Chemometric Study of Excitation–Emission Matrix FluorescenceData: Quantitative Analysis of Petrol–Kerosene Mixtures

O. DIVYA and ASHOK K. MISHRA*Department of Chemistry, Indian Institute of Technology Madras, Chennai-36, India

Products of petroleum crude are multifluorophoric in nature due to the

presence of a mixture of a variety polycyclic aromatic hydrocarbons

(PAHs). The use of excitation–emission matrix fluorescence (EEMF)

spectroscopy for the analysis of such multifluorophoric samples is gaining

progressive acceptance. In this work, EEMF spectroscopic data is

processed using chemometric multivariate methods to develop a reliable

calibration model for the quantitative determination of kerosene fraction

present in petrol. The application of the N-way partial least squares

regression (N-PLS) method was found to be very efficient for the

estimation of kerosene fraction. A very good degree of accuracy of

prediction, expressed in terms of root mean square error of prediction

(RMSEP), was achieved at a kerosene fraction of 2.05%.

Index Headings: Excitation–emission matrix fluorescence; EEMF; Multi-

variate calibration; N-way partial least squares regression; N-PLS.

INTRODUCTION

Multivariate methods have received wide interest recentlybecause of their ability to handle large datasets with time-efficient procedures for analysis. Spectroscopic techniques inconjunction with multivariate calibration are also receivingmuch attention because they make the analysis fast andefficient.1–4 Application of multivariate methods to fluores-cence spectroscopic data is an efficient method for the analysisof complex samples.5 The nature of certain complex multi-fluorophoric samples is such that their emission spectra keepchanging with excitation and vice versa, thus making theanalysis very difficult. Fluorescence spectroscopy is known forits advantages with respect to selectivity, sensitivity, and itsnondestructive nature. However, conventional fluorescencecannot give reliable and good data from complex multi-fluorophoric samples. Techniques such as synchronous fluo-rescence spectroscopy (SFS)6 and excitation–emission matrixfluorescence (EEMF)7 are being used for the analysis ofmultifluorophoric systems. The EEMF technique, which is alsocalled total fluorescence spectroscopy, allows plotting ofemission intensities at all combinations of excitation andemission wavelengths in a single three-dimensional (3D) grapheither as a contour or as a topographical surface diagram. Themeasurement is done by holding the excitation wavelengthconstant and scanning the emission wavelength over the regionof interest. Repeating this process at progressively higherexcitation wavelengths yields spectral data that can bepresented as a three-dimensional surface, where the x-axis isthe emission wavelength, the y-axis is the excitation wave-length, and the z-axis is the fluorescence intensity. The EEManalysis provides a ‘‘fingerprint’’ consisting of a 3D emission–excitation intensity diagram. This ‘‘fingerprint,’’ along withmultivariate calibration, can be used to obtain qualitative and

quantitative information about the multifluorophores present inthe sample.5

The products of petroleum crude, i.e., petrol, kerosene, anddiesel, are highly fluorescent due to the presence of polycyclicaromatic hydrocarbons (PAHs). This complex mixture ofPAHs absorbs and emits at different wavelengths, and hencevarious energy degrading interactions occur within the system.In spite of their complexities and differences, petroleumproducts show systematic behavior with respect to fluorescenceparameters.8

Much cheaper domestic fuel, such as kerosene, has beenused as an adulterant in heavy duty fuels such as diesel andpetrol, and this practice is very common particularly inSoutheast Asia.9,10 It causes the degradation of engineperformance and leads to the emission of gases produced fromimproper combustion, which is a major cause of environmentalpollution. Most of the methods based on physical propertiessuch as density, flash point, viscosity, odor-based methods,11

ultrasonic techniques,12 titration techniques,13 and opticaltechniques14 suffer limitations in terms of the sensitivity andaccuracy in determining adulteration levels. Patra et al.15–17

have extensively studied the properties of petroleum productsby the EEMF method. This study motivated us to apply certainmultivariate calibration methods to EEMF data of petroleumproducts for a quantitative analysis.

A chemometric approach to the EEMF spectra of diesel–kerosene mixtures has recently been found to be successful.18

The analysis of a diesel–kerosene mixture is much easierbecause diesel contains larger amounts of higher-memberfused-ring PAHs than kerosene, which imparts very differentEEMF spectral features to both substances. Petrol and keroseneboth contain intermediate or lower-ring PAHs. Hence, althoughthe EEMF spectra of pure petrol and pure kerosene lookdifferent, monitoring small additions of kerosene to petrol is arather difficult task.17 In this study, using N-way partial leastsquares regression (N-PLS), which is the common multivariatetool to analyze three-way data (data cuboids), we havedeveloped a calibration model with EEMF spectral data ofpetrol–kerosene mixtures. The reliability and robustness of themodel can be understood from the root mean square error(RMSE) values.

THEORY

The data generated by EEMF (fluorescence emission spectrameasured at several excitation wavelengths for severalsamples) methods has three dimensions (emission wavelength,excitation wavelength, and fluorescence intensity). EEMFmeasurements give output in the form of a matrix, and a seriesof data matrices obtained from multiple samples make up adata cube, which provides a three-way data set. The multi-waydata analysis method N-PLS19 can then be applied to this data.

Received 19 January 2008; accepted 1 April 2008.* Author to whom correspondence should be sent. E-mail: mishra@iitm.ac.in.

Volume 62, Number 7, 2008 APPLIED SPECTROSCOPY 7530003-7028/08/6207-0753$2.00/0

� 2008 Society for Applied Spectroscopy

Partial Least Squares Regression. Partial least squaresregression (PLSR) analysis is one of the most widely usedmultivariate calibration methods. It is based on the latentvariable decomposition of two blocks of variables X and Y,which may contain spectral and concentration data, respec-tively.20 These matrices can be simultaneously decomposedinto a sum of f latent variables, as shown in Eq. 1:

X ¼ TPT þ E ¼X

tf pTf þ E ð1aÞ

Y ¼ UQT þ F ¼X

uf qTf þ F ð1bÞ

in which T and U are the score matrices, P and Q are theloading matrices for X and Y, and E and F are the residualmatrices. The two matrices are correlated by the scores T andU, for each latent variable (Eq. 2):

uf ¼ bf tf ð2Þ

where bf is the regression coefficient for the f latent variable.The matrix Y can be calculated from uf by applying Eq. 3:

Y ¼ TBQT þ F ð3Þ

N-way Partial Least Squares Regression. For multi-waycalibration, N-PLS has recently received more attention than

other tools. The general N-PLS was introduced by Bro.21

N-PLS is the extension of the PLS regression model to multi-way data. The aim of the three-way PLS algorithm is todecompose the three-way array Xijk into a set of triads. A triadconsists of a score vector, t, related to the first mode, and twoweight vectors, wJ and wK, related to the other two modes. Themodel is given by Eq. 4:

Xijk ¼X

tif w Jjf w K

kf þ eijk ð4Þ

where eijk contains the residues. These vectors are calculated tohave the maximum covariance with the unexplained part of thedependent variable, y.

Quality of Estimates and Predictions. The best choice formeasuring the merits of a calibration is root mean square error(RMSE), which is a measure of the variabilty of the differencebetween the predicted and reference value for a set of samples.It is defined as

RMSE ¼

Xn

i¼1

ðypred � yrefÞ2

n

2

6664

3

7775

1=2

ð5Þ

where ypred and yref are the predicted and reference values,respectively, of sample i in the calibration set (or validation setor prediction set) and n is the number of samples used. RMSECis the root mean square error of calibration, which explainshow good the model is. The ability of a model to predict a newsample is expressed in terms of the root mean square error ofcross-validation (RMSECV), which explains the ruggedness ofthe model. When the model is applied to a new set of data it ispossible to calculate a root mean square error of prediction(RMSEP), provided the reference values for the new data setare known.22

EXPERIMENTAL

Apparatus. Fluorescence spectra were obtained on a HitachiF-4500 spectrofluorimeter. For EEM fluorescence measure-ment, the scan speed was 240 nm/s and the photomultipliertube (PMT) voltage was fixed at 700 V. The band pass forexcitation and emission monochromators was kept at 5 nm.The excitation source was a 150 W Xenon lamp. The excitationand emission matrix fluorescence was collected in theexcitation wavelength range of 250–590 nm with an intervalof 10 nm and in the emission wavelength range of 260–600 nmwith an interval of 10 nm, respectively. Conventional right-angle geometry was used for all the measurements becausesuch a geometry has been shown to give better spectralresolution.23

Reagents and Procedure. Petrol and kerosene werecollected from authorized local vendors in Chennai. Sampleswith different relative fractions of kerosene (in % v/v) in petrolwere prepared by adding the appropriate volumes of neatkerosene to neat petrol. Thirty-five (35) samples were used inthe calibration set and eight samples were used in theprediction set, with varying amounts of petrol and kerosene.Relative kerosene fraction (in % v/v) in the samples variedfrom 0% to 100%. For 0% to 10% and 90% to 100% kerosenefraction in petrol, the interval maintained was 1%, and for therange 10% to 90% the interval was 5%.

FIG. 1. Excitation–emission matrix fluorescence (EEMF) spectra of (A) neatpetrol and (B) neat kerosene.

754 Volume 62, Number 7, 2008

The multivariate model was developed for the entireconcentration range (0–100% kerosene fraction in petrol).The model thus developed can be applied to monitor thekerosene adulteration in petrol. The possible range ofadulteration for the adulterer to derive any benefit will be1–10% kerosene fraction in petrol, and hence a greater numberof samples were used in this concentration range.

Data Processing. Preprocessing of the raw data obtainedfrom the fluorimeter and all calculations were performed with aPentium 4 Personal Computer using the algorithms from thePLS_Toolbox 3.5 written in MATLAB language.24

RESULTS AND DISCUSSION

Excitation–Emission Matrix Fluorescence Spectra ofPetrol and Kerosene. The excitation–emission matrix fluo-rescence spectra of neat petrol and neat kerosene are shown inFig. 1. Contour diagrams of the synthetic adulterated mixturesconsisting of various amounts of petrol and kerosene are givenin Fig. 2. The spectra for kerosene occupies a high-energyregion in the matrix and the petrol spectra extend into a lower-energy region. Apart from that, the contour map of kerosenespreads within a short range, whereas for petrol it spreads

towards the lower-energy range. This is due to the presence of

many higher-member polycyclic aromatic compounds in

petrol. It is clear that the contour plots of neat petrol and neat

kerosene look different, and the addition of kerosene fraction to

petrol shifts the contour maxima to shorter wavelengths. This

concentration-dependent shift of fluorescence in petroleum

fuels has already been reported.25–30 Wang and Mullins studied

the fluorescence lifetimes of a series of crude oils at various

concentrations and they observed that, in highly concentrated

crude oils, short-wavelength light is absorbed by (small)

molecules with a large bandgap, and the excitation energy is

transferred to (large) molecules that have a small bandgap.25

The net effect of this energy transfer from small chromophores

to large fluorophores is a red shift of the fluorescence. Ralston

et al. reported a decrease in quantum yields of crude oils with

increasing wavelength excitation due to the exponential

increase of internal conversion with decreasing HO-LU

gap.26 Downare and Mullins have noted that energy transfer

produces large red shifts in the fluorescence emission spectra

for long-wavelength excitation of different crude oils.27 The

concentration-dependent wavelength shifts in three-dimension-

FIG. 2. EEMF contour spectra of petrol and kerosene mixtures.

APPLIED SPECTROSCOPY 755

al fluorescence spectra of petroleum samples were investigatedin detail by Smith et al.28 and Patra et al.29,30

For petrol, the excitation and emission maxima are at 350and 400 nm, respectively, and for kerosene the excitation andemission maxima are at 330 and 340 nm, respectively.

It is very difficult to carry out quantitative identification anddiscrimination of 1–10% kerosene adulterated petrol from thespectra, since their spectral features do not show significantvariation. Excitation–emission fluorescence measurementsobtained in the form of a three-way data set were arranged insuch a way so as to get the data in the form of a data cube, andthis was then used as the input for the multivariate analysis.

N-partial Least Squares Regression Analysis. The datawere arranged in a three-way array (35 3 35 3 35), composedof 35 samples, 35 excitation wavelengths, and 35 emissionwavelengths. The percentage variance explained by the firstfour latent variables is maximum (99.8%) when the N-PLScalibration method is applied on the fluorescence data of thepetrol–kerosene system. RMSEC and RMSECV valuesobtained are 2.12 and 2.70, respectively, which is optimumand shows that the model can be best fitted with four latentvariables. The results are given in Table I. The low RMSEvalue obtained with four latent variables indicates therobustness of the model. The scores plot developed by theN-PLS calibration is shown in Fig. 3, which is an efficientmodel as it clearly discriminates samples having a greaterpetrol fraction from those having a greater kerosene fraction.

To evaluate the accuracy and reliability of the model, thereference values of concentration are plotted against the

concentration values predicted by the model. Figure 4 shows

the predictive ability of the N-PLS method, which gives a very

good correlation of 0.999.

The Unfold-Partial Least Squares Model. Excitation–

emission fluorescence data, which was arranged in the form of

a data cube having dimension of 35 3 35 3 35, were unfolded

by combining the spectral modes. Hence, a matrix with

dimensions of 35 3 1225 was obtained and PLS was calculated

on the unfolded matrix. A minimum value of RMSECV was

obtained for four components; hence, the model was developed

using four factors and the scores plot is given in Fig. 5.

RMSECV and RMSEC values obtained for the unfold-PLS

model with various factors are given in Table II. It clearly

shows that with four latent variables, the maximum variance

can be explained. The scores plot efficiently differentiates the

TABLE I. Percentage variance explained and RMSEC and RMSECVvalues obtained for the N-PLS model developed with various latentvariables.

No of components 1 2 3 4 5

Percentage variance (%) 73.88 97.67 99.50 99.80 99.86RMSECV — 4.98 4.57 2.70 2.67RMSEC — 4.43 3.79 2.12 1.70

FIG. 3. Score plot of the four-component N-PLS model calculated from the 353 35 3 35 cube of EEMF spectral data.

FIG. 4. Reference concentrations versus the predicted concentrations ofpetrol–kerosene mixtures in the calibration dataset developed by the N-PLSmethod.

FIG. 5. Score plot of the unfold-PLS method developed on the EEMF datamatrix (35 3 1225).

756 Volume 62, Number 7, 2008

samples having a greater petrol fraction from those having agreater kerosene fraction.

Figure 6 shows the predictive ability of the unfold-PLSmethod. A very good correlation coefficient of 0.998 wasobtained when the reference values of concentration are plottedagainst the concentration values predicted by the unfold-PLSmodel.

The concentrations of eight samples represented as theprediction dataset having varying kerosene fractions werepredicted using the developed model, and the values obtainedshow a very good agreement with the reference values, asshown in Table III.

Patra et al. used excitation–emission matrix spectralsubtraction fluorescence (EEMSS) for the quantitative analysisof petrol–kerosene mixtures, whereas analysis by EEMfingerprint alone is not sufficient to carry out the analysis.17

But in the EEMSS method, the sequence of procedure is longand drawn out, involving a much longer time for themanagement of a large body of data. This work shows thatthe analysis of petrol–kerosene mixtures became much simplerand more time efficient when multivariate tools are used onEEMF spectral data.

The fluorescence spectra of a particular petroleum productsuch as petrol or kerosene is expected to vary with the sourceof the petroleum crude and with the refining process.Therefore, a batch-to-batch variation of the fluorescencespectral response is expected. A universally applicable model,therefore, cannot be built. The objective of this work is mainlytowards methodology development for EEMF data of petrol–

kerosene samples using chemometric methods. A veryelaborate and exhaustive dataset has been used for the purpose.However, for practical applications, optimization of dataacquisition and use of optimized datasets for model makingcan considerably reduce the time needed for model creation.

CONCLUSION

Application of multivariate methods to EEMF data impartsadditional sensitivity, time efficiency, and advantage in theanalysis of petrol–kerosene mixtures over EEMF data alone, asseen from some of the work done previously in our lab.15–17

The present work shows that the N-PLS model developed withthe EEMF spectral data of petrol–kerosene samples appears tobe efficient. The developed method is able to sense very lowkerosene additions to petrol with a prediction error of 2%kerosene fraction. The accuracy of the predictive ability is dueto the incorporation of one more dimension in the data, whichincludes additional information regarding the concentration.The accuracy of the method, as evaluated from the root meansquare error of prediction (RMSEP) value, is found to be verygood.

ACKNOWLEDGMENTS

One of the authors, O. Divya, acknowledges the Council of Scientific andIndustrial Research (CSIR) New Delhi for the SRF scholarship. The authorsgratefully acknowledge the Council of Scientific and Industrial Research, NewDelhi, for the financial support to carry out this work. The authors would like tothank Mr. V. Venkataraman, Scientist (CEERI Centre, CSIR complex,Chennai), for valuable suggestions and help.

1. S. Macho and M. S. Larrechi, Trends Anal. Chem. 21, 799 (2002).2. A. Sakudo, Y. Suganuma, T. Kobayashi, T. Onodera, and K. Ikuta,

Biochem. Biophys. Res. Commun. 341, 279 (2006).3. H. Winning, F. H. Larsen, R. Bro, and S. B. Engelsen, J. Magn. Reson.

190, 26 (2008).4. N. Dupuy and F. Douay, Spectrochim. Acta, Part A 57, 1037 (2001).5. J. Christensen, L. Norgaard, R. Bro, and S. B. Engelsen, Chem. Rev. 106,

1979 (2006).6. J. B. F. Lloyd, Nature Phys. Sci. 231, 64 (1971).7. J. H. Rho and J. L. Stuart, Anal. Chem. 50, 620 (1978).8. D. Patra and A. K. Mishra, Trends Anal. Chem. 21, 787 (2002).9. V. P. Bhatnagar, J. Acoust. Soc. India. 9, 19 (1981).

10. M. S. Bahari, W. J. Criddle, and J. D. R. Thomas, Analyst (Cambridge,U.K.) 115, 417 (1990).

11. A. A. Gupta, K. K. Swami, A. K. Misra, A. K. Bhatnagar, and P. K.Mukhopadhaya, Hydrocarbon Technol. 15, 137 (1992).

12. V. P. Bhatnagar, J. Acoust. Soc. India 9, 19 (1981).13. M. S. Bahari, W. J. Criddle, and J. D. R. Thomas, Analyst (Cambridge,

U.K.) 115, 417 (1990).

TABLE II. Percentage variance explained and RMSECV and RMSECvalues obtained for the unfold-PLS model developed with various latentvariables.

No of components 1 2 3 4 5

Percentage variance (%) 87.09 99.68 99.88 99.99 99.99RMSECV — 11.93 4.49 2.80 2.04RMSEC — 11.34 4.02 2.26 1.54

FIG. 6. Reference concentrations versus the predicted concentrations ofpetrol–kerosene mixtures in the calibration dataset obtained by the unfold-PLSmethod.

TABLE III. Reference values and predicted values of the relativekerosene fraction present in the mixture using N-PLS and the unfold-PLSmodel.

Reference value(relative kerosene

fraction (in % v/v))

Predicted value (relativekerosene fraction (in % v/v))

N-PLS Unfold PLS

0 �5.47 �3.658 7.21 6.50

15 16.34 15.4240 39.94 39.8765 63.44 64.9480 79.60 81.0892 94.065 94.5598 97.88 95.14

RMSEP Value 2.05 2.21

APPLIED SPECTROSCOPY 757

14. S. Roy, Sens. Actuators, B 55, 212 (1999).15. D. Patra and A. K. Mishra, Anal. Chim. Acta 454, 209 (2002).16. D. Patra and A. K. Mishra, Indian J. Chem., Sect A 40, 374 (2001).17. D. Patra and A. K. Mishra, Appl. Spectrosc. 55, 338 (2001).18. O. Divya and A. K. Mishra, Anal. Chim. Acta 592, 82 (2007).19. L. Stahle, Chemom. Intell. Lab. Syst. 7, 95 (1989).20. H. Martens and T. Naes, Multivariate Calibration (John Wiley and Sons,

Chichester, 1989), p. 2.21. R. Bro, J. Chemom. 10, 47 (1996).22. R. Kramer, Chemometric Techniques for Quantitative Analysis (Marcel

Dekker, New York, 1998).

23. D. Patra and A. K. Mishra, Analyst (Cambridge, U.K.) 125, 1383 (2000).24. B. M. Wise, N. B. Gallagher, R. Bro, and J. M. Shaver, PLS_Toolbox 3.5

(Eigenvector Research, 2005).25. X. Wang and O. C. Mullins, Appl. Spectrosc. 48, 977 (1994).26. T. D. Downare and O. C. Mullins, Appl. Spectrosc. 49, 754 (1995).27. C. Y. Ralston, X. Wu, and O. C. Mullins, Appl. Spectrosc. 50, 1563

(1996).28. G. C. Smith and J. F. Sinski, Appl. Spectrosc. 53, 1459 (1999).29. D. Patra, K. L. Sireesha, and A. K. Mishra, J. Sci. Indust. Res. 59, 300

(2000).30. D. Patra and A. K. Mishra, Talanta 53, 783 (2001).

758 Volume 62, Number 7, 2008