پديدآورندگان :
Hashemi Nasab Fatemeh Sadat Department of Chemistry, Sharif University of Technology, Tehran, Iran , Parastar Hadi h.parastar@sharif.edu Department of Chemistry, Sharif University of Technology, Tehran, Iran;
كليدواژه :
Oil fingerprinting , Gas chromatographic , Crude oil , Source identification , Pattern recognition , Chemometrics , Principal Component Analysis , Hierarchical Cluster Analysis , Partial Least Square Discriminant Analysis and Counter propagation Artificial Neural network
چكيده فارسي :
Spills of oil and related petroleum products in the marine environment can have serious biological and economic impacts. There are many oil tankers on the surface of the sea and 0.68х109 kg crude oil spilled into soil per year [1]. Therefore, oil fingerprinting plays a basic role in source identification of oil spill. Gas chromatographic (GC-MS and GC-FID) and spectroscopic (fluorescence and FT-IR) methods are among the most frequently used techniques for oil fingerprinting. However, the huge amount of produced data are complex and to be analyzed by chemometric methods [2]. Chemometrics provides many tools for pattern recognition, calibration and optimization that can increase the speed of the analysis and allow for more extensive use of the available data in this field. Examples of used chemometric methods for identification of crude oils source are Principal Component Analysis (PCA), parallel factor analysis (PARAFAC), Hierarchical Cluster Analysis (HCA), Partial Least Square Discriminant Analysis (PLS-DA) and Counter propagation Artificial Neural network (CPANN) [3, 4]. The purpose of this study was to provide an update of the state-of-the-art of oil fingerprinting techniques to demonstrate the use of a rapid, inexpensive and useful technique for distinguishing between crude oils. In this regard, a fractionation method based on SARA test [5] was used to divide nine crude oils (obtained from Sharif upstream petroleum institute) into aliphatic (saturate), aromatic, resin and asphaltene. It is necessary to remove the asphaltene fraction before proceeding with chromatography. After fractionation, three fractions of nine oil samples were analyzed by GC-FID and GC-MS. The obtained fingerprints were baseline corrected, aligned and auto-scaled and then analyzed using unsupervised classification methods of PCA and HCA. Evaluation of PCA score plot (explaining 93.69% of variance accounted for three PCs) showed that aromatic fractions belong to three classes and result of HCA with Ward’s method confirmed that. The clustering results of aliphatic and resin fractions also showed the presence of 3 classes but due to their different composition, classes were not the same. The results of unsupervised classification were then used a starting point for supervised classification methods of PLS-DA and CPANN. The results of PLS-DA analysis for aromatics showed best discrimination compared to other fractions. The procedure required an initial variable selection step using the variable important in projection (VIP) followed by a PLS-DA model. The optimum number of LVs for aromatic fraction (explaining 93.51 % and 95.87% of variance accounted for three LVs for X- and y-block, repectively) in PLS model was determined using R2 cross-validation (leave one out cross-validation). The figures of merit for classification of oils were R2Cal(0.975 0.972 0.930), R2CV(0.712 0.913 0.581), RMSEC(0.0786 0.0699 0.1245), SEN(1.000 1.000 1.000), and Specificity(1.000 1.000 1.000) . The CPANN top map of aromatic fractions represented 3 different classes that confirmed the results obtained by PCA score plot, HCA and PLS-DA. The classification figures of merit for CPANN were ERcal(0), ERCV(0.417), accuracy cal(1.000), accuracy cal(0.556), SEN(1.000 1.000 1.000) , Specificity(1.000 1.000 1.000). Finally, it is concluded that aromatic fractions were appropriate for identification of crude oils sources.
