Fuzzy Logic and Neuro-fuzzy Modelling of Diesel Spray

Penetration
S.H.Lee, R.J.Howlett, S.D.Walters and C.Crua
Intelligent Systems & Signal Processing Laboratories,
Engineering Research Centre, University of Brighton,
Moulsecoomb, Brighton, BN2 4GJ, UK.
Email: [email protected], [email protected],
[email protected] & [email protected]
Abstract. This paper describes a comparative evaluation of two fuzzy-derived
techniques for modelling fuel spray penetration in the cylinders of a diesel internal
combustion engine. The first model is implemented using conventional fuzzy-based
paradigm, where human expertise and operator knowledge were used to select the
parameters for the system. The second model used an adaptive neuro-fuzzy
inference system (ANFIS), where automatic adjustment of the system parameters is
effected by a neural networks based on prior knowledge. Two engine operating
parameters were used as inputs to the model, namely in-cylinder pressure and air
density. Spray penetration length was modelled on the basis of these two inputs. The
models derived using the two techniques were validated using test data that had not
been used during training. The ANFIS model was shown to achieve an improved
accuracy compared to a pure fuzzy model, based on conveniently selected
parameters.

Introduction

In a diesel engine, the combustion and emission characteristics are influenced by fuel
atomisation, nozzle geometry, injection pressure, shape of inlet port, and other factors. In
order to improve air-fuel mixing, it is important to understand the fuel atomisation and
spray formation processes. Researchers have investigated the characteristics of the spray
behaviour, formation and structure for the high-pressure injector by experimental and
theoretical approaches in order to improve the combustion performance and reduce
exhaust emissions. However, further detailed studies of the atomisation characteristics and
spray development processes of high-pressure diesel sprays are still relevant.
Intelligent systems, software systems incorporating artificial intelligence, have shown
many advantages in engineering system control and modelling. They have the ability to
rapidly model and learn characteristics of multi-variant complex systems, exhibiting
advantages in performance over more conventional mathematical techniques. This has led
to them being applied in diverse applications in power systems, manufacturing,
optimisation, medicine, signal processing, control, robotics, and social/psychological

sciences [1, 2]. Fuzzy logic is a problem-solving technique that derives its power from its
ability to draw conclusions and generate responses based on vague, ambiguous,
incomplete and imprecise information. To simulate this process of human reasoning it
applies the mathematical theory of fuzzy sets first defined by Zadeh, in 1965 [3]. Fuzzy
inference is the process of formulating a mapping from a given input value to an output
value using fuzzy logic. The mapping then provides a basis from which decisions can be
made, or patterns discerned. It has been proved that the system can effectively express
highly non-linear functional relationships [4]. Fuzzy inference systems (FIS) have been
successfully applied in fields such as automatic control, data classification, decision
analysis, expert systems and computer vision.
The Adaptive Neuro-Fuzzy Inference System (ANFIS), developed in the early 90s by
Jang [5], combines the concepts of fuzzy logic and neural networks to form a hybrid
intelligent system that enhances the ability to automatically learn and adapt. Hybrid
systems have been used by researchers for modelling and predictions in various
engineering systems. The basic idea behind these neuro-adaptive learning techniques is to
provide a method for the fuzzy modelling procedure to learn information about a data set,
in order to automatically compute the membership function parameters that best allow the
associated FIS to track the given input/output data. The membership function parameters
are tuned using a combination of least squares estimation and backpropagation algorithm
for membership function parameter estimation. These parameters associated with the
membership functions will change through the learning process similar to that of a neural
network. Their adjustment is facilitated by a gradient vector, which provides a measure of
how well the FIS is modelling the input/output data for a given set of parameters. Once
the gradient vector is obtained, any of several optimisation routines could be applied in
order to adjust the parameters so as to reduce error between the actual and desired outputs.
This allows the fuzzy system to learn from the data it is modelling. The approach has the
advantage over the pure fuzzy paradigm that the need for the human operator to tune the
system by adjusting the bounds of the membership functions is removed.
Many of the combustion problems are exactly the types of problems and issues for
which an AI approach appears to be most applicable and has the potential for making
better, quicker and more accurate predictions than traditional methods. The increasing
availability of advanced computer equipment and sensory systems, frequently results in
the production of large amounts of information-rich data, and there are often inadequate
means of analysing it so as to extract meaning. The aim of this investigation was to apply
intelligent systems tools and techniques to achieve an improved ability to analyse large
complex data sets generated during engine research in a semi-automated way. An
intelligent paradigm was created based on a fuzzy logic inference system combined with
conventional techniques.

2
2.1

Methods
Pure Fuzzy Logic Model

Fuzzy logic provides a practicable way to understand and manually influence the
mapping behaviour. In general, fuzzy logic uses simple rules to describe the system of
interest rather than analytical equations, making it easy to implement. An advantage, such
as robustness and speed, fuzzy logic method is one of the best solutions for system
modelling and control. A FIS contains three main components, the fuzzification stage, the
rule base and the defuzzification stage. The fuzzification stage is used to transform the socalled crisp values of the input variables into fuzzy membership values. Then, these
membership values are processed within the rule-base using conditional if-then
statements. The outputs of the rules are summed and defuzzified into a crisp analogue
output value. The effects of variations in the parameters of a FIS can be readily
understood and this facilitates calibration of the model.
The system inputs, which in this case are the cylinder pressure and the air density, are
called linguistic variables, whereas high and very high are linguistic values which are
characterised by the membership function. Following the evaluation of the rules, the
defuzzification transforms the fuzzy membership values into a crisp output value, for
example, the penetration depth. The complexity of a fuzzy logic system with a fixed
input-output structure is determined by the number of membership functions used for the
fuzzification and defuzzification and by the number of inference levels. A fuzzy system of
this kind requires that knowledgeable human operate initialise the system parameters e.g.
the membership function bounds. The operator must then optimise these parameters to
achieve a required level of accuracy of mapping of the physical system by the fuzzy
system. While the visual nature of a fuzzy system facilitates the optimisation of the
parameters, the need for it to be accomplished manually is a disadvantage.
2.2

ANFIS Model

ANFIS largely removes the requirement for manual optimisation of the fuzzy system
parameters. A neural network is used to automatically tune the system parameters, for
example the membership function bounds, leading to improved performance without
operator invention. In addition to a purely fuzzy approach, an ANFIS was also developed
for the estimation of spray penetration because the combination of neural network and
fuzzy logic enables the system to learn and improve its performance based on past data.
The neuro-fuzzy system with the learning capability of neural network and with the
advantages of the rule-base fuzzy system can improve the performance significantly and
can provide a mechanism to incorporate past observations into the classification process.
In a neural network the training essentially builds the system. However using a neuro-

fuzzy scheme, the system is built by fuzzy logic definitions and then it is refined using
neural network training algorithms.

Experimental Work

A large collection of spray data are generated using the Ricardo Proteus test engine.
These data comprised images depicting the spray patterns of diesel injection processes,
under selected conditions of relative pressure, nozzle size and type and in-cylinder air
temperature. The images representing time-varying spray under each relative pressure
condition were examined and processed using a thresholding technique whereby each
image representing the instant of maximum penetration length was then determined,
yielding a maximum penetration value which could be linked with its corresponding
relative pressure across the injector. The collected maximum spray penetration values and
corresponding relative pressures then formed a labelled data to be modelled by the FIS as
shown schematically in Figure 1.
Image database
Greyscale
images

Image t=0
Image t=1
Image t=2

Threshold to
monochrome

Pure Fuzzy Logic

Inference
System/Adaptive
Neuro-fuzzy
Inference System
(ANFIS)

Image t=n

Fig. 1. Schematic diagram of FIS modelling

Monochrome
images

Extract maximum
penetration length

Max. penetration value

Optimised
model

Diesel spray
model

3.1

Pre-processing

Raw penetration lengths were plotted against time under each relative pressure and
density condition. Polynomial fitting was employed to produce best fitted curves where
maximum penetration values can be depicted. These were combined into a vector with
which to train the ANFIS as shown in Table 1.
Table 1: Training data sets and results
Data set
1
2
3
4
5
6
3.2

Parameters
Relative pressure (MPa)

60
60
100
100
160
160

Density (kg/m3)

14
35
14
35
14
35

Measured penetration (mm)

53
32
52
38
54
36

Pure Fuzzy Inference Model

Figure 2 illustrates the fuzzy sets which were used in the pure fuzzy logic inference
system. There were two stages in the inference model, the in-cylinder pressure and the air
density; both stages are described in detail. The pressure and density range from 60MPa 160MPa and 14kg/m3 - 42kg/m3 respectively. Both in-cylinder pressure and air density
fuzzy sets used generalised bell-shaped membership functions for classes low, medium
and high. It was empirically selected based on the features of all data under consideration
although in many cases membership functions are fixed and somewhat arbitrarily chosen.
The process was carried out by examining the ranges of all data sets to determine where
the majority of points were located. The functions were also created to have an
approximately equal amount of overlap between each membership curve. Experimental
adjustment of the limits of the membership classes enabled the response of the model to
be tailored to the experimental output from the experimental data.
The rule structure is essentially predetermined by the users interpretation of the
characteristics of the input parameters in the model. The contents of these rule-base and
membership functions undertake many modifications as part of the process of heuristic
optimisation and in many cases it is a continuing process. Examples of the rules initially
contained in the rule-base for the pure fuzzy model are shown in Table 2.

Table 2. Fuzzy rule-base

IF Pressure = Low AND Density = Low THEN Penetration = Large
IF Pressure = Low AND Density = Med THEN Penetration = Small
IF Pressure = Low AND Density = High THEN Penetration = Small
IF Pressure = Med AND Density = Low THEN Penetration = Medium
IF Pressure = Med AND Density = Med THEN Penetration = Very Large
IF Pressure = Med AND Density = High THEN Penetration = Very Small
IF Pressure = High AND Density = Low THEN Penetration = Large
IF Pressure = High AND Density = Med THEN Penetration = Medium
IF Pressure = High AND Density = High THEN Penetration = Very Small

Fig. 2. Fuzzy sets

Fig. 3. Pure fuzzy logic model surface

plot

The fuzzifed values for the outputs of the rules were classified into membership sets
similarly to the input values. While the output membership functions may be trapezoidal
or triangular, in this case, an output singletons were used which has a compact form and
computationally efficient representation. The fuzzy output singletons were defuzzified to
a crisp value of penetration depth by means of the widely-used centre of gravity method.
The control surface in Figure 3 shows the crisp value of penetration depth at different
combinations of in-cylinder pressure and air density using a pure fuzzy logic model. Each
of these intersection points indicates the differing predicted value of spray penetration
depth, which is determined by the design of fuzzy sets, rule-base and membership
functions. The surface plot acts as a practical means of determining the output needed for
each combination of input parameters.

3.3

Neuro-fuzzy Model

A FIS was devised using Matlab based application, ANFIS. A neuro-adaptive

learning technique facilitated the learning of information about a data set by the fuzzy
modelling procedure, in order to compute the membership function parameters that best
allow the associated FIS to track the given input/output data rather than choosing the
parameters associated with a given membership function arbitrarily.
A Matlab programme was generated and compiled; The pre-processed input/output
spray vector matrix which contained all the necessary representative features was used to
train the FIS. Figure 4 shows the structure of the ANFIS; a Sugeno FIS was used in this
investigation. Figure 5 shows the fuzzy rule architecture of the FIS which consisted of 9
fuzzy rules. During training in ANFIS, 6 sets of pre-processed data were used to conduct
180 cycles of learning. Figure 6 shows the final membership functions under two different
air input conditions derived by training the generalised bell-shaped membership function.

Fig.
4.
structure

The

ANFIS

model

Fig. 5. Fuzzy rule architecture of the

generalised bell-shaped membership
function

Fig. 6. Fuzzy sets

Fig. 7. Surface plot showing

relationship between input and output
parameters

Results and Discussion

Table 3 shows the predicted penetration length obtained from the ANFIS. Figure 7
depicts a three-dimensional plot that represents the mapping from relative pressure and air
density to spray penetration length. As the relative pressure and air density increases, the
predicted penetration length increases in a non-linear piecewise manner, this being largely
due to non-linearity of the characteristic of the input vector matrix derived from the raw
image data. This assumes that these raw image data are fully representative of the features
of the data that the trained FIS is intended to model. However the data are inherently
noisy and training data may not always faithfully represent all the features of the data that
should be presented to the model. Therefore, the accuracy of the model will be adversely
affected under such circumstances.
4.1

Model Validation

The data in Table 3 was used to determine how well the FIS model could predict the
penetration length corresponding to various values of pressure and density. Figure 8
shows scatter plot of the measured and FIS modelled penetration length utilising six sets
of testing data. These two diagrams demonstrate that the predicted values are close to the
experimentally-measured values, as many of the data points fall very close to the diagonal
(dotted) line, indicating good correlation. Figure 9 shows similar comparisons between the
FIS-modelled and measured values of the penetration length using the same testing data.
Clearly the model created by ANFIS has a better agreement than the pure fuzzy logic
model. The correlation coefficient also suggested identical findings.
Table 3. Testing data and results
Data
set
1
2
3
4
5
6

Parameters
Relative pressure (MPa)

60
60
100
100
160
160

Penetration (mm)
Density (kg/m3)

28
40
28
40
28
40

Measured

33
35
40
29
40
30

Correlation coefficient

Pure Fuzzy Paradigm

30
28
40
23
39
21
0.971

ANFIS

33
35
41
29
40
30
0.997

Heated air

Heated air
45

Pure Fuzzy Logic

45

ANFIS
40
35
30

Pure Fuzzy Logic

ANFIS

35
30
25

25

20

20
20

25

30

35

40

45

Penetration measured (mm)

Fig. 8. Scatter plot of measured

penetration and predicted penetration
4.2

Measured

40
P e n e tra tio n (m m )

P e n e tra tio n m o d e l le d (m m )

50

Data set

Fig. 9. Comparisons between predicted

and measured penetration

Discussion

The ANFIS is a non-linear computational method that has potential for modelling
complex systems with unclear input to output relationships due to its ability to combine
fuzzy logic and system identification techniques in a hybrid manner. This type of system
has several advantages when assigned to applications in which only partial knowledge of
the system characteristics are known, as is typically the case with engineering systems.
Additionally, the ANFIS can rapidly identify important characteristics of the data, which
is an important and useful feature of models used for estimation purposes in IC engines
research. In the experiment, we have used an ANFIS to predict changes in diesel spray
penetration depth as a potential means to monitor impending changes in combustion
chamber and fuel injector design. As an initial step toward modelling and prediction with
an ANFIS for this particular application, it has proven very useful for short-term
prediction of penetration depth using engine operating parameters as the input.
The correlation coefficient reflects a models ability to predict the output based on the
input used. While both models performed fairly well and approximated the output
function to a reasonable extent, the ANFIS model exhibited improved performance in this
respect. Pure fuzzy logic models were conveniently constructed whilst the ANFIS
performed well in cases where the input to output relationships become more complex.

Conclusions

This paper demonstrated that fuzzy and neuro-fuzzy techniques can be used to model
diesel fuel spray penetration for an internal combustion engine, leading to convenient and
quick investigation on the effect of penetration length under different operating
parameters, including in-cylinder pressure, density, air temperature, etc. The pure fuzzy

logic and neuro-fuzzy system, ANFIS employed in this work are quick and robust. It has
been applied to sets of pre-processed raw diesel engine spray data and successfully
compared. The pure fuzzy logic model employed simple calibrated membership functions
and nine optimised rules to represent a diesel spray input/output mapping whilst the
neuro-fuzzy model has based on a total of six sets of experimental image data which were
used for training the FIS. Both devised models were validated by comparing the predicted
results against the experimental data. The correlation coefficient of the penetration length
estimated by ANFIS is 0.997. The pure fuzzy logic model has a smaller figure of 0.971
which suggested a poorer correlation with this model.
These fuzzy models set an example of how intelligent technique can be used in diesel
spray modelling. The system is very conductive to improvement and adjustment and it can
be fine-tuned and improved over time when more engine operating parameters become
available. Moreover, these techniques and idea can conveniently be extended to, and be
invaluable for, other combustion systems such as modelling and emission predictions in:
boilers, furnaces and incinerators. Also, for internal combustion engines, potential
applications include modelling and control of: spark ignition engines and gas engines.

References
1. Kalogirou S.A. Applications of artificial neural-networks for energy systems, Appl.
Energy 67 (2000) pp.1735
2. Xu K., Luxmoore A.R., Jones L.M., Deravi F., Integration of neural networks and
expert systems for microscopic wear particle analysis, Knowledge-Based Systems 11
(1998) pp.213227
3. Zadeh L.A. Fuzzy sets, Information Control 8 (1965) pp.338353
4. Wang L.X. Fuzzy systems are universal approximators. Proceedings of the IEEE
International conference on Fuzzy Systems, (1992) pp.1163-1170
5. Jang J. ANFIS: Adaptive network-based fuzzy inference systems, IEEE Transactions
on Systems, Man, and Cybernetics 23, (1993) pp.665-685
6. Howlett R.J., de Zoysa M.M., Walters S.D. and Howson P.A. Neural Network
Techniques for Monitoring and Control of Internal Combustion Engines, Int.
Symposium on Intelligent Industrial Automation 1999
7. Baba N. and Sato K. 1998. A Consideration on the Learning Algorithm of Neural
Network, Knowledge-Based Intelligent Electronic system (KES98), South Australia,
(1998) pp.7-12
8. Harris C.J., Brown M., Bossley K.M., Mills D.J., Ming Feng. Advances in Neurofuzzy
Algorithms for Real-time Modelling and Control, Engineering Applications of
Artificial Intelligence, Vol.9, Issue 1, (1996) pp. 1-16
9. Masters, T. Practical Neural Network Recipes in C++. Academic Press. London,
1993
10. Heywood J.B., Internal Combustion Engine Fundamentals. McGraw-Hill, 1988