Sie sind auf Seite 1von 12

* Corresponding author. Tel.

: #44-1232-274-147; fax: #44-1232-


661-729.
Control Engineering Practice 8 (2000) 519}530
A comparison of idle speed control schemes
M. Thornhill, S. Thompson*, H. Sindano
School of Mechanical and Manufacturing Engineering, Queen's University of Belfast, Ashby Building, Stranmillis Road, Belfast BT9 5AH, UK
SAGEM, Research and Development, Birmingham, UK
Received 2 December 1998; accepted 25 October 1999
Abstract
This paper examines the idle speed regulation control problem in multi-point spark ignited petrol engines. Several possible
solutions are presented, including proportional plus integral control, fuzzy logic control, adaptive fuzzy logic control, adaptive fuzzy
logic control in conjunction with Smith prediction and dynamic matrix control. All of the controllers are compared in simulation and,
where possible, on a production vehicle. The performance measures used for comparison purposes were mean-square error and
maximum error. It is shown that there are several possible alternatives to the existing proportional plus integral control used on the
air bypass valve of production vehicles. 2000 Elsevier Science td. All rights reserved.
Keywords: Idle speed regulation; PI control; Fuzzy logic control; Adaptive fuzzy logic; Adaptive fuzzy logic with Smith prediction; Dynamic matrix
control; Mean-square error; Maximum error
1. Introduction
Idle speed control (ISC) represents one of the most
generic and basic automotive control problems confront-
ing automotive control researchers and practitioners.
Engine idle speed refers to engine operation under closed
throttle conditions; on average, vehicles consume about
30% of their fuel in city driving during idling (Jurgen,
1995). It is therefore important to try and optimise ve-
hicle and powertrain operation at idle, especially with
respect to the often con#icting requirements of improved
fuel economy, good noise, vibration and harshness qual-
ity, and reduced emissions.
The complete ISC problem, in order of importance,
encompasses three di!erent operational phases:
1. Idle speed regulation. At idle the engine speed is never
truly constant (see Section 1.1 and Fig. 8). Therefore,
the controller must be capable of maintaining engine
speed close to the set point (selected target idle speed
value) with as little deviation as possible (the engine
must run smoothly). Essentially, the better the idle
speed regulation the lower the selected idle speed
value can be set, subject to some allowance for distur-
bance rejection and entry/exit to idle speed. Ideally,
idle speed regulation would be performed solely by the
air bypass valve.
2. Rejection of known disturbances. Idle speed distur-
bance rejection tests are a popular test to perform in
an engine test cell. Most modern dynamometers per-
mit load adjustments and the data logging equipment
permits appropriate traces to be collected. In a vehicle
the disturbances are due to electrical loads (switching
on of air conditioning, window heating, lighting, etc.)
or for a vehicle with power steering, low-speed
manoeuvring. These are events which, when they oc-
cur, may cause the engine to stall or mis"re. Typically,
the solution to the problem requires several feed for-
ward control loops using accessory load information,
and other ad hoc compensation schemes for temper-
ature, barometric pressure and other environmental
conditions. Often these would act on both the air
bypass valve position and the spark advance.
3. Entry and exit from/to idle speed. Very little research
has been reported in this area. However, any step
change from a high-speed level to the set point idle
speed can cause a speed undershoot and the engine to
stall. One solution is to provide a smooth target
transition speed that the controller is required to follow.
Further, the order of these operational phases also
indicates the order of increasing di$culty in performing
0967-0661/00/$- see front matter 2000 Elsevier Science Ltd. All rights reserved.
PII: S 0 9 6 7 - 0 6 6 1 ( 9 9 ) 0 0 1 9 0 - 2
Nomenclature
ABV air bypass valve
B(s) predicted plant output
BTDC before top dead centre
J
G
width of triangular membership function l for
input i
DMC dynamic matrix control
learning rate
vector of fuzzy basis function values
e error signal
E(s) error signal in the frequency domain
EMS engine management system
f (x) non-linear function
FJG
G
fuzzy set l
G
for input i
#ops #oating point operations
FIR "nite impulse response
FLC fuzzy logic control
Gc(s) controller transfer function
Gp(s) plant transfer function
G(s) delay-free plant model
I integral
ISC idle speed control
k
G
integral gain
k
N
proportional gain
l rule number

$
JG
G
membership grade of fuzzy set FJG
G
max maximum function
M number of rules
MSE mean-square error
ME maximum error
n number of inputs to fuzzy logic system or
number of sample points (MSE and ME)
N
G
speed (rpm) at sample point i
N
PCDG
reference speed (rpm) at sample point i
vector of output membership function values
PI proportional, integral.
PRBS pseudo-random binary sequence
R(s) reference signal in the frequency domain
rpm revolutions per minute
SA spark advance
SP Smith prediction
t time
tri triangular membership function
system delay
u controller output
vars. variables
wJ weight of rule l
x system state variable
x system state vector
x J
G
centre of triangular membership function l for
input i
y plant output
y J output membership function l value
>(s) output signal in the frequency domain
vehicle tests and in obtaining controller performance
measures. Since this paper sets out to compare several
possible idle speed controllers only the baseline regula-
tion tests are considered in detail.
In order to satisfy the baseline regulation problem,
current ISC strategies use both spark and air control;
even though spark control is required for other functions.
When only air control is used to maintain idle speed it is
di$cult to achieve satisfactory performance. The argu-
ment is that the geometry of the inlet manifold introduces
a delay in the air to speed loop. Therefore, since it can be
demonstrated that changes in spark advance will pro-
duce rapid changes in speed, then this delay should be
compensated for using spark control.
Typically, a production ISC strategy includes PI con-
trol for the air loop and proportional feedback control
for the spark loop. Since the primary objective of the ISC
system is to track a constant desired speed by manipula-
ting the air bypass valve (ABV) the integral portion of the
PI controller is considered to be the core of the ISC
strategy.
Controller gains are normally tuned by hand, using
trial and error, to obtain best performance. This method
of tuning can be inaccurate since the idle speed system is
non-linear and the engine characteristics change over its
lifetime. When combined with short term environmental
changes, or changes in fuel type, it can be seen that a "xed
control scheme cannot be optimal at all times.
This paper presents a comparison of several ISC
schemes, all of which use the air bypass valve (ABV) as an
actuator. The control schemes that are compared are:
(1) PI control, this is the standard production control-
ler.
(2) Fuzzy logic, as it is capable of controlling non-linear
systems.
(3) Adaptive fuzzy logic, as it is capable of controlling
time-varying, non-linear systems.
(4) Adaptive fuzzy logic with Smith prediction, as it is
capable of controlling time varying, non-linear sys-
tems with time delays.
(5) Dynamic matrix control (DMC), as it is also capable
of dealing with systems with time delays.
All of these control schemes can be designed without
the need for a formal mathematical model and tuning, for
the non-adaptive schemes, can be performed on-line.
520 M. Thornhill et al. / Control Engineering Practice 8 (2000) 519}530
Fig. 1. Inputs, outputs and disturbances to the idle speed system.
When used, the spark controller remains unchanged
from that supplied on the production vehicle. Air}fuel
ratio control is enabled for all tests and is necessary for
the entry to idle speed tests.
Three sets of tests were performed in order to compare
these control schemes for the regulation of idle speed:
(1) Simulation with constant spark advance. (This simu-
lation tests the system integrity, i.e. in the event of
spark control failure what is the residual ISC).
(2) Engine test with constant spark advance.
(3) Engine test with proportional spark control. Al-
though spark control may not appear necessary for
the regulation tests, it may be essential for entry and
exit from/to idle speed and therefore is included in
the regulation tests.
Two performance measures were used to compare the
control schemes, namely mean-square error (MSE) and
maximum error (ME). The MSE gives a feel for how
smooth the engine idles as perceived by the driver of the
vehicle. The ME determines the largest deviation from
the target idle speed. At low target idle speeds a large ME
can cause the engine to stall. Due to engine management
system (EMS) limitations on the vehicle tested, the DMC
algorithm and the adaptive fuzzy logic controller with
Smith prediction could not be tested.
1.1. The idle speed system
Fig. 1 shows a representation of the idle speed system
with the physical inputs and outputs shown.
Of the factors which a!ect engine speed, the engine
controller only has control over the fuel, air, re-circulated
exhaust gas and spark timing. The other factors a!ecting
engine speed are either part of the engine design or are
a function of atmospheric conditions.
In this paper, the vehicle tested had a spark ignition
engine with port fuel injection, a stepper motor that
admits auxiliary air, variable spark timing and exhaust
gas re-circulation all controlled by the engine manage-
ment system which had a sample time of 0.032 s. As usual
with ISC the engine speed was controlled using a combi-
nation of spark and air control so that the engine speed
tracks a reference idle speed. The reference idle speed is
pre-set by the engine designer at a speed that gives
smooth running of the engine while maintaining accept-
able fuel economy. Engine fuel control is reserved for
emission control purposes. In all tests exhaust gas re-
circulation was disabled and the fuel control enabled.
For the constant spark advance tests, the spark advance
was held constant at 103 BTDC.
An internal combustion engine running at idle speed is
characterised by various problems (Abate & DiNunzio,
1990; Nishimura & Ishii, 1986; Washino, Nishiyama
& Ohkubo, 1986; Hrovat & Sun, 1996):
1. A trend to trigger speed oscillations with a period of
around 2}5 s in fuel-injected engines. This is caused by
a positive reaction that is inherent in the engine and is
related to its functioning mode and construction
characteristics.
2. A drop in the number of revolutions because of the
application of a braking torque. There may be various
reasons for this torque, such as switching on of an
electrical load, switching on of air conditioning system
or activation of the power steering.
3. Poor running of the engine as it is operating in a low-
speed and low-torque area for which the design has
not been optimised. Particularly in the case in which
there is a tendency to lower the target engine rotation
speed (deliberately or because of the above-mentioned
braking torques), this leads to bad combustion of the
fuel or failure to burn the fuel. In view of the low value
of the torques involved, this irregularity of operation,
if not appropriately compensated, causes #uctuations
in engine speed with high-frequency spectrum charac-
teristics with a period more or less aligned to that of
engine phase duration and random distribution. This
disturbance is therefore di!erent from that outlined in
problem 1 above. The problem described here is
heightened further when the engine is cold so that
friction is greater and combustion more irregular.
4. The idle speed of an engine is also subject to slowly
varying changes due to changes in operating condi-
tions or engine condition. These slow changes may be
due to the conditions under which the engine operates,
like ambient temperature and pressure, fuel quality,
lubricant temperature, etc. In addition, the condition
of repair of the engine and the state of wear change
slowly over the life of an engine.
5. Induction to power stroke delay, this delay is variable
and is a function of engine speed. This delay is cited as
being the dominant dynamic e!ect in the idle speed
control problem.
In summary, the idle speed system is non-linear and
time varying with variable time delays.
M. Thornhill et al. / Control Engineering Practice 8 (2000) 519}530 521
Fig. 2. Input membership functions used in the fuzzy controller.
Fig. 3. Output membership functions for the fuzzy controller.
2. Control schemes tested
The following control schemes were tested for their
ability to regulate the idle speed to a value of 864 rpm in
the presence of the non-linearities described in Section
1.1. The parameters in each control scheme were used
both in simulation and on the engine tests.
2.1. PI control
The gains used here were the production controller
gains of 0.03 on both the proportional and integral gains.
The production controller uses the formula
u"k
N
e#k
G

e dt, (1)
to calculate the controller output.
2.2. Fuzzy logic control
The fuzzy logic controller demonstrated here (Takagi
& Sugeno, 1985) is the simplest type of fuzzy controller in
that there is only one input and one output. The only
input to the fuzzy controller was the error between en-
gine speed and idle reference speed. Five fuzzy sets were
de"ned on this input universe of discourse (see Fig. 2).
The "ve linguistic fuzzy sets de"ned are large negative
(ln), small negative (sn), zero (ze), small positive (sp) and
large positive (lp). These "ve fuzzy sets provide complete
coverage of the input space (i.e. for every value of speed
error at least one fuzzy set has a non-zero value), for
values less than !200 rpm (relative to the set point),
ln"1 and for values greater than 200 rpm, lp"1.
As there was only one input to the fuzzy controller and
"ve fuzzy sets for this input, there were only "ve control
rules.
Rule base 1 shows the rules used for the fuzzy control-
ler, the antecedent part of the rule refers to the input
membership functions shown in Fig. 2, and the conse-
quent part of the rule refers to the output membership
functions shown in Fig. 3.
Rule base 1. Rules for the fuzzy controller.
IF error is lp THEN change in stepper position is outmf1
IF error is sp THEN change in stepper position is outmf2
IF error is ze THEN change in stepper position is outmf3
IF error is sn THEN change in stepper position is outmf4
IF error is ln THEN change in stepper position is outmf5
Defuzzi"cation obtains a crisp numerical value from
fuzzy inputs, the Takagi}Sugeno fuzzy system has been
used exclusively throughout this paper. For a real-valued
input vector x"(x

,
2
, x
L
)', the output y( x ) of Takagi
and Sugeno's fuzzy system is a weighted average of the
yJ's
y( x )"
+
J
wJyJ
+
J
wJ
, (2)
where the weight wJ implies the overall truth value of the
premise of rule 'J' for the input and is calculated as
wJ"
L

$
J
G
(x
G
). (3)
The output from the defuzzi"cation stage is the "nal
crisp output of the fuzzy logic controller. Fig. 3 shows
that the output membership functions range from !0.5
to 0.5 steps, while the ABV can only move in discrete
steps. However a #oating point value of ABV is constant-
ly added to or subtracted from in the controller, this
#oating point value is then rounded to the nearest integer
to give an integer number of ABV steps.
522 M. Thornhill et al. / Control Engineering Practice 8 (2000) 519}530
Fig. 4. The overall scheme of direct adaptive fuzzy control.
Fig. 5. Input membership functions used for engine speed input.
Note that the output of the fuzzy controller is the
change in stepper position. That is, the stepper position is
constantly added to or subtracted from and this produces
integral action in the controller.
2.3. Adaptive FLC
Non-adaptive fuzzy logic controllers rely on experts to
de"ne the input/output membership functions and rules.
A more optimal controller would be one that adapts to
errors between desired output and actual output to
achieve `optimala control.
There are many adaptive fuzzy logic schemes that
adapt input/output membership functions or the rule
base in the fuzzy logic controller (Procyk & Mamdani,
1979; Lee, 1991). The method used in this paper is taken
from Wang (1994), since Wang's direct adaptive fuzzy
controller is demonstrated to have excellent tracking
properties (even for non-linear systems) and so was
chosen as an appropriate control strategy.
In this application, the output membership functions
are adapted so that the fuzzy logic controller is tuned to
minimise the error between engine speed and the idle
reference speed. That is the controller is a direct adaptive
fuzzy controller (Wang, 1994), with adaptation of output
membership functions only (see Fig. 4). The parameters
of the block labelled fuzzy controller in Fig. 4 are de-
scribed in Section 2.3.1.
Fig. 4 shows a supervisory control scheme as part of
adaptive fuzzy logic. Supervisory control is used by
Wang (1994) to maintain stability of the controller during
learning. Supervisory control switches on when the signal
to be controlled has exceeded a pre-speci"ed value that
the designer has set. To enable analysis of `controller
learninga these blocks have been neglected during the
design of direct adaptive fuzzy idle speed controllers.
The formula for the fuzzy controller is as follows:
u(x)"
+
J
y J[L
G
tri(x
G
!x J
G
(),J
G
)]
+
J
[L
G
tri(x
G
!x J
G
,J
G
)]
. (4)
In the fuzzy logic system used, only the yJ (output
membership function) values are adapted using a least-
mean-squares algorithm. Again, the output membership
function values correspond to ABV positions.
The input membership function could also be adapted
but this would require a non-linear adaptation scheme
(such as back-propagation), which would necessitate ad-
ditional computation time. Careful selection of the input
membership functions was required so that the input
space was fully covered, see Fig. 5.
The adaptive law used is based on a least-means-
squares formulation.
Q "e( x ). (5)
The fuzzy basis function vector is de"ned as follows:
'J
2
JL '( x )"
L
G

$
JG
G
(x
G
)
K
J
2
KL
JL
(L
G

$
JG
G
(x
G
))
. (6)
The learning rate is a crucial tuning factor in this
system. A large value of learning rate causes the control-
ler to learn quickly but may also lead to instability
during learning, a smaller value of learning rate causes
slower but more accurate learning.
2.3.1. Parameters of adaptive FLC
Only one input to the fuzzy controller is used, namely
engine speed. Fig. 5 shows the input membership func-
tions used for the input engine speed, these membership
functions provide total coverage of the input space and
were selected to be most dense around 700 rpm, which is
below the constantly desired idle speed.
M. Thornhill et al. / Control Engineering Practice 8 (2000) 519}530 523
Fig. 6. The Smith predictor arrangement.
Five rules were obtained from the output from each
fuzzy set. One output membership function was assigned
to each rule. These output membership functions were
given initial values of 30 (steps) at the start of each
simulation, but could be given any value. Weighted aver-
age defuzzi"cation was used, see Eq. (2). During simula-
tion the position of each output membership function
was adapted to minimise the error between engine speed
and idle reference speed at each simulation time step.
2.4. Adaptive FLC with Smith prediction
Adaptive fuzzy logic control with Smith prediction was
tested as adaptive FLC alone resulted in unstable control
which was thought to be due to signi"cant delays in the
air}speed control loop.
Smith proposed the control scheme shown in Fig. 6
(Dutton, Thompson & Barraclough, 1997). Ignoring the
dashed lines, what happens in Fig. 6 is that a delay-free
model of the plant is used to generate the output signal
which would exist if the delay were absent (assuming
a good model). This delay-free signal is then used in the
usual feedback loop (via B(s)), instead of the plant output.
To help account for errors in the delay-free model, the
delay itself is also modelled, and used to generate what
should be a model of the actual plant output, including
the delay e!ect. The dashed lines show how this is then
compared with the actual output, >(s), so that a model-
ling error is also fed back into the control loop via B(S).
In this way, the e!ects of errors in the model of G(s) are
reduced.
The stability of the closed-loop system is identical to
that obtained if the plant's time delay had originally been
ignored. However, the closed-loop transfer function con-
tains a time delay in its numerator (since G
N
(s)"
G(s)eOQ) and this will modify the response of the system
compared with that of the time delay free plant. How-
ever, this technique depends upon an accurate plant
model (particularly, an accurate model of the delay
(Dutton et al., 1997)).
To achieve adaptive fuzzy logic control with Smith
prediction, the adaptive fuzzy logic controller described
in Section 2.4 is inserted in the block labelled G
A
(s) in
Fig. 6. One of the advantages of fuzzy logic is that it is
model-free, however the Smith predictor arrangement
negates this advantage in that it requires a model. For the
simulation and practical tests, the models used in the
Smith predictor were the same experimentally obtained
FIR model, see Section 4, with the (assumed) delay trun-
cated for the delay-free model.
In the actual system the dynamics of the system vary
over the input space and so no one model would be
optimal. A method to overcome this problem would be
to use an adaptive LMS "lter (Widrow & Wallach, 1996),
so that the model for the Smith predictor would be
optimal at all times. This however requires more com-
putational load as the system model is being adapted at
all times.
2.5. Dynamic matrix control
Dynamic matrix control (Cutler & Ramaker, 1980) is
a time-domain-based control algorithm. The basic idea is
to use a time-domain step response model of the process
to calculate the future changes in the manipulated vari-
able required to minimise some performance index.
DMCis included in order to provide a widely used and
generally accepted alternative controller. Note that the
step response used for calculating the controller matrix is
obtained from the FIR plant model. Therefore, the simu-
lation results should give the best possible control. In
practice there would be some plant-model mismatch and
an inevitable deterioration in the expected performance.
3. Performance measures
To compare the performance of the di!erent controller
schemes, the following performance measures were used:
(1) The mean-square error (MSE):
MSE"
L
G
(N
G
!Nref
G
)`
n
. (7)
Eq. (7) is only used with idle speed regulation tests, i.e.
the dynamic response at steady state. In real terms the
MSE gives an indication of the smoothness of idle as
apparent to the driver of the vehicle e.g. the MSE for the
production idle speed controller is 203 rpm` which feels
smooth to the driver of the vehicle. For this study it will
be assumed that an acceptable upper bound on the MSE
will be 203 rpm`.
(2) The maximum error (ME):
ME"max(N!Nref ). (8)
This performance measure is used only with idle speed
regulation tests. The maximum error performance
measure indicates the band around the idle reference
524 M. Thornhill et al. / Control Engineering Practice 8 (2000) 519}530
Fig. 7. Finite impulse response on air-to-speed loop.
Fig. 8. Idle speed regulation, PI control, P"0.03, I"0.03.
Table 1
Comparison of control schemes in simulation with constant spark
advance
Type of control scheme Mean-square error
(rpm`)
Maximum
error (rpm)
PI (P"0.03, I"0.03) Unstable Unstable
Fuzzy 140 40
Adaptive FL Unstable Unstable
Adaptive FL#SP 11 14
DMC 3 5
speed within which the idle speed is maintained. In the
case of low idle speed, this band is critical as a large drop
from the reference speed could cause the engine to stall.
The production idle speed controller produced a max-
imum error value of 50 rpm for the test shown, this value
is used as the acceptable upper bound.
4. Simulation
4.1. Idle speed model for simulation
For the purposes of testing the adaptive fuzzy logic
algorithm, a "nite impulse response (FIR) 6.4 s in length
was calculated using pseudo random binary sequence
(PRBS) tests on the air bypass valve (ABV) and measur-
ing the resulting speed output (Thornhill, 1998). Note
that this is a "nite impulse response since it is limited to
a "nite number of samples de"ned over a "nite range of
time intervals.
The calculated FIR is shown in Fig. 7. The y-axis in
this "gure has been scaled to give the change in speed for
an impulse of magnitude 1 step on the AB valve stepper
motor. The idle speed oscillations described in Section
1 can be seen to have a period of about 2s, these oscilla-
tions do not die away with time.
This FIR was used in Matlab/Simulink simulations as
the vector in a discrete "lter block.
4.2. Comparison of control schemes using MSE and ME
Fig. 8 shows a typical idle speed regulation test result
on an engine for PI control of the ABV and spark
advance with P"0.03 for both loops and I"0.03 for
the ABV loop. The top graph of Fig. 8 shows the speed
response, the middle graph shows the ABV motion and
the bottom graph shows the spark advance. Similar tests
were performed for all the aforementioned control
schemes. The MSE and ME were calculated for each test
and are tabulated in Table 1.
Table 1 shows the MSE values calculated from idle
speed regulation tests in simulation. The MSE values of
PI control and adaptive FLC control are not shown as
unstable control resulted in simulation. Instability result-
ed for PI control as the gains were tuned for operation in
conjunction with spark control. The instability that oc-
curred without spark control demonstrates poor loop
integrity. Instability in the case of adaptive FLC was
found to be a result of the system time delay in the inlet
manifold air to speed loop. The MSE values for adaptive
FLC with Smith prediction (SP) and DMC are excellent
and well within the speci"ed 203 rpm` value. From these
simulation results it appears that fuzzy logic control,
adaptive fuzzy logic with SP and DMC possess good
control integrity, i.e. acceptable idle control will result
even in the presence of constant spark advance.
M. Thornhill et al. / Control Engineering Practice 8 (2000) 519}530 525
Table 2
Comparison of computation and memory requirements for each con-
trol algorithm in simulation
Type of control scheme Floating point
operations
Number of
variables
PI (P"0.03, I"0.03) 10 3
Fuzzy 31 32
Adaptive FL 52 34
Adaptive FL#SP 1045 438
DMC 374204 6802
Table 3
Technical speci"cations for the test vehicle
Engine 1.8i 16 valve
Capacity (cc) 1761
Cylinders 4 in line
Bore and stroke (mm) 83;81.4
Max power DIN (hp/rpm) 112/5500
Max torque (lb ft/rpm) 114/4250
Cooling system Liquid with electric cooling fans
Fuel system Multipoint electronic fuel injection
Transmission 5-speed manual
Table 1 also shows the maximum error values cal-
culated from simulation. The trends observed in the MSE
"gures are repeated. The PI control and adaptive FLC
are unstable control and the other results are well within
the speci"ed upper 50 rpm ME bound for smooth
running.
4.3. Comparison of memory requirements and computation
time
Due to the limitations of the vehicles engine manage-
ment system microprocessor, memory and computation
time requirements were calculated for each control
scheme using MATLAB. Table 2 shows the comparison
of memory and computation time for each controller.
Computation time refers to the number of Floating Point
OPerations (#ops) required by each control algorithm
during each sample interval.
The memory requirements shown in Table 2 are taken
from simulation studies based upon the simulation re-
sults shown above.
It should be noted that the number of variables can be
increased or decreased for some of the above control
schemes. Therefore, the values quoted are the ones that
produced the simulation results reported in the previous
simulation sections.
4.4. Summary of simulation results
The results of the simulation work support the litera-
ture (Hrovat & Sun, 1996) in that it shows that the
existing PI control used in production automobiles can
be improved upon. The advantage to be gained from
improved ABV control is that less control e!ort will be
required from spark advance (SA) which in turn, can be
set at the MBT value and so o!er better fuel economy
and improvements in idle stability. The trade-o! is the
need for additional EMS memory and an increase in
#ops.
Surprisingly, these simulation studies have shown that
Wang's adaptive FLC is not suitable for idle speed con-
trol. However when used in conjunction with Smith
prediction this technique has been shown to produce
good performance indices (see Table 1). Elsewhere, it has
been shown that this adaptive FLC with Smith predic-
tion, when used for idle speed control, is robust to model-
ling errors in the delay model used for Smith prediction
(Thornhill, 1998). This suggests that the closed-loop sys-
tem is bene"ting from some phase advance in addition to
delay compensation (Santacesaria & Scattolini, 1993).
On the basis of these tests the DMC results appear
superior to all the other controllers. This is not surprising
in that there was an exact match between the plant and
the model. However, at the moment the memory require-
ments and number of #ops required would not make the
DMC a practical alternative.
5. Engine tests
The following engine test results, taken from test ve-
hicles, were performed on fully warmed up engines. For
the tests, which included spark advance, the existing
production spark controller was used. This consisted of
a simple proportional controller with the proportional
gain set at 0.03.
Due to memory and speed limitations of the engine
management system (EMS), it was impossible to test the
adaptive fuzzy logic controller with Smith prediction or
the DMC controller.
In all the tests the objectives are to "nd for
each controller the performance measures detailed in
Section 3.
5.1. The test vehicle
The test vehicle was a standard production vehicle
commonly found in Europe and having the technical
speci"cations given in Table 3. A Sagem s2000 Develop-
ment Engine Management System was used for the de-
velopment of all controllers.
5.2. Comparison of control schemes using MSE and ME
Table 4 shows the MSE values calculated from engine
tests using constant spark advance. This table shows
526 M. Thornhill et al. / Control Engineering Practice 8 (2000) 519}530
Table 4
Comparison of control schemes during engine testing with constant
spark advance
Type of control scheme MSE ME
PI (P"0.03, I"0.03) 42575 487
Fuzzy 815 83
Adaptive FL Unstable Unstable
Table 5
Comparison of control schemes during engine testing in conjunction
with proportional spark control
Type of control scheme MSE ME
PI (P"0.03, I"0.03) 203 50
Fuzzy 176 50
Adaptive FL (start) 363 47
Adaptive FL (end) 168 39
Fig. 9. Simulation responses for disturbance rejection of the fuzzy
controller with constant spark advance.
exactly the same trends as the simulations (Table 1) with
the exception that marginally stable control is achieved
for PI control of the ABV. Adaptive FLC with Smith
prediction and DMC are not shown, as it was not
possible to code the control algorithms due to EMS
limitations.
Table 4 also shows the maximum error values cal-
culated from engine tests. The same trends as seen in
simulation (Table 1) are shown here with the exception
that PI control does not go unstable but remains margin-
ally stable, this is probably due to the use of a linear model
in simulation to represent the nonlinear idle speed system.
Table 5 shows test results from the aforementioned
control schemes in conjunction with the spark control
scheme provided with the production vehicle tested. No
simulation results for these forms of control are available.
Although the adaptive FLC controller went unstable
with constant spark advance, once tuned it outperformed
the production controller. The di!erence between the
adaptive FLC value for MSEbetween the start and end is
50 s of adaptation. However, there is no guarantee that
this value would continue to improve (or stability be
maintained) but this form of control looks promising, as
no on-line tuning is required.
Table 5 also shows the maximum error values cal-
culated from engine tests. The values shown are very
similar with the lowest maximum error value shown
taken from the end of the Adaptive FLC test.
5.3. Disturbance rejection and entry to idle tests
This paper is primarily concerned with the comparison
of the ABV element of an IS controller. However for
completeness it seems appropriate to brie#y consider
disturbance rejection and entry to idle, which addition-
ally require spark advance control and various feed for-
ward loops. In all of the engine tests performed there was
no attempt to retune the proportional spark advance
controller supplied with the vehicles EMS or modify any
of the feed forward loops.
On the vehicle the disturbance rejection tests proved
problematic. Since all the controllers performed well it
proved impossible, from the available traces, to identify
with any certainty the initiation of a disturbance or the
resulting transient.
The simulation model is only valid for a constant
spark advance set at 10" BTDC. Therefore the only
controller that could be tested using simulation was the
fuzzy logic controller. Fig. 9 shows the fuzzy logic con-
trollers ability to reject a sudden drop in speed of
100 rpm, and Fig. 10 shows its ability to track a reference
signal into idle. Using linear controllers, if the engine
speed is suddenly required to drop from some high level
to idle (step input) then the speed will invariably under-
shoot unless the controller is heavily overdamped. To
avoid this some smooth curve, like the decaying reference
speed shown in Fig. 10, is normally produced and the
systems ability to track this curve examined.
Fig. 11 shows the response of the production vehicles
PI controller on entry to idle. In this test the throttle was
depressed and released suddenly. The speed response is
then allowed to decay naturally to 1500 rpm, at which
speed the idle reference speed starts and decays linearly
to the nominal idle speed of 864 rpm. A similar test was
performed with the fuzzy logic controller and produced
the responses shown in Fig. 12.
Some comments on the disturbance rejection and
entry to idle tests:
1. In the experimental test the PI controller when com-
bined with the proportional spark advance controller
M. Thornhill et al. / Control Engineering Practice 8 (2000) 519}530 527
Fig. 10. Simulation responses for entry to idle of the fuzzy controller
with constant spark advance.
Fig. 11. Engine responses for entry to idle of the PI controller.
Fig. 12. Engine responses for entry to idle of the fuzzy controller.
clearly performs well. This is to be expected since the
system was tuned for this controller combination.
However, the results would suggest that this controller
is at the limits of its development in that in order to
achieve the indicated performance loop integrity has
been compromised (that is failure in the spark loop,
resulting in a constant spark advance, may cause idle
speed regulation instability).
2. The fuzzy logic controller, Figs. 9 and 10, appears to
have good disturbance rejection and tracking proper-
ties even without spark advance control. This suggests
that a production ISC using a fuzzy logic controller
would be less dependent on spark advance and have
better loop integrity.
3. Although the simulation results shown in Fig. 10 have
been produced by the linear FIR model described in
this paper the results have been checked using a neural
network model validated over the engines operational
range. Both models show similar trends.
4. The experimental results for entry to idle using the
fuzzy logic controller, Fig. 12 are inconclusive. The
problem appears to be that the air bypass valve was
not resetting itself and therefore most of the control
e!ort is being provided by the spark advance loop.
Similar results and a similar problem were encoun-
tered when testing the adaptive fuzzy logic controller.
Note that the simulation results in Fig. 10 have as-
sumed that the air bypass valve is reset.
5.4. Summary of engine test results
Due to excessive memory and speed requirements (see
Table 2), it was found that some of the controllers could
not be tested on the engine.
The controllers tested were PI, fuzzy, and adaptive
fuzzy. Observations made during the tests and presented
in this paper indicate that:
1. The idle speed system at open loop naturally oscil-
lates; the frequency and amplitude of speed oscilla-
tions are dependent on the engine speed.
528 M. Thornhill et al. / Control Engineering Practice 8 (2000) 519}530
Table 6
Comparison of controllers from simulation results
Best control Adaptive Local
memory
Human
knowledge
Comp.
simplicity
Model
required
Integrity of
control
PI 5 1
Fuzzy 3 Y 2 Y
Adaptive FLC 6 Y Y Y 4
Adaptive FLC#SP 2 Y Y Y 5 Y Y
DMC 1 6 Y Y
2. Of the control schemes tested on the vehicle with
constant spark advance, only fuzzy control provided
acceptable idle speed control (as predicted by simula-
tion), from the point of view of idle speed regulation.
Given that the fuzzy logic controller is stable with
constant spark advance it was surprising that the
adaptive FLC was not.
3. Of the control schemes tested in conjunction with
spark control, all algorithms performed well with
fuzzy control and adaptive fuzzy control outperform-
ing the production controller in the baseline regula-
tion tests (see Table 5).
4. In simulation, fuzzy control, adaptive FLC with Smith
prediction and DMC showed integrity of control with
regard to failures in the spark loop that would cause
a constant spark advance signal.
5. The results from engine testing showed good agree-
ment in their general trends with the simulation re-
sults, giving con"dence in those simulations that could
not be tested on the vehicle e.g. adaptive FLC with
Smith prediction and DMC.
6. Discussion
As only some of the proposed controllers could be
tested on the vehicle, and the simulation and test results
correlated well, the following comparison is based on
simulation results. The simulation performance of each
controller, summarised in the above sections, are com-
pared in Table 6 under the following headings:
Best control: The controllers are ranked from best (1) to
worst (6) based on simulation results of ABV control.
Adaptive: All controllers that possess adaptation char-
acteristics are marked with Y.
Local memory of controller: All controllers that possess
local memory (Thomas & Armstrong-HeH louvry, 1995)
are marked with Y. This corresponds to adaptive con-
trollers only.
Incorporation of human knowledge: Whether or not
human rules can be included in the controller (e.g. fuzzy
logic).
Computational simplicity/ memory required: Controllers
are ranked from the simplest (1) to the most complex (6).
Model required: If the controller requires a model of the
system, it is marked with an Y.
Integrity of control loops: Whether the control with
constant spark advance is acceptable (Y).
Based upon Table 6, DMC control is cited as the best
control scheme based on the mean-square-error and
maximum error values calculated from simulation (see
Table 1). However in a real system, adaptive FLC with
SP would be the preferred control scheme, see advant-
ages of adaptive FLC with SPlisted in Table 6 that DMC
does not possess. Fuzzy control would also be a viable
alternative to the production controller (particularly as
computational simplicity is required with the current
EMS).
Further, fuzzy logic is a more natural scheme to deal
with non-linear systems due to its ability to approximate
non-linear functions to arbitrary accuracy (Wang, 1994).
The local memory of adaptive fuzzy logic allows very
di!ering control e!orts to be applied at di!erent areas of
the input space without a!ecting the control learned
elsewhere, this prevents unlearning of previous informa-
tion as di!erent regions of the input space are entered. It
was found during training of adaptive fuzzy logic systems
that it was necessary to use triangular input membership
functions as Gaussian input membership functions
spread over the entire input space causing unlearning of
information through leakage e!ects (i.e. every input
membership functions was always "red to some degree).
The local memory property ensured that the control
quality always improved over time. In conjunction with
a temperature input, the entire control surface for all
engine-operating temperatures could be learned adap-
tively.
Based on test results obtained from the engine with
constant spark advance, fuzzy control achieved the best
performance indices (see Table 1). This result however is
based solely on the particular choice of output member-
ship functions that were used for testing.
For test results in conjunction with spark control, the
adaptive FLC outperformed all other controllers and
was improving throughout the engine test. This was
a surprising result in that the same controller was unsta-
ble with constant spark advance (see Table 5).
M. Thornhill et al. / Control Engineering Practice 8 (2000) 519}530 529
7. Conclusions
E Five control schemes (PI control, fuzzy control, adap-
tive FLC, adaptive FLC with SP, Dynamic Matrix
Control (DMC)) have been tested in simulation and
compared in terms of performance for idle speed regu-
lation. The results of these comparisons are shown in
Table 1 using mean square error and maximum error
as performance measures.
E Application of adaptive fuzzy logic to control of the
ABV is presented. Use of adaptive fuzzy logic and
a constant spark advance results in unstable control
for all controller learning rates.
E Control of the ABV using adaptive fuzzy logic in
conjunction with proportional (P) control of the spark
advance results in superior control of idle speed regu-
lation relative to the production controller on engine
tests.
E Actual vehicle results comparing fuzzy logic, conven-
tional PI control and adaptive FLC (with constant
spark advance and with proportional spark control)
are presented and shown to broadly agree with the
simulation results.
E A novel control scheme based on adaptive fuzzy logic
and Smith prediction is introduced, this scheme shows
excellent promise in simulation.
E Using state of the art EMS technology, it was deter-
mined that DMC and adaptive FLC with SP could
not be programmed in this instance due to excessive
memory and speed requirements.
Acknowledgements
The authors wish to thank SAGEM, Birmingham for
supporting this research project.
References
Abate, M., & DiNunzio, V. (1990). Idle speed control using optimal
regulation. SAE paper no. 905008.
Cutler, C. R., & Ramaker, B. L. (1980). Dynamic matrix control:
A computer control algorithm. 86th national A.I.Ch.E. meeting, San
Francisco, CA, USA.
Dutton, K., Thompson, S., & Barraclough, B. (1997). The art of control
engineering. England: Addison-Wesley.
Jurgen, R. (1995). Automotive electronics handbook. New York, NY:
McGraw-Hill.
Lee, C. C. (1991). A self-learning rule based controller employing
approximate reasoning and neural net concepts. International
Journal of Intelligent Systems, 6, 71}93.
Nishimura, Y., & Ishii, K. (1986). Engine idle stability analysis and
control. SAE paper 860415.
Procyk, T. J., & Mamdani, E. H. (1979). A linguistic self-organising
controller. Automatica, 15, 15}30.
Santacesari, C., & Scattolini, R. (1993). Easy tuning of Smith predictor
in presence of delay uncertainty. Automatica, 29, 1595}1597.
Takagi, T., & Sugeno, M. (1985). Fuzzy identi"cation of systems and
its applications to modelling and control. IEEE Transactions on
Systems, Man, and Cybernetics, SMC-15(1), 116}132.
Thomas, D., & Armstrong-HeH louvry, B. (1995). Fuzzy logic control
* A taxonomy of demonstrated bene"ts. Proceedings of the IEEE,
83(3), 407}421.
Thornhill, M. (1998). Idle speed control of spark ignited engines. PhD.
thesis, The Queen's University of Belfast.
Wang, L. (1994). Adaptive fuzzy systems and control, design and stability
analysis. Englewood Cli!s, NJ: Prentice-Hall, ECS Professional.
Washino, S., Nishiyama, R., & Ohkubo, S. (1986) A fundamental study
for the control of periodic oscillation of SI engine revolutions. SAE
paper no. 860411.
Widrow, B. E., & Wallach, E. (1996). Adaptive inverse control.
Englewood Cli!s, NJ: Prentice-Hall, ECS Professional.
530 M. Thornhill et al. / Control Engineering Practice 8 (2000) 519}530

Das könnte Ihnen auch gefallen