Fuel 104 (2013) 798–804

Contents lists available at SciVerse ScienceDirect

Fuel

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

Study of motor oil adulteration by infrared spectroscopyand chemometrics methods

M. Bassbasi a, A. Hafid a, S. Platikanov b, R. Tauler b, A. Oussama a,⇑a Laboratory of Applied Spectrochemistry and Environmental, Faculty of Sciences and Technics of Beni Mellal, University Moulay Soulymane, Moroccob IDAEA-CSIC, Jordi Girona, 18-26 Barcelona, Spain

h i g h l i g h t s

" Adulteration of motor oils is investigated by FT-IR and chemometrics." PCA, PLS2-DA and PLSR make possible to discriminate among three types of oils." Models predict the amount of oil adulteration with very high accuracy.

a r t i c l e i n f o

Article history:Received 10 January 2012Received in revised form 21 May 2012Accepted 29 May 2012Available online 19 June 2012

Keywords:Motor oilAuthenticationAdulterationATR-MIRPLS-DA

0016-2361/$ - see front matter � 2012 Elsevier Ltd. Ahttp://dx.doi.org/10.1016/j.fuel.2012.05.058

⇑ Corresponding author.E-mail address: oussamaabdelkhalek@yahoo.fr (A.

a b s t r a c t

Fourier transform infrared spectroscopy (FTIR) coupled to chemometrics techniques was used to inves-tigate high quality motor oils samples adulterated with lower quality oils, like used oils and standard oils.The results showed that Partial Least Squares (PLS) models based on infrared spectra were a suitable ana-lytical method for predicting adulteration of high quality motor oils in the concentration range from 0% to36% (w/w), with prediction errors lower than 3% (w/w). Partial Least Squares Discriminate Analysis(PLS2-DA) gave good classification results with 100% correct class prediction, in the spectral range of1800–600 cm–1 and concentration range of 0–20% w/w for the two tested oil adulterants in their binarymixtures with the high quality oil. The proposed method can be employed for quality monitoring andcontrol and rapid screening analysis of adulterated motor oils.

� 2012 Elsevier Ltd. All rights reserved.

1. Introduction

Worldwide directives and legislation for environmental qualityhave established a set of lower levels for airborne pollutant emis-sions from combustion. Nowadays, motor manufacturers and fuelcompanies are investing large amounts of resources in green tech-nologies in order to fulfil the increasingly strict emission require-ments. On the other hand, merchandise profit goals using cheapsubstituent products instead of more expensive higher technolog-ical petroleum products has arisen the question about qualityauthentication. Motor oils, as products of petroleum refinery, arefundamental for optimum car performance [1]. Among the majorfunctions of motor oils are to protect the engine from many phys-ically and chemically related malfunctions like heating, corrosionand contamination [1,2]. As a part of the motor combustion, themotor oils qualities have become increasingly important, not onlyas motor lubricants, but also because of the impact of the combus-tion emissions emitted in the environment.

ll rights reserved.

Oussama).

Many products of the petroleum industry, such as heavy crudeoils, gasoline, diesel fuel, jet fuel and others are investigated toascertain for their quality and authenticity, using different chro-matographic and spectroscopic methods [3,4]. Instrumental analy-sis techniques like gas chromatography, high performance liquidchromatography, NMR and Mass spectrometry have been widelyused for this purpose, however, these techniques are rather expen-sive, time-consuming, they require skilled operators and even theycan have a high environmental impact.

Infrared spectroscopy has always had a significant place in lu-bricant analysis to characterize qualitatively its different constitu-ents. Near Infrared Spectroscopy (NIRS) and Mid InfraredSpectroscopy (MIRS) coupled to chemometrics methods have beenshown to be powerful techniques for authentication of petrol [5]due to their simple application and robustness, fast performanceand cheap sample preparation. With the advent of FTIR spectros-copy, the possibility of developing quantitative methods of lubri-cant analysis is facilitated. It is because of the inherentspectroscopic power of FTIR instruments, as well as of the ad-vances in sample handling techniques and of the more availabilityof new chemometrics methods, that quantitative analysis using

Table 1Motor oil characteristics of the three motor oils used in the study (H, N and U).

Method of analysis: Norm NF EN ISO 3104

Kinematic viscosity at40 �C

Kinematic viscosity at100 �C

High quality oilH

88.36 mm2/g 14.07 mm2/g

Standard oil N 90.29 mm2/g 11.68 mm2/gUsed oil U 26.88 mm2/g 5.78 mm2/g

M. Bassbasi et al. / Fuel 104 (2013) 798–804 799

Mid Infrared Spectroscopy MIRS has been greatly facilitated [6].Despite the fact that motor oils are also among the key productsof petroleum industry [7,8], there are not many papers dealingwith adulteration studies of oil products [9]. In recent study hasbeen shown that coupling NIRS with chemometrics is suitable forquantitative assessment of food adulteration of virgin olive oilswith low quality vegetable oils [10].

Preliminary studies from Balabin et al. [3,11] have revealed theimportance of quality control of motor oil adulteration by means ofNIR. A large number of chemometric techniques (linear and nonlin-ear methods like SIMCA, PLS, KNN, MLP and SVM) were used [11]to evaluate the origin of motor oils according to their base stock(synthetic, semi synthetic, and mineral) and to their kinematic vis-cosity at low (SAE 0 W, 5 W, 10 W, 15 W) and high temperatures(SAE 20, 30, 40, 50). Moreover, the possible differentiation andclassification between different types of commercial motor oilsusing IR Spectroscopy and mathematical data analysis procedureshas been shown by Zieba-Palus et al. [9]. PLS quantification hasbeen applied to NIR reflectance spectroscopy for the determinationof motor oil contamination in sandy loam [12].

To continue and increase the existing knowledge about motoroil authentication procedures, as a key factor in the identificationof possible product adulterations, in this work we propose theuse of FTIR coupled to chemometrics methods. The goal of thisstudy is to propose a new procedure to replace traditional labora-tory methods, which are relatively slow, inaccurate, and requirethe use of expensive polluting chemicals by newer methods basedon the coupling of IR spectroscopy and chemometrics methods. Inparticular, FTIR spectroscopy coupled to chemometrics methods,like PCA, PLSR and PLS2-DA, will be tested for discriminationand/or classification of different oil types and for quality assess-ment and quantification of possible oil adulteration in an easy,accurate and fast way. Therefore, the main objective of this workis to develop and to propose a simple analytical method, basedon FTIR and chemometrics methods to determine and to quantifythe possible adulteration of highly valued motor oil using lowquality oil alternatives.

2. Materials and methods

2.1. Sampling

In these work, three classes of motor oils have been investi-gated: high quality high priced motor oil (H, synthetic 10W40 mo-tor oil), standard commercial low-cost motor oil (N, commercial oilsold without any indication of its origin or its quality) and finally,used motor oil (U, also sold without being authenticated). Thesethree types of oils were purchased from local market in Moroccoand are assumed as being different types and brands. Table 1shows the motor oils characteristics.

MIR spectra of pure oil samples and of their binary combinationin mixtures were measured and used for data analysis. In total, 111samples were prepared in two sets. A first set (included in Study A)included 61 samples and it was analyzed in order to conduct a clas-sification study about the adulteration of a high quality motor oil(H) with the other two lower quality oil substituent (N and U).The second data set (included in Study B) was composed of 50samples and it was used for quantification of adulteration of thehigh quality oil (H) produced by the standard cheap motor oil (N).

2.2. Classification study (Study A)

Sixty samples were prepared in binary mixtures using differentbinary combinations of the investigated motor oils. The investi-gated adulteration range was 0–20% in weight (w/w). One first

subset of 20 samples was prepared where the high quality oil (H)was adulterated with the standard cheap motor oil (N) and notedas NH oil mixtures group. A second group of 20 samples was com-posed of binary mixtures between the high quality oil (H) and theused oil (U) and noted as the UH oil mixtures group. And a thirdgroup of 20 samples was prepared mixing the standard oil (N) withthe used oil (U) and noted as the UN oil mixtures group. All thesesamples were analyzed by MIR and their corresponding spectra aregiven in Fig. 1b. Also the pure spectrum of high quality motor oilwas included in the data. These 61 samples were preliminary ex-plored in order to discriminate among these three oil mixturesand the pure H spectrum. These 61 samples were further randomlysubdivided into 2 new subsets of 46 and 15 samples. The firstgroup of 46 samples (the pure spectrum of high quality was in-cluded in this data subset) was used for the calibration of the clas-sification model; and the second group of 15 selected samples wasused to validate the model externally.

2.3. Quantitative study of the adulteration of high quality motor oil(Study B)

Another set of 50 oil mixture samples were prepared for thequantitative study. The selected samples were prepared in binarymixtures (NH) of the standard cheap motor oil (N) and the highquality oil (H), at percentages ranging from 1% to 36% (w/w) of con-centration of the standard cheap motor oil. Also pure samples of thetwo motor oils (N and H) types were considered in the study. Thestandard motor oil (N) was chosen as the adulterant of the highquality oil because of its easy availability at a considerably reducedprice compared to that of the high quality oil (H). During the modelcalibration, a 0% of adulteration was assumed for the spectrum ofthe pure sample of high quality oil (H) and a 100% of adulterationwas assumed for the pure spectrum of the standard cheap oil (N).The study was repeated considering 3 different calibration scenar-ios in respect to the included adulterant concentration ranges, i.e.1–36% (only binary mixtures, NH), 0–36% (pure high quality oiland binary mixtures, H and NH) and 0–100% (pure high qualityoil, pure standard cheap oil and binary mixtures, H, N and NH).

Three different spectral ranges data were also tested for each ofthe above mentioned calibration ranges, i.e. the full spectral range4000–600 cm�1, the spectral range 3000–600 cm�1; and the spec-tral range 1800–600 cm�1.

The 50 samples used for the quantitative study were also fur-ther randomly subdivided into two subsets, one subset for themodel calibration (38 samples) and another subset for the modelexternal validation (12 samples). All samples were kept and ana-lyzed under similar conditions by MIR.

2.4. Acquisition of MIR spectra

A Bruker Vector 22 instrument equipped with a DTGS detector,Globar (MIR) source, and KBr separator was used to record thespectra of all oil mixtures and of pure motor oils. Measurementswere taken within the range 4000–400 cm�1 at a resolution of4 cm�1. All the experimental work reported in this paper has been

Fig. 1. (a) ATR-MIR spectra of the three pure oil samples (included in Study A) in the spectral range 4000–600 cm�1. In blue, pure standard cheap motor oil (NN), in green,pure high quality motor oil (HH) and in red, used oil (UU); (b) ATR-MIR spectra of the three type binary mixture samples (included in Study A) in the spectral range 4000–600 cm�1; (c) ATR-MIR absorbance spectra in the spectral range 1800–600 cm�1of the 50 oil NH mixture samples and pure N (standard cheap motor oil) and pure H (highquality motor oil) spectra used for PLSR modeling and quantification (included in Study B).

800 M. Bassbasi et al. / Fuel 104 (2013) 798–804

conducted in the laboratory of environmental and applied Spectro-chemistry, from the University of Beni Mellal in Morocco.

3. Chemometrics

The starting point of chemometrics analysis is the preliminaryspectral pretreatment before the development of an optimal cali-

bration model. Usually first or second derivatives (using Gap-Nor-ris and Savitzky-Golay algorithms) [13,14] or the application ofspectra normalization (Standard Normal Variate, SNV) and others,are used as spectral initial pretreatment [15].

For the preliminary spectra exploration study, FTIR spectra ofthe first set of 60 samples set (Study A) was initially investigatedusing PCA [16]. This projection method extracts information about

1 For interpretation of color in Figs. 1–3, the reader is referred to the web version othis article.

M. Bassbasi et al. / Fuel 104 (2013) 798–804 801

the latent (hidden) structures of the data set. It transforms a largenumber of correlated original variables (FTIR spectra) into a smal-ler number of uncorrelated, orthogonal variables explaining maxi-mum variance, called principal components (PCs) [16]. Samples(motor oil mixtures) are projected on these principal componentsgiving the samples scores. Plots of these sample scores and of thecorresponding variable loadings in principal components allowthe interpretation of the main sources of data variance. The mainadvantage of using PCA is that it reduces the dimensionality ofthe problem (number of variables) but retaining most of the origi-nal variability in the experimental data and filtering noise and ofminor irrelevant sources of variance. Therefore, PCA allows for asimpler interpretation of variance sources in a particular data set.

Since the spectral data (independent variables, data matrix X)also contains rich quantitative information, a calibration model(Study B) is proposed to extract quantitative information and topredict the percentage (response variable, y) of oil adulterationin oil mixtures (relative amounts of N oil type in H oil type). In thiswork, the Partial Least Squares Regression (PLSR) [17] method hasbeen used to develop such a calibration model. PLSR attempts tomaximize the covariance between X and y data blocks. PLSsearches for the factor subspace most congruent to both datablocks, and its predictions are usually better than using other mul-tilinear regressions methods such as MLR. A new matrix of weights(reflecting the covariance structure between the X predictors and yresponse variables) is calculated and can provide rich factor inter-pretation and information for each latent variable independently.

For the classification studies, the Partial Least Squares discrim-inate analysis method, PLS-DA, [6] usually is applied. This tech-nique finds the components or latent variables whichdiscriminate as much as possible between two different groupsof samples from their FT-IR spectra (X block) according to theirmaximum covariance with a target class defined in the y datablock. It attempts to describe whether a spectrum of a sample be-longs or not to a particular class, consisting of zeros and ones.According to the number of simultaneously regressed y vectorstwo different PLS-DA approaches are possible. In case of only oneclass is modeled at a time the method is the ordinary PLS1-DA.When several classes are simultaneously modeled at the sametime, the PLS2-DA modified method can be used [18]. For the clas-sification study (Study A) in this work PLS2-DA, [6] was used.

The selection of optimal number of components in PCA and oflatent variables in PLSR is done using the lowest prediction errorin cross validation (leaving-out-one sample at a time) related tothe PRESSk, the sum of squares prediction error for the modelwhich includes k factors (components), and optimal prediction ofy values for the external validation samples not included in the cal-ibration step. The model giving the lowest relative prediction er-rors in external validation is finally chosen.

3.1. Figures of merit

Quality assessment of the obtained results was discussed bycomparison of different figures of merit like the coefficient ofdetermination of model fitting or R2; Root mean squared error ofcalibration or of external validation/prediction (RMSEC andRMSEP), calculated as follows:

RMSEC ¼

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiXn

i¼1

ðyi � yiÞ2

n

vuuut;

where the yi are the values of the model predicted concentrationsand yi are the actual measured (experimental) values, when calibra-tion samples are included in the development of the model and n isthe number of samples. RMSEC is a measure of how well the model

fits experimental concentrations. RMSEP is calculated exactly asRMSEC except that the estimates are now the values from externalvalidation samples. RMSEP is a measure of how well the model willmake predictions.

In this work the calculation of relative prediction errors of con-centrations in percentage is also given, for both calibration andprediction steps and they are calculated as follows:

Rel: error in % ¼

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiXn

i¼1

ðyi � yiÞ2

Xn

i¼1

ðyiÞ2

vuuuuuuut� 100;

Quality assessment of the obtained results is discussed by comparisonof predicted values vs. measured values, both for calibration and forvalidation data sets. All chemometrics calculations were performedusing The Unscrambler software (version 10.1, Camo, Norway).

4. Results

4.1. Spectral features

In Fig. 1a and b, all experimental spectra are dominated by fourclusters of spectral bands. The first group of spectral bands are ob-served in the spectral region between 2850 cm�1 and 2920 cm�1

arising from symmetric and asymmetric stretching vibrations ofmethylene (ACH2) and methyl (ACH3) groups [19]. The peak ataround 2360 cm�1 is due to the presence of fluctuations indepen-dent from sample composition [20]. The second group of bands in-cludes the spectral region around 1750 cm�1, arising from thevibration of the carbonyl group (AC@O) and it is characteristicfor the pure standard cheap oil sample. The third cluster of bandsis in the spectral region between 900 cm�1 and 1400 cm�1, whichare attributed to the vibrations of (ACH2), (ACH3) and aliphatic(ACOCA) groups. This area shows certain variations among thepure spectra of all studied motor oils. A remarkable variation is ob-served in the peak heights around 1163 cm�1 for the standardcheap sample (Fig. 1a), proved to be related to the stretching vibra-tion of the AC@O ester groups [21]. The last cluster of spectralbands is observed in the region of 723 cm�1 and is due to vibra-tions of cis AHC@CHA [22].

4.2. Exploration and classification (Study A) of motor oil binarymixtures by Mid Infrared Spectroscopy (MIR)

Principal component analysis with full cross validation was ap-plied to the first data set of 61 classification samples (see Materialsand Methods section) exploring the full acquired spectral range(4000–600 cm�1). In this study, PCA has been used as an exploratorytool to investigate the major trends in the whole collected spectraand to figure out the factors that give the possibility to discriminateamong pure samples and binary mixtures of motor oil samples.

Although several data pretreatments were tested, like first andsecond derivatives, and multiplicative scatter correction, the SNVmethod was finally found to be the one giving the best resultsand it was then applied as a pretreatment procedure before PCAmodeling.

The PCA model with three components already explained 98%of the total data variance (PC1 captured 85% and PC2 captured11% of the variance respectively). PC1 vs. PC2 scores plot of thespectra of the first data set given in Fig. 2a, distinguished three ma-jor clusters of samples (in blue, red and green color1s). The score of

f

Fig. 2. (a) PCA scores plot (PC1 vs. PC2) for the analysis of the spectra (4000–600 cm�1 range) of oil mixture samples in the adulteration range 0–20% w/w; (b) PCA loadingsplot on PC1 for the analysis of the spectra (4000–600 cm�1 range); (c) PCA scores plot (PC1 vs. PC3) for the analysis of the spectra (4000–600 cm�1 range) of oil mixturesamples in the adulteration range 1–20% w/w; (d) PCA loadings plot on PC3 for the analysis of the spectra (4000–600 cm�1 range). HH sample is the spectrum of pure highquality motor oil; NH samples are binary mixtures of high quality motor oil adulterated with different amounts of standard cheap motor oil; UH samples are binary mixturesof high quality motor oil adulterated with different amounts of used motor oil; UN samples are binary mixtures of standard cheap motor oil adulterated with used motor oil;

802 M. Bassbasi et al. / Fuel 104 (2013) 798–804

the high quality oil HH (in black) did not differentiate from thosebinary mixtures of NH group where the concentrations of the N adul-terant (standard cheap oil) are very low. However on the PC1 vs. PC3scores plot (Fig. 2c) already was possible to discriminate betweenthe high quality oil’s score and all scores of NH group. The FirstPC1 differentiates the samples according to the motor oil quality.Decreasing oil quality is in the direction from negative to positivescores on PC1, in agreement with having the HH (black) and NHGroup (blue) in the very left part of the plot (more negative scores),and the lowest quality oil mixtures (UN green) at the more right partof the plot (more positive scores). Mixtures between used and highquality oils (UH) are at the middle position of the PC1 axis, betweenthe two other oil mixture groups. PC2 distinguishes among differentconcentration levels of adulteration, without considering a particularadulterant. The information on PC3 assumes the effect of the type ofthe adulterant in the high quality motor oil and also very well com-plete the information gathered on the first PC. Both components incombination make possible adulteration detection.

Moreover, exploring the loadings plots of PC1 and PC3 (Fig. 2band d) it is possible to conclude that the decreasing the overallquality of the oil (from HH > NH > UH > UN) or in adulteration,the FTIR spectra changes in a particular manner. Spectral bands,from the FTIR spectrum of high quality motor oil, in the regionaround 1750 cm�1 (the vibration of the carbonyl groups AC@O),1147 cm�1 (most probably ACH2, (ACH3) and aliphatic (ACOCA)groups, around 700 (AHC@CHA, cis) cm�1 will show higher load-ings. In smaller extent will be affected the pure spectrum FTIR ofthe high quality oil in the area of 2940–2850 cm�1 with decreasingof the oil quality or in adulteration. The area close to 2360 cm�1

has not to be assumed because it is due to the presence of fluctu-ations independent from sample composition [20].

Once proved that the spectral range of 1800–600 cm�1 is themost informative wavenumber range for discrimination among

pure samples and binary mixtures of motor oil samples, the nextstep is to build a classification model using PLS2-DA. This PLS2-DA model was built considering the spectral range 1800–600 cm�1 as X variables, while the y variables were associated withthe three different oil mixture classes and the pure high quality oilsample (one different y variable for each oil class, with 1 or 0depending on whether it belongs or not to the considered datagroup). The model obtained in this way was able to discriminateamong the three oil mixture classes (UN, UH and NH) and the pureHH, as it can be seen from the PLS2-DA scores plot in Fig. 3, wherenow the clusters are better distinguished than PCA clusters inFig. 2. The first PLS2 latent variable (LV1 explains 46% of y varianceand 87% of X variance) discriminates between the oil mixtures hav-ing the high quality oil from the low-standard combination of oilmixtures. The third PLS2 latent variable explains a rather high per-centage of variance of y (29%) based only on a small amount of var-iance in X (1%), related again with the amount and type of oiladulteration. High positive scores on LV3 are related with smallamounts of adulterant, and high negative scores on LV3 are relatedwith large amounts of adulterant.

Table 2 shows the calculated figures of merit of the results ob-tained by the PLS2-DA model using the 46 calibration samples sub-set (see Material and Methods section). A high correlation betweenmeasured and predicted classes (R2 around 0.99 in all cases) andlow prediction errors (RMSEC between 0.04 and 0.05) were ob-tained. When the previously obtained PLS2-DA model was appliedto classify the 15 oil validation samples subset (see Material andMethods section) a high correlation coefficient (R2 0.99) and alow prediction error (RMSEP) between 0.04 and 0.05 were also ob-tained (similar than for the spectra calibration set).

Table 3 shows that for the 15 validation samples (5 from everyclass), a 100% correct classification was achieved, i.e. all mixture oilspectra of the validation data set were correctly matched to the

Fig. 3. PLS2-DA scores plot (LV1 vs. LV3) in the analysis of the spectra (1800–600 cm�1 range) of oil mixture samples in the adulteration range 1–20% w/w. HH, NH, UH andUN as in Fig. 2 caption.

Table 3Results of the classification of motor oil binary samples for the prediction of samplesnot used in calibration external validation data set.

Samples Classesa

NH UH UN

ECH NH-Pred NH-Ref UH-Pred U.H-Ref UN-Pred UN-Ref

NH41 1,05 1 0,00 0 �0,05 0NH47 0,98 1 0,01 0 0,01 0NH48 1,00 1 0,09 0 �0,09 0NH53 1,07 1 �0,10 0 0,03 0NH55 1,02 1 �0,07 0 0,05 0UH11 �0,04 0 1,01 1 0,02 0UH13 0,13 0 0,90 1 �0,02 0UH17 �0,01 0 0,98 1 0,03 0UH3 0,01 0 1,03 1 �0,04 0UH7 0,03 0 1,01 1 �0,04 0UN11 �0,04 0 0,04 0 1,00 1UN15 0,02 0 0,03 0 0,95 1UN19 �0,05 0 0,09 0 0,96 1UN5 �0,01 0 �0,03 0 1,04 1UN9 �0,01 0 0,00 0 1,01 1

a NH, UH and UN are defined in Fig. 2 caption. Values close to one meanbelonging to the class, values close to zero mean not belonging to the class.

M. Bassbasi et al. / Fuel 104 (2013) 798–804 803

three corresponding classes. In Table 3, the predicted values by thePLS2-DA model are always very close to 1. These results confirmthat the predictive ability of the developed PLS2-DA model wasvery good. Therefore, it was concluded again that FTIR spectros-copy together with the application of chemometrics PLS2-DAmethod can be used to discriminate adulterated motor oils.

4.3. Quantitative determination of the amount of adulteration of highquality motor oil with a cheap oil substituent (Study B)

Table 4 summarizes the figures of merit obtained for the differ-ent PLSR models tested in this work. All of them were built using 2LVs. For brevity in this discussion, only the relative errors (in%) inthe prediction of adulteration of the high quality oil (H) by thestandard oil (N) for the external validation mixture oil samplesare given. The best data pretreatment resulted to be again meanc-entering in combination with SNV method, with relative predictionerrors in the content of the adulterant (N oil) of 2.6–4% in theexternal validation data sets. Predictions were better if the purehigh quality oil spectrum was included in the calibration data set(0% of oil adulteration). On the contrary, when pure cheap oil spec-tra were included in the calibration data set (100% adulteration),relative prediction errors were worse (starting at 6% of error). Mostprobably this fact was due to the extrapolation of the model due tomissing values in the calibration data set in the range of 36–100%of adulteration and possible to the presence of spectra non-linear-ity associated to these higher concentration ratios. Further explo-ration would be needed to ascertain what happens at largerconcentration ratios of adulterants and to find the way to modelthem. However this is of less interest for the purposes of this work,

Table 2Figures of merit achieved by PLS2-DA discrimination of the three different types ofmotor oil mixture samples (NH, UH and UN).

Classesa Figures of meritb

LVs R2c

RMSEC RMSEP

NH 4 0.990 0.0464 0.0451UH 4 0.989 0.0486 0.0554UN 4 0.987 0.0532 0.0407

a Investigated classes by PLS-DA. NH, UH and UN are defined in Fig. 2 caption.b Reported model figures of merit: LVs-number of Latent Variables; R2

c – R-squarein calibration; RMSEC-Root Mean Squared Error in Calibration; RMSEP-Root MeanSquared Error in external validation.

since higher adulteration concentrations ratios are seldom presentas in practice in commercial oil samples in the market.

Finally, in relation to the spectral range more useful for theinvestigation of oil adulteration, it appeared that the lowest rela-tive prediction errors were obtained for the spectral range 1800–600 cm�1 (Fig. 1c) The models assuming both (4000–600 cm�1

and 3000–600 cm�1) spectral ranges showed more or less similarprediction results strongly related to the investigated concentra-tion calibration range. The best predictive PLSR model was estab-lished at the calibration range of 1–30% (for the adulteration ofhigh quality motor oil with cheap motor oil), when meancenteringand SNV pretreatments were applied to the 1800–600 cm�1 spec-tral range.

5. Conclusions

FTIR spectroscopy coupled to chemometrics methods has beenshown to be a simple, useful and powerful technique for the detec-tion of adulteration of high quality motor oils with cheaper or/andused oil substituent. Both PCA and PLS2-DA chemometrics

Table 4Figures of merit obtained in PLSR modeling under different conditions for calibration, internal cross validation and external validation of motor oil mixturesa.

Wavenumber (cm�1)c 0–100%b 0–36% 1–36%

Data pretreatment LVs R2C

RMSEC RMSEP RE% LVs R2C

RMSEC RMSEP RE% LVs R2C RMSEC RMSEP RE%

Meancentering + SNV 4000–600 2 0.99 1.87 1.55 6.3 2 0.99 0.84 0.7 2.9 2 0.99 0.73 0.7 2.93000–600 2 0.99 1.9 1.69 6.9 2 0.99 1.07 1.04 4.3 2 0.99 1.08 1.08 4.41800–600 2 0.99 2.43 2.53 10 2 0.99 0.74 0.65 2.7 2 0.99 0.71 0.63 2.6

a Reported model figures of merit: LVs-number of Latent Variables; R2c – R-square in calibration; RMSEC-Root Mean Squared Error in calibration; RMSEP-Root Mean

Squared Error in external validation; RE%-relative error of prediction in% for external validation data.b Invsetigated calibration range.c Investigated spectral range.

804 M. Bassbasi et al. / Fuel 104 (2013) 798–804

methods were able to discriminate accurately to what class be-longs a particular adulterated motor oil sample (in binary mix-tures) in the concentration range 0–20% w/w. Both methods (PCAand PLS-DA) could be used as rapid multivariate techniques formotor oil adulteration assessment.

PLSR calibration models, developed in the 0–36% concentrationcalibration range w/w gave less than 3% relative error for externalvalidation samples. We can conclude therefore, that FTIR spectra ofmotor oils could be properly modeled by PLSR using meancenter-ing and SNV as initial data pretreatments. This study has alsounderlined that the spectral region 1800–600 cm�1 was especiallyuseful for good predictions of motor oil adulterant.

This investigation concluded that combining ATR-MIR withchemometrics can be used for motor oil adulteration detection.Coupling both, ATR-MID and chemometrics, guarantee good pre-diction of oil adulterated concentrations with minimal samplepreparation, robustness, high-throughput and minimum back-ground training. Also this approach can be used in routine workfor quality control strategies allowing a rapid authentication orquantification of adulteration problems in the field of motor oilsindustry.

Acknowledgements

The authors are thankful for the financial help for Project Coop-eration between Morocco and Spain, AECID B/031059/10(extension).

References

[1] Yee FC, Yunus RM, Sin TS. 2nd World Engineering Congress Sarawak, Malaysia[22.7.2002].

[2] Al-Ghouti MA, Al-Atoum L. Virgin and recycled engine oil differentiation: aspectroscopic study. J Environ Manage 2009;90:187–95.

[3] Balabin R, Safieva R. Motor oil classification by base stock and viscosity basedon near infrared (NIR) spectroscopy data. Fuel 2008;87:2745–52.

[4] Silva A, Silva A, Ribeiro J, Martins F, Da Silva F, Silva C. Chromatographic andspectroscopic analysis of heavy crude oil mixtures with emphasis in

nuclear magnetic resonance spectroscopy: a review. Anal Chim Acta2011;707:18–37.

[5] Al-Ghouti MA, Al-Degs YS, Amer M. Application of chemometrics and FTIR fordetermination of viscosity. Talanta 2010;81:1096–101.

[6] De Luca M, Terouzi W, Ioele G, Kzaiber F, Oussama A, Oliverio F, et al.Derivative FTIR spectroscopy for cluster analysis and classification of moroccoolive oils. Food Chem 2011;124:113–8.

[7] Speight JG, Ozum B. Petroleum refining processes. New York, NY: MarcelDekker Inc; 2002.

[8] Meyers RA. Handbook of petroleum refining processes. New York,NY: McGraw-Hill Professional; 2004.

[9] Zieba-Palus J, Koscielniak P, Lacki M. Differentiation of used motor oils on thebasis their IR spectra with application of cluster analysis. J Mol Struct2001;596:221–8.

[10] Christy A, Kasemsumran S, Du Y, Ozaki Y. The detection and quantification ofadulteration in olive oil by near infrared spectroscopy and chemometrics. AnalSci 2004;20:935–40.

[11] Balabin R, Safieva R, Lomakina E. Near-infrared (NIR) spectroscopy for motoroil classification: From discriminant analysis to support vector machines.Microchemical J 2011;98:121–8.

[12] Stallard BR, Garcia MJ, Kaushik S. Near-IR reflectance spectroscopy for thedetermination of motor oil contamination in sandy loam. Appl Spectrosc1996;50:334–8.

[13] Norris K, Williams P. Optimization of mathematical treatments of raw near-infrared signal in the measurement of protein in hard red spring wheat. I.Influence of particle size. Cereal Chem 1984;61:158–65.

[14] Savitzky A, Golay M. Smoothing and differentiation of data by simplified leastsquares procedures. Anal Chem 1964;36:1627–39.

[15] Esbensen K. Multivariate data analysis – In practice. 5th ed. Oslo, Nor: CAMOProcess AS; 2002.

[16] Jolliffe IT. Principal component analysis. New York, NY: Springer; 1986.[17] Geladi P, Kowalski B. Partial least-squares regression: a tutorial. Anal Chim

Acta 1986;185:1–17.[18] Sjöström M, Wold S, Söderström B. PLS discriminant plots. In: Gelsema E, Kanal

L, editors. Pattern recognition in practice II. Amsterdam: Elsevier; 1986.[19] Guillén M, Ruiz A, Cabo N. Study of the oxidative degradation of farmed

salmon lipids by jeans of Fourier transform infrared spectroscopy. Influence ofsalting. J Sci Food Agr 2004;84:1528–34.

[20] Iñón FA, Garrigues JM, Garrigues S, Molina A, de la Guardia M. Selection ofcalibration set samples in determination of olive oil acidity by partial leastsquares-attenuated total reflectance-Fourier transform infrared spectroscopy.Anal Chim Acta 2003;489:59–75.

[21] Vlachos N, Skopelitis Y, Psaroudaki M, Konstantinidou V, Chatzilazarou A,Tegou E. Application of Fourier transform-infrared spectroscopy toedible oils.Anal Chim Acta 2006;573–74:459–65.

[22] Al-Alawi A, van de Voort FR, Sedman J. New FTIR method for determination offree fatty acids in oil. J Am Oil Chem Soc 2004;81:441–6.