1. Introduction

Polycyclic aromatic hydrocarbons (PAHs), a group of semi-volatile organic compounds consisting of two or more fused aromatic rings, are generally formed during incomplete combustion of carbon, pyrosynthesis or pyrolysis of hydrocarbons. They can pollute nature water by oil spills, direct industrial effluents, municipal discharges and condensing of PAH vapors at ambient temperature [1]. PAHs have the confirmed carcinogenic and mutagenic properties, and therefore are recognized as priority contaminants to be detected despite low solubility in aquatic environment [2].

High performance liquid chromatography (HPLC) and gas chromatography (GC) are standard and effective methods for determining PAHs with low solubility in water [3–5]. Nevertheless, they generally involve complex processes such as extraction, cleaning up and chromatographic determination, etc. All these made them laborious, tedious, time-consuming and therefore unable to perform real time and in situ analysis. Instead, three dimensional fluorescence spectra detection technique has been employed as a rapid and nondestructive method to in situ monitor PAHs in water [6–8]. However, the broad fluorescence band and the overlaps within fluorescence spectra of object analytes from the same chemical family significantly degrade the selectivity of this method. Fortunately, this is resolved by the use of advanced chemometric tools, especially second-order multivariate calibration methods. Parallel factor analysis (PARAFAC), the most popular second-order calibration method, is capable of predicting the concentration and extracting the spectral profile of object analytes directly in the presence of unknown interferents [9]. Renee et al. accurately quantified trace pesticides and polycyclic aromatic hydrocarbons (PAHs) by PARAFAC [6]. In addition, there are two other promising alternatives, i.e., multi-way partial least square (N-PLS) [10] and unfolded partial least square (U-PLS) [11]. They are more flexible to cope with spectra data than PARAFAC [12]. By coupling with a separate procedure known as residual bilinearization (RBL) they also show the second-order advantage [13]. Francis et al. demonstrates the feasibility of these methods in determining polycyclic aromatic hydrocarbons in edible oils [14].

Although the aforementioned methods could be used to determine PAHs in the natural water to certain degree, however, it is still difficult to quantify the concentrations of PAHs because of their low solubility, different fluorescence intensities and the complex matrix interferents from water. The present work aims at evaluating and comparing the applicability of three methods, i.e., PARAFAC and the combination of RBL with either N-PLS or U-PLS, for predicting PAHs concentrations in river water and reservoir water with standard addition.

2. Theory

2.1 Three-way PARAFAC

An excitation–emission matrix spectrofluorometer can generate a J × K data matrix, where J and K are the number of emission wavelength and excitation wavelength, respectively. By ‘stacking’ the matrixes obtained from I samples, a three-way array X with dimensions I × J × K could be produced. For this data array a trilinear model can be expressed as [9, 15]:

2.2 N-PLS/RBL

Multi-way partial least square (N-PLS) is an extension of the partial least square regression to multi-way data. Combined with calibration data array X (I_cal × J × K) and vector of calibration concentrations y (I_cal × 1), the N-PLS algorithm decomposes the multi-way array X into a score vector T (I_cal × A) and two weight loadings, W^J (J × A) and W^K (K × A) where I_cal, J, K and A is the number of calibration samples, emission wavelengths, excitation wavelengths and the number of latent factors respectively. The parameter A can be selected by leave one-out cross-validation [16]. A vector of regression coefficients v (size A × 1) is also obtained. If no unexpected components occurred in the test sample Xu, the analyte concentration is estimated according to:

In the presence of unexpected constituents in test sample, the sample score is unsuitable for analyte prediction. The residual of the N-PLS model of the test sample signal (S_p) will be abnormally large in comparison with the typical instrumental noise level:

A separate procedure called residual bilinearization (RBL) is used to solve this problem, which aims at minimizing the residual computed while fitting the sample data to the sum of the relevant contributions:

During the RBL procedure, the loadings are kept constant at the calibration values, and t_u is varied until the final RBL residual error S_u is minimized using a Gauss–Newton procedure:

2.3 U-PLS/RBL

In the unfolded partial least square method (U-PLS), the calibration data array X (I_cal × J × K) is first vectorized into I_cal-th vectors with size JK × 1 of each, and then together with the vector of calibration concentrations y (I_cal × 1) a usual PLS model is built. A set of loadings P (JK × A) and weight loadings W (JK × A) are obtained, as well as regression coefficients v (A × 1). If no unexpected components occurred in the unknown sample Xu, v could be employed to estimate the analyte concentration according to Eq. (2). When unexpected constituents occur in unknown sample Xu, a similar procedure as that of N-PLS, residual bilinearization (RBL), is used to solve this problem.

3. Experiment and data processing

3.1. Apparatus

Fluorescence measurements were performed on a Hitachi F-7000 spectrofluorometer equipped with a 150 W xenon lamp and connected to a PC microcomputer. The slit band widths of excitation and emission monochromators were both fixed at 5 nm. The voltage of photomultiplier detector was set at 600 V and the scan rate at 12000 nm min^–1. The excitation-emission fluorescence matrices were recorded with excitation wavelengths EX in range of 240-350 nm and emission wavelengths EM 290-500 nm at a 2-nm interval.

3.2. Reagents and solutions

Phenanthrene (PHE), pyrene (PY), anthracene (AN) and fluorene (FLU), acenaphthene (ACE) and fluoranthene (FLA), among the most abundant PAHs in the dissolved phase of natural Water, are selected as object analytes not only because of their spectra bands overlaps, but also their different fluorescent intensity [Figs. 1(a)-1(f)], which make their concentration prediction difficult. They were bought from Aladdin Co. and used without further purification. Stock solutions (100 mg L⁻¹ for each PAH) were prepared by dissolving appropriate PAHs in HPLC-grade ethanol, and then stored in a dark flask at 4°C for use. The working solutions of PHE, PY, AN, FLU ACE and FLA were prepared by diluting the stock solutions to 100 μg L⁻¹ using deionized water.

 

Fig. 1 Three-way fluorescence spectra of (a) PHE 8.0μg L⁻¹, (b) PY 8.0μg L⁻¹, (c) AN 8.0μg L⁻¹, (d) FLU 8.0μg L⁻¹, (e) ACE 8.0μg L⁻¹, and (f) FLA 8.0μg L⁻¹

Download Full Size | PDF

3.3. Calibration and validation samples

A ten-sample set of PAHs mixture was prepared using deionized water and used as calibration set. The following ten samples prepared using deionized water was designed as validation set. The concentrations of each added PAH were in the range 1 – 10 μg L⁻¹.

3.4 Real samples

Two test sets with eight samples of each set were prepared with river water and reservoir water, respectively. River water and reservoir water were sampled from Nanfei River and Dongpu reservoir in Hefei (China), and filtered through 0.45 μm filter. The sampled waters were analyzed with high performance liquid chromatography and had no obvious detected object analytes before adding PAHs. Figures 2(a) and 2(c) presents the three fluorescence spectra of sampled river and reservoir water, and Figs. 2(b) and 2(d) the river and reservoir water adding the same object analytes. Obviously different water background offers different interferents.

 

Fig. 2 The three fluorescence spectra of sampled natural water: (a) river water, (b) river water adding PAHs, (c) reservoir water, (d) reservoir water adding PAHs. The added PAHs in river and reservoir water are the same and as follows: PHE 1.0 μg L⁻¹, PY 1.2 μg L⁻¹, AN 2.0 μg L⁻¹, FLU 1.8 μg L⁻¹, ACE 1.8 μg L⁻¹ and FLA 1.0 μg L⁻¹

Download Full Size | PDF

3.5 Data processing

After removing the Raman and Rayleigh scatterings by using EEMCAT under MATLAB environment [17], the acquired fluorescence matrices were arranged in a 38 × 106 × 56 three-way data set where 38 was the number of samples, 106 the emission wavelengths and 56 the excitation wavelengths.

4. Results and discussion

In order to evaluate the performance of PARAFAC, N-PLS/RBL and U-PLS/RBL, the similarity coefficient, root mean square error of prediction and relative error of prediction are calculated. The similarity coefficient of three dimensional fluorescence spectra is calculated according to [18]

The root mean square error of prediction (RMSEP) and the relative error of prediction (REP) are calculated according to [19]

Firstly the experimental three-way data set is fitted by the three-way PARAFAC according to Eq. (1) by combining the data matrix of each sample needed for analysis with all the calibration samples. Prior to performing three-way PARAFAC, the number of components is determined as six for validation samples, as seven for reservoir samples and as eight for river samples by the core consistency test and the PARAFAC sum of squared errors (SSE) [17, 20]. Spectra profiles obtained from the three-way PARAFAC model with non-negativity restrictions are shown in Fig. 3. The fluorescence spectra of each PAH are clearly distinguished [Figs. 3(a) and 3(b)]. The resolved spectra (colorized lines) fit the corresponding reference spectra (black lines) with similarity coefficients (r) all above 0.935 (Table 1). All those demonstrated six analytes can be recognized correctly by PARAFAC although different extents of overlaps occur among the bands. Furthermore, the concentrations of PAHs can also be predicted by concentration score loadings of validation and test samples, and the linear regression lines which resulted from concentration score loadings against real concentrations of calibration samples. Based on the linear regression curves, it is easy to predict the concentrations of PAHs in validation and test samples. The obtained results are collected in Tables 2, 3 and 4. In validation samples, the results for AN are good with REP 1% whereas those for ACE are poor with REP 13% (Table 2).

 

Fig. 3 The resolved spectra (colorized lines) and corresponding reference spectra (black lines) of object analytes: (a) emission spectra, (b) excitation spectra. The λex (nm)/λem (nm) are as follows:250/364, 240/372, 250/380, 264/304, 290/322 and 286/460 for PHE, PY, AN, FLU, ACE and FLA, respectively.

Download Full Size | PDF

Table 1. Similarities between resolved and reference spectra.

View Table | View all tables in this article

Table 2. Statistical results for the quantification of the target PAHs in validation samples by three-way PARAFAC, N-PLS/RBL and U-PLS/RBL.

View Table | View all tables in this article

Table 3. concentrations predicted by three-way PARAFAC, N-PLS/RBL and U-PLS/RBL for each PAH of test samples prepared with river water.

View Table | View all tables in this article

Table 4. Concentrations predicted by three-way PARAFAC, N-PLS/RBL and U-PLS/RBL for each PAH of test samples prepared with reservoir water.

View Table | View all tables in this article

Then N- and U-PLS are also implemented. In order to estimate the numbers of latent variables for both N- and U-PLS, leave-one-sample-out cross-validation is performed over the calibration set. The optimum latent variables are estimated according to the criterion of Haaland and Thomas [16]. Then the predicted concentrations obtained with the validation sets are calculated with Eq. (2) without RBL because no interferences are added in these samples. The statistical results for validation set by N- and U-PLS are also listed in Table 2. Except for FLA, PARAFAC, N- and U-PLS achieve comparable results with the same values of root mean square error of prediction (RMSEP) and relative error of prediction (REP). N- and U-PLS both yield good predictions for FLA with relative error of prediction (REP) equal to 4%. U-PLS shows an improvement over PARAFAC and N-PLS with high sensitivity (SEN) and low limit of detection (LOD).

In the presence of unexpected interferents, it is necessary to exploit the second-order advantage which is provided by the RBL procedure [14]. Hence, for each test sample, the number of RBL components should be tuned, in addition to the latent variables of calibration set, which is the same as that employed during analysis of the validation samples. The required number of RBL components in test sample is established as one, two, three or four depending on the corresponding analyte and sample. This conclusion is reached by monitoring the change in the residual parameter S_u [Eq. (6)] as a function of increasing RBL components. The added and found concentrations of each PAH in different test sets are also listed in Tables 3 and 4. In comparison, U-PLS/RBL yields good predictions with relative error (REP) less than or equal to 6% for PHE, PY, AN and FLU in reservoir and river samples (Tables 3 and 4). The relative errors (REP) for ACE and FLA are 19% and 35% in river samples (Table 3). The relative errors (REP) for ACE and FLA are 12% and 30% in reservoir samples (Table 4). However, upon considering the complexity of the system, the latter values are acceptable. In contrast, poor predictions for ACE and FLA are observed when PARAFAC and N-PLS are applied with relative errors (REP) greater than 20%. FLA was the worst predicted analyte which may be because of its very low fluorescent intensity.

In order to demonstrate the stable performance of three methods, the same experiment was implemented in the next day. The results are listed in Table 5. U-PLS/RBL still have good performance in concentration prediction with relative error of prediction (ERP) less than or equal to 32% for river water and 24% for reservoir water (Table 5).

Table 5. The next day results of object PAHs in different water predicted by three-way PARAFAC, N-PLS/RBL and U-PLS/RBL.

View Table | View all tables in this article

5. Conclusion

In this work, PARAFAC, N-PLS/RBL and U-PLS/RBL are used to determining six kinds of PAHs in river and reservoir water. The feasibility of these methods is also evaluated. The results demonstrate that in comparison with PARAFAC and N-PLS/RBL, U-PLS/RBL have obvious advantage on predicting the concentrations of our tested six PAHs, especially of compounds with low fluorescent intensity. But we should note that PARAFAC cannot only resolve spectra of different components but also predict concentrations. It is suggested that the combination of PARAFAC and U-PLS/RBL may be better choice to resolve spectra and determine concentrations of PAHs in in situ monitoring water quality.