Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 153 (2016) 108117

Contents lists available at ScienceDirect

Spectrochimica Acta Part A: Molecular and Biomolecular

Spectroscopy
journal homepage: www.elsevier.com/locate/saa

Chemometrics-assisted excitationemission uorescence analytical data

for rapid and selective determination of optical brighteners in the
presence of uncalibrated interferences
Ali Gholami , Saeed Masoum, Atefeh Mohsenikia, Saleheh Abbasi
Department of Analytical Chemistry, Faculty of Chemistry, University of Kashan, Kashan, Iran

a r t i c l e

i n f o

Article history:
Received 21 April 2015
Received in revised form 23 July 2015
Accepted 4 August 2015
Available online 10 August 2015
Keywords:
Molecular uorescence
Second-order chemometric methods
Optical brighteners
Laundry powder
Environmental samples

a b s t r a c t
This study describes a novel approach for the simultaneous determination of CBS-X and CXT as widely used optical brighteners in household detergent, by combining the advantage of the high sensitivity of molecular uorescence, and the selectivity of second-order chemometric methods. The proposed method is assisted by secondorder chemometric analyses employing the PARAFAC, SWATLD and APTLD that help us to determine CBS-X
and CXT in laundry powders and environmental samples, through the unique decomposition of the three-way
data array. Proposed method can provide the extraction of relative concentrations of the analytes, as well as
the spectral proles. This approach achieves the second-order advantage and in principle could be able to overcome the spectral uncalibrated interference problems in the determination of CBS-X and CXT at the ng g1 level.
By spiking the known concentrations of these compounds to the real samples, the accuracy of the proposed
methods was validated and recoveries of the spiked values were calculated. High recoveries (90.00%113.33%)
for the spiked laundry powders and real environmental samples indicate the present method successfully
faces this complex challenge without the necessity of applying separation and preconcentration steps in environmental contaminations.
2015 Elsevier B.V. All rights reserved.

1. Introduction
Nowadays, variety types of laundry detergents such as laundry powder, laundry soaps and liquid laundry are produced in the whole of the
world. Laundry detergents contain several chemical components, including surfactant, optical brightener, bleaching agent and builder.
Optical brighteners (OBs) or uorescent whitening agents (FWAs in
the detergent industry) are synthetic chemical compounds that stick to
the surface of the clothing or fabrics. These particles make fabrics seem
cleaner by absorbing ultraviolet light (usually 340370 nm), and reemitting it as blue light (typically 420470 nm), thus they enhance
the intensity of visual reection and strength the optical impression of
whiteness and brightness. Fluorescent whitening agents that are used
in detergents are moderately water-soluble organic compounds with a
high afnity for cellulosic material. When they adsorb to the textiles,
their strong blue uorescence eliminate the yellowish cast of white fabrics and make them look whiter [1].
Optical brighteners can be classied, based on their structure and
properties. A number of uorescent compounds such as quinolone, coumarin, benzoxazole, and distyrylbiphenyl are used as uorescent whitening agents. All OBs are highly substituted ring (aromatic) structures
Corresponding author.
E-mail address: agholami@kashanu.ac.ir (A. Gholami).

http://dx.doi.org/10.1016/j.saa.2015.08.012
1386-1425/ 2015 Elsevier B.V. All rights reserved.

and contain many double bonds that can be activated by UV light [2].
These components are used in textiles, papers, plastics and synthetic bers, plus widespread use in some medical, chemical, cosmetics and petroleum applications. Around 80% of FWAs that are used in paper, textile
and detergent formulations are based on the stilbene derivatives.
Benzenesulfonic acid, 2,2([1,1-biphenyl]-4,4-diyldi-2,1-ethenediyl)
bis-disodium salt or Tinopal CBS-X (Fig. 1-A) and disodium 4,4-bis[(4anilino-6-morpholion-1,3,5-trizin-2-yl)amino]stilbene-2,2-disulphonate
or CXT (Fig. 1-B) are most widely used in detergent industries [3].
CBS-X is used in household detergents at concentration levels
between 0.05% and 0.15% and undergoes photo degradation N50% in 12
months. CXT is used in detergents at concentration levels between 0.02%
and 0.10% and undergoes photo degradation N 70% in 28 days [4,5].
Residues of OBs are left on our clothes, and they can cause allergic reactions when absorbed by skin. Detergent industries due to the ubiquitous use of optical brightener are important potential sources for the
pollution of natural waters and ecosystem. Laundry wastewater is the
largest supplier of optical brighteners to wastewater systems. Therefore,
detection of optical brighteners in aquatic systems can help water municipalities and avoid increased anthropogenic input that may greatly
affect ecosystems. When OBs enter in the water system, they cause a potential hazard to aquatic life. This pollution remains in wastewater for
long periods and negatively affects water quality, animal and plant
life. Brighteners are highly susceptible to adsorption. They are removed

A. Gholami et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 153 (2016) 108117

109

Fig. 1. (a) Structure of CBS-X , (b) structure of CXT (c) Excitation spectrum of CBS-X and CXT and (d) Emission spectrum of CBS-X and CXT.

from surface and ground water by adsorption into the soil and tend to
accumulate on sewage sludge. This allows researchers to assess the effectiveness of natural cleansing of wastewater by determining brightener concentrations in ground or surface waters [6].
The increasing public health concerns have encouraged us to consider the content and distribution of FWAs in laundry detergent as a quality
control parameter. The amount of FWAs in textile and paper products is
usually determined by direct spectrophotometric method [711]. These
methods do not give individual concentration of FWAs and measure
only total FWA concentrations. Environmental samples usually contain
a number of organic species at relatively high concentrations, which can
interfere with the optical brightener's quantication.
Since 1976, high performance liquid chromatography (HPLC) has been
applied to separate and determine FWAs in detergents and environmental
samples [1218]. Reversed-phase HPLC with ion-pair chromatography

(IPC) by uorescence detection has been applied for the separation of

mixture of FWAs in paper materials and infant clothes. HPLC method
needs to expensive solvent and instrumentation. Environmental samples
frequently contain several organic compounds in high concentration,
which can interfere with quantication of optical brightener.
Hyphenated method such as combining HPLC with mass spectroscopy has also been reported to determine FWAs in detergents [19], paper
materials and infant clothes [20]. However, ion suppression is one of the
most undesirable problems affecting these techniques during quantication and a separation step was needed to remove the possible matrix
effect [21]. Furthermore, the expenses of LC/MS systems are high, operation system is laborious and unavailable in many commercial laboratories. Simultaneous determination of detergent ingredients was carried
out by combining spectroscopy method with chemometric techniques
such as partial least square (PLS) [22,23].

110

A. Gholami et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 153 (2016) 108117

Analysis of complex matrices into their pure-component contributions with unknown components without separation is a fundamental
problem in analytical chemistry. Spectrum overlapping was resolved
by multivariate quantication methods such as classical least squares
(CLS), inverse least squares (ILS) and partial least square (PLS). PLS and
related chemometric methods need a large number of calibration and
test sets and also for providing a robust regression model, all analytes
and interferences have to be in the calibration set at the appropriate concentration level. Analysis and quantication of analytes in complex matrices are achievable without any previous knowledge about their
chemical substances by chemometric methods. The philosophy behind
second order or higher order data analysis is quantication of analytes
even in the presence of unmodeled or uncalibrated interferences and
this property, in fact, is known as second order advantage [24].
Environmental samples contain different organic materials such as
surfactants, chromophoric compounds, synthetic colors and poly hydrocarbon aromatic (PAH) which have been intense overlapping with the
analytes.
Unfortunately, PLS and related chemometric methods do not exhibit
the second order advantage, therefore all components in the complex
samples must be known, but identication of all components in complex matrices is impossible and time consuming.
Second order multivariate calibration methods were successfully applied for three-way array of data that is obtained by stacking the data
matrices of different samples under each other.
Second-order data is usually brought from hyphenated instruments,
such as an excitationemission matrix (EEM) of spectrouorometer and
high-performance liquid chromatography with a diode array detector
(HPLC-DAD), gas chromatography with a mass detector (GC-MS), and
absorbance spectra-pH data [25,26].
Aim of this study is applying uorescence spectroscopy to routinely
determine FWAs in laundry detergent, wastewater and agricultural soil
samples. Fluorescence spectroscopy as an analytical technique has the
advantage of selectivity and sensitivity in comparison with other spectroscopic methods.
An improved way to analyze the data is exciting the sample at different excitation wavelengths to obtain the emission spectra. Fluorescence
data array, X, can be produced by stacking a series of excitationemission
spectra for each sample to create an excitation emission matrix (EEM).
Each EEM is a matrix and by incorporating several samples, a three-way
array (sample emission excitation) is obtained as a landscape that
covers the total area of uorescence. This three-way array can be processed by any of the second-order algorithms such as parallel factor analysis (PARAFAC), alternating penalty trilinear decomposition (APTLD) and
self-weighted alternating trilinear decomposition (SWATLD). The presentation of data in an EEM format is a good way of visualizing the different uorophores in the sample. But apart from the possibility of
characterizing distinct peaks, it is difcult to unravel the individual
uorophores in the sample [27]. The present work attempts to demonstrate the combination of EEM array in a photoluminescence system
assisted by three-way calibration method for simultaneous determination of optical brighteners with strongly overlapping spectra in the presence of uncalibrated interferences. The proposed method is the rst
application of multi-way methods in the multi-dimensional analysis of
the optical brighteners in laundry detergent and environmental samples.
2. Theory
2.1. PARAFAC modeling
PARAFAC is usually employed for modeling trilinear data arrays. The
PARAFAC model can be written as following equation.

where xijk is the intensity of the ith sample at the jth variable (emission
mode) and at the kth variable (excitation mode), F is the total number of
responsive components, aif is the relative concentration of component f
in the ith sample, bjf is the signal intensities at emission wavelength j
and ckf is the signal intensities at excitation wavelength k in each dimension for component f.
Stacking a series of sample concentration matrices can make up a
data cube X; the algorithm decomposes the cube X into three matrices
A, B and C with elements aif, bjf, ckf, respectively. The algorithm proceeds
to minimize the sum of squares of the residuals, eijk, in the model using
alternating lease squares (ALS) until convergence is reached [28].
2.2. APTLD modeling
Alternating penalty trilinear decomposition (APTLD) is applied for
the decomposition of three-way data arrays. Performance of this algorithm is based on utilizing alternating penalty factors which minimizes
three new least squares-based constrained objective functions. The
values of penalty factors p, q and r should be chosen before implementation of the algorithm. In additions, this method is insensitive to increasing number of component by choosing a large number of penalty
factors, thus avoiding the complexity of determining a correct component number for the model which is intrinsic in the PARAFAC algorithm.
It has the second-order advantage for determination of analytes in the
presence of unknown interference. PARAFAC can give degenerate solutions when two components of the same mode have the signals with
high correlations. This problem called two-factor degeneracy. The
APTLD algorithm can avoid the two-factor degeneracy problem and
the proposed algorithm can overcome the slow convergence, which is
difcult to handle for the traditional PARAFAC algorithm.
The common algorithm for the APTLD method is as follows:
1. Obtain the accurate number of species.
2. Randomly initialize B and C by choosing suitable penalty factors p, q
and r.
3. Compute A from B and C using least-squares regression.
4. B and C were computed in the same way as matrix A in step 3.
5. Provide column wise normalization by scaling B and C.
6. Repeat steps 24 until a stopping criterion is satised [29,30].
2.3. SWATLD modeling
A self-weighted alternating trilinear decomposition (SWATLD) derived from the alternating trilinear decomposition (ATLD) algorithm.
The SWATLD algorithm is based on a least-squares scheme, in which diagram operation is added to each loss function, making it more stable and
exible, in other words, an overestimation of the number of factors will
not affect on meaningful obtained results by SWATLD. There are different
features for SWATLD such as decreasing the computation time in each iteration for optimizing procedure and being insensitive to the excess estimated factors in the model, the unique optimizing approach and
quantication of analytes even in the presence of unknown interferences.
This algorithm is based on the following procedure:
1.
2.
3.
4.
5.

Estimate the number of chemical components (chemical rank).

Initialize loading matrixes B and C randomly or by other procedures.
Estimate A from B and C by least-squares regression.
Estimate B and C in the same way as matrix A in step.
Update A, B and C according to steps 34, until the certain stop criterion has been reached [31,32].

3. Experimental section
3.1. Materials

xi jk f 1 if bjf ck f ei jk

All solutions were prepared from analytical grade. Ethanol, methanol,

and acetonitrile were provided from Merck (Darmstadt, Germany). OBs

A. Gholami et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 153 (2016) 108117

were purchased from Ciba-Geigy Company (Switzerland). The three

commercial laundry powders were purchased from the local market.
One wastewater sample was obtained from a detergent factory. Two agricultural soil samples were gathered near the industrial area.
3.2. Instrumentation
A Perkin-Elmer LS 55B Luminescence spectrophotometer equipped
with a xenon lamp was used for the uorimetric measurements. All
measurements were done in a 10 mm 10 mm quartz cuvette.
The chromatography analyses were performed with highperformance liquid chromatography (HPLC) Agilent 1200 series model
with an ultravioletvisible detector that was set to 340 nm. The analytical column was ODS with 15 cm length. Mobile phase is 27.5:12.5:60
acetonitrile/methanol/water containing 0.05 M ammonium acetate
with ow rate 1 ml/min.
3.3. Software
PARAFAC analysis was performed using the N-way toolbox in
MATLAB environment provided by Rasmus Bro [33]. APTLD and
SWATLD analysis were done using the MVC2 toolbox that is provided
by Olivieri [34].
3.4. Preparation of standards and treatment of real samples
To select the best conditions for the resolution of the binary
mixtures, we investigated that the solubility of optical brighteners in
ethanol is more than water in the room temperature, therefore uorescence intensity increases in this solvent. Therefore, all of stock solutions
were prepared using ethanol as solvent at room temperature. The Stock
solutions of CBS-X (10,000 mgkg1) and CXT (10,000 mgkg1) were
prepared by dissolving 0.8800 g CBS-X and CXT to nal volume of
100 ml with ethanol. Ten samples with the binary mixture of CBS-X
and CXT were prepared by appropriate dilution of the stock solutions
with ethanol. The concentrations of these species are listed in
Table S1. The four test samples that contain different amounts of
analytes (CBS-X and CXT) were prepared for validating the employed
proposed methods are listed in Table S2. Excitation and emission spectra of pure CBS-X and CXT in ethanol solvent were shown in (Fig. 1-C)
and (Fig. 1-D), respectively. Excitation wavelengths between 250 and
380 nm with 1 nm intervals were changed and emission wavelength between 300 and 550 nm were recorded (Fig. S1). Maximum excitation
wavelengths for CBS-X and CXT are 342 and 350 nm, respectively and
maximum emission wavelengths for CBS-X and CXT are 418 and
428 nm, respectively.
In order to investigate the second-order advantages, three laundry
powder samples and three environmental samples were investigated.
Three laundry powder and two soil samples were prepared by dissolving 1.000 g of these samples in ethanol and dissolved samples were ltered through lter paper to remove solid sediments and diluting to
volume of 100.0 ml with ethanol. About 1.000 g of wastewater sample
was dissolved and diluted to volume of 25.0 ml with ethanol.
3.5. Procedure and data acquisition
The rst step in the three way methods is providing second order trilinear data. To generate uorescence landscapes, excitation wavelengths between 250 and 380 nm with 5 nm intervals (27 excitation
wavelengths) and emission wavelength between 300 and 550 nm
with 0.5 nm intervals (502 emission wavelengths) were used. Thus, a
three-way 27 502 10 data array was collected. All measurements
were made on the calibration set of ten samples of the binary mixture
of CBS-X and CXT. The scan rate was 1500 nm min1 and the best intensity of uorescence was obtained if the excitation and emission slit
width were set at 11 and 8 nm, respectively.

111

4. Result and discussion

4.1. Data pretreatment
Fluorescence landscape is inuenced by scattering effects such as
rst and second order Rayleigh and Raman scattering. The Rayleigh
scattering band can be strong in intensity depending on the excitation
wavelength and will affect the measured uorescence intensity when
the emission wavelength is close to the excitation wavelength. First
and second order Rayleigh and Raman scattering do not conform to
the required tri-linearity and data analysis is often complicated due to
these scattering. The Rayleigh and Raman proles are not operatively
modeled using three-way chemometric models. The scattering may
have unfavorable effects on the resolved spectra and score plot when
such landscapes are decomposed by using three-way analyses especially if the peaks from the analytes are close to the Rayleigh scattering
lines. Consequently, the model fails to converge when three way algorithms are used to the analysis of three-way data array without removing the scattering.
The effect of Rayleigh scattering was compensated by subtracting
the measurement matrix of a scattering blank from each of the samples
prior to analysis. While this may be successful for idealized laboratory
analyses, for real samples with saturated Rayleigh scattering and
Raman-active species this approach may be insufcient. The most common way to handle this problem is to set this value to missing in analysis [3537]. To overcome this problem, the EEM of blank was subtracted
from the data. The values consistent with the scattering effect were set
to NaN (not-a-number, a variable that can be handled by Matlab) which
refers to missing data and then replaced by their corresponding estimates. The use of missing data to replace scattering light has been successfully applied in this research.
4.2. Number of estimated component
The fundamental point of the PARAFAC analysis is the choice of the
proper number of factors (or components), which is achieved by tting
trilinear model based on data using various components. Selection of
the number of components in known or prepared samples is based on
the previous information of the system and utilizes the analysis of the
residual model. The choice of the number of components is very difcult
in the case of natural samples that there is no prior knowledge of the
system. There are several available tools to provide information to evaluate the correct number of components such as standard deviation of
residual (SD), principle component analysis (PCA), single vector decomposition (SVD) and three way cross validation (CV).
In this work, to evaluate model stability with correct factors, residual
analysis and cross validation methods were applied.
The residual model is based on the best linear t of the adjusted line
in the calibration cases. In this method, considerable decrease is observed upon an increase in the number of components. The number of
component is estimated, where standard deviation is stable at the
noise level.
The cross validation strategies rely on keep out parts of the data
away from the model evolution which is applied as the test samples, develop an N-PLS or TUCKER models from new data and predict the test
samples by these models and nally the predicted values are compared
with the real ones. By calculating the squared variances between predicted and observed values, predictive residual sums of squares
(PRESS) is obtained, which is a measure of the predictive power of the
tested model [38]. The process is repeated with N+1 factor until steep
decrease or a global minimum in the PRESS is found.
The obtained results by the mentioned procedure present the total
number of constituents is three (one unknown interference not included in the calibration set besides two OBs) for laundry powder samples 1
and 2. The number of components attained by mentioned methods for
laundry powder sample 3 is four (Fig. S2), therefore, there are two

112

A. Gholami et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 153 (2016) 108117

unexpected components not included in the calibration set. The total

number of components for agricultural soils 1, 2 and wastewater sample
are four, ve and six, respectively. Results of these models for determining the number of components are shown in Table 1.
4.3. Application of three-way analysis on the laundry powder data
The DTLD/GRAM (direct trilinear decomposition/generalized rank
annihilation method) was applied as an iterative optimization procedure of APTLD, SWATLD and PARAFAC. When the following stop criterion reaches a certain threshold ( = 106), the optimization process is
nished.
"
#
SSRM SSRM1
SSRM1

Recovery factor 100  CRecovered =CTrue

SSR is the residual sum of squares and M is the current iteration

number. Test samples for validating of the employed algorithms before
analysis of real samples were used. Table S2 shows that the obtained results are nearly the same as those of true ones with acceptable percentage errors and illustrates that three algorithms provide an accurate
resolution of the substances that have been seriously overlapping spectra when the correct component number is used in the computation.
The power of PARAFAC, APTLD, and SWATLD could not be completely
appreciated until its capacity to analyzed trilinear dataset in presence
of interfering species is demonstrated. At this point, it is necessary to
make data array that each unknown real sample separately was joined
to the standards for achieving the second order advantages. The data arrays of 11 mixture samples, which the rst ten samples were used as
calibration samples and the last one was joined to a calibration set as
unknown sample. The size of analyzed data array was (11 502 27).
Simultaneous quantitative analyses were provided by three-way
methods in real samples. Table 2 indicates the resolved concentrations
of unknown samples using APTLD, SWATLD and PARAFAC.
The spectral proles of emission and excitation extracted by
PARAFAC algorithm on the EEM data of laundry sample 3 are shown
in Fig. 2. The comparison of the excitation and emission proles
Table 1
Obtained results from different methods in determining the number of components.

Laundry powders

Samples 1 and 2

Sample 3

Agriculture soils

Samples 1

Sample 2

Wastewater

provided by the PARAFAC model with the experimental ones shows

an acceptable agreement. For SWATLD and the proposed APTLD with
p=q=r=1020 , these results indicate that tow algorithms give satisfactory resolutions for excitation and emission spectral proles for all samples which are very similar to the obtained loadings from PARAFAC
algorithm.
For evaluation of accuracy of the proposed methods to analysis of the
real samples, known concentration of two OBs (CBS-X and CXT) were
spiked in the real samples and the data array is obtained by stacking
the standard set and spike samples, then recoveries for each spiked
sample were calculated. Comparison of the predicted concentrations
and recoveries provided by proposed algorithms shows a good predictive ability towards the spiked laundry powder samples by three algorithms (Table 3).

Samples 1

Number of
components

SD of
residual

CV

1
2
3
4
1
2
3
4
5
1
2
3
4
5
1
2
3
4
5
6
1
2
3
4
5
6
7

3.2
2.3
1.7
1.7
3.7
2.6
1.8
1.3
1.3
7.8
5.4
2.4
1.7
1.7
16.0
8.0
5.0
2.3
1.5
1.4
13
8.6
5.9
2.8
2.3
2.0
2.0

200.3
100.1
50.2
49.2
20.2
10.5
6.3
1.2
1.2
49.8
12.5
8.1
2.1
2.1
45.5
12.3
9.2
5.8
1.1
1.1
30.0
11.1
8.7
8.5
5.4
1.4
1.4

Where CRecovered is the measured concentration after adding known

concentration of the analyte to the real sample and CTrue is the expected
concentration [39].
4.4. Analysis of the environmental samples using chemometric resolution
methods
Determinations of the environmental samples with complex matrices were performed for evaluating the capability of the proposed
methods. After processing the three way data, excitation and emission
loading plots that provide by PARAFAC for agricultural soil samples
are very similar to the obtained loading from APTLD with p = q =
r = 1020 and SWATLD algorithm. Fig. 3 shows the spectral proles of
emission and excitation that are extracted by PARAFAC algorithm for
agricultural soil sample 1. For SWATLD and the proposed APTLD with
p = q = r = 1020, these results indicate that tow algorithms give satisfactory resolutions for excitation and emission spectral proles for two
soil samples which are very similar to the obtained loadings from
PARAFAC algorithm. By comparing the obtained results, it was found
that simultaneous determination of CBS-X and CXT is possible in the
presence of interferences.
The obtained results of PARAFAC for wastewater sample are rather
poor, and conrm the potentiality of the three-way methods such as
APTLD and SWATLD for the analysis of these complex samples for the
determination of optical brighteners. Fig. 4 shows excitation and emission spectral prole of wastewater sample that are provided by
SWATLD. The loading plots of APTLD are similar to prediction values
of SWATLD and these result are in good agreement with pure spectra
of these components.
Table 2
Quantitative analysis of CBS-X and CXT in the real samples.

Laundry powders

Sample 1

Sample 2

Sample 3

Agriculture soils

Sample 1

Sample 2

Wastewater

Sample 1

Methods

CBS-X%

CXT%

PARAFAC
APTLD
SWATLD
PARAFAC
APTLD
SWATLD
PARAFAC
APTLD
SWATLD
PARAFAC
APTLD
SWATLD
PARAFAC
APTLD
SWATLD
PARAFAC
APTLD
SWATLD

5.24 102
5.37 102
5.37 102
0.00
0.00
0.00
5.02 102
5.10 102
5.10 102
1.31 104
1.21 104
1.21 104
3.31 105
3.40 105
3.40 105
NA
1.08 105
1.08 105

1.13 101
1.14 101
1.14 101
1.09 101
1.02 101
1.02 101
0.00
0.00
0.00
2.78 104
2.86 104
2.86 104
7.88 105
5.73 105
5.73 105
NA
3.07 105
3.07 105

A. Gholami et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 153 (2016) 108117

113

Fig. 2. Obtained emission (A) and excitation (B) loadings from PARAFAC algorithm for laundry powder sample 3.

Table 3
Obtained recoveries of analytes in laundry powder samples.

Sample 1

Methods

Added CBS
(ng g1)

Found CBS
(ng g1)

Expected
CBS
(ng g1)

Recovery (%)

Added CXT
(ng g1)

Found CXT
(ng g1)

Expected
CXT
(ng g1)

Recovery (%)

PARAFAC

0.00
1.00
2.00
0.00
1.00
2.00
0.00
1.00
2.00
0.00
1.50
3.00
0.00
1.5
3.00
0.00
1.5
3.00
0.00
1.0
2.0
0.0
1.0
2.0
0.00
1.0
2.0

5.96
6.64
7.88
6.11
6.86
7.69
6.11
6.86
7.69

1.70
3.22

1.37
2.89

1.37
2.89
5.71
6.89
7.05
5.80
6.83
7.81
5.80
6.83
7.81

6.96
7.96

7.11
8.11

7.11
8.11

1.50
3.00

1.50
3.00

1.50
3.00

6.71
7.71

6.80
7.80

6.80
7.80

95.40
98.99

96.48
94.82

96.48
94.82

113.33
107.33

91.33
96.34

91.33
96.34

102.68
91.43

100.44
100.12

100.44
100.12

0.00
5.00
10.00
0.00
5.00
10.00
0.00
5.00
10.00
0.00
5.00
10.00
0.00
5.00
10.00
0.00
5.0
10.0
0.00
10.0
15.0
0.00
10.0
15.0
0.00
10.0
15.0

12.85
17.45
24.16
12.97
17.59
24.53
12.97
17.59
24.53
24.83
28.83
34.62
23.21
30.00
36.50
23.21
30.00
36.50

17.85
22.85

17.97
22.97

17.97
22.97

29.83
34.83

28.21
33.21

28.21
33.21

10.0
15.0

10.0
15.0

10.0
15.0

97.75
105.73

97.88
106.79

97.88
106.79

96.64
99.39

106.34
109.90

106.34
109.90

90.00
92.26

108.00
99.46

108.00
99.46

APTLD

SWATLD

Sample 2

PARAFAC

APTLD

SWATLD

Sample 3

PARAFAC

APTLD

SWATLD

9.00
13.84

10.80
14.92
0.0
10.80
14.92

114

A. Gholami et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 153 (2016) 108117

Fig. 3. Emission (A) and excitation (B) loadings that are provided by the ve-component PARAFAC model for agricultural soil sample 2.

The obtained recoveries on the spike environmental samples are

shown in Table 4. The results suggest that three proposed methods
are suitable for determination of studied components in two soil samples but the PARAFAC (with suitable constraints) did not give acceptable results for wastewater sample compare with APTLD and SWATLD.
The extreme spectral overlap between analyte species with uncalibrated interference is the main reason that complicates the analysis by
PARAFAC. It should be considered that when more than one uncalibrated component exists in unknown samples, the obtained result by
PARAFAC may not be accurate because the proles of these components
may consist of linear combinations of their true spectra. In contrast to
APTLD and SWATLD, PARAFAC can give degenerate solutions when
two components have the signals with high correlations and show the
same shaped loading vectors.
4.5. Comparison of the performance of three algorithms
Analytical performance, speed of convergence, number of iteration
sensitivity to a number of components and model prediction
ability are features that are considered in multivariate calibration
models.
The DTLD/GRAM and best tting model were applied to begin the iterative optimization procedures of three algorithms. The stability of
models for obtaining unique result was done to check with three replications run but using the best tting model did not demonstrate any noticeable difference compared to DTLD in improving the t. Either best

tting model or DTLD in all samples can repeatedly give almost identical
results for three algorithms in all samples. Because a best-tting model
needs more iteration setups and consumes an extended computation
time, DTLD is used as an initial value estimation that provides less iteration. By using DTLD as an initial value and choosing a suitable number
of component, the PARAFAC, SWATLD and APTLD can provide unique
results (Table S3).
Number of iterations and recoveries were calculated and listed in
Table S4 by using four and ve as component numbers with DTLD as
an initial value for one of the laundry powder samples. The average of
the iteration number for APTLD (IT = 45) is same as SWATLD and less
than PARAFAC (IT = 210). The results disclose that the APTLD and
SWATLD algorithms not only resolve the proles accurately but also
overcome the slow convergence. Comparison between proposed
methods reveals that any clear changes were not observed in SWATLD
and APTLD by increasing number of components in contrast to
PARAFAC. SWATLD and APTLD are insensitive to the estimated component number, therefore avoiding the difculty of determining a correct
component number for the models, which is an initial and important
stage in the PARAFAC algorithm model.
The degree of overlapping between the proles of all components in
the data matrix is appropriate so these three algorithms can provide a
same reasonable result if a sufcient number of components be selected. Three algorithms can predict the reasonable result and have good
agreement with each other for laundry powder and agricultural soil
samples.

A. Gholami et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 153 (2016) 108117

115

Fig. 4. Obtained emission (A) and excitation (B) loadings from SWATLD for wastewater sample.

The performance of APTLD depends on the choice of penalty factors

p, q, r. Choosing very small p, q and r values (such as p = q = r = 106
or 104) makes situation, which is similar to the PARAFAC algorithm
such as a large number of iterations, sensitivity to an additional number
of component. This algorithm provides better performance with large
penalty factors and in this situation is insensitive to an excess number
of components and avoids slow convergence that diminishes the calculation time in each iteration. However, altering p, q and r from 108 to
1020 does not make noticeable difference in the quality of the value
for recovery factor and give acceptable results (Table S5). The large penalty factors indicate that APTLD is faster than PARAFAC and make APTLD
close to SWATLD. The results of APTLD are the same as SWATLD for penalty factors set to 1020. In this paper for decreasing computation time
p = q = r = 1020 was selected.
Table 5 reports the analytical gures of merit of the three
calibration models in the real sample. Sensitivity (SEN) was dened
as the slope of the calibration line. The precision for each analyte was
considered in terms of concentration according to the following
equation:
s
XI
^i 2
y y
i1 i
:
Precision SEN
I2
1

In this formulation, yi and i are the loadings for the given analyte
assessed from the developed algorithms and the loadings obtained
from the calibration line, respectively, and i is the number of used calibration samples.
Limit of detection (LOD) was calculated directly from the calibration
curve [40]. LOD was calculated as 3.3/SEN, where is the estimation of
random error in the y direction and equal to:

s
XI
^ i 2
y y
i1 i
I2

Limits of detection were found between 0.721.53 ng g1 and 4.40

5.58 ng g1 for CBS-X and CXT, respectively, with acceptable precision
in different samples.
Table S6 indicates that the obtained results by the presently proposed method using PARAFAC, APTLD and SWATLD, for determination
of the laundry powder samples are similar to those obtained by HPLC.
Fig. S3 shows HPLC chromatograms of two laundry powder samples
for determination of CBS-X and CXT. Determination of optical
brighteners in environmental samples needs to very low limit of
detection and elimination of matrix effect. Lower limit of detection
may be achieved by concentrating the extract by suitable solvent
(preconcentration), then a proportion of the extract must be added to

116

A. Gholami et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 153 (2016) 108117

Table 4
Recoveries of CBS-X and CXT mixtures in environmental samples.

Agriculture soil 1

Methods

Added CBS
(ng g1)

Found CBS
(ng g1)

Expected CBS
(ng g1)

Recovery (%)

Added CXT
(ng g1)

Found CXT
(ng g1)

Expected CXT
(ng g1)

Recovery (%)

PARAFAC

0.00
1.00
2.00
0.00
1.00
2.00
0.00
1.00
2.00
0.00
1.00
2.00
0.00
1.00
2.00
0.00
1.00
2.00
0.00
1.00
2.00
0.00
1.00
2.00
0.00
1.00
2.00

5.99
6.82
7.40
5.54
6.61
7.43
5.54
6.61
7.43
3.77
4.46
5.42
3.87
4.67
5.63
3.87
4.67
5.63
5.06
4.29
4.25
4.92
5.61
6.98
4.92
5.61
6.98

6.99
7.99

6.54
7.54

6.54
7.54

4.77
5.77

4.87
5.87

4.87
5.87

6.06
7.06

5.92
6.92

5.92
6.92

97.56
92.61

101.08
98.54

101.08
98.54

93.50
93.93

95.89
95.91

95.89
95.91

71.00
60.22

94.76
95.08

94.76
95.08

0.00
5.00
10.00
0.00
5.00
10.00
0.00
5.00
10.00
0.00
5.00
10.00
0.00
5.00
10.0
0.00
5.00
10.0
0.00
5.00
10.00
0.00
5.00
10.00
0.00
5.00
10.00

12.68
16.98
22.43
13.03
18.68
25.12
13.03
18.68
25.12
8.96
13.57
17.74
6.52
12.31
17.67
6.52
12.31
17.67
0.0
13.87
33.39
13.98
18.20
24.20
13.98
18.20
24.20

17.68
22.68

18.03
23.03

18.03
23.03

14.96
18.96

11.52
16.52

11.52
16.52

5.00
10.00

18.98
23.98

18.98
23.98

96.04
98.89

103.60
109.07

103.60
109.07

90.70
93.56

106.85
106.96

106.85
106.96

275.45
333.90

95.89
100.91

95.89
100.91

APTLD

SWATLD

Agriculture soil 2

PARAFAC

APTLD

SWATLD

Wastewater

PARAFAC

APTLD

SWATLD

the clean-up column (with appropriate pH) and eluted the column with
a suitable solvent and nally analysis could be performed by HPLC. This
chromatography method is imprecise, laborious, need to toxic and expensive pure organic solvent.
As a result, there is much interest to improve the uorometric method that requires inexpensive and simple-to-operate equipment for analysis of OBs in real samples. One of the major shortcomings of
uorometric analysis of OBs is interferences from natural organic

Table 5
Figures of merit obtained from the multivariate calibration procedures in real samples.
Samples
Laundry powders

PARAFAC

APTLD

SWATLD

Agriculture soils

PARAFAC

APTLD

SWATLD

Wastewater

PARAFAC

APTLD

SWATLD

Sensitivity
Precision
Detection limit
Sensitivity
Precision
Detection limit
Sensitivity
Precision
Detection limit
Sensitivity
Precision
Detection limit
Sensitivity
Precision
Detection limit
Sensitivity
Precision
Detection limit
Sensitivity
Precision
Detection limit
Sensitivity
Precision
Detection limit
Sensitivity
Precision
Detection limit

CBS-X

CXT

792.6
0.22
0.72
669.2
0.22
0.73
669.2
0.22
0.73
650
0.28
0.92
651.8
0.39
1.28
651.8
0.39
1.28

532.69
0.46
1.53
532.69
0.46
1.53

120
1.37
4.53
133.2
1.5
4.94
133.2
1.5
4.94
104.5
1.69
5.58
105
1.66
5.48
105
1.66
5.48

97.8
1.33
4.40
97.8
1.33
4.40

matter, that this problem could be solved with three-way chemometric

resolution techniques.
5. Conclusion
In this research, a rapid, simple and selective method based on three
way analysis of the excitationemission uorescence data was developed for determination of CBS-X and CXT as optical brighteners in the
presence of unknown interferences in laundry powders and environmental samples. The method is assisted by second-order chemometric
analyses employing the PARAFAC, SWATLD and APTLD that help us to
model these real samples in the presence of interferences, through the
unique decomposition of the three way data array. These methods
allow the extraction of relative concentrations of the analytes, as well
as the spectra proles and enable to handle the direct interference effect
of a complex matrix.
By spiking known concentration of CBS-X and CXT to the real samples, the accuracy of the proposed methods was validated and recoveries of the spiked value were calculated. The obtained results show that
these algorithms can be used for the direct determination of CBS-X
and CXT at the ng g1 without need for separation or preconcentration
step before analysis in real samples. Also comparison of trilinear second
order calibration algorithms has been carried out for APTLD, PARAFAC
and SWATLD. PARAFAC algorithm converged slowly but can afford reasonable and unique result; if the correct number of component was selected. We should be careful in using PARAFAC where two chemical
interference species are present with a high degree of spectral overlapping with analyte. In this situation, PARAFAC cannot predict reasonable
concentration prole of components and can't afford acceptable recoveries. The APTLD algorithm can avoid the two-factor degeneracy problem due to penalty constraints that holds the very fast convergence.
Indeed, the performance of APTLD is very stable when the value of the
penalty factors is greater than 108 and close to SWATLD. Both of these
algorithms are also insensitive to component number and hold very
fast convergence speed and provide reasonable results when there is a
high degree of spectral overlapping between analyte and interferences
in real samples.

A. Gholami et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 153 (2016) 108117

Acknowledgments
The authors are grateful to the University of Kashan for supporting
this work by grant no. 363018/2.
Appendix A. Supplementary data
Supplementary data to this article can be found online at http://dx.
doi.org/10.1016/j.saa.2015.08.012.
References
[1] J.A. Field, T.M. Field, T. Poiger, H. Siegrist, W. Giger, Water Res. 32 (1998) 19391947.
[2] O.W. Neiditch, J. Am, Oil Chem. Soc. 58 (1981) 162165.
[3] Ullmans Encyclopedia of Industrial Chemistry, 2001 Electronic release. 6th edition,
Wiley-VCH Interscience, 2001 (http://www.wiley-vch.de/contents/ullmann/ull_
10518.htm).
[4] Ciba Specialty Chemicals, Inc, Human and Environmental Risk Assessment on Ingredients of European Household Cleaning Products: Fluorescent Brightener FWA-1,
2004 (http://www.heraproject.com/RiskAssessment.cfm).
[5] Ciba Specialty Chemicals, Inc, Human and Environmental Risk Assessment on Ingredients of European Household Cleaning Products: Fluorescent Brightener FWA-5,
2004 (http://www.heraproject.com/RiskAssessment.cfm).
[6] C.J. Anastasiou, Optical Brightener Detection for Tracking Wastewater Contributions
to Coastal, Estuarine and Freshwater Systems, Department of Environmental Protection Watershed Management Program, Florida, 2007 130.
[7] Chinese National Standard (CNS) No. 12899, Identication of Fluorescent Brightening Agents Classes on Textiles, 1991 (http://www.webstore.jsa.or.jp/webstore/Com/
FlowControl.jsp).
[8] C.R. Ganz, C. Liebert, J. Schulze, S. Stensby, J. Water Pollut, Control Fed. 47 (1975)
28342849.
[9] A.W. Burg, C.J. Kensler, M.W. Rohovsky, Environ. Control 7 (1977) 91120.
[10] H. Hu, Z. He, Y. Ni, H. Zhang, Y. Zhou, Pulp Pap. 110 (2009) 3741.
[11] W. Coas, P.G. Desideri, L. Lepri, J. Chromatogr. 322 (1985) 363370.

[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]

117

J.L. Jasperse, P.H. Steiger, J. Am, Oil Chem. Soc. 69 (1992) 621625.
D. Kirkpatrick, J. Chromatogr. 139 (1977) 168173.
B.P. Mcpherson, N. Omelczenko, J. Am. Oil Chem. Soc. 57 (1980) 388391.
G. Calabro, P. Gurro, G. Micali, Analyst 109 (1984) 155158.
J.A. Field, T.M. Field, T. Poiger, W. Giger, Environ. Sci. Technol 30 (1996) 22202226.
W. Giger, J.M.A. Stoll, Anal. Chem. 69 (1997) 25942599.
W.H. Ding, W.C. Shu, J. Chromatogr. A 1088 (2005) 218223.
W.H. Ding, W.C. Shu, J. Chin. Chem. Soc. 56 (2009) 797803.
D.L. DuVal, K. Miyajima, I. Ogura, J. Am. Oil Chem. Soc. 72 (1995) 827833.
H.C. Chen, W.H. Ding, J. Chromatogr. A 1108 (2006) 202207.
L. Carolei, I.G.R. Gutz, Talanta 66 (2005) 118124.
S.H. Ahmadi, J. Azad, K. Kargosha, M. Mansourian, Talanta 75 (2008) 589593.
M.M. Reis, S.P. Gurden, A.K. Smilde, M.M.C. Ferreira, Anal. Chim. Acta 422 (2000)
2136.
R. Bro, P. Geladi, A.K. Smilde, J. Wiley, Multi-Way Analysis with Applications in the
Chemical Sciences, Wiley Online Library, Hoboken, NJ, 2004.
I.D. Wilson, U. Brikman, J. Chromatogr. 1000 (2003) 325335.
R. Bro, Anal. Chim. Acta 500 (2003) 185194.
Y. Akhlaghi, M.R. Hormozi-Nezhad, M. Kompany-Zareh, Anal. Chem. 84 (2012)
66036610.
D.M.F.Y. Ding, L.Q. Hu, H.L. Wu, A.L. Xia, R.Q. Yu, J. Chemom. 19 (2005) 6576.
L.Q. Hu, S.F. Li, H.L. Wu, A.L. Xia, R.Q. Yu, S.H. Zhu, J. Chemom. 21 (2007) 133144.
Z.P. Chen, J.H. Jiang, H.L. Wu, Y. Li, R. Yu, Q. Yu, Chemom. Intell. Lab. Syst. 52 (2000)
7586.
Z.P. Chen, H.L. Wu, R.Q. Yu, J. Chemom. 15 (2001) 439453.
http://www.models.kvl.dk/source.
A.C. Olivieri, H.L. Wu, R.Q. Yu, Chemom. Intell. Lab. Syst. 96 (2009) 246251.
A. Afkhami, M. Bahram, R. Bro, C. Stedmon, J. Chemometr. 20 (2006) 99105.
C.M. Andersen, A. Rinnan, Chemom. Intell. Lab. Syst. 76 (2005) 9199.
R. Bro, G. Tomasi, Chemom. Intell. Lab. Syst. 75 (2005) 163180.
R. Bro, K. Kjeldahl, H.A.L. Kiers, A.K. Smilde, Anal. Bioanal. Chem. 390 (2008)
12411251.
Practical Guide to the Analytical Validation of Body Fluid Chemistry Testing, Mayo
Medical Laboratories, 2013.
R.C.R. Beck, L.C. Cichota, R.B. Friedrich, A. Ravanello, C.M.B. Rolime, Quim Nova 321
(2009) 10521054.