Analytica Chimica Acta 671 (2010) 27–35

Contents lists available at ScienceDirect

Analytica Chimica Acta

journal homepage: www.elsevier.com/locate/aca

Gasoline classification using near infrared (NIR) spectroscopy data: Comparison

of multivariate techniques
Roman M. Balabin a,∗ , Ravilya Z. Safieva b , Ekaterina I. Lomakina c
a
Department of Chemistry and Applied Biosciences, ETH Zurich, 8093 Zurich, Switzerland
b
Gubkin Russian State University of Oil and Gas, 119991 Moscow, Russia
c
Faculty of Computational Mathematics and Cybernetics, Lomonosov Moscow State University, 119992 Moscow, Russia

a r t i c l e i n f o a b s t r a c t

Article history: Near infrared (NIR) spectroscopy is a non-destructive (vibrational spectroscopy based) measurement
Received 8 April 2010 technique for many multicomponent chemical systems, including products of petroleum (crude oil) refin-
Received in revised form 6 May 2010 ing and petrochemicals, food products (tea, fruits, e.g., apples, milk, wine, spirits, meat, bread, cheese,
Accepted 7 May 2010
etc.), pharmaceuticals (drugs, tablets, bioreactor monitoring, etc.), and combustion products. In this paper
we have compared the abilities of nine different multivariate classification methods: linear discriminant
analysis (LDA), quadratic discriminant analysis (QDA), regularized discriminant analysis (RDA), soft inde-
Keywords:
pendent modeling of class analogy (SIMCA), partial least squares (PLS) classification, K-nearest neighbor
Discriminant analysis (LDA, QDA, RDA)
Petroleum (crude oil)
(KNN), support vector machines (SVM), probabilistic neural network (PNN), and multilayer perceptron
Biofuel (biodiesel, bioethanol, (ANN-MLP) – for gasoline classification. Three sets of near infrared (NIR) spectra (450, 415, and 345 spec-
ethanol–gasoline fuel) tra) were used for classification of gasolines into 3, 6, and 3 classes, respectively, according to their source
Soft independent modeling of class analogy (refinery or process) and type. The 14,000–8000 cm−1 NIR spectral region was chosen. In all cases NIR
K-Nearest neighbor method spectroscopy was found to be effective for gasoline classification purposes, when compared with nuclear
Support vector machine magnetic resonance (NMR) spectroscopy or gas chromatography (GC). KNN, SVM, and PNN techniques
Probabilistic neural network for classification were found to be among the most effective ones. Artificial neural network (ANN-MLP)
Near infrared spectroscopy
approach based on principal component analysis (PCA), which was believed to be efficient, has shown
much worse results. We hope that the results obtained in this study will help both further chemomet-
ric (multivariate data analysis) investigations and investigations in the sphere of applied vibrational
(infrared/IR, near-IR, and Raman) spectroscopy of sophisticated multicomponent systems.
© 2010 Published by Elsevier B.V.

1. Introduction ing is important to monitor the quality of products (e.g., gasoline

Near infrared (NIR) spectroscopy is a non-destructive measure-
ment technique for many chemical compounds, including products
of petroleum refining and petrochemicals, food products (tea,
fruits, milk, meat, etc.), pharmaceuticals (drugs, tablets, bioreac-
tor monitoring, etc.), combustion products, and many other [1–4].
It has proved its efficiency for laboratory and industrial applica-
tions [1,2]. Many current studies are focused on NIR sensor design
for on-line [3] or portable operation [4].
One of the possible usages of near infrared spectroscopy is in a
petroleum industry [5–12], because of the fact that many products
of petroleum refining and petrochemicals consist of hydrocarbons,
whose content can be estimated by NIR method. Near infrared spec-
troscopy (Tables 1 and 2) is useful to overcome many limitations,
especially in a complicated real process, where on-line measur-
∗ Corresponding author. Tel.: +41 44 632 4783.
E-mail address: balabin@org.chem.ethz.ch (R.M. Balabin).
production) [13].
Multivariate statistics techniques have boosted the use of NIR
instruments [13,14]. Only methods of chemometrics are able to
process enormous amounts of sophisticated experimental data that
are provided by NIR technique. Many calibration methods were
used for gasoline quality prediction [5–12,15]. Linear techniques,
“quasi-nonlinear” techniques, and “truly” nonlinear techniques –
all of them were used for gasoline properties and quality coeffi-
cients prediction [5–12].
Despite the fact that gasoline (or other products of petroleum
refining) classification is also a practically important task, not many
papers are devoted to this problem [11,16–18]. In this paper we
have tried to evaluate the efficiency of different classification meth-
ods for gasoline classification by source and type.
Gasoline identification by source (refinery) is an important
factor for both quality control and identification of gasoline adul-
teration. This method is based on the difference in crude oil that
results in chemical difference of gasolines from variant refineries.
Gasoline type is needed for classification of gasolines by quality

0003-2670/$ – see front matter © 2010 Published by Elsevier B.V.

doi:10.1016/j.aca.2010.05.013
28 R.M. Balabin et al. / Analytica Chimica Acta 671 (2010) 27–35

Table 1
Gasoline sample sets.

Set A Set B Set C

a b
Classification by Source (refinery ) Source (process ) Type
Number of samples 150 117 115
Number of classes 3 6 3
Classes Refinery 1; Refinery 2; Refinery 3 Straight-run; reformate; catalysate; isomerizate; Normald ; Regular; Premium
hydrocracking gasoline; mixturec
Distribution of samplese 50; 50; 50 30; 12; 15; 13; 12; 35 55; 45; 15
a
Refinery 1, Orsknefteorgsintez (Russia); Refinery 2, Kirishinefteorgsintez (Russia); Refinery 3, Astrakhan’gazprom (Russia).
b
For details one can consult Refs. [46,47].
c
All gasoline fractions (straight-run, reformate, catalysate, isomerizate, hydrocracking) are presented in mixtures.
d
Normal (Russia) – regular gasoline with octane number of 80.
e
Distribution of samples is given according to the classes above.

(and price). The process of gasoline production defines the chemi-
cal composition of gasoline fraction. This difference is an important
factor for gasoline mixing.
In this paper three gasoline sample sets were used to com-
pare prediction capabilities of linear discriminant analysis (LDA),
quadratic discriminant analysis (QDA), regularized discriminant
analysis (RDA), soft independent modeling of class analogy
(SIMCA), partial least squares (PLS) classification, K-nearest neigh-
bor (KNN), support vector machines (SVM), probabilistic neural
network (PNN), and multilayer perceptron (MLP). All classifications
are based on gasoline near infrared (NIR) spectroscopy data.
2. Brief descriptions of the classifiers
In this section we have tried to give some basic remarks about
the methods used. The references on detailed papers were also
provided.
a given measurement is from a given class is the likelihood ratio
test.
Using QDA each of the covariance matrices is estimated sep-
arately, which requires a larger sample than that used in LDA to
reach the same level of reliability in the estimations and hence in
the predictions. QDA details can be found in Refs. [19–21].
2.3. Regularized discriminant analysis (RDA)
Regularized discriminant analysis (RDA) can be called a “com-
promise” between LDA and QDA.
RDA introduces a double type of bias in the estimation of the
covariance matrices of each class with the aim of stabilizing the
prediction in the case of a deficient estimation of their elements.
This is a very frequent problem encountered with real data because
the estimations of the matrices are very sensitive to the presence of
different data. For a detailed analysis one can consult Refs. [21,22].

2.1. Linear discriminant analysis (LDA) 2.4. Soft independent modeling of class analogy (SIMCA)

Among many possible techniques for data classification, linear Soft independent modeling of class analogy (SIMCA) is a method
discriminant analysis (LDA) is a commonly used one. Linear dis- for supervised data classification that requires a training data set
criminant analysis (LDA) is used to find the linear combination of consisting of samples (or objects) with a set of attributes and their
features which best separate two or more classes of object or event. class membership. The term “soft” refers to the fact the classifier
The resulting combinations may be used as a linear classifier, or can identify samples as belonging to multiple classes and not nec-
more commonly in dimensionality reduction before later classifi- essarily producing a classification of samples into non-overlapping
cation. This method maximizes the ratio of between-class variance classes.
to the within-class variance in any particular data set thereby guar- In order to build the classification models, the samples belong-
anteeing maximal separability. Details about LDA can be found in ing to each class need to be analyzed using principal components
Refs. [19–21]. analysis (PCA). For a given class, the resulting model then describes
a line, plane or hyperplane. For each modeled class, the mean
orthogonal distance of training data samples from the line, plane or
2.2. Quadratic discriminant analysis (QDA)
hyperplane (calculated as the residual standard deviation) is used
to determine a critical distance for classification.
Quadratic discriminant analysis (QDA) is closely related to LDA
New observations are projected into each PC model and the
(see above). Unlike LDA however, in QDA there is no assumption
residual distances calculated. An observation is assigned to the
that the covariance of each of the classes is identical. When the
model class when its residual distance from the model is below
assumption is true, the best possible test for the hypothesis that
the statistical limit for the class. The observation may be found to
belong to multiple classes and a measure of goodness of the model
Table 2 can be found from the number of cases where the observations are
NIR experimental parameters.
classified into multiple classes. Details about SIMCA model building
Set A Set B Set C and usage can be found in Ref. [23].
Spectral range (cm−1 ) 14,000–8000 14,000–8000 13,500–8500
Sample volume (cm3 ) 100 100 100 2.5. Partial least squares (PLS) regression and classification
Optical path (cm) 10 10 10
Spectra resolution (cm−1 ) 16 8 8 Partial least squares (PLS) regression is an extension of the mul-
Number of scans 64 64 72
Cell material Quartz Quartz Quartz
tiple linear regression model.
Time of one measurement 2–3 3–4 2–3 PLS method has found widespread use for multivariate analysis
(min) of spectral data. The main goal of PLS usage is calibration model
Number of measurements 3 3–5 3 building, but this technique can also be applied for classification
(per sample)
purposes. Some information about the idea of PLS classification can
No spectra preprocessing was used. be found in Section 3.6.3 and Refs. [24–27].
 R.M. Balabin et al. / Analytica Chimica Acta 671 (2010) 27–35 29

Table 3
Building parameters for K-nearest neighbor (KNN) classification model.

Weighting Formulaa Parameters Number of neighbors (K)

b
None – – K = 1–50
Weighting 1 w(xi ) = exp(−D(xi , p)) – K = 1–35
−q
Weighting 2 w(xi ) = D(xi , p) q = 1–5 K = 1–35
1
Weighting 3 w(xi ) = 2 ε = 0.001–10 K = 1–40
D(xi ,p) +ε
−1
Weighting 4  i , p) + 1)
w(xi ) = (D(x – K = 1–35
1, D(xi , p) ≤ d0
Weighting 5 w(xi ) = d0 = 0.001–10 K = 1–40
0, D(xi , p) > d0
  2 L  
xi,l − pl .
L
a
Two types of distances D(xi ,p) between vectors xi and p were used: D(xi , p) = xi,l − pl and D(xi , p) =
l=1 l=1
b
“Voting” KNN.

2.6. K-Nearest neighbor (KNN)
K-Nearest neighbor (KNN) method was first introduced by Fix
and Hodges [28].
In this method a distance (Euclidean or other) is assigned
between all points in a data set. The data points, K-closest neigh-
bors (K being the number of neighbors), are then found by analyzing
(sorting) the distance matrix. The K-closest data points are then
analyzed to determine which class label is the most common among
the set (“voted” KNN). The most common class label is then assigned
to the data point being analyzed.
Other variants of KNN (“weighted” KNN) are also possible. In
these methods different neighbors have different “weight” – ability
to influence on class of sample – according to their distances from
the sample. Some remarks about the possible types of functions for
weighting are given in Table 3.
For KNN classifier, it is necessary to have a training set which
is not too small, and a good discriminating distance. KNN performs
well in multi-class simultaneous problem solving. There exists an
optimal choice for the value of the parameter K, which brings to the
best performance√ of the classifier. This value of K is often approxi-
mately close to N [31,32] (see also Table 3). For a detailed analysis
one can consult Refs. [29–32].
2.7. Support vector machines (SVM)
Support vector machines have been developed by Vapnik [33].
Support vector machines (SVM) are based on some “beautifully
simple ideas” [31] and provide a clear intuition of what learning
from examples is all about. SVM are also showing high perfor-
mances in practical applications [31,34–37].
In very simple terms, an SVM corresponds to a linear method in
a very high dimensional feature space that is nonlinearly related to
the input space. Even though we think of it as a linear algorithm in a
high dimensional feature space, in practice, it does not involve any
computations in that high dimensional space. By the use of kernels
(see Table 4), all necessary computations are performed directly in
the input space [38–40].
Support vector machines map input vectors to a higher
dimensional space where a maximal separating hyperplane is con-
structed. Two parallel hyperplanes are constructed on each side of
the hyperplane that separates the data and maximizes the distance
between the two parallel hyperplanes. An assumption is made that
the larger the margin or distance between these parallel hyper-
planes the better the generalization error of the classifier will be.
Details about SVM classifier can be found in Refs. [31,34–40].
2.8. Probabilistic neural network (PNN)
The probabilistic neural network (PNN) was developed by
Specht [41,42]. The probabilistic neural network provides a gen-
eral solution to pattern classification problems by following an
approach developed in statistics, called Bayesian classifiers. PNN
uses a supervised training set to develop distribution functions
within a pattern layer. These functions, in the recall mode, are used
to estimate the likelihood of an input feature vector being part of
a learned category, or class. The learned patterns can also be com-
bined, or weighted, with the a priori probability, also called the
relative frequency, of each category to determine the most likely
class for a given input vector. For a detailed analysis one can consult
Refs. [41,42].
2.9. Multilayer perceptron (MLP)
The ideology of multilayer perceptron (MLP) usage for classifi-
cation problem is similar to PLS classification (see Section 3.6.7 and
Ref. [24]). The main dissimilarity between MLP and PLS is the ability
of MLP to solve nonlinear problems [43,44]. For a detailed analysis
one can consult Refs. [24,44].
3. Experimental
3.1. Sample sets
Three gasoline sample sets (Sets A–C) were used in our work to
evaluate efficiency of different classification methods. Information
about them can be found in Table 1. All gasolines are unleaded Rus-
sian gasolines without additives; they were provided by refineries.
Table 1 also provides the information about gasoline classes and
distribution of samples over them. Property prediction (using MLP)
of some of these gasolines was already provided in Refs. [5,8].
3.2. Apparatus and experimental parameters
The NIR spectra were acquired with a Near-IR FT Spectrome-
ter InfraLUM FT-10 (LUMEX, Russia) fitted with special sampler for

Table 4
Building parameters for support vector machine (SVM) classification model.

Kernel Formula Kernel parameters Cost parameter C

Linear K(xi , xj ) = xi xj + 1 – C = 0.001–100

p
K(xi , xj ) = (xi xj
Polynomial + 1)
2
p = 2–8 C = 0.001–100
xi −xj
Gaussian K(xi , xj ) = exp −  = 0.01–10 C = 0.005–50
2
30 R.M. Balabin et al. / Analytica Chimica Acta 671 (2010) 27–35

Fig. 1. Example of gasoline near infrared (NIR) spectra. Maximal transmittance is

5 a.u.

liquids (see Table 2 in Ref. [5]). The spectra were acquired at room
temperature (20–23 ◦ C). Experimental parameters (for all data sets)
are shown in Table 2.
InfraLUM FT-10 (as well as most NIR-spectrometers) has no
thermostatic control. In order to compensate this shortcoming, the
Fig. 2. Optimization of partial least squares (PLS) classification model: dependence
background spectrum was taken before and after the measurement of classification error (see Eq. (1)) on ı-parameter (see Section 3.6.3). Example
of each spectrum; then, the averaged background spectrum was parameters: Set A; LV = 14. Using different sample set division the accuracy of ı-
subtracted from the sample spectrum. This allowed obtaining an parameter can be estimated as ±0.1 (95% confidence interval).
analytical signal with satisfactory accuracy and precision (Fig. 1).
The instrument calibration (wavelength and transmittance values) this goal. The results are reported about the most effective data
was performed using four pure hydrocarbons: toluene, hexane, reduction technique.
benzene, and isooctane. It was repeated at least once per day to
ensure the experimental setup stability and data accuracy (repro- 3.6. Optimization of classifiers
ducibility).
To compare different classification models their best results
3.3. Software and computing should be obtained. The result of model usage depends on the
model parameters (see below). We have used wide ranges of mod-
MATLAB 7.4 was used as standard software for classification els’ parameters to achieve the best results.
methods realization. The following toolboxes were used:
3.6.1. Discriminant analysis (LDA, QDA, RDA)
• MATLAB Statistics Toolbox.
The optimal number of principal components (PC; see Section
• MATLAB Support Vector Machine Toolbox.
3.5) was found for each method of discriminant analysis: lin-
• MATLAB Neural Network Toolbox.
ear discriminant analysis (LDA), quadratic discriminant analysis
• N-way Toolbox for MATLAB.
(QDA), and regularized discriminant analysis (RDA). Principal com-
• PLS Toolbox Version 4.0.
ponent ranges were 1–50, 1–50, and 1–40 for LDA, QDA, and RDA,
respectively. The parameter ˛ ∈ [0;1] that controls the RDA model
Standard programs of these Toolboxes were modified and
complexity was also optimized [21,22].
extended by Balabin (see also Refs. [5,6,8,9]).

3.4. Model efficiency estimation

To characterize prediction ability (efficiency) of created classi-

fication model, the error of cross-validation (E) was used
Nwrong N0 − Nright
E= = (1)
N0 N0
where Nwrong and Nright refer to the number of wrongly and rightly
classified samples; N0 is the total number of samples in validation
set.
Five-fold cross-validation was used to evaluate the model’s effi-
ciency. It was checked that validation set consists of all gasoline
classes. Other variants of cross-validation procedure (10-fold, 7-
fold, etc.) were checked and found to produce almost identical
results.

3.5. Spectra reduction

In order to create effective and robust classification model, spec-

tra data (that have up to 780 independent variables) should be
reduced. Two common ways of data reduction, spectra averaging
and principal component analysis (PCA), were used by us to achieve
Fig. 3. Optimization of partial least squares (PLS) classification model: dependence
of classification error (see Eq. (1)) on number of latent variables (LV). Example
parameters: Set A; ı = 0.7. Since the LV is an integer number the expected accuracy
is ±1.
 R.M. Balabin et al. / Analytica Chimica Acta 671 (2010) 27–35 31

3.6.4. K-Nearest neighbor (KNN)

Different types of K-nearest neighbor (KNN) method (voted and
weighted by different weighting functions) were used to get the
best result (Table 3).

3.6.5. Support vector machines (SVM)

Different types of kernel functions (and corresponding param-
eters) were checked for optimization of SVM classification model
(Table 4). Cost parameter C [31,34–36] was also optimized (Table 4).

3.6.6. Probabilistic neural network (PNN)

Optimal PNN spread value (the distance an input vector must be
from a neuron’s weight vector to be 0.5 [45]) was found in the range
of 10−3 to 10−7 (Fig. 4). The number of input neurons for PNN was
equal to the number of PC (1–50) or the size of averaged spectra
(152–780).

3.6.7. Multilayer perceptron (MLP)

Fig. 4. Optimization of probabilistic neural network (PNN) classification model:
dependence of classification error (see Eq. (1)) on spread value (see Section 3.6.6).
Example parameters: Set B; PC = 39. Using different sample set division the accuracy
of the spread value can be estimated as ±0.013 (95% confidence interval).
3.6.2. Soft independent modeling of class analogy (SIMCA)
The optimal number of principal components (PC) or the optimal
level of spectra averaging was found for soft independent modeling
of class analogy (SIMCA) method. Principal components range was
from 1 to 30; spectra averaging up to 4 times were applied.
Artificial neural network (ANN) analysis of spectral data was also
performed with multilayer perceptron (MLP). Similar to PLS anal-
ysis, each class was coded by N binary strings (see Section 3.6.3). A
three-layer MLP (structure: PC–NN–N) with input, hidden, and out-
put layers was used. Each weight of the bond that connects hidden
and input units or output and hidden units was trained by back-
propagation. The prediction of the class of individual samples was
performed using N predicted output scores (see also PLS). Training
parameters for multilayer perceptron are shown in Table 5. Opti-
mal value of parameter ı (ı = 0.3–0.9) was also found for each data
set.

3.6.3. Partial least squares (PLS) 4. Results and discussion

For PLS classification the class (one of N) of each sample was
coded by N binary strings. In other words, class j was represented 4.1. Gasoline classification by refinery (Set A)
as
⎛ ⎞ Table 6 presents the results of Set A classification. The error of
classification (according to Eq. (1)) varies greatly from 0% for SVM
⎝0 . . . 0, 1, 0 . . . 0⎠
    and PNN methods to 11–14% for LDA, QDA, RDA, and SIMCA. The
j−1 N−j average errors of classification for different refineries are close: 7,
9, and 7% for Refineries 1–3, respectively.
The N predicted scores (by ordinary PLS method) were used to The methods can be separated into the following groups
predict the class of each sample. If the j-score (normalized) was (according to the error of classification): highly effective methods
more than or equal to parameter ı, the predicted class of the sample (KNN, SVM, and PNN), methods of medium effectiveness (MLP and
was j. The optimal value of parameter ı (ı = 0.3–0.9) and the optimal PLS), and methods of low effectiveness (SIMCA and discriminant
number of latent variables (LV = 1–25) were found for each data set analysis methods). For these three groups the average errors are
(Figs. 2 and 3). 1 ± 2, 9 ± 1, and 13 ± 1%, respectively.

Table 5
Training parameters of multilayer perceptron (MLP).

Set A Set B Set C

Algorithm Constant learning rate Constant learning rate Constant learning rate
Minimised error function Mean squared error (MSE) Mean squared error (MSE) Mean squared error (MSE)
Learning Supervised Supervised Supervised
Initialisation method Random Random Random

Transfer functions
Input layer –a – –
Hidden layer Hyperbolic tangent (tanh) Hyperbolic tangent (tanh) Hyperbolic tangent (tanh)
Output layer Hyperbolic tangent (tanh) Hyperbolic tangent (tanh) Hyperbolic tangent (tanh)

Learning rate 0.001 0.002 0.02

Maximal number of epochb 100,000 100,000 15,000
Early stopping 5-Fold cross-validation 5-Fold cross-validation 5-Fold cross-validation
Number of training iterations 250 250 200
Number of input neurons (principal components) 1–25 1–25 1–20
Number of hidden neurons 1–25 1–25 1–20
a
No transfer function is used.
b
Was never reached during the training process.
32 R.M. Balabin et al. / Analytica Chimica Acta 671 (2010) 27–35

Table 6
Classification of gasolines in sample Set A: Gasoline source – refinery.

Refinery 1 Refinery 2 Refinery 3 Total

Nright /N0 E Nright /N0 E Nright /N0 E Nright /N0 E

LDA 45/50 10% 40/50 20% 46/50 8% 131/150 13%

QDA 45/50 10% 41/50 18% 44/50 12% 130/150 13%
RDA 46/50 8% 44/50 12% 44/50 12% 134/150 11%
SIMCA 44/50 12% 41/50 18% 44/50 12% 129/150 14%
N-PLS 46/50 8% 46/50 8% 44/50 12% 136/150 9%
MLP 44/50 12% 46/50 8% 48/50 4% 138/150 8%
KNN 48/50 4% 50/50 0% 48/50 4% 146/150 3%
SVM 50/50 0% 50/50 0% 50/50 0% 150/150 0%
PNN 50/50 0% 50/50 0% 50/50 0% 150/150 0%

E is the error of 5-fold cross-validation (Eq. (1)); Nright is the number of rightly classified samples; N0 is the total number of samples in validation set. Methods: LDA, linear
discriminant analysis; QDA, quadratic discriminant analysis; RDA, regularized discriminant analysis; SIMCA, soft independent modeling of class analogy; N-PLS, multilinear
partial least squares regression; KNN, K-nearest neighbor; SVM, support vector machines; PNN, probabilistic neural network; MLP, multilayer perceptron.

One should note that with highly effective methods of classi-
fication (group one) gasoline producers (refineries) can easily be
distinguished by NIR spectra.
4.2. Gasoline classification by process (Set B)
The results of Set B classification are shown in Table 7. The error
of classification varies from 1% for PNN methods to 35% for LDA. The
average errors of classification for different processes are different
(from 8 to 31%). They are: 20, 30, 25, 29, 31, and 8% for straight-run
gasoline, reformate, catalysate, isomerizate, hydrocracking gaso-
line, and their mixture, respectively. One can note that classification
error for gasoline mixtures is much lower.
By effectiveness of Set B classification the methods can be sepa-
rated into three groups (the same as in Section 4.1): highly effective
methods (KNN, SVM, and PNN), methods of medium effectiveness
(MLP and PLS), and methods of low effectiveness (SIMCA and dis-
criminant analysis methods). For these three groups the average
errors are 33 ± 2, 19 ± 1, and 4 ± 3%, respectively.
Average classification error for Set B (20%) is much higher than
for Set A (8%). So the classification of gasoline by processes is a much
more sophisticated task. The greater number of classes (6 instead
of 3) and lower “samples/classes” ratio (20 instead of 50) should
also be taken into account.
Confusion matrixes for Set B classification by MLP, KNN, and
PNN are presented in Tables 8–10. From these matrixes one can
conclude that the main problem of effective classification model
building is connected with confusion of pure gasoline fractions f
and their mixtures. It is an obvious problem because of the fact
that gasoline mixtures have different compositions and sometimes
can consist of one or two fractions (e.g., reformate + isomerizate).
One should note that with highly effective methods of classifi-
cation model building (SVM and PNN) gasoline fractions (produced
by different processes) can easily be distinguished.
4.3. Gasoline classification by type (Set C)
Effectiveness of different classification models for Set C is shown
in Table 11. The range of classification errors for Set C is from 0%
(for PNN) to 13% (for RDA). The average errors of classification
for different gasoline types are 6, 6, and 16 for Normal, Regular,
and Premium, respectively. The great value of Premium classifica-
tion error can be explained by a lower number of samples 15 (in
comparison with 55 and 45 for Normal and Regular).
Classification model can also be grouped by the effectiveness of
Set C classification. KNN, SVM, and PNN form the first (the most
effective) group with classification error of 1 ± 1%; other models
(LDA, QDA, RDA, SIMCA, PLS, and MLP) form another group with
classification error of 11 ± 2%.
Average classification error for Set C (8%) is equal to the error
for Set A and is much lower than the error for Set B (20%). So the
classification of gasoline by type is an easy task (first of all for KNN,
SVM, and PNN methods).
With highly effective methods of classification model building
(KNN, SVM, and PNN) gasolines of different types can easily be
discriminated by NIR spectroscopy.
4.4. Final remarks
The methods presented above (Sections 4.1–4.3) show that nine
multivariate methods used in this study can be separated into three
groups, according to their classification accuracy. The first group
(“highly accurate methods”) includes KNN, SVM, and PNN methods;
the second group (“methods of medium effectiveness”) consists of
MLP and PLS models; the third group (“methods of low effective-
ness”) includes SIMCA and discriminant analysis methods (LDA,
RDA, and QDA). Only the last data set (Set C) does not give any
difference for the last two groups. It means that gasoline classifi-
cation according to its type can be effectively done only by rather

Table 7
Classification of gasolines in sample Set B: Gasoline source – process.

Straight-run Reformate Catalysate Isomerizate Hydrocracking gasolineMixture Total

Nright /N0 E Nright /N0 E Nright /N0 E Nright /N0 E Nright /N0 E Nright /N0 E Nright /N0 E

LDA 20/30 33% 6/12 50% 8/15 47% 7/13 46% 5/12 58% 30/35 14% 76/117 35%
QDA 20/30 33% 6/12 50% 8/15 47% 6/13 54% 5/12 58% 32/35 9% 77/117 34%
RDA 20/30 33% 8/12 33% 9/15 40% 7/13 46% 5/12 58% 31/35 11% 80/117 32%
SIMCA 21/30 30% 8/12 33% 9/15 40% 8/13 38% 6/12 50% 30/35 14% 82/117 30%
N-PLS 24/30 20% 8/12 33% 12/15 20% 8/13 38% 10/12 17% 32/35 9% 94/117 20%
MLP 26/30 13% 9/12 25% 10/15 33% 10/13 23% 10/12 17% 31/35 11% 96/117 18%
KNN 28/30 7% 9/12 25% 15/15 0% 12/13 8% 10/12 17% 35/35 0% 109/117 7%
SVM 28/30 7% 11/12 8% 15/15 0% 12/13 8% 11/12 8% 35/35 0% 112/117 4%
PNN 30/30 0% 11/12 8% 15/15 0% 13/13 0% 12/12 0% 35/35 0% 116/117 1%

See comments for Table 6.

 R.M. Balabin et al. / Analytica Chimica Acta 671 (2010) 27–35 33

Table 8
Multivariate perceptron (MLP) classification: confusion matrix for Set B.

Number Straight-run Reformate Catalysate Isomerizate Hydrocracking gasoline Mixture

Straight-run 30 26 0 0 0 0 4
Reformate 12 0 9 0 1 0 2
Catalysate 15 0 2 10 1 0 2
Isomerizate 13 0 0 0 10 0 3
Hydrocracking gasoline 12 0 0 0 0 10 2
Mixture 35 1 1 0 0 2 31

Columns, real class; rows, predicted class.

Table 9
K-Nearest neighbor (KNN) classification: confusion matrix for Set B.

Number Straight-run Reformate Catalysate Isomerizate Hydrocracking gasoline Mixture

Straight-run 30 28 0 0 0 0 2
Reformate 12 0 9 1 0 0 2
Catalysate 15 0 0 15 0 0 0
Isomerizate 13 0 0 0 12 0 1
Hydrocracking gasoline 12 0 0 0 0 10 2
Mixture 35 0 0 0 0 0 35

Columns, real class; rows, predicted class.

Table 10
Probabilistic neural network (PNN) classification: confusion matrix for Set B.

Number Straight-run Reformate Catalysate Isomerizate Hydrocracking gasoline Mixture

Straight-run 30 30 0 0 0 0 0
Reformate 12 0 11 0 0 0 1
Catalysate 15 0 0 15 0 0 0
Isomerizate 13 0 0 0 13 0 0
Hydrocracking gasoline 12 0 0 0 0 12 0
Mixture 35 0 0 0 0 0 35

Columns, real class; rows, predicted class.

Table 11
Classification of gasolines in sample Set C: Gasoline type.

Normal Regular Premium Total

Nright /N0 E Nright /N0 E Nright /N0 E Nright /N0 E

LDA 50/55 9% 40/45 11% 11/15 27% 101/115 12%

QDA 51/55 7% 40/45 11% 10/15 33% 101/115 12%
RDA 48/55 13% 41/45 9% 11/15 27% 100/115 13%
SIMCA 51/55 7% 40/45 11% 12/15 20% 103/115 10%
N-PLS 50/55 9% 42/45 7% 12/15 20% 107/115 10%
MLP 50/55 9% 42/45 7% 13/15 13% 110/115 9%
KNN 55/55 0% 45/45 0% 14/15 7% 114/115 1%
SVM 55/55 0% 44/45 2% 15/15 0% 114/115 1%
PNN 55/55 0% 45/45 0% 15/15 0% 115/115 0%

See comments for Table 6.

sophisticated classification algorithms: KNN, SVM, and PNN meth-
ods.
With respect to its effectiveness the classification methods can
be arranged in the following order:
PNN ≥ SVM ≥ KNN  ANN-MLP ≥ PLS > SIMCA ≥ RDA ≈ LDA ≈ QDA.
It should be noted that – except for accuracy – calibration
method is also characterized by simplicity for investigator (com-
prehensibility of main algorithms, availability and price of software,
etc.) and by volume of required calculations (capacity of comput-
ers for realization, time and price of a model creation, etc.) [5,8].
With respect to these parameters, the above-stated methods can
be grouped into two well-separated groups (see also Ref. [5]):
(i) according to computation time needed: resource demanding
(PNN, SVM, and ANN-MLP) and resource non-demanding (KNN,
PLS, SIMCA, and discriminant analysis methods);
(ii) ease of use: sophisticated (PNN, SVM, MLP, and PLS) vs. simple
(KNN, SIMCA, RDA, LDA, and QDA).
ANN training is more than 103 times time-consuming than
MLR or PLS model building approach [5,8]. These data should
also be considered when making a decision on advantages of one
or another method for processing of spectral and reference data.
One should note that these expenses are of no importance for
the end-consumer (e.g., refinery laboratory). No real difference in
calculation time of created model is observed. The same can, in
principle, be said about “ease of use”: if the models are just applied
by the end-user no difference is observed. The obstacles can be
met during model building where understanding of mathematical
principles, on which this or that model is based, is needed [5].
5. Conclusions
The results of different classification model applications are
shown in Fig. 5. We have examined different approaches to clas-
sify gasolines according to their source (refinery and process) and
type. We can conclude that:
34 R.M. Balabin et al. / Analytica Chimica Acta 671 (2010) 27–35
Fig. 5. Results of gasoline classification by different methods. LDA, linear discriminant analysis; QDA, quadratic discriminant analysis; RDA, regularized discriminant analysis;
SIMCA, soft independent modeling of class analogy; PLS, partial least squares regression; KNN, K-nearest neighbor; SVM, support vector machines; PNN, probabilistic neural
network; MLP, multilayer perceptron. PNN error of classification for Set A and C is equal to zero. SVM classification error for Set A is equal to zero.
(1) Three groups of classification methods were identified: KNN,
SVM, and PNN; MLP and PLS; SIMCA and discriminant anal-
ysis methods (LDA, RDA, and QDA). Each group has different
classification accuracy (see Section 4.4 for details).
(2) The probabilistic neural network (PNN) method is the most
effective way of gasoline classification using near infrared (NIR)
data. This method can be recommended for industrial gasoline
analysis.
(3) K-Nearest neighbor (KNN) and support vector machines (SVM)
classifiers have also shown adequate results. One should note
that since KNN technique is much easier for understanding and
computing, it can also be recommended for gasoline NIR spectra
analysis.
The choice of classification model depends on the researcher’s
knowledge: KNN can be recommended for people who are not
advanced in mathematics and PNN can be used by advanced
researchers. It should be noted that KNN software can easily be
written by a person with basic programming skills, while PNN
method is in need of special software (e.g., MATLAB Neural Network
Toolbox) or advanced programming skills.
Alternatively, for an advanced researcher, the choice of a method
depends strongly on the data structure and the aim of the study.
Therefore, the advanced researcher should have a clear strategy of
how to explore and/or model the data, to know when and why a
given method could be applied. Our data presented above can help
in the strategy choice.
We hope that the results obtained by us will help both fur-
ther chemometric investigations [48–50] and investigations in
the sphere of vibrational (infrared, near infrared, and Raman)
spectroscopy [51–59] of multicomponent systems. The results
presented herein can help in the rapid and accurate classi-
fication (or analysis) of biofuels (e.g., bioalcohols/alcohol fuel,
ethanol–gasoline fuel, cellulosic ethanol, bioethers, algae fuel),
products of petroleum refining (liquid petroleum gas, gasoline,
naphtha, kerosene/jet aircraft fuels, diesel fuel, (marine) fuel oils,
lubricating and industrial oils, paraffin wax, asphalt and tar, and
petroleum coke) and petrochemicals (olefins and their precursors,
aromatic hydrocarbons: e.g., benzene or mixed xylenes). The use
of near infrared spectroscopy in other fields of analytical (and/or
quantum [49,60–62]) chemistry, such as pharmaceutical (drug)
quality control, food quality control (e.g., green/black tea), active
pharmaceutical ingredient (API)/pharmakon analysis of tablets,
can be enhanced by application of modern methods of multi-
variate data analysis, including artificial neural networks as well
as other machine learning methods (data mining, pattern recog-
nition, and adaptive control) within the framework of Bayesian
statistics.
References
[1] B. Osborne, T. Fearn, Near Infrared Spectroscopy in Food Analysis, Wiley, New
York, 1986.
[2] F. Chauchard, R. Cogdill, S. Roussel, J.M. Roger, V. Bellon-Maurel, Chem. Intell.
Lab. Syst. 7 (2004) 141.
[3] P. Fayolle, J.-M. Roger, V. Steinmetz, L. Dusserre-Bresson, V. Bellon, Sensoral
98 Colloque international sur les capteurs de la qualite des produits agro-
alimentaires, in: ENSAM-Cemagref-INRA, Montpellier, France, 1998, p. 533.
[4] T. Hyvarinen, E. Herrala, J. Malinen, P. Niemla, NIR analysers can be miniature,
rugged and handheld, in: K. Hildrum, T. Isaksson, T. Naes, A. Tandberg (Eds.),
Near Infrared Spectroscopy, Bridging the Gap Between Data Analysis and NIR
Applications, Ellis Horwood, London, 1992, p. 1.
[5] R.M. Balabin, R.Z. Safieva, E.I. Lomakina, Chemometr. Intell. Lab. Syst. 88 (2007)
183.
[6] R.M. Balabin, R.Z. Safieva, J. Near Infrared Spectrosc. 15 (2007) 343.
[7] R.M. Balabin, R.Z. Syunyaev, S.A. Karpov, Fuel 86 (2007) 323.
[8] R.M. Balabin, R.Z. Safieva, E.I. Lomakina, Chemometr. Intell. Lab. Syst. 93 (2008)
58.
[9] R.M. Balabin, R.Z. Syunyaev, S.A. Karpov, Energy Fuels 21 (2007) 2460.
[10] N. Pasadakis, V. Gaganis, C. Foteinopoulos, Fuel Process. Technol. 87 (2006) 505.
[11] K. Brudzewski, A. Kesik, K. Kołodziejczyk, U. Zborowska, J. Ulaczyk, Fuel 85
(2006) 553.
[12] K. Brudzewski, S. Osowski, T. Markiewicz, J. Ulaczyk, Sens. Actuators B: Chem.
113 (2006) 135.
[13] J. Yoon, B. Lee, C. Han, Chemom. Intell. Lab. Syst. 64 (2002) 1.
[14] D.A. Burns, E.W. Ciurczak, Handbook of Near-Infrared Analysis, CRC Press, 2001.
[15] D. Donald, D. Coomans, Y. Everingham, D. Cozzolino, M. Gishen, T. Hancock,
Chemom. Intell. Lab. Syst. 82 (2006) 122.
[16] Q.-K. Zhang, L.-K. Dai, Control Instrum. Chem. Ind. 32 (2005) 53.
[17] T. Otto, R. Saupe, R. Bruch, U. Fritzsch, V. Stock, T. Gessner, N. Afanasyeva, Proc.
SPIE – Int. Soc. Opt. Eng. 4491 (2001) 234.
[18] M. Kim, Y.-H. Lee, C. Han, Comput. Chem. Eng. 24 (2000) 513.
[19] A.I. Belousov, S.A. Verzakov, J. von Frese, Chemom. Intell. Lab. Syst. 64 (2002)
15.
[20] D. Michie, D.J. Spiegelhalter, C.C. Taylor (Eds.), Mach. Learn., Neural and Statis-
tical Classification, http://www.amsta.leeds.ac.uk/∼charles/statlog/.
[21] M.S. Sánchez, L.A. Sarabia, Chemom. Intell. Lab. Syst. 28 (1995) 287.
[22] J.H. Friedman, J. Am. Stat. Assoc. 84 (1989) 165.
[23] S. Wold, M. Sjostrom, SIMCA: a method for analyzing chemical data in terms of
similarity and analogy, in: B.R. Kowalski (Ed.), Chemometrics Theory and Appli-
cation, American Chemical Society Symposium Series 52, American Chemical
Society, Washington, DC, 1977, pp. 243–282.
[24] Y. Tominaga, Chemom. Intell. Lab. Syst. 49 (1999) 105.
[25] S. Wold, A. Ruhe, H. Wold, W.J. Dunn III, SIMA J. Sci. Stat. Comput. (1984) 735.
[26] P. Geladi, B.R. Kowalski, Anal. Chim. Acta 185 (1986) 1.
[27] W.G. Glen, W.J. Dunn III, D.R. Scott, Tetrahedron Comput. Methodol. 2 (1989)
349.
[28] E. Fix, J.L. Hodges, Int. Stat. Rev. 57 (1989) 238.
[29] B.V. Dasarathy (Ed.), Nearest Neighbor (NN) Norms: NN Pattern Classification
Techniques, 1991.
[30] Shakhnarovish, Darrell, Indyk (Eds.), Nearest-Neighbor Methods in Learning
and Vision, The MIT Press, 2005.
[31] S.R. Amendolia, G. Cossu, M.L. Ganadu, B. Golosio, G.L. Masala, G.M. Mura,
Chemom. Intell. Lab. Syst. 69 (2003) 13.
 R.M. Balabin et al. / Analytica Chimica Acta 671 (2010) 27–35
[32] K. Fukunaga, Introduction to Statistical Pattern Recognition, 2nd ed., Academic
Press, Boston, MA, 1990.
[33] V.N. Vapnik, The Nature of Statistical Learning Theory, Springer-Verlag, New
York, 1995.
[34] T. Joachims, Text categorization with support vector machines: learning with
many relevant features, in: Proceedings of the 10th European Conference
on Machine Learning (ECML), Springer-Verlag, Berlin, Heidelberg, New York,
1998.
[35] S. Dumais, et al., Proceedings of the Conference on Information and Knowledge
Management, 1998.
[36] M.A. Hearst, B. Scholkopf, S. Dumais, E. Osuna, J. Platt, IEEE Intell. Syst. 13 (1998)
18.
[37] Y. Yao, P. Frasconi, M. Pontil, Proceedings of the 3rd International Con-
ference of Audio- and Video-Based Person Authentication AVBPA, 2001,
p. 253.
[38] N. Cristianini, J. Shave-Taylor, An Introduction to Support Vector Machine (and
Other Kernel-Based Learning Methods), Cambridge Univ. Press, Cambridge, UK,
2000.
[39] V.N. Vapnik, Statistical Learning Theory, Wiley, New York, 1998.
[40] M. Pontil, A. Verri, Properties of support vector machines, Neural Comput. 10
(1998) 955.
[41] D.F. Specht, IEEE International Conference on Neural Networks, 1988, p. 525,
doi:10.1109/ICNN.1988.23887.
[42] D. Specht, Neural Networks 3 (1990) 109.
[43] M.H. Hassoun, Fundamentals of Artificial Neural Networks, 1st ed., MIT Press
Cambridge, MA, USA, 1995.
35
[44] S. Haykin, Neural Networks: A Comprehensive Foundation, Macmillan College
Publishing Company, NY, 1994.
[45] H. Demuth, M. Beale, M. Hagan, Matlab Neural Network Toolbox 5, User’s Guide,
www.mathworks.com/access/helpdesk/help/pdf doc/nnet/nnet.pdf.
[46] J.G. Speight, B. Ozum (Eds.), Petroleum Refining Processes, 1st ed., CRC, 2001.
[47] R.A. Meyers, Handbook of Petroleum Refining Processes, 3rd ed., McGraw-Hill
Professional, 2003.
[48] R.M. Balabin, R.Z. Safieva, Fuel 87 (2008) 2745.
[49] R.M. Balabin, E.I. Lomakina, J. Chem. Phys. 131 (2009) 074104.
[50] R.M. Balabin, R.Z. Syunyaev, J. Colloid Interface Sci. 318 (2008) 167.
[51] M. Blanco, J. Coello, H. Iturriaga, S. Maspoch, J. Pagès, Chemometr. Intell. Lab.
Syst. 50 (2000) 75.
[52] R.M. Balabin, J. Phys. Chem. A 113 (2009) 4910.
[53] R.M. Balabin, J. Phys. Chem. A 113 (2009) 1012.
[54] R.M. Balabin, J. Phys. Chem. Lett. 1 (2010) 20.
[55] R.Z. Syunyaev, R.M. Balabin, I.S. Akhatov, J.O. Safieva, Energy Fuels 23 (2009)
1230.
[56] R.M. Balabin, J. Disper. Sci. Technol. 29 (2008) 457.
[57] R.Z. Syunyaev, R.M. Balabin, J. Disper. Sci. Technol. 28 (2007) 419.
[58] Y. Roggo, P. Chalus, L. Maurer, C. Lema-Martinez, A. Edmond, N. Jent, J. Pharm.
Biomed. Anal. 44 (2007) 683.
[59] R. M. Balabin, Phys. Chem. Chem. Phys., 12 (2010) in press, doi:10.1039/
b924029b.
[60] R.M. Balabin, Chem. Phys. 352 (2008) 267.
[61] R.M. Balabin, J. Chem. Phys. 129 (2008) 164101.
[62] R.M. Balabin, J. Phys. Chem. A 114 (2010) 3698.