You are on page 1of 8

Fuel 81 (2002) 1243±1250

www.fuel®rst.com

A neural network approach to the prediction of diesel fuel lubricity q


D.M. Korres, G. Anastopoulos, E. Lois*, A. Alexandridis, H. Sarimveis, G. Bafas
Department of Chemical Engineering, National Technical University of Athens, Iroon Polytechniou 9, Zografou Campus 15700, Athens, Greece
Accepted 12 December 2001; available online 8 February 2002

Abstract
The continuous sulfur reduction in diesel fuel has resulted in poor fuel lubricity and engine pump failure, a fact that led to the development
of a number of methods that measure the actual fuel lubricity level. However, lubricity measurement is costly and time consuming, and a
number of predictive models have been developed in the past, based mainly on various fuel properties. In the present paper, a black box
modeling approach is proposed, where the lubricity is approximated by a radial basis function (RBF) neural network that uses other fuel
properties as inputs. The HFRR apparatus was used for lubricity measurements. In the present model, the variables used included the diesel
fuel conductivity, density, kinematic viscosity at 40 8C, sulfur content and 90% distillation point, which produced the smallest error in the
validation data. q 2002 Elsevier Science Ltd. All rights reserved.
Keywords: Lubricity; Diesel; Neural networks

1. Introduction A problem that arises in all the modeling work, is that


although a relation between lubricity and a number of other
The measurement of fuel lubricity has become very fuel properties has been proven to exist, it is not easy to
signi®cant for the re®neries, since the introduction of low express it in the form of a single equation. The objective in
sulfur diesel fuels. This is due to the fact that fuels with this work was to develop a neural network black box model,
lower sulfur content tend to have poor lubricity [1]. able to predict the lubricity of diesel fuels, using as inputs
However, direct measurement of lubricity is a rather costly other fuel properties that can be easily measured.
and time consuming procedure, rendering essential the Neural networks are powerful modeling tools that have
development of methods that can predict fuel lubricity with- the ability to identify underlying highly complex relation-
out having to run bench tests. ships from input±output data only [13]. Jamialhamadi and
Several researchers have attempted to relate fuel lubricity Javadpour [14], used a neural network to correlate rock
with some physical and chemical properties of low sulfur permeability with the depth and the porosity of an oil reser-
diesel fuels. Wei and Spikes [2] attempted to correlate lubri- voir. Neural networks have also been used successfully in
city with fuel properties such as cloud point, viscosity, the static and dynamic modeling of diferrent complex
sulfur content, di-aromatics and poly-aromatics content. chemical engineering systems such as distillation columns
Lacey and Westbrook [3] surveyed lubricity of fuels [15] and chemical reactors [16,17]. In the present work, a
purchased by military installations using the S BOCLE. special type of neural network that employs radial basis
They concluded that lubricity of unadditized fuel was direc- functions (RBF) in the hidden layer, has been utilized.
tionally correlated with di- and poly-aromatic content, and Compared with the other neural network architectures,
with viscosity to a lesser degree. Wang and Cusamo [4] RBF networks possess certain advantages that make them
employed SL BOCLE to evaluate lubricity of low sulfur very popular. The most important of these advantages are
diesel fuels and found that the lubricity of low sulfur fuels the simplicity of the network structure, and the high speed of
is determined primarily by the viscosity and the di-aromatic convergence during the training phase [18].
content [5±12]. In the present model, the variables eventually used
included the diesel fuel conductivity, density, kinematic
* Corresponding author. Tel.: 130-1-772-3190; fax: 130-1-619-7750.
viscosity at 40 8C, sulfur content and 90% distillation
E-mail address: elois@orfeas.chemeng.ntua.gr (E. Lois). point, which produced the smallest error in the validation
q
Published ®rst on the web via Fuel®rst.comÐhttp://www.fuel®rst.com data.
0016-2361/02/$ - see front matter q 2002 Elsevier Science Ltd. All rights reserved.
PII: S 0016-236 1(02)00020-0
1244 D.M. Korres et al. / Fuel 81 (2002) 1243±1250

networks (RNNs) and the radial basis function networks


(RBFs). As far as data categorization is concerned,
Kohonen's self organizing maps (SOMs) is a typical
neural networks methodology.
The development of a neural network black box model
consists of two stages: primarily, a set of input±output data
is used to `train' the neural network. The objective is to tune
the free parameters of the network, so that it can `learn' the
underlying input±output relationship. In the next stage, a
second set of input data is fed to the network, and
the model is validated based on the accuracy of the new
predictions.
The typical structure of an RBF neural network can be
seen in Fig. 1. It consists of three layers, namely the
input, hidden and output layers. Each layer consists of
a number of neurons (nodes). The nodes in the input
layer are used only to pass the input data to the hidden
Fig. 1. Typical structure of an RBF neural network. layer. No calculations are performed in the input layer
nodes, and the connections between the input and the
output layer are not weighted.
2. Radial basis function (RBF) neural networks The hidden layer contains k nodes, which apply a
nonlinear transformation on the input variables. A graphical
2.1. Neural network topology representation of a hidden layer node is depicted in Fig. 2.
More speci®cally, each node j ˆ 1,2,¼,k has a center value
Neural networks are powerful mathematical algo- cj, where cj is a vector whose dimension is equal to the
rithms, which mimic the human brain's ability to number of inputs to the node. For each new input vector
learn. Generally, neural networks are used by engineers x ˆ [x1,x2,¼,xN], the Euclidean norm of the difference
either for building black box models when ®rst princi- between the input vector and the node center is calculated
ples equations cannot be easily developed, or for clus- as follows:
tering and categorization of data. In all types of
arti®cial neural networks, the basic structure consists v
u N  2
of a number of interconnected processing units, called uX
nj …x† ˆ cj 2 x ˆ t xi 2 cj;i …1†
neurons. The calculations that each neuron performs, iˆ1
along with the way that the neurons are interconnected,
determine a particular neural network type. The most The output of the hidden layer nodes is determined by a
common types used for model building are the feedfor- nonlinear function, called the `activation function' of the
ward neural networks (FNNs), the recurrent neural network. A typical selection for the activation function is

Fig. 2. Structure of a hidden layer node.


D.M. Korres et al. / Fuel 81 (2002) 1243±1250 1245

Table 1 of the Euclidean distances between each cluster center cj


Standard test methods and the training points assigned to the particular cluster:
Fuel property Measurement method k X
X m

Eˆ Bji cj 2 xi …5†
Lubricity CEC F-06-A-96 jˆ1 iˆ1
Conductivity ASTM D-2624
Density at 15 8C ASTM D-1298 where m is the number of training examples that will be
Flash point ASTM D-93 presented to the network, and Bji is a k £ m membership
Viscosity at 40 8C ASTM D-445
Sulfur content ASTM D-5453
matrix which contains only 1's and 0's as elements, with
Cloud point ASTM D-2500 exactly one 1 per column. A detailed description of the
Cetane index ASTM D-4737 algorithm would be out of the scope of this paper. The
Pour point ASTM D-97 interested reader could refer to Leonard et al. [19] and
Distillation ASTM D-86 Darken et al. [20].
The ®nal step of this phase is to determine the width s of
the Gaussian function: the Gaussian activation function. For every node j the width
! s j is calculated via:
n2  X 
f … n† ˆ 2 2 …2† 1 p 1=2
s sj ˆ cj 2 ch 2 …6†
p hˆ1
where the width s is a real constant. Thus, the output of the
where c1,c2,¼,cp are the p nearest node centers to the hidden
hidden layer node j is:
node j. Usually the parameter p is kept at low values.
z j ˆ f …nj † …3† Having the network structure at hand, the second phase
consists of determining the connection weights w. This is
A set of synaptic weights wj, j ˆ 1,2,¼, k is applied to the done through the use of multiple linear regression, where the
connections between the hidden and the output layer. The vector of the real output data Y ˆ [y1,y2,¼,ym] is regressed
nodes in the output layer serve only as summation units, on the outputs of the hidden nodes z:
which produce the ®nal output of the network. Considering  
a one-output network, Fig. 1, the overall output y^ will be: w ˆ inv zT z zT YT …7†
X
k
After the completion of phases (A) and (B), all the par-
y^ ˆ wj zj …4†
jˆ1
ameters of the RBF network are fully determined, and the
network is ready to be used as a predictor. Each time a new
input vector x is available, the corresponding output of the
2.2. Training methodology hidden layer nodes z is calculated via Eqs. (1) and (3) and
then the predicted value for y is obtained through the use of
In contrast to the usual neural network training proce- Eq. (4).
dures which are completed in one phase, the standard train-
ing methodology of the RBF networks consists of two
distinct phases: 3. Experimental procedure

(A) Calculation of the hidden layer parameters. In order to evaluate the relationships between lubricity
(B) Determination of the connection weights between the and other diesel fuel properties, 76 fuel samples were
hidden layer and the output layer. sourced from Petrola Hellas Re®nery and from Aspropyrgos
Re®nery. The samples were tested for lubricity, conduc-
In the ®rst phase, the network structure (i.e. the number of tivity, density, sulfur content, kinematic viscosity, ¯ash
hidden nodes) must be selected. This is a key parameter that point, distillation properties, pour point and cloud point.
affects mostly the prediction capabilities of the network. A The standard methods used for the determination of the
small number of nodes may result in a network that will not various fuel properties are given in Table 2, and the results
be able to approximate well the original system, while a are shown in Tables 2 and 3. All the samples had sulfur
large network may lead to the over®tting phenomenon levels less than 0.05% by weight. The average density of
where we observe small prediction errors for the data that the fuels was 845 kg/m 3, ¯uctuating between 836 and
are used in the training phase, but large errors when new 859 kg/m 3, while the cetane index varied in the range
data are used as inputs to the network. Then, the positions of 46.6±70.4, the average value being 54. The kinematic visc-
the node centers are determined. There are a lot of methods osity of the fuels ranged between 2.90 and 5.02 cSt at 40 8C,
to perform this task, but the k-means clustering is the most the average being 4.1 cSt. The ®nal boiling point of the fuels
common algorithm. The objective of this algorithm is to ®nd was between 320 and 386 8C, with an average value of
a set of k cluster centers that minimizes E, which is the sum 346 8C. Finally, the 90% recovery temperature of the fuels
1246
Table 2
Fuel properties (training set)

a/a WS 1.4 Conductivity Cetane Density Sulphur Kinematic IBP T10 T50 T90 FBP Flash Cloud Pour
(Mm) (pS/m) index (gr/cm 3) (ppm) viscosity (8C) (8C) (8C) (8C) (8C) (8C) point point
(40 8C, cSt) (8C) (8C)

1 527 0 54.9 0.841 247 3.93 159 241 278 334 340 81 2.2 210.3
2 511 0 54.6 0.843 341 5.02 175 238 285 333 336 100 1.7 23.1
3 467 3 54.5 0.836 308 4.08 178 230 271 320 322 84 0.6 210.3

D.M. Korres et al. / Fuel 81 (2002) 1243±1250


4 455 3 55.1 0.836 324 4.30 178 231 273 329 334 84 0 213
5 490 2 54.1 0.839 240 4.25 181 233 274 328 333 87 2.2 23
6 530 0 54.0 0.839 244 4.37 180 231 275 330 335 95 1.7 23
7 529 0 54.0 0.841 254 4.66 159 235 276 333 335 91 3 0
8 492 2 55.0 0.839 260 4.16 163 234 278 327 330 86 1 25
9 489 2 50.8 0.841 425 4.40 164 221 266 324 326 84 2 23
10 577 0 55.7 0.836 399 4.26 175 235 274 326 332 91 3 22
11 478 2 54.0 0.838 387 4.37 184 231 273 325 334 93 4 0
12 393 10 56.2 0.832 461 4.50 183 229 272 327 334 88 4 0
13 488 2 54.6 0.838 436 4.38 187 234 275 328 333 92 5 0
14 551 1 50.4 0.841 403 4.57 168 222 267 306 328 91 6 23
15 481 1 51.1 0.840 413 4.30 164 220 267 324 328 97 3 21
16 494 1 50.0 0.841 414 4.40 185 227 260 314 321 97 4 1
17 498 1 50.3 0.839 382 4.23 162 222 258 322 324 88 3 21
18 486 2 52.0 0.838 412 4.26 166 222 268 314 320 90 3 21
19 370 8 53.6 0.842 496 3.31 187 226 282 349 379 81 2 28
20 450 5 70.4 0.807 483 3.07 154 242 276 341 369 84 22 24
21 301 18 55.9 0.856 540 3.25 205 260 314 365 383 74 10 8
22 311 15 59.4 0.845 420 4.50 225 262 301 354 378 73 5 25
23 333 13 46.6 0.859 484 3.72 185 215 284 372 385 79 23 210
24 480 2 56.4 0.838 415 4.46 163 219 290 342 373 84 2 21
25 383 11 53.4 0.843 434 3.00 165 220 287 348 371 75 2 26
26 420 7 48.6 0.853 455 3.11 154 238 274 319 345 77 23 210
Table 3
Fuel properties (validation set)

a/a WS 1.4 Conductivity Cetane Density Sulphur Kinematic IBP T10 T50 T90 FBP Flash Cloud Pour
(?m) (pS/m) index (gr/cm 3) (ppm) viscosity (8C) (8C) (8C) (8C) (8C) (8C) point point
(40 8C, cSt) (8C) (8C)

1 503 0 54.0 0.841 244 3.86 183 232 279 336 342 99 2.8 27
2 451 3 56.5 0.836 314 4.25 195 231 281 336 340 84 21 213
3 459 3 52.4 0.836 315 4.23 161 221 264 320 323 85 21.1 210.9
4 449 3 56.0 0.836 302 4.19 179 232 278 325 335 84 0.6 211.4
5 487 2 52.0 0.842 262 4.83 175 232 270 330 331 90 5 1
6 519 1 53.3 0.837 234 4.07 183 228 269 319 331 87 23 26
7 522 0 53.9 0.842 301 4.96 175 235 280 337 338 88 5 3
8 499 1 53.4 0.840 286 4.51 177 235 272 324 329 90 2 23
9 537 0 53.8 0.839 390 4.17 170 230 274 334 335 87 3 21
10 461 2 53.5 0.838 260 4.31 179 228 272 324 334 83 2 23
11 453 3 54.7 0.838 283 4.11 159 232 276 332 338 87 4 23
12 466 2 52.2 0.838 412 4.27 168 224 267 324 326 86 2 23
13 526 1 68.2 0.808 425 4.53 179 232 272 322 330 96 6 0
14 513 1 54.6 0.836 433 4.28 180 234 270 325 330 94 4 21

D.M. Korres et al. / Fuel 81 (2002) 1243±1250


15 491 2 53.3 0.838 415 4.25 179 228 271 323 331 90 5 23
16 466 2 52.7 0.839 362 4.45 180 227 270 320 325 88 1 23
17 447 3 54.6 0.838 409 4.45 182 226 280 322 335 96 5 21
18 452 4 56.7 0.832 430 3.95 183 230 273 321 331 93 4 21
19 549 0 55.9 0.838 233 4.43 198 237 280 317 338 92 5 0
20 380 7 54.1 0.838 472 4.51 179 228 275 329 337 96 6 21
21 473 2 51.5 0.837 399 4.21 170 221 264 312 325 89 4 0
22 511 1 53.4 0.840 424 4.30 179 227 276 328 332 88 5 23
23 449 3 58.9 0.840 417 4.31 215 253 290 346 356 93 4 23
24 522 0 56.4 0.831 416 4.48 152 230 271 322 324 97 5 3
25 501 1 52.5 0.840 387 4.38 204 229 271 320 322 91 4 0
26 470 2 54.1 0.841 429 4.52 182 231 281 331 334 92 7 1
27 462 3 56.2 0.841 352 4.37 175 233 291 331 336 93 2 0
28 492 1 54.5 0.841 159 4.42 183 230 283 332 337 92 3 21
29 521 0 53.2 0.838 402 4.30 186 228 271 324 335 89 3 21
30 494 1 49.5 0.838 415 4.46 158 220 255 315 326 90 3 21
31 469 2 56.4 0.838 401 4.33 166 222 290 345 370 92 8 24
32 458 3 55.7 0.839 420 4.48 168 224 289 346 368 94 7 24
33 462 3 53.3 0.845 474 4.43 167 223 288 344 367 88 7 24
34 453 3 56.2 0.838 414 4.66 167 221 291 345 369 96 7 24
35 475 2 55.3 0.840 422 4.26 164 220 292 343 372 86 2 0
36 488 2 56.3 0.838 422 4.38 165 222 291 347 371 88 5 0
37 311 17 53.6 0.844 455 3.47 166 225 292 345 370 70.5 2 26
38 421 12 50.0 0.849 480 2.97 159 235 275 320 353 62 210 212
39 388 7 53.4 0.847 410 3.60 167 225 295 355 376 72.3 3 27
40 379 9 53.8 0.843 490 3.47 171 218 291 349 375 75 21 24
41 362 10 49.5 0.854 520 3.27 190 229 284 352 371 72 23 29
42 506 1 49.8 0.855 430 3.22 192 232 287 356 374 72 24 29
43 371 11 52.3 0.844 514 3.20 185 224 280 347 375 82 21 25
44 320 17 53.8 0.846 450 3.68 177 227 294 351 374 56 3 28
45 395 7 52.3 0.842 480 3.19 183 221 277 345 372 84 21 25
46 454 4 56.4 0.841 416 3.36 188 249 285 313 329 76 29 212
47 391 12 51.4 0.852 360 3.44 176 230 293 350 376 81 5 24
48 456 9 47.5 0.861 418 2.90 174 228 292 349 375 69 2 25

1247
49 387 8 47.5 0.858 422 3.12 188 227 282 350 369 74 4 29
50 376 10 59.7 0.852 451 3.50 164 269 320 368 386 76 4 29
1248 D.M. Korres et al. / Fuel 81 (2002) 1243±1250

Table 4 4. Results and discussion


RBF network training parameters

Number of hidden layer nodes 15 The data set was split into two subsets: a training set of 26
Training sum of squared errors (SSE) 5762 samples (Table 2) and a validation set of 50 samples (Table
Training time 5s 3). The ®rst set was used to train the RBF network and the
second to validate the neural network predictive capabil-
was between 306 and 372 8C, with an average value of ities. The model used 15 hidden nodes as shown in
333 8C. Table 4. In the same table the training time and the sum
All tribological experiments were carried out by employ- of squared errors (SSE) between the experimental data and
ing an HFRR apparatus (high frequency reciprocating rig), the neural network predictions are shown. It must be empha-
according to CEC F-06-A-96 method [21]; the test tempera- sized that by increasing the number of hidden layer nodes,
ture was 60 8C, and the volume of fuel sample used was the network can approach the training data to any given
2 ml. Relative humidity was kept between 49 and 56%, accuracy. However, it is usually not desirable to achieve a
while the mean ambient temperature in the laboratory was very low SSE in the training phase, since this can lead to the
24 8C. The lubricating ef®ciency of the fuels was estimated over®tting phenomenon, which was described earlier. The
by measuring the average wear scar diameter (WSD) of the results obtained by the training set are depicted in Fig. 3,
spherical specimen, using a photomicroscope. The wear where a 95% con®dence limit was set.
scars quoted were corrected to give the wear scar diameter Not all of the measured diesel fuel properties were used
values at a pressure of 1.4 kPa. for the development of the network. In particular, it was

Fig. 3. Experimental vs. predicted lubricity values (training set).

Fig. 4. Experimental vs. predicted lubricity values (validation set).


D.M. Korres et al. / Fuel 81 (2002) 1243±1250 1249

Fig. 5. Univariate analysis of fuel lubricity vs. conductivity.

found through trial and error that using the conductivity, city with other diesel fuel properties, using a fuel matrix of
density, kinematic viscosity, sulfur content and 90% recov- 76 fuel samples. The samples were tested for lubricity,
ery temperature, the methodology produced the smallest conductivity, density, sulfur content, kinematic viscosity,
error in the validation data. Indeed, a 5 input±1 output ¯ash point, distillation properties, pour point and cloud
network was set up using the above mentioned variables point. A black box modeling approach was used, where
as inputs and the lubricity as output. The predicted lubricity the diesel fuel lubricity was approximated by an RBF
values by this neural network, along with the experimental network, that used other diesel fuel properties as inputs. It
values are depicted in Fig. 4 and correspond to a correlation was shown that the proposed neural network was easy to use
coef®cient R 2 ˆ 0.94. It can be seen that more than 90% of and predicted the lubricity of the diesel fuel samples, with
the predicted lubricity values lie within the 95% con®dence good accuracy. In particular, it was found through trial
limit. and error that using the conductivity, density, kinematic
Apart from serving as a predictor, the neural network viscosity, sulfur content and 90% recovery temperature,
model provides useful information regarding the importance the methodology produced the smallest error in the
of each input variable in determining fuel lubricity. This validation data.
information can be obtained through univariate analysis, The predicted lubricity values by this neural network,
where the effect of each input variable on the output is exam- gave a correlation coef®cient R 2 ˆ 0.94, with more than
ined separately. More speci®cally, the values of fuel lubricity 90% of the predicted lubricity values to lie within the
are calculated, while the domain of the particular input vari- 95% con®dence limit.
able is scanned and all the remaining input variables are kept
at constant values (usually the means of the training data).
The results can be visualized using the univariate analysis Acknowledgements
graphs, where the lubricity values versus each particular
input variable are plotted. The produced plots show not We wish to thank Mrs Ch. Papachristou, director of the
only the importance of each input variable, but also the direc- Petrola Re®nery Laboratory for supplying us with most of
tion in which it affects the fuel lubricity. By applying univari- the fuel samples used for this study.
ate analysis on the neural network model that was developed
in this work, we can conclude that the most critical parameter
References
for the determination of fuel lubricity is the fuel conductivity,
while the rest of the input parameters have a less signi®cant [1] Miura M, Ikeda T, Takizawa H, Yoshida H, Ikebe H. SAE Paper
effect. The univariate analysis graph for fuel conductivity, 972895, 1997.
which is shown in Fig. 5, reveals the signi®cance of this [2] Wei D, Spikes H. Wear 1986;111:217.
particular factor and shows clearly the direction in which [3] Lacey PI, Westbrook SR. SAE Paper 950248, 1995.
the fuel lubricity is affected. [4] Wang JC, Cusamo CM. SAE Paper 952564, 1995.
[5] Lacey PI, Lestz SJ. SAE Paper 920823, 1992.
[6] Lacey PI, Lestz SJ. SAE Paper 920824, 1992.
[7] Cameron F. SAE Paper 981362, 1998.
5. Conclusions [8] Nikanjam M, Henderson PT. SAE Paper 920825, 1992.
[9] Tucker RF, Stradling RJ, Wolveridge PE, Rivers KJ, Unbbens A. SAE
In the present work, an effort was made to correlate lubri- Paper 942016, 1994.
1250 D.M. Korres et al. / Fuel 81 (2002) 1243±1250

[10] Octane Week, April 14 1997, XII, p. 15. [15] MacMurray JC, Himmelblau DM. Computers & Chemical Engineer-
[11] Meyer K, Stolz U, Rehbein P. A Tribological Approach to Determine ing 1995;19:1077.
the Friction and Wear Properties of New Environmentally Benign [16] Al-Enezi G, Elkamel A. Petroleum Science and Technology
Diesel Fuels in Conjunction with Wear Mechanisms in Critical 2000;18:407.
Parts of Diesel Injection Equipment. 21st Leeds±Lyon Symposium [17] Helena Cristina IL, Maciel Filho A, Maciel Filho R. Computers &
on Tribology `Lubricants and Lubrication': Leeds, UK, 1994. Chemical Engineering 1998;22(Supplement 1):S981.
[12] Ashton CPS, Thomson MW, Dixon S. The Relationship between [18] Sarimveis H. Systems Analysis Modeling Simulation 2000;38:555.
Conductivity and Lubricity in `New' European On Road Diesel [19] Leonard JA, Kramer MA. IEEE Control Systems 1991;31:31.
Fuels. 6th International Conference on Stability and Handling of [20] Darken C, Moody J, IEEE INNS Int J Conf On Neural Networks;
Liquid Fuels, Vancouver, BC, Canada, October 13±17, 1997. Proceedings, 1990, Vol. II.
[13] Haykin S. Neural networks. 2nd ed. Prentice Hall International, [21] CEC F-06-A-96: Measurement of Diesel Fuel Lubricity ± Approved
1999. Test Method. HFRR Fuel Lubricity Test.
[14] Jamialahmadi M, Javadpour FG. Journal of Petroleum Science and
Engineering 2000;26:235.

You might also like