You are on page 1of 5

Real Time Supervision of Diesel Engine Injection

with RBF-based Neural Networ


S. LEONHARDT*, N. GAO* and V. KECMAN**

e-mail: LEOC3IRT.RT.E-TECHNIK.TH-DARMSTADT.DE

**
" Department of Control Engineering Department of Electronics
Technical University of Darmstadt, Germany School of Technology, Heilbronn, Germany
FAX : ++(49) 6151-293445

Abstract
This paper deals with real time supervision of turbo Diesel combustion
engines based on acquisition and evaluation of cylinder pressure signals.
The idea is to subtract a "reconstructed" towed pressure from the fired
pressure signal which results in a contrast amplification. From the resul-
ting difference pressure signal, significant features are extracted. By means
of a RBF (radial basisfinctions) neural network, these features are map-
ped on injected fuel mass and injection angle. By comparison with the
corresponding specified input signals, injection faults can de detected.
With the presented method, it becomes possible to separate fuel mass
problems from injection angle failures. The concept has been implemented
on a dynamic engine test stand.

1. Introduction is very important. Proper fuel injection is a crucial pro-


blem as injection pressure is very high resulting in a
Diesel engines are widely used in automotive industry. heavy load for all parts involved.
In engine design, low fuel comsumption, low emission So far, several methods for fault detection of engine
and long-term durability are currently key issues. injection have been developed. One way is to monitor
From a systems engineering viewpoint, a Diesel the flywheel speed signal, see [l] - [3].By this procedu-
engine has two inputs (fuel mass m, and injection angle re, it is possbile to detect severe fuel mass problems or
e,) while the torque T&, soot and the undesired exhaust cylinder misfire. However, for proper monitoring of the
gases form output signals. Note that engine speed n is a injection angle (which is very important to control emis-
function of torque and load dynamics and thus not an sions), more accurate methods are required. Thus, in this
independent variable. Fig. 1 gives the structure of a paper, we will present a concept which is based on cy-
typical car engine. linder pressure analysis. While having the disadvantage
of needing additional sensors, this method allows to
extract detailed information on the combustion process.
The concept is supported by measurements obtained
from a dynamic test stand at TU Darmstadt running a
1600 ccm Dieael engine (4 cylinderg), [4].

2. Fault Detection and Diag~osis

During the last decade, there has been extensive research


activity on theory and applications of fault diagnosis for
dynamic systems. A growing number of references is
available, see e.g. [SI - [7]. Fig. 2 shows an overal state
Fig. 1 : structure of Turbo-charged Diesel engine of the art scheme of knowledge based fault detection and
diagnosis. In this concept, the four main tasks may be
For optimal driving of a Diesel engine (i.e. with maxi- subdivided in (1) fault detection by analytic and ( 2 )
mum efficiency and low exhaust emission), on-line heuristic symptom generation, (3) symptom integration
control and diagnosis of the Diesel combustion process and (4) fault diagnosis.
2128
Analytic

Hauristi

Fig. 2 : knowledge based fault and diagnosis scheme, from [7]

In this aper, w will focus on fault detection with ana- engines, several approaches for enb-.ie control and dia-
lytic symptom generation, especially on contrast amplifi- gnosis based on pressure information have already been
cation, feature extraction and reconstruction of input proposed including air-to-fuel ratio estimation and spark
variables by using a radial basis function (RBF) neural advance control by Powell and co-workers (see [8] for
network, Fig. 3. references), spark advance control by pressure ratios [9]
and cylinder knock control, [lo]. There are fewer results
on Diesel engines, see e.g. [I 11.

3.1 Constructing the "Difference Pressure"


The idea of the presented approach is to subtract a (re-
constructed) "towed pressure signal" from the measured
cylinder pressure signal.

In the ideal non-firing case, pB,, of an engine is (al-


most) symmetric to Top Dead Centre (TDC). Fuel injec-
tion changes that signal significantly, see Fig. 4.
Fig. 3 :overall concept of injection montoring

By comparison of the specified (desired) input variables


with the reconstructed signals, a fault detection leading
to symptoms like "burned fuel mass lower than expec-
ted" or "injection too early" may be performed. Howe-
ver, details on this last and easy step will be omitted in
this paper.

3. Feature Extraction I
-150 -100 60 0 50 100 150
crank shaft angle [deg. a. TDC]
Combustion is a complicated, nonlinear and stochastic
process. The most detailed information on motor condi- Fig. 4 :cylinder pressure curves, 2400 rpm, mB = 10 mg, eSB
= 16 deg. b. TDC
tion and combustion performance can be extracted by
computing the combustion heat curve. However, real
time conditions are difficult to match due to high com- Note that only the pressure difference carries information
putational costs. In addition, very precise pressure sen- on input variables m, or eSB. In a signal theoretical sen-
sors would be required. se, using Apcy, results in a contrast amplification. The
Therefore, it was decided to avoid energy computa- proposed method has some similarities with the pressure
tions and use the pressure signal directly. For IC (Otto) ratio algorithm published by Matsekunas ([SI). However,

2129
the advantage of using ApLylinstead of the pressure ratio + the centre of gravity (ec,~t,)
is that one becomes independent of inevitable sensor- BDC BDC
offset. In addition, the contrast amplification is higher.
Since the method is based on signal analysis rather than
18Apd0 [ Ap'de
Bbc BbC
energy computations, even minor static nonlinearities of e, = BDC 9 x c = BDC (4)
the pressure sensor would be acceptable. Some theoreti-
cal considerations on the method have already been pu- BDC BDC
blished, [ 121.
One remaining problem is to construct the towed the peak difference pressure (epeak,
nPak)
pressure. In our approach, we used the cylinder pressure + the secant length Bi
signal prior to injection and reflected it around TDC.
The missing segment was filled by a numerical function
approximation
for Ap(6i) 2 Ap, with i = 4, 10, 16, ... Fig. 7 - 9 show
feature maps at 2100 rpm. Note that there is some de-
pendency on engine speed, but less than for uncharged
engines.
. . . . .

for 8, +8,D > 18 I. Parameters a and b were estimated


on-line by a least squares fit. Fig. 5 shows measured and
reconstructed towed pressure curves at 2400 rpm.
. . ... ... ... ... ..
. .

i h . s . . ~ a l
_..............i....... i. .. 50 15 20 25 30 35 40 45 50
rf! 40 ... .
..
.
.. 8, [deg. a. TDC]
Fig. 7 :centre of gravity map of Ap+ 2100 rpm

5 0 1
. . . . . . . 6 ..........:............j.......p5.ing.......j ...........j. ........
l , , , . , . l . I .

100 0 -100
crank shaft angle [deg. a. TDC]
Fig. 5 : reconstructed and measured towed pressure

The error between measured signal and reconstruction


was usually not more than 5%.

3.2 Feature Extraction


For data reduction, we intended to characterize Apqr by em[deg. a. TDC]
numeric values. Several features come to mind, see Fig. Fig. 8 :Ape, maximum location map, 2100 rpm
6.

50- I. :. I.. :.. I.. I.. I.. I..


45 1...... ........ .......:......
~

. ~ f ....... f .......f ....... ........


. ~

TDC
crank shaft angle [deg. a. TDC]
. . . . . . . .
PI6 -;6 -;4 -12 -io 4 4 d -2
Fig. 6 : ApJe) and extracted features [deg. b. TDC]
Fig. 9 : secant length e,, at 2100 rpm

2130
4. Symptom Generation 5. Experimental Results

Generally, artifical neural networks perform some sort of 5.1 Constant Speed Operation
nonlinear mapping. While perceptron type networks are The RBF net was trained with the data setup
especially suited for discrete mappings (e.g. classifica- mB imgl : 5, 10, 15, 20, 25
tion, see [13]), RBF networks may be used for conti- 8, [deg. b. TDC] 4,7, 10, 13, 16
nuous mappings, [14], [15]. This type of network recent- n [rpm1 1200, 1500, 1800, ..., 3600
ly drew a lot of interest for application in control and Reconstruction performance was tested with
identification, see e.g. 1161. A RBF network is a two- mB [mgl 7, 12, 17, 22
layer net in which the first layer has radial basisfinct- ,e [deg. b. TDCI 6,10,14
ions while the second consists of linear functions, Fig. [rpml 1600, 2000, 2400, 2800
10.
n1 By comparing reconstructed fuel mass and injection
angle with the specified values for reference, the predic-
tion performance can be evaluated, Fig. 11 and 12.

mut Hidden Layer output


Fig. 10 : topology of RBF network

Neural networks learn by adapting their weights when -a 2


confronted with certain training data (reference vectors). I..
Assume that the mapping to be learned is R""+ R"*and " injection dge [deg.b. TDC]
that n, reference vectors are available. Let Fig. 11 :prediction performance for mB,2400 rpm
ai = -
1 (X-X,,,>'.( w"')"(x-x,&,) (6)
2

be the scalar argument of the i-th radial basis function


(Gauss curve) given by

i = 1 .. n,. Note that by computation of a, one actually


evaluates the squared geometric distance of X to the
different reference vectors XRpWhile the weight matrix
WI" is specified in advance (the geometric interpretation
of this matrix is "shaping the variance of the Gauss
Fig. 12 :prediction performance for ,e, 2400 rpm
curves"), W1'] is adjusted during learning. Details on the
learning procedure are omitted due to space restrictions,
but can be found in e.g. [16], [17]. The resulting map-
ping is given by
5.2 Dynamic Operation
The Diesel engine was driven with a simulated vehicle
load on the dynamic engine test stand ([4]). Specified
It was found that using five input and two output varia- fuel mass and injection angle were measured. As an ex-
bles, i.e. In,, 8, 8,,eak,np& n ) + ImB, gave good ample, an acceleration experiment including gear shifting
performance. Other combinations of features are under from 1st gear up to 5th gear is presented in Fig. 13 und
investigation. 14. Note that the reconstructed inputs usually were wit-
hin 10 % deviation. The "noise" on the reconstructed
signals is mostly due to pressure oscillations within the
EDC and the statistics of the combustion.
2131
Acknowledgments
The contributions of Lothar Niemietz ([17]) and Ralf
fuel mass [mg] engine speed [rp"
Schwarz ([ 1811 to this project are gratefully acknowl-
edged.

15 0 References
10
5 Brown, T.S. and Neilf, W.S., "Determinationof Engine Cylinder
0 Pressure from Crankshaft Speed Fluctuations", SAE Paper No.
0 20 40 60 80 100 920463, 1992.
Rizzoni, G., "Diagnosis of Individual Cylinder Misfires by
time [sec] Signature Analysis of Crankshaft Speed Fluctuations", SAE Paper
No. 890884, 1989.
Fig. 13 : reconstruction of fuel mass mB
Fiihrer, J., Sinsel, S. and Isermann, R., " Erkennung von
Ziindaussetzern aus Drehzahlsignalen mit Hilfe eines
Frequenzbereichverfahrens", 13. Tagung Elektronik im
Kraftfahrzeug, Essen, Germany, 1993
injection angle [deg. b. TDC] Leonhardt, S., Schmidt, C., Voigt, K., Isermann, R., "Real-
engine speed [rpm1
time Simulation of Drive Chaines for Use in Dynamical Engine
Test Stands", American Control Conference, Chicago, IL, June
-6 3000 24-26th, 1992.
-8 2000 Patton, R., Frank, P.M. and Clark, R., "Fault Diagnosis in
-10 1000 Dynamic System - Theory and Applications", Prentice Hall Int.,
-12 0 Hemel Hempstead, 1989.
Frank, P.M., "Advancesin Observer-basedFault Diagnosis", Int.
-14
Conf. on Fault Diagnosis (Tooldiag' 93), Toulouse, France, April
5 - 7th, 1993.
0 20 40 60 80 100
Isermann, R.,"Integration of fault detection and diagnosis
time [sec] methods", IFAC Symposium SAFEPROCESS '94, Helsinki,
Finland, June 1994.
Powell, J.D., "Engine Control using Cylinder Pressure: Past,
Fig. 14 : reconstruction of injection angle eSs Present, and Future", Journal of Dynamic Systems, Measurement,
and Control, Vol 115, pp. 343 - 350, June 1993.
Matsekunas, F.A., "Engine Combustion Control with Ignition
The software required 12 msec overall computation time Timing by Pressure Ratio Measurement", U.S. Patent No. 4,62-
per crankshaft rotation and cylinder running on T800 2,939, Nov 18th, 1986.
transputers (32 bit microcontrollers) in OCCAM. Sawamoto, K., Kawamura, Y., Kita, T. and Matsushita, K.,
"Individual Cylinder Knock Control by Detecting Cylinder Pres-
sure", SAE Paper No. 871911, 1987.
Reddy, P.R., Krishna, D.M, Mallan, K.R.G. and Ganesan, V.,
6. Conclusion "Evaluation of Combustion Parameters in Direct Injection Diesel
Engines - an easy and relieable Method, SAE Paper No, 930605,
A fault detection scheme for monitoring combustion 1993.
quality and proper injection pump operation has been Ludwig, C., Leonhardt, S., Ayoubi, M. and Isermann, R.,
"Measurementand Monitoring of Pressure Curves in Diesel Engi-
presented. A method to obtain a difference pressure nes", American Control Conference, San Francisco, CA, June 2-
signal was proposed. Significant features were extracted 4th, 1993
for data reduction purposes. By using a RBF neural Lippmann, R.P., "An Introduction to Computing with Neural
network, these features were mapped back on the Nets", IEEE ASSP Magazine, No. 4, pp. 4 - 22, 1987
specified (desired) input values allowing comparisons. If Moore, B., "Theory of Networks for Learning", SPIE
Applications of Artificial Neural Networks, Vol 1294, pp. 22-30,
deviations higher than the inherent modelling error occur 1990.
(more than & 1.5 mg or +. 1 deg.), symptoms like "fuel Poggio, T. and Girosi, F., "Networks for approximation and
mass lower than specified" or "fuel angle too early" may learning", Proceedings of the IEEE, Vol. 78, No 9, pp. 1481-
be generated andlor deviations may be displayed 1497, Sept. 1990.
quantitatively. Sanner, R.M. and Slotine, J.-J., "Stable Adaptive Control and
Recursive Identification using Radial Gaussian Networks", 30th
Note that the presented concept may also be viewed Conf. on Decision and Control, Brighton, England, Dec. 1991.
as a engine observer which uses states (pressures) to Niemietz, L., "Vergleich von Verfahren zur Rekonstruktion
estimate inputs (fuel mass and injection angle). physikalischer Parameter an einem Viertelfahrzeug",
In this application, RBF-type neural networks proved to Diplomarbeit, TU Darmstadt, 1994.
Schwarz, R., "Echtzeitfehlerdiagnose am Turbo-Dieselmotor",
be a feasible tool for continuous mappings. As an
Studienarbeit, TU Darmstadt, 1994
alternative, CMAC-type neural networks may be used.
Research on this topic is in progress.
2132

You might also like