Sei sulla pagina 1di 10

ASPECTS REFERRING TO THE

INTERNAL COMBUSTION ENGINES FUNCTIONALITY




Univ. Prof. eng. Ion COPAE PhD
Military Technical Academy, Bucharest email: ioncopae@hotmail.com



Abstract
Certain procedures and methodologies will be presented in this paper, which can be
used in studying the internal combustion engine. These procedures are useful to identify,
manage and diagnose the behaviour of the internal combustion engine.
So we are going to use the Artificial Intelligence (AI) algorithms such as: neural
network, fuzzy logic, neuro-fuzzy algorithms, and vector support machine. Also we are going
to highlight the importance of the entropic, complex theories and their specific procedures
that find their use in studying the internal combustion engine. We are also going to explain
how the extremal and symbolic analysis and multivariable statistics analysis will find their
way into better understanding the processes of the internal combustion engine.
We are going to present procedures that investigate the internal combustion engine
using fractals and fractal calculus, time analysis, frequency analysis, and also time-frequency
analysis. We will use in the paper elements of time-scale calculus, vector calculus, matrix
calculus and also elements of tensor calculus. All these methods are used in this paper to
investigate the engines dynamics and the fuel saving methods available for the todays
engines. Also the immense possibilities offered by MatLab software will be presented. This
software is used to investigate the engines functionality based on the three inseparable
components: identifying controlling diagnosing the engine.

Keywords
car engine, information theory, multivariate statistics, tensor calculus, extreme value
analysis, on-board computer

Since the new technologies were implemented into the construction of the
automobiles, the nowadays engines have become in their own turn, complex systems. In
order to study these engines you need to turn to new algorithms and methods. The ECU
(Engine Control Unit) installed on-board offers important information which can be read and
stored and afterwards used to study the internal combustion engines processes [4].
The automobiles engines were designed, at the beginning, based only on theoretical
knowledge available at that time. We had limited practical data to use. The technical
approaches in studying classical or electronically controlled engine were conditioned by the
technical level available. Most improvements were made only in theoretical certain areas that
relate to the engines processes. These improvements paved the way towards innovating and
powerful new tools, that permit us to communicate with the cars ECU and even control the
way it functions. We can now make important breakthroughs in both theoretical and practical
areas. Turning to new methodologies in studying the internal combustion engine we will also
turn to knowledge which is tackled by many other disciplines.
Studying the internal combustion engine has become much more systemic and also it
has reached a very high scientific level. In order to enhance the automobiles performances or
to reduce the vehicles emissions, the engines fast dynamic processes were studied in much
more detail.
Today, studying the internal combustion engine involves the following:
the studies are based on practical data and lead to relevant results; starting from the
experimental data we are establishing mathematical models that describe the internal
processes which, in their turn, are used as base for theoretical approaches that explain the
phenomena in detail. Of course the end purpose is to enhance the vehicles overall
performance.
the engines functionality is studied in an interdisciplinary manner. There are
concepts, procedures and algorithms specific for different disciplines such as mathematics,
system and signal theories, functional analysis, electronic control, mathematical statistics,
computer programming, information theory etc.
the engines functionality is also studied in a systemic way, considering the terrains
and drivers influences, using concepts and algorithms from systemic and electronic control
theories. The mathematical patterns that describe the processes allow us to set the values for
maximum performance (described by the engine when it functions at maximum load) but
most important it allows us to set the values for the partial regimes, which are most common
during the engines exploitation.
tackle issues that never before were discussed in the engines usual area of
discussions. These problems were mainly discussed when investigating different systems
dynamic and they refer to: variance analysis, sensibility analysis, correlation analysis,
coherent analysis, bispectral frequency analysis, robust analysis, spectral dynamic,
regressions, neuronal networks, genetic algorithms, fuzzy assemblage, neuro-fuzzy
algorithms, bootstrap techniques, Bayesian statistics, fractals and fractional calculus,
extremal analysis, symbolic analysis, vectorial analysis, matrix and tensor analysis etc.
apply all three techniques of data processing: frequency analysis, time analysis and
time-frequency analysis. On each case the processing algorithms diversified and became
more and more accurate. There are very many ways from where we can chose an
investigating procedure that allows us to understand and explain the details of the
phenomena.
use high scientific level algorithms and analysis procedures revealed by new
mathematical approaches of systems and signal theories. The fast improvement in numeric
calculus and computers also raises the technical scientific level. This level is also completed
by methods of investigation from AI and genetics areas.
to study dynamics and engines fuel consumption we can use new procedures and
algorithms that allow us to make these studies at the same time.
it uses a unitary approach of the engines functionality involving three inseparable
elements: identifying (establishing the mathematical patterns based on experimental data),
controlling and diagnosis.
Certain examples will be presented in the following pages, that best explain the
possibilities offered by todays theoretical and experimental internal combustion engine
study. So, from the vectorial calculus we know that the scalar product of two vectors is given
by the relation:
cos u v u v u =

(1)

where u is the smallest angle between vectors ( 0 u t s s ); if we know the two vectors then,
resulting from the anterior relation, we can determine the angle between these two.

cos ; 1 cos 1
u v
u v
u u

= s s


(2)

We can notice that the cosine value of the scalar product is situated in the same area
as the correlation coefficient ( 1 1 s s ); for this reason the scalar product is frequently
used to evaluate the indirect measurement of the linear correlation between vectors. To this
purpose, fig.1 shows the values of the scalar product between two vectors, described in the
figure, specific for Tacuma car. The vectors were obtained following a 50 experimental data
recording tests.



Fig. 1. Vector calculus

The graphs from fig.1 clearly show, among other things, the existence of some
unlinear correlations between certain values (best revealed by the differences obtained in the
7
th
and 9
th
tests). This has implications on setting the mathematical patterns that best describe
the internal combustion engine processes.
Information Theory is also applied to study the engines functionality; information
represents the fundamental concept in prediction and it is characterised by a probability
distribution p [2;6]. This way, Hartley has defined the information contained in n events
i
x X e with the help of the following relation:

2
( ) log ( )
i i
I x p x = (3)

As we can see from this relation, if ( ) 1 p x = , than ( ) 0 I x = ; this means that the
event that occurs once contains no information. In order to characterise the uncertainty
occurred during an event we can use the entropy concept: Shannon himself, the man who first
introduced this concept, used the term of uncertainty. The entropy represents the product
between probability and information throughout the all the n events
i
x X e :

1
( ) ( ) ( )
n
i i
i
H X p x I
=
=

x
i
(4)
or considering relation (3):
2
1
( ) ( )log ( )
n
i
i
H X p x p
=
=

x (5)

As mentioned before, the entropy represents the measurement of uncertainty; in
thermodynamics the entropy defines a systems disorder. We deduct from here that when the
entropy has a higher value, the uncertainty as well has a higher value, and this way the
prediction is situated at a lower value. Observing a systems evolution in time, we can say
that its entropy is maximum when the system is at a static level. So, a systems dynamics
described by a reduced entropy, ensures a better prediction then its static state.
Lets take two variables X and Y that have a joint probability density p(x,y). In this
case, the joint entropy of the two variables (also called comentropy) is calculated by a similar
relation with (5):

2
( , ) ( , )log ( , )
x y
H X Y p x y p x y =

(6)

Another useful concept is mutual information, described by the following relation:

( ; ) ( ) ( ) ( , ) I X Y H X H Y H X Y = + (7)

It represents a measurement of how much does the uncertainty of variable X is
reduced if we know variable Y. This is a concept that allows us to measure how does the
prediction level increases by evaluating how the uncertainty level decreases. The higher
values for mutual information, the uncertainty level is lower thus the prediction level is
higher.
For exemplification in fig.2 are presented the entropy H values and the mutual
information I
xy
for the observed parameters mentioned in the 50 tests which were carried out
on a Tacuma car. We can see, from fig.2a, that the engine speed n has the highest entropy.
The throttles position , has the lowest entropy. This always happened for all 50
experimental tests. So we can say that the first parameter offers a better prediction of
outcome results if we use a self regressive mathematical model (using only the engines
speed and different regressors).
Fig.2b illustrates engine torque and engine speed parameters, which explains the
mutual information concept. We can see that the mutual information offered by the pair of
parameters which were mentioned earlier has higher values than the mutual information
between throttles position and engine torque. If we establish two mathematical models for
engines functionality like M
e
=f(n) and M
e
=f(), we can say that the first model will have a
better prediction over the outcome of the engines torque than the second model.



Fig. 2. Values of entropy and mutual information

So if we want to establish mathematical models that best describe the engines
functionality and at the same time offer a high degree of precision, we need to choose those
parameters that have the highest mutual information. These parameters are called relevant
parameters.
As an example, lets establish a mathematical model that will give us the horary fuel
consumption C
h
(as a resulting/dependeble parameter) taking into account two factorial
parameters (independent parameters); the factorial parameters are spark timing |; inlet air
pressure p
a
, injection duration t
i
, engine speed n and the throttles position . Fig.3 shows in
intersections the parameters name and also entropys value and on the arches we can reed
the mutual information I
xy
.
As we can see from fig.3, the first two relevant parameters are engine speed (mutual
information is 2,3547 bits) and throttles position (mutual information is 1,3152 bits); so we
can now say that the highest precision is assured by C
h
=f(n, ) mathematical model.
Also we can see from the same fig.3 that if we wish to use a mathematical model
based on three independent parameters we need to turn to the C
h
=f(n, , p
a
) model because
the mutual information between inlet air pressure and horary fuel consumption is 1,2249 bits.
Finally, from the same figure we can see that the highest mutual information value is shared
by the spark timing and the engine speed (3.9093 bits) and the lowest mutual information is
shared by injection duration and spark timing (0,82094 bits).
Accordingly to what weve established, the highest precision is ensured by a
mathematical model which gives us the horary fuel consumption and takes into account the
engines speed and throttles position.


Fig. 3. Graph with values of entropy and mutual information

Indeed, from fig.4a we can see that the description offered by the mathematical value
is almost identical when compared with the actual data collected from the ECU.



Fig. 4. Generalized nonlinear mathematical models
As we can see from fig.4a the mathematical model is:

5 2 7 2
0,156 0, 0889 0, 00012 3,11 10 3, 6 10
h
C n

= + + + n

(8)

This model is valid for all 50 tests weve had available (obviously nonlinear, we can
say that this is a generalised mathematical model).
In fig.4b, there was a mathematical model, as a comparison, that took into account the
throttles position and the inlet air pressure; as expected the mathematical model gives an
error of 10,4% when compared to the real values collected from ECU. This error is an
unacceptable value.
We can study the functionality of the internal combustion engine using concepts from
multivariable statistics. Multivariable statistics, also known as spatial statistics or
geostatistics, is a part of statistics that deals with large sets of data that vary during time
classic statistics - but also in space this is where the two last concepts were introduced. The
first concept, used more frequently, comes from the fact that it deals with different type of
data or multiple parameters [3;7].
Multivariable statistics began being applied in the automotive industry when the
electronic control was first introduced. This was almost a necessity because the on-board
computer had to operate with large sets of data collected from all sensors installed by the
manufacturer. Multivariable statistics turns to multiple correlation and regressions, cluster
(group) analysis, discriminant analysis, principal components and factor analysis and it uses a
classification and establishes datas pattern [1;5;8].
Te emphasise this last statement, we can see in fig.5 how the classification algorithms
and datas pattern establishment is used when studying the internal combustion engine; the
graphic shows how some inlet pressure losses in the case of a Cielo Executive engine, divides
into classes. It reveals that, specific inlet pressure losses can be established by classification
and by data pattern recognition; in other words we can recognize and isolate different engine
malfunctions, in this case the range of engines speed (intentionally provoked pressure losses
that lead to engines speed alteration).



Fig. 5. Data pattern and classification
Tensor analysis also lets us emphasize certain aspects of engines functionality. This
analysis (also known as multimodal analysis) is an extension from matrix analysis which in
its turn is an extension of vector analysis. In fact their names come from the fact that vector
analysis deals with vectors, matrix analysis deals with matrix and tensor analysis deals with
tensors [9;10;11].
As an example, the dynamic study is shown in fig.6 which took into account the
position of the air dumper plates position, the engines speed and its power. The parameters
were measured during 10 tryouts which were carried out on a Tacuma vehicle. The criteria
for analysis is the 2 norm, marked as L
2
(because the 2 norm square it represents the energy).
In fig.6a we have the results from tensor analysis (all parameters were taken into account at
the same time) and in fig.6b we have the results from vector analysis (the parameters were
taken into account separately). From these two representations we can see that tensor analysis
gives us a different classification from vector analysis. The most suitable classification is that
of tensors analysis (T6, T5, T8, T1, T2 etc.), because it is analysing the systems behaviour
considering the variation of all three parameters at same time. This fact isnt the case for
vector analysis.



Fig. 6. Dynamic study

In many practical situations, the average values for the engines parameters have less
importance then their extreme values. So we are more interested in the engines maximum
power, maximum fuel consumption or maximum noise it produces etc. In other words we are
much more interested in the maximum and minimum values that our investigated parameters
will show.
Classic statistics can not be applied in the extreme cases we have mentioned because
probability theory relies on operating with statistical average and this is the reason we need to
turn to extreme value analysis. This theory is based not on the usual classic distribution but
on the probability asymptotic distribution. Because the extreme values are particular cases,
the theory relies on very rare events. The German mathematician Gumbel set the basis for
this theory.
To exemplify, in the fig.7 are presented the experimental values and the extreme
values for a Tacuma engines level of noise.
Fig.8 shows a detail with all three anterior sets of data.



Fig. 7. Experimental data and extreme values of the noise level




Fig. 8. Experimental data and extreme values of the noise level, detail
Extreme value theory relies on the generalised distribution defined by probability
density, for a certain parameter x:

1 1/ 1/
1
( , , , ) 1 exp 1
x x
f x


o
o o o

(

| | | |
= + +
(
| |
\ . \ .
(

(9)

where is the positioning parameter, o is the scale parameter and is the shape
parameter.
To study the engine today means that we can turn to different processes and
algorithms that are useful in dynamic systems and signal theories. The scientific levels are
very high and it allows us to better understand the intimacy of the fast dynamic processes that
occur inside the internal combustion engine.

References
1. Balakrishnama S. Linear Discriminant Analysis. Institute for Signal and
Information Processing, Mississippi State University, 1995
2. Blahut R. Principles and Practice of Information Theory. Addison-Wesley,
Cambridge MA, 1988
3. Clarke B. Analysis of Multivariate Time Series Data. Colorado State University,
2003
4. Copae I., Lespezeanu I., Cazacu C. Dinamica autovehiculelor. Editura ERICOM,
Bucureti, 2006
5. Duda R. Pattern Classification, Wiley-Interscience, New York, 2001
6. Gray R. Entropy and information theory. Stanford University, New York, 2007
7. Hytyniemi H. Multivariate Statistical Methods in Systems Engineering. Report
112, Helsinki University of Technology, 1998
8. Jain A. Statistical Pattern Recognition. Departament of Computer Science and
Engineering, Michigan State University, 1999
9. Martin Carla D. The Rank of a Tensor. James Madison University, 2006
10. Moravitz Martin Carla D. Tensor Decomposition Workshop Discussion Notes
American Institute of Mathematics (AIM) Palo Alto, CA. Cornell University, 2004
11. Smirnov A.V. Introduction to Tensor Calculus. McGraw-Hill, New York, 2004
12. Welling M. Extreme Components Analysis. Department of Computer Science,
University of Toronto, 2002

Potrebbero piacerti anche