Foods are often contaminated by mineral oils, which have negative effects on human health. Mineral oils are complex mixtures of hydrocarbons derived from petroleum. They are absorbed by the body, stored in various organs, and may cause microgranulomas in several tissues, which are extremely harmful to the human body. In 2008, a research on mineral oil in food reviewing the toxicity and occurrence of this kind of contamination was developed by the European Food Safety Authority (EFSA)^[1,2]. Although data on mineral oil contamination are available only for a limited number of food categories, a significant contribution to the human daily intake can be ascribed to vegetable oils, which have been reported to be frequently contaminated with amounts of mineral oils around 0.02−0.08g·kg−1^[3]. In 2008, nearly 100,000 tons of imported Ukrainian sunflower seed oil was contaminated with more than 1000mg·kg−1 of mineral oil in the European Union. Even though the source of the contamination was never officially confirmed, it is probable that mineral oil was added as a fraud^[4]. The occurrence of mineral oils in foodstuffs is related to different sources: environmental contamination^[5,6], food-grade mineral oils widely used for different purposes in many processed foods^[7], and food contact materials^[8]. Along with the above mentioned contamination sources, some unscrupulous traders illegally adulterate vegetable oils in order to reduce production costs. Thus, in order to protect consumers from the hazards of mineral oil, the control of mineral oil in vegetable oils is of major importance, and food testing laboratories must ensure testing speed in order to provide a timely warning to the producers and relevant authorities in the case of a crisis. For these reasons, laboratories should be provided with rapid, sensitive, and reliable analytical and cost effective methods for the determination of mineral oil.

A large number of analytical methods for the determination of mineral oils in edible oils have been proposed over the years, such as liquid–liquid–gas chromatography (GC)^[9], GC–flame ionization detection (FID)^[10], and liquid chromatography (LC-GC-FID)^[11]. These methods not only require advanced laboratory facilities but also involve elaborate processing steps and consume a great deal of time^[12], for instance, the removal of olefins is necessary before subjection of the sample to LC-GC-FID. Mineral oil was studied using Raman spectroscopy by de Jong et al.^[13], but the detection limit of mineral oil is about 0.25% (=2500mg/kg), which is much higher than the maximum limit of 50 mg/kg that is defined by the European Union for mineral oil in edible oil. Thus, some rapid, convenient, and accurate quantitative detection methods are needed. With the combination of modern chemometrics and instrumentation, near-infrared (NIR) spectroscopy is widely applied for rapid, accurate, low-cost, and non-destructive analysis in multiple industries. NIR spectroscopy has been revealed as promising for classification and quantitative determination of various materials, including oil quantity^[14], olive oil^[15,16], energy drinks^[17], soybean oil^[18], and biodiesel/diesel blends^[19,20].

The objective of this study was to develop a rapid and accurate method based on NIR spectroscopy integrated with chemometrics for rapid quantitative detection of mineral oil contamination in vegetable oil. The NIR spectra of mineral oil/vegetable oil blends were acquired in the spectral range of 910–2150 nm. The original spectrum is preprocessed by direct orthogonal signal correction (DOSC). The successive projections algorithm (SPA) was employed for selecting effective wavelengths in the calibration set after the DOSC preprocessing. Based on the effective wavelengths, prediction models were set up using two different modeling approaches—partial least squares (PLS) and principal component regression (PCR). After the models were established, the prediction set was then analyzed in order to estimate the actual predictive capability of the established models.

The vegetable oil sample in this study was rapeseed oil. Two brands (B1 and B2) of rapeseed oil were used during the period from February to May 2019 within the framework of a monitoring program for the control of mineral oil in food products. The mineral paraffin oil “Paraffin viscous PH Eur, BP, USP” [CAS 8012-95-1] was from Merck (Darmstadt, Germany). B1 brand rapeseed oil samples were adulterated with mineral oil at eight different concentrations: 1, 0.5, 0.2, 0.1, 0.08, 0.05, 0.03, and 0.01g·kg−1, respectively. The samples were divided into a calibration set and a prediction set. The calibration set was used to establish models for quantitative detection of mineral oil contamination in vegetable oil, and the prediction set was used as an external test set for validating the actual prediction ability of the models. We took 25 samples for each concentration in the B1 brand. Fifteen samples were for calibration and ten for prediction. The other brand (B2) of rapeseed oil sample was adulterated with 0.8g·kg−1 and 0.1g·kg−1 mineral oil for prediction, and we obtained 10 samples of 0.8g·kg−1 and 0.1g·kg−1, respectively. That is to say, we also validated the model with a new concentration (concentration not present in the calibration set) from another brand of rapeseed oil. The sample information in the calibration and prediction sets is shown in Table 1.

NIR spectrometer (Zeiss MCS 611NIR 2.2 WR) was used for spectral acquisition between 910 nm and 2150 nm. The NIR spectrum of each sample was obtained by taking the average of 33 scans. The spectra of rapeseed oil/mineral oil blends were obtained in the range from 910 to 2150 nm with 4 nm spectral resolution, and the average spectrum was used in modeling the data. The time required to achieve a spectral measurement was 30 s.

NIR spectra are often influenced by instrumental variation and measurement conditions, such as background noise and baseline shift. Therefore, the spectra are often pretreated before calibration. A relatively new preprocessing technique, DOSC, was introduced by Westerhuis et al.^[21]. This technique has better ability to subtract the irrelevant information compared to the previous methods. The DOSC was performed using MATLAB (R2018a, the MathWorks Inc., USA). The spectrum after being pretreated (XDOSC) was used to do the following analysis.

Figure 1 presents the transmittance^[22] curves in the range of 910–2150 nm, collected from the mineral oil and rapeseed oil (B1 brand). The two average curves were obtained using MATLAB 2018a and Origin 8.0. Mineral oils are mixtures of liquid hydrocarbons derived from petroleum. In detail, the C-H first overtone in hydrocarbon appears at about 1800 nm. It can be seen from Fig. 1 that there is an absorption peak in the transmission curve of mineral oil at 1814 nm. Rapeseed oils are made up of esters. The C=O keys second overtone in esters appears at about 1910–1950 nm. Figure 1 shows that rapeseed oil has an absorption peak at 1930 nm. Moreover, it can be seen from the figure that the transmittance curves of mineral oil and rapeseed oil are quite different, preliminarily indicating that pure rapeseed and mineral oils can be discriminated.

Fig. 1. Transmittance spectra of rapeseed oil (B1 brand) and mineral oil.

下载图片 查看所有图片

Principal component analysis (PCA) is a multivariate projection method, widely used in chemometrics to compress high-dimensional data into a lower-dimensional space with a minimal loss of information^[23]. It is designed to extract and display essential information from a data set and forms new variables called principal components (PCs). In the study, PCA was conducted on the full-spectrum data. The PCA was performed using the Statistical Product and Service Solutions (SPSS). All of the spectra of the calibration set sample were evaluated by PCA. Figure 2 presents the PC2×PC1 score plot resulting from the application of PCA. It is noteworthy that the first PC represents 90.47% of the data variance, while the second PC represents 7.51%. As can be seen, there is no overlap between different concentrations of mineral oils, indicating that the NIR spectrum conveys appropriate information for the classification task. This observation shows that the NIR spectra together with PCA can realize the classification of different concentrations of mineral oil in rapeseed oil.

Fig. 2. PC2×PC1 score plot for the overall set of 120 rapeseed–mineral oil samples.

下载图片 查看所有图片

Spectral data are known to possess variable (wavelength) dimensions with redundancy among contiguous variables^[24]. Elimination of irrelevant variables can predigest calibration modeling and improve results in terms of accuracy and robustness^[25]. The SPA was originally proposed by Araújo et al. in the context of multivariate calibration^[26,27]. In this study, SPA was applied to select the effective wavelengths that have the greatest contribution to the quantitative detection of mineral oil. SPA was carried out in a MATLAB environment (R2018a, the MathWorks Inc., USA). According to previous studies, effective wavelengths might be equally or more efficient than full wavelengths because they carry the most important information relevant to determination. In this study, 36 variables were selected as the effective wavelengths from the full spectra (910–2150 nm) for the quantitative detection of mineral oil contamination in vegetable oil, as shown in Table 2.

Once the effective wavelengths were selected, the spectral data sets were then reduced to a matrix with a dimension of s×t, where s is the number of samples (s=120), and the number of variables t is reduced from 311 to 36 wavelengths (the number of effective wavelengths). In this study, the models were established using two calibration algorithms, PLS^[28,29] and PCR^[30], based on the effective wavelengths. In the PCR and PLS models, root-mean-square error (RMSE) and correlation coefficient (R) are the two most important parameters. RMSE is closer to zero, whereas R is closer to one, implying the model’s better prediction ability. Table 3 presents the results of different calibration models. The best performance was obtained by DOSC-SPA-PCR, which provides a satisfactory calibration, RMSE 0.005, and R better than 0.99. DOSC and SPA made a significant contribution to the model, because most of the unrelated information in the original spectra has been removed. The results of DOSC-SPA-PLS and DOSC-SPA-PCR models are shown in Figs. 3(a) and 3(b). The obtained DOSC-SPA-PLS and DOSC-SPA-PCR models are shown as follows: $$\begin{gathered} YPLS=−0.19784X910nm−0.00317X930nm+0.20723X1066nm+0.09309X1614nm \\ +0.21298X1674nm+0.52350X1690nm+4.28272X1706nm−2.75074X1710nm \\ +0.59575X1734nm+1.34289X1746nm−0.44838X1750nm−10.30360X1754nm \\ −0.20254X1782nm+0.48721X1790nm−1.21490X1814nm−0.44141X1850nm \\ −0.25346X1866nm+0.49816X1898nm−0.02357X1906nm−0.13418X1922nm \\ +0.42399X1930nm−0.10942X1970nm−0.28962X2026nm−0.53677X2030nm \\ +0.04219X2050nm+0.30724X2066nm−0.03143X2078nm+0.00698X2090nm \\ −0.31297X2098nm−0.21812X2106nm+0.16126X2110nm+0.44248X2114nm \\ +0.14626X2122nm+0.07488X2126nm+0.88256X2134nm+0.26406X2150nm \\ +0.63877, \end{gathered}$$$$\begin{gathered} YPCR=−0.50496X910nm+0.15351X930nm+0.30558X1066nm+0.09631X1614nm \\ +0.72671X1674nm+0.36840X1690nm+2.62579X1706nm−2.08881X1710nm \\ +1.60547X1734nm+1.32047X1746nm−0.62717X1750nm−6.49843X1754nm \\ +2.42693X1782nm−0.62806X1790nm−1.81212X1814nm−0.58636X1850nm \\ −0.90724X1866nm+1.73341X1898nm−0.87935X1906nm−0.25609X1922nm \\ +0.61519X1930nm+0.30046X1970nm−0.71081X2026nm−0.26458X2030nm \\ +0.47511X2050nm+0.38152X2066nm−0.72060X2078nm+0.27836X2090nm \\ −0.35282X2098nm+0.39210X2106nm−0.08173X2110nm+0.14682X2114nm \\ −0.45689X2122nm+0.31962X2126nm+0.73204X2134nm+0.24402X2150nm \\ +0.73951, \end{gathered}$$where Xinm is the transmittance at the wavelength of i nm, and YPLS and YPCR are the predicted values.

Fig. 3. Calibration results of (a) DOSC-SPA-PCR and (b) DOSC-SPA-PLS. Fifteen samples per concentration were used for modeling, and there were 120 samples in total.

下载图片 查看所有图片

We used samples from the prediction set for the B1 brand for evaluating the prediction ability of the two models: DOSC-SPA-PLS and DOSC-SPA-PCR. There are 80 spectra for prediction. Figure 4 shows the predicted results of the two models. Different RMSE of prediction (RMSEP) values can be calculated: DOSC-SPA-PLS 0.008 and DOSC-SPA-PCR 0.006, respectively, while the correlation coefficients (R_p) values are 0.985 and 0.990, respectively. After that, we used 0.8g·kg−1 and 0.1g·kg−1 from another brand of rapeseed oil (B2) to validate the ability of the models. There are 10 samples of 0.8g·kg−1 and 0.1g·kg−1 in the prediction set, respectively. By the standards of the oil manufacturer, for the concentration of 0.8g·kg−1, the threshold can be 0.05. It means 0.8±0.05g·kg−1 can be considered as 0.8g·kg−1. Likewise, 0.1±0.05g·kg−1 can be considered as 0.1g·kg−1. Figure 5 presents the results of DOSC-SPA-PCR and DOSC-SPA-PLS for the concentration of 0.8g·kg−1 and 0.1g·kg−1 mineral oil for B2 rapeseed oil. As can be seen from Fig. 5, in the DOSC-SPA-PCR model, the prediction results of 0.8g·kg−1 and 0.1g·kg−1 are with 90% and 100% accuracy, respectively, while in the DOSC-SPA-PLS model, the prediction results of 0.8g·kg−1 and 0.1g·kg−1 both are with a 90% accuracy. All of the results together demonstrate that DOSC-SPA-PCR could be considered as the better method for quantitative detection of mineral oil contamination in vegetable oil.

Fig. 4. Prediction results of (a) DOSC-SPA-PCR and (b) DOSC-SPA-PLS. Ten samples per concentration were used for prediction, and there were 80 samples in total.

下载图片 查看所有图片

Fig. 5. Prediction results of (a) DOSC-SPA-PCR and (b) DOSC-SPA-PLS for B2 rapeseed oil.

下载图片 查看所有图片

In conclusion, NIR spectroscopy exhibited good performance as a rapid, accurate, sensitive, reliable, and cost-effective method for quantitative detection of mineral oil contamination in vegetable oil. Compared with previous studies, this study realized the rapid (about 30 s) quantitative detection of mineral oil contamination in vegetable oil with a limit of detection of 0.01g·kg−1. The results of the transmittance curves and PCA indicate that the NIR analysis technique can realize the quantitative detection of mineral oil contamination in vegetable oil. Then, out of 311 wavelengths, only 36 wavelengths were selected by SPA as the effective wavelengths after DOSC preprocessing. After that, DOSC-SPA-PCR and DOSC-SPA-PLS analyses were performed in order to obtain a model suitable for the mineral oil rapid quantification. Relatively high R (0.998) and low RMSE (0.005) were obtained by the DOSC-SPA-PCR model compared with DOSC-SPA-PLS. We used two brands of rapeseed oil to validate the ability of the two models. A satisfactory accuracy with R and RMSE of prediction by the DOSC-SPA-PCR method of 0.990 and 0.006 for the B1 brand was obtained, and the accuracy of 0.8g·kg−1 and 0.1g·kg−1 for the B2 brand in the prediction set was 90% and 100%, respectively. The study indicates that the NIR spectroscopy combined with DOSC-SPA-PCR is appropriate for rapid quantitative detection of mineral oil contamination in vegetable oil with a limit of detection of 0.01g·kg−1. The results may be useful for the development of a miniature spectrometer in mineral detection.