors and unknown con-
compensator or faster than desired via the early switching com- tributions from the evaporative canister system.
pensator. To allow for a small range of switching periods to For simplicity, the feedback control law defines the state as
produce zero control error, a deadzone is applied to the desired the difference between and to arrive at
switching period signals. The late switching compensator
makes both proportional and integral corrections if more than
the desired number of events have passed since the last output (8)
MEYER et al.: AIR-TO-FUEL RATIO SWITCHING FREQUENCY CONTROL FOR GASOLINE ENGINES 639

The relative composition of the exhaust stream as measured by


the pre-catalyst EGO sensor can be defined as

(9)

where is the largest possible plant delay and is the th co-


efficient of the finite impulse response model of the system, both
of which depend on the trapped air mass per cylinder and
the engine speed . This relationship follows from [20] and
accounts for the transport lag of the exhaust system, the mixing
phenomena and the sensor dynamics. It should be noted that this Fig. 2. Sine wave with parameters of interest highlighted.
approximation is only used to prove the stability of the control
system; the performance of the AFR control does not depend
on the accuracy of this model. The stability analysis presented Provided , it directly follows that
in this paper only depends upon the fact that the plant delay is
bounded by a maximum value of events. (15)
The total control action taken by the AFR switching period
controller can be represented with the following: or equivalently

(10) (16)
where is the proportional contribution from the late switching
compensator, is the integral contribution from the late when . Built on this last relationship, the
switching compensator, and is the integral contribution from integral action of the early switching compensator is governed
the early switching compensator. The proportional correction by
of the late switching compensator is defined as
(17)
(11)
where is the gain of the early switching compensator and
satisfies
and the integral correction is defined as

(12)
(18)
where is the proportional gain of the late switching com- where is a calibration parameter.
pensator and is the integral gain of the late switching com-
pensator. Both and are calibration parameters which are III. STABILITY AND CONTROL GAIN SELECTION
smooth functions of the operating conditions (manifold pressure
The control laws within an AFR switching period controller
and engine speed).
are all represented in the EQR domain allowing air estimation
One of the most unique characteristics of an AFR switching
errors and fueling errors, including those resulting from the fuel
period control structure is the use of trigonometry within the
dynamics compensator, to be treated as a single additive EQR
early switching compensator to relate the switching period to
disturbance. In this domain, the engine system behaves like a
the average EQR offset. Let be the average EQR offset
single input (EQR of the combustion mixture), single output
error defined by
(relative richness/leanness of the exhaust gas mixture at the
EGO sensor) system. A rigorous analysis of the closed-loop
(13) behavior of the combustion mixture EQR trajectory has been
performed. The fundamental results of this analysis are now re-
ported in the form of three theorems. The corresponding proofs
In the presence of a constant offset error that has a magnitude
to these theorems are provided in Appendices A–C. In what
less than the dither amplitude, the number of events between
follows, define as the minimum value of the plant delay
EGO switches for the open loop system can be directly related
and as the maximum value of the plant delay. Also let
to the average EQR offset error using trigonometry. Consider
represent a special set defined by
the height parameter and length parameter depicted in Fig. 2
for the general sinusoid function . For all values (19)
of , the height can be related to the length using
the trigonometric identity which satisfies the invariant property that
(14) (20)
640 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 3, MAY 2013

and has the additional property that all trajectories within ensure steady-state tracking, the dither period must satisfy (27).
cross the origin (which physically corresponds to The best performance, however, is achieved when the dither pe-
the switching of the EGO voltage) at least every events. riod is as close to this bound as possible. Having a small but
Theorem 3.1: For all disturbance trajectories which satisfy finite value for ensures that the system converges to the de-
sired dither signal in finite time, rendering the controller more
(21) practical. Even in the presence of low frequency disturbances,
it is still possible for the system to reach this set. Since the size
and of also dictates the size of the dither maximum offset error,
it should be chosen as small as possible. For the four cylinder
(22)
engine used to validate this control structure, the largest plant
, there exists a scalar and an integer delay is 11 events. Based on these considerations, was chosen
such that the closed-loop trajectory of an AFR switching period to be 0.5 events and was selected to be 25 events.
control system remains bounded and converges to the special Selection of the control gain represents a trade off between
invariant set in finite time. robustness and steady-state performance. To reject disturbances
Theorem 3.2: For any constant disturbance, there exist a as fast as possible, the value of should be maximized. To re-
scalar produce the sinusoidal dither in the presence of a constant distur-
bance, the control gain must satisfy the upper bound described
(23) in (26). It is important to note that can be selected as a func-
tion of the operating conditions and the bounds described in (26)
and a corresponding scalar need not be satisfied at every operating condition. For the oper-
ating regions where (26) is satisfied, the system will converge
(24) to the desired dither waveform. For the operating regions where
(26) is not satisfied, the desired waveform may not be repro-
such that the closed-loop trajectory of an AFR switching period duced; however, Theorem 3.2 guarantees that the EQR will still
control system remains bounded and converges to the invariant remain close to stoichiometry. This ensures that reasonable per-
set in finite time. formance would be achieved if the engine were operated in such
Theorem 3.3: For any constant disturbance, the closed-loop a region for an extended period of time.
trajectory of an AFR switching period control system converges Based on the observed performance of the production AFR
to a sinusoidal signal with a frequency and amplitude equal to controller, the design target for disturbance rejection was the
the frequency and amplitude of the imposed periodic signal with removal of a 0.10 EQR error within 150 events. The value of
an offset error (distance to the EQR setpoint) that is bounded required to achieve this performance can be
by identified by dividing the two quantities. Assuming this value
for , the dither amplitude was selected to be 0.0125 to satisfy
(25) (26). With all of the key dither parameters identified, the value
of as well as the gains scheduled values of and were cal-
provided that ibrated in simulation. The calibration procedures are outlined in
[15]. The optimal value of was found to be 0.477. The cali-
(26) bration surfaces for and are shown in Fig. 3 along with
a summary of the key design parameters. Within respect to the
(27) surface plot, all of the operating conditions below the white
line meet the requirements described in Theorem 3.3. Because
and
steady-state operation is unlikely to occur in the operating re-
(28) gions above this line, disturbance rejection performance is more
important and thus larger values of were selected.
The combination of these three theorems is quite powerful.
Since Theorem 3.1 ensures that for any bounded disturbance
the closed-loop system trajectory remains bounded, the stability IV. EXPERIMENTAL VALIDATION
properties in Theorem 3.2 and Theorem 3.3 can be extended
to all time intervals where the disturbance remains constant. In Using a chassis dynomometer, the performance of the
practice, air and fuel estimation errors are essentially constant switching period controller was validated with a 2.4 liter,
provided the operating conditions change sufficiently slow. As four-cylinder gasoline engine with variable intake and exhaust
the engine speed and manifold pressure equilibrate after a load valve timing. The switching period controller, including the air
change transient, for example, the lumped disturbance variable and fuel estimators, used the same sensor set as the production
converges to a constant value. Once this occurs, the system be- vehicle. To quantify the controller performance, a pre-catalyst
haves as it would for a constant disturbance and therefore re- UEGO sensor was installed in the same location as the pre-cata-
mains close to the origin while switching at regular intervals. lyst EGO sensor. Since this engine has four independent inputs
Theorem 3.3 is also very practical because it helps dictate (engine speed, intake manifold pressure, intake cam timing and
the values of the dither parameters and the control gains. To exhaust cam timing), it would be difficult to accurately capture
MEYER et al.: AIR-TO-FUEL RATIO SWITCHING FREQUENCY CONTROL FOR GASOLINE ENGINES 641

Fig. 3. Calibration values for the AFR switching period controller.

Fig. 4. EQR response comparison: 0.10 EQR disturbance at event 250 and 770.

the combined effect of these inputs with a simple control-ori- tain a near stoichiometric AFR. As demonstrated in Fig. 5, the
ented model of the air path system. However, the trapped fresh distributions of the pre-catalyst EQR values achieved by the two
air mass is effectively predicted with a volumetric efficiency controllers appear similar. Further analysis of these distributions
based estimator. Even though the volumetric efficiency varies reveals that the AFR switching period controller reduces the oc-
as a function of the inputs, this relationship can be captured currence frequency of high amplitude EQR errors.
with a four dimensional lookup table. Similarly, a third-order To quantify this reduction, the percentage of occurrences in
gain scheduled fuel dynamics compensator is used to correct which the measured pre-catalyst EQR falls within an expanding
for the fuel dynamics. This compensator performs quite well range of EQR values is provided in Table I. Both controllers are
especially during steady-state operation. able to keep the AFR within EQR from stoichiometry
As a preliminary means of validation, the disturbance rejec- for approximately two thirds of the drive cycle. When consid-
tion performance of the AFR switching period controller was ering errors as large as 0.05, the robustness benefits of an AFR
tested and compared to the production controller. For this test a switching period controller can already be observed. The occu-
positive 0.10 EQR step disturbance was applied and then later rance frequency of EQR errors larger than 0.05 is 10.6% for the
removed resulting in a negative 0.10 EQR step disturbance. As AFR switching period controller and 12.3% for the production
demonstrated in Fig. 4, both controllers have approximately the controller. The AFR switching period controller also reduces
same disturbance rejection rate. As desired, the AFR switching the occurrence frequency of EQR errors larger than 0.10 (2.4%
period controller was able to reject a disturbance of this mag- versus 3.9%) and EQR errors larger 0.15 (0.9% versus 1.6%).
nitude in approximately 150 events. Unlike the production con- When the response of the two controllers are compared
troller, however, the AFR switching period controller produces side-by-side, the performance advantages of the AFR switching
virtually no overshoot once stoichiometry is reached. period controller are readily apparent. Fig. 6 compares the per-
To further quantify the effectiveness of the AFR switching formance of the AFR switching period controller to the
period controller, an FTP-72 drive cycle was run using both the production AFR controller for the first maneuver of the FTP-72
production AFR controller and the AFR switching period con- drive cycle. This data was collected in the event domain;
troller. To effectively control the AFR, disturbances introduced therefore, the datasets do not perfectly align. As shown in the
by transient operation must be quickly rejected in order to main- top plot, the engine speed trajectory matches well between
642 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 3, MAY 2013

Fig. 5. Distribution of the pre-catalyst EQR for an FTP-72 drive cycle.

TABLE I proven. For all bounded disturbances, an AFR switching pe-


DISTRIBUTION OF THE PRE-CATALYST EQR OVER AN FTP-72 DRIVE CYCLE riod control system remains stable. Additionally, the maximum
number of events between output switches is at worst bounded
by an affine function of the maximum disturbance magnitude.
As the dynamics of a disturbance decay, the AFR trajectory
asymptotically approaches a desired dither signal. Moreover,
this tracking performance is achieved without relying on an es-
timate of the plant delay other than the maximum plant delay
bound. The stability analysis presented in this paper also pro-
vides guidelines for choosing the amplitude and period of the
dither signal as well as the early and late switching control gains.
the two drive cycles. The middle plot of Fig. 6 compares the The bulk of AFR controllers which have appeared in open
pre-catalyst EQR traces during this portion of the FTP cycle. To literature strive to control the average AFR but allow for un-
exemplify the real world disturbance rejection performances of controlled oscillations. With a switching period controller, how-
the two controllers, a region of data has been highlighted in the ever, the actual AFR trajectory is controlled. This additional
lower left plot of Fig. 6. A gear shift occurs at the first part of level of control helps to ensure that the conversion efficiency
this region introducing a significant dynamic disturbance. Both of the TWC is maximized. In the next iteration of this work, a
controllers are able to recover from this type of disturbance, but corresponding post-catalyst AFR controller is being developed.
the AFR switching period controller produces lower peak EQR It has been shown in previous studies that the period and ampli-
errors. Compared to the production controller, the variance in tude of periodic oscillations around stoichiometry have a strong
the pre-catalyst EQR achieved by the AFR switching period influence on the conversion efficiency of a TWC. As shown in
controller was also reduced by a factor of 1.48 for this span of the stability theorems, the disturbance rejection performance is
events. After the disturbance converges to a constant value, the maximized when the period of the dither signal is minimized;
AFR switching period controller is able to rapidly converge to however, the optimal dither period in terms of tailpipe-out emis-
the reference dither signal. The steady-state performance of the sion may be different. In future work, the choice of the dither
AFR switching period controller is further highlighted in the signal with respect to emission performance will be studied.
lower right plot of Fig. 6.
APPENDIX A
PROOF OF THEOREM 3.1
V. CONCLUSION
First it should be noted that the two disturbance trajectory
The switching period control architecture presented in this bounds are reasonable and would be satisfied for any properly
paper extracts amplitude information from a binary EGO sensor calibrated set of air and fuel estimation algorithms. When the
by analyzing the switching period. With this amplitude informa- controller is first activated, after the system has reached its nom-
tion, both steady-state and transient AFR errors can be rapidly inal operating temperature, the late switching compensator takes
rejected with a minimum amount of overshoot. More impor- action until the output switches. Let represent the event
tantly, these robust performance benefits have been analytically when first changes sign and represent the event when
MEYER et al.: AIR-TO-FUEL RATIO SWITCHING FREQUENCY CONTROL FOR GASOLINE ENGINES 643

Fig. 6. EQR response comparison over a FTP-72 drive cycle.

the first EGO switch occurs. These two events can be related to . Therefore, the change in over this event span
each other with simplifies to

(29)

Until the first EGO switch occurs, the control law can be rep- (33)
resented with
Using the disturbance rate of change bound, the EQR error can
(30) be bounded by

. Consider the EQR system model described


in (8). The change in over the span of events can be (34)
described with and

(35)
.
If such that
(31)
(36)
Because the dither signal is periodic
, then the system will converge to the
(32) origin monotonically (with respect to a time step of )
644 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 3, MAY 2013

during this time frame. For this to occur, the late switching con- If , then the system could potentially continue to
trol action must overpower the effect of any disturbance (high drift away from the origin even after the late switching compen-
gain feedback). The control input must satisfy sator activates. Under the worst case scenario, the disturbance
could continue to grow until it reaches either its upper or lower
(37) maximum bound. Once this bound is reached, however, the late
switching compensator will force the system back to the origin.
. Comparing this condition to the con- At the instance in which the late switching compensator acti-
trol law described in (30), it is clear that monotonic conver- vates, the disturbance is not more than away from its
gence with respect to a time step of can be guaranteed maximum magnitude bound, and the system can drift away from
when the late switching control gain is larger than the max- the origin at a maximum rate of . Combining these
imum rate of change of the disturbance (i.e., ). When observations, the maximum drift error must be bounded
, the disturbance can change faster than the controller by
can compensate. Because the magnitude of the disturbance is
bounded by , the disturbance magnitude cannot increase (43)
indefinitely. Even though the system may non-monotonically
converge to the origin, the controller ensures that the system
It should be noted that when , the drift error term is
reaches the origin. The maximum number of events required to
always zero.
reach the origin can be bounded by
The maximum total error is therefore the summation
if of the overshoot error bound, the accumulated error bound, the
(38) dither contribution bound and the drift error bound as in
otherwise
(44)
where .
Regardless of the size of the late switching control gain, the where could be either due to the late switching compensator
transport delays which characterize the system will cause the or the early switching compensator. Because of the time varying
late switching compensator to overshoot stoichiometry . nature of the disturbance, the corrective action made by the early
Under the worst case scenario, the effect of the applied control switching compensator could be in the wrong direction. The
action will appear in the output after events. The magni- overshoot error made by the early switching compensator
tude of the maximum overshoot caused by the late switching is therefore bounded by
compensator must, therefore, satisfy
(45)
(39)
The value of is simply the maximum of the two overshoot
After reaching the origin, the late switching compensator will errors or
start acting if the output has not switched after
events. During this time, the disturbance could be forcing (46)
the system away from the origin at a maximum rate of .
Defining as Based on this analysis, the maximum time interval between
EGO switches can also be bounded. The plant delay, the wait
(40) period before the late switching control activates and the time
required to eliminate the maximum error all affect this bound.
Because the size of the maximum error depends on whether
is it clear that before the late switching compensator activates, , the maximum number of events between output
the system can accumulate an error as large as provided switches also depends on this condition. When , the
. If is used to represent the maximum error maximum error can be eliminated at a minimum rate of
that can accumulate before the late switching compensator . Conversely when , the disturbance can
activates, then must therefore satisfy change by as much as thereby increasing the time re-
quired to eliminate the error to . The value of
(41)
is therefore
An error this large can only be accumulated after
events without switching. The dither also can cause the system (47)
to move away from the origin. If the sign of changes on event
, then the maximum contribution from the dither de- if and
pends on the value of the dither signal at the event according to
(48)
(42)
otherwise. Therefore, all trajectories converge to the set
Since the dither can take a maximum value of , this expression in finite time and subsequently remain inside
can be bounded from above by . this set.
MEYER et al.: AIR-TO-FUEL RATIO SWITCHING FREQUENCY CONTROL FOR GASOLINE ENGINES 645

APPENDIX B the sign of cannot change when so


PROOF OF THEOREM 3.2 .
Theorem 3.2 can be directly derived from (38), (44), and (47) Knowing that the maximum overshoot contribution from the
for the case when and . For a constant dither is at most , the maximum total error can now be
disturbance, the drift error, the accrued error and the overshoot bounded by
error caused by the early switching compensator are all zero.
The maximum number of events required to reach the origin is (59)
therefore bounded by
Although the late switching compensator can cause a peak over-
(49) shoot error as large as , only the contribution of the
integral part of the controller is maintained. The maximum sus-
tained overshoot error caused by the late switching com-
Likewise, the maximum total error is
pensator is bounded by
(50)
(60)
and the maximum number of events between output switches
once the origin is reached is bounded by Theorem 3.2 already guarantees that the origin will be
reached in finite time. If the closed-loop controller were de-
(51) activated after the first EGO switch at event , then
such that

(61)
APPENDIX C
PROOF OF THEOREM 3.3 If the peak overshoot caused by the switching period controller
is less than the dither amplitude, then the periodic dither compo-
To prove this theorem, it will first be shown that all trajecto-
nent alone will cause another EGO switch and this switch would
ries of the system enter and remain in the invariant set
occur after no more than events. This condition is met
. Until now, the contribution of the dither has
when
been bounded conservatively by . From (42), it is clear that
the bound on can be reduced to if corresponding
(62)
to a sign change in , the value of the dither signal satisfies

(52) which is always true provided that the constraints in (26) and
(27) are satisfied.
or equivalently The overall governing control law satisfies
(53) (63)
It is sufficient to show that when
so that such that
(54)
(64)
the sign of does not change. Define as the event when
This means that when the AFR switching period controller
(55) is active it will force the system trajectory toward the origin
unless changes sign (the origin was reached) between events
Using as a reference, the sign of must not change within the and . The maximum amount of overshoot accrued
interval . This is ensured during this span of events is . One can conclude that with the
when controller active, the next EGO switch event will still occur
within events. Therefore, all trajectories converge to
(56) within finite time and subsequently remain
inside this set.
holds for this interval. Substituting in for the change in , this
Now the local behavior of the system within the set
inequality reduces to
can be analyzed. This local analysis relies on the
assumption that every late switch is followed by an early switch
(57)
and early switch is followed by a late switch. This assumption
is met if the following four conditions hold.
. As long as 1) The complete effect of the control action taken by the late
switching compensator is seen in the output before the next
(58) late EGO switch.
646 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 3, MAY 2013

2) The control action taken by the early switching compen- The closed-loop system can also be represented by
sator influences the output before the following late EGO
switch. (72)
3) The control action taken by the late switching compensator
which simplifies to
first appears in the output before the following early EGO
switch. (73)
4) The proportional gain of the late switching compensator is
not larger than the minimum sustained overshoot caused Neither the late nor the early switching compensator will take
by the late switching compensator. control action unless the offset error causes a switching period
Although omitted here, it can be shown that all four of these error larger than events. Recall that (15) relates the number
conditions are automatically satisfied when (27) and (28) are of events between output switches to the average offset error.
true. For the closed-loop system, the sustained error, rather than the
Consider a new time scale represented by the time index average error, dictates the number of events between output
that increments every early EGO switch. With this reference, switches. The relative lateness of a late EGO switch can be re-
the order of these corrections is late, then early. For readability, lated to the sustained error using
each variable in this time scale is distinguished with an overline.
The state variable represents the sustained error which is equal (74)
to the sum of the constant disturbance and the value held in
the integrator of the controller. This analysis is only valid for
trajectories inside the set . which simplifies to
In the new time scale, the state equation is simply
(75)
(65)
Define the relative earliness of an EGO switch as
where is the combined integral contributions from the early
and late switching compensators. The initial conditions for this (76)
time scale are dependent on the sustained overshoot caused by
the late switching compensator according to The instantaneous sustained error can also be related to the rel-
ative earliness of an early EGO switch according to
(66)

Therefore, the set of all initial conditions for is (77)

(67) which simplifies to

Because only the integral contributions affect the sustained (78)


offset error, the control law in this time scale can be represented
as If
(68) (79)
where represents the relative lateness of an output switch in
the manner then either the late or the early switching compensator will take
action depending on the situation; otherwise, neither controller
(69) will act. The system will converge to the invariant set
defined as
identifies when an early switch occurs according to (80)
if
(70) if such that that satisfy
otherwise

and is the value of the sustained offset error when the


early switch occurs. Because of the plant delay, not all of the (81)
late switching compensation affects the measurement before the
early switching occurs. Therefore the value of is between and if that satisfy
and as described by
(82)
(71)
Clearly (82) holds, because neither controller makes any correc-
where . tions when the offset error is this small.
MEYER et al.: AIR-TO-FUEL RATIO SWITCHING FREQUENCY CONTROL FOR GASOLINE ENGINES 647

As long as neither compensator applies control action which and amplitude exactly and an error between the desired and
is larger than , the bounds described in (81) are satisfied. achieved offset that satisfies (25).
For the late switching compensator this means that the following
must be true:

(83) REFERENCES

. The trigonometric relationship containing


[1] D. Hamburg, J. Cook, W. Kaiser, and E. Logothetis, “An engine-dy-
shown in (75) can be bounded by namometer study of the A/F compatibility between a three-way catalyst
and an exhaust gas oxygen sensor,” SAE, Warrendale, NJ, Tech. Rep.
830986, Apr. 1983.
(84) [2] M. Defoort, D. Olsen, and B. Willson, “The effect of air-fuel ratio con-
trol strategies on nitrogen compound formation in three-way catalysts,”
Int. J. Engine Res., vol. 5, pp. 115–122, 2004.
. Since is constrained by (26), it [3] A. Lie, J. Hoebink, and G. Marin, “The effects of oscillatory feeding
follows that of CO and on the performance of a monolithic catalytic converter
of automobile exhaust gas: A modeling study,” Chem. Eng. J., vol. 53,
no. 757, pp. 47–54, 1993.
(85) [4] L. Mianzo, H. Peng, and I. Haskara, “Transient air-fuel ratio pre-
view control of a drive-by-wire internal combustion engine,” in Proc.
Amer. Control Conf., 2001, pp. 2867–2871.
and therefore (83) must hold. [5] J.-M. Kang and J. Grizzle, “Nonlinear control for joint air and fuel
management in a SI engine,” in Proc. Amer. Control Conf., 1999, pp.
By selecting in this manner, the late switching compen- 17–23.
sator will decrease the magnitude of the offset at every time step [6] K. Muske, J. P. Jones, and E. M. Franceschi, “Adaptive analytical
without changing the sign of the error. As a result can be model-based control for SI engine air-fuel ratio,” IEEE Trans. Con-
trol Syst. Technol., vol. 16, no. 4, pp. 763–768, Jul. 2008.
bounded by [7] R. Zope, K. Grigoriadis, J. Mohammadpour, and M. Franchek, “Air-
fuel ratio control of spark ignited engines with TWC using LPV tech-
niques,” in Proc. ASME Dyn. Syst. Control Conf., 2009, pp. 897–903.
(86) [8] R. Zope, J. Mohammadpour, K. Grigoriadis, and M. Franchek, “Ro-
bust fueling strategy for an SI engine model as an linear parameter
varying time-delayed system,” in Proc. Amer. Control Conf., 2010, pp.
The magnitude of the correction made by the early compensator 4634–4639.
is . As a result, system eigenvalues have a lower bound [9] Y.-W. Kim, G. Rizzoni, and V. Utkin, “Automotive engine diagnosis
of and control via nonlinear estimation,” Control Syst. Mag., vol. 18, no.
5, pp. 84–99, Oct. 1998.
[10] M. Won, S. Choi, and J. Hendrick, “Air-to-fuel ratio control of spark-
(87) ignited engines using Gaussian network sliding control,” IEEE Trans.
Control Syst. Technol., vol. 6, no. 5, pp. 678–687, Sep. 1998.
[11] S. Wang and D. L. Yu, “A new development of internal combustion
and an upper bound of engine air-fuel ratio control with second-order sliding mode,” J. Dyn.
Syst., Meas., Control, vol. 129, no. 757, pp. 84–99, 2007.
[12] R. Möller, C. Onder, L. Guzzella, M. Votsmeier, and J. Gieshoff, “Anal-
(88) ysis of a kinetic model describing the dynamic operation of a three-way
catalyst,” Appl. Catalysis B: Environmental, vol. 70, pp. 269–275, Jan.
2007.
. [13] J. C. P. Jones and R. A. Jackson, “Potential and pitfalls in the use of
dual exhaust gas oxygen sensors for three-way catalyst monitoring and
In summary, as long as satisfies (26) and satis- control,” Proc. Inst. Mechan. Eng. D, J. Automobile Eng., vol. 217, pp.
fies (27), all trajectories of the original system converge to the 475–488, 2003.
[14] A. Vosz, S. Midlam-Mohler, Y. Guezennec, and S. Yurkovich, “Ex-
invariant set within a finite number of perimental investigation of switching oxygen sensor behavior due to
events as bounded by (49). Once inside the invariant set exhaust gas effects,” in Proc. Int. Mechan. Eng. Congr. Expo., 2006,
, the EGO sensor will switch twice every dither pe- pp. 381–388.
[15] J. Meyer, S. Yurkovich, and S. Middlam-Mohler, “Architectures for
riod (one late switch and one early switch) as long as the period phase variation compensation in AFR control,” in Proc. Amer. Control
of the dither signal satisfies (27) and the late switching propor- Conf., 2010, pp. 1447–1452.
tional gain satisfies (28). The combined effect of the disturbance [16] S. Liu, K. Dudek, and R. Shafto, “A phase and frequency locking loop
for engine air-fuel ratio control,” in Proc. Amer. Control Conf., 2008,
and the integral compensation one event prior to the late EGO pp. 326–330.
switch strictly decreases as long as the EQR offset has a magni- [17] J. Meyer, S. Rajagopalan, S. Midlam-Mohler, S. Yurkovich, and Y.
tude larger than . When the offset is equal to or less than , the Guezennec, “Application of an exhaust geometry based delay predic-
tion model to internal combustion engines,” in Proc. Dyn. Syst. Control
offset error remains constant . Therefore the offset error Conf., 2009, pp. 557–563.
converges to the invariant set . Inside this set, the system [18] C. F. Aquino, “Transient A/F control characteristics of a 5 liter central
trajectory simplifies from (8) to fuel injection engine,” SAE, Warrendale, PA, Tech. Rep. 810494, 1981.
[19] E. Curtis, C. Aquino, and D. Trumpy, “A new port and cylinder wall
wetting model to predict transient air/fuel excursions in a port fuel in-
(89) jected engine,” presented at the Int. Spring Fuels Lubrications Meeting,
Dearborn, MI, 1996, 961186.
[20] J. Meyer, S. Yurkovich, and S. Middlam-Mohler, “A model based esti-
where and . Therefore, the EQR of the mator for cylinder specific air-to-fuel ratio corrections,” ASME J. Dyn.
combustion mixture reproduces the dither signal in frequency Syst., Meas. Control, vol. 133, no. 3, pp. 1–14, 2011.
648 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 3, MAY 2013

Jason Meyer received the B.S., M.S., and Ph.D. de- and Control Systems Laboratory (Simon & Schuester, 1998). He was with
grees in mechanical engineering from The Ohio State Ohio State University for 27 years, from 1984 to 2011, where he was active in
University, Columbus, in 2007, 2008, and 2011, re- research and educational activities of the Center for Automotive Research.
spectively. Prof. Yurkovich was a recipient of the John R. Ragazzini Award in Control
He is a Research Associate with the Department Education of the AACC in 2008, the IEEE CSS Distinguished Member Award
of Mechanical Engineering, The Ohio State Univer- in 2000, and an IEEE Third Millennium Medal.
sity Center for Automotive Research, Columbus. His
primary expertise is in the field of modeling and con-
trol of dynamic systems. His research is focused on
automotive applications including internal combus- Shawn Midlam-Mohler received the Ph.D. degree
tion engines, aftertreatment systems and battery en- in mechanical engineering from The Ohio State Uni-
ergy storage systems. versity, Columbus, in 2005.
He is a research Scientist and Adjunct Assistant
Professor with the Department of Mechanical and
Aerospace Engineering, The Ohio State University
Stephen Yurkovich (M’84–F’01) holds the Lewis Center for Automotive Research, Columbus. His
Beecherl, Jr., Distinguished Chair in Engineering at primary research field is in the area of advanced
the University of Texas at Dallas, Richardson, where automotive powertrain systems with a focus on en-
he is Department Head of Systems Engineering. His ergy efficiency and emissions control. Although his
research has focused on the theory and applications primary focus is in the area of internal combustion
of control systems, in numerous application areas. engines, he also has research activities in the area of hybrid electric, electric and
He has been an author on more than 200 publica- plug-in hybrid electric vehicles. Beyond the field of automotive technology, he
tions in journals, edited volumes, and conference has a broader interest in the topic of sustainable energy.
proceedings, and has authored and co-authored the
books Control Laboratory (Kendall/Hunt, 1992),
Fuzzy Control (Addison-Wesley-Longman, 1998),

Anda mungkin juga menyukai