Original papers
Rapid Determination of Degradation of Frying Oil Using Near-Infrared Spectroscopy
2014 年 20 巻 2 号 p. 217-223
詳細
2014 年 20 巻 2 号 p. 217-223
The degradation of frying oil was determined using near-infrared (NIR) spectroscopy and partial least-squares (PLS) regression. One hundred and fifty six samples of frying oil (104 in a calibration set and 52 in a validation set) were obtained after use in an actual potato frying process. NIR transmission spectra of the samples were acquired directly using glass test tubes (13 mm dia.) and a NIR spectrometer. Calibration models with very high accuracy were developed for predicting acid value (AV) and total polar compounds (TPC) using PLS regression with full cross-validation. The coefficients of determination for calibration (R2) and standard error of cross-validation (SECV) were 0.99 (SECV: 0.17 mgKOH/g) and 0.98 (SECV: 1.25%) for AV and TPC, respectively. The accuracy of the NIR calibration models was tested using the validation set, yielding values for the root mean square of the prediction (SEP) of 0.17 mgKOH/g and 1.04% for AV and TPC, respectively. The results demonstrate that frying oils can be successfully monitored to a very high accuracy using NIR spectroscopy combined with glass tubes of 13 mm diameter as cells.
Deep-fried foods are popular because of their unique flavor and textural characteristics, which are closely connected to the quality of the oil in which the food is fried (Blumenthal, 1991). During frying, oil is subjected to prolonged periods of heating at high temperatures in the presence of air and water. This leads to a wide range of complex chemical reactions, such as thermal oxidation, hydrolysis and polymerization (White, 1991, Clark and Serbia, 1991). The compounds generated from these chemical reactions not only have negative effects on flavor but also add undesirable constituents to fried foods (Tyagi and Vasishtha, 1996). Therefore, quality control for frying oil and a rapid technique for analysis are both very important.
Various quality attributes are used to evaluate the quality of frying oil—acid value (AV) or free fatty acid (FFA) value, total polar compounds (TPC), or total polar materials (TPM)—using traditional chemical methods. Most standard analytical methods for oil analysis are expensive, require lengthy sample preparation times and often depend on advanced instrumentation (Fritsch, 1981, Orthoefer, 1988, Osawa et al., 2007). Thus, physical methods based on colorimetric reactions, refractive index, density and viscosity have also been proposed (Fritsch, 1981, Tous, 1968, Bansal et al., 2010). These methods are relatively easy to use but are not suitable for monitoring chemical changes in frying oils during deep-fat frying. Techniques such as high-performance liquid chromatography (HPLC) and gas chromatography (GC) have also been used for identifying adulterants, but these are expensive and time-consuming (Andrikopoulos et al., 2001).
NIR spectroscopy is a rapid, reliable and non-destructive technique that is widely used for quality and process control and for quantitative characterization of various products without the use of reagents or solvents (Chen et al., 2007). Previous studies have reported the determination of FFA and TPM in heated soy-based oils with added water by NIR spectroscopy using a quartz cuvette of 2 mm path length combined with data analysis using partial least squares (PLS) and forward stepwise multiple linear regressions (Ng et al., 2007). Yavari et al. (2009) also used visible/NIR spectroscopy combined with PLS techniques to predict viscosity, AV and TPC values in frying oils. These studies demonstrated the ability of NIR spectroscopy to successfully monitor the thermal degradation of edible oils, but were limited either to oils heated in the presence of pure water or to blends of hydrogenated and non-hydrogenated oils. Thus, these studies did not realistically evaluate the oils used to fry foods.
Ng et al. (2011) used NIR spectroscopy with a quartz cuvette of 2 mm path length to predict the TPM and FFA contents of soy-based frying oils actually used to fry foods. Ogutcu et al. (2012) also used NIR spectroscopy with PLS techniques to predict the viscosity, FFA and TPM values of oil samples generated during a dough frying process. Relatively high correlations between wet chemical analysis and NIR spectroscopy data were reported. However, these studies were limited to samples of commercially refined oils which were used at only one constant temperature in the frying process. This would in itself generate a strong correlation between NIR predictions and analytically derived values for AV and TPC from the frying oil samples, and thus would not result in the development of a practical NIR model for predicting AV or TPC values. Moreover, because a standard quartz cuvette with a 2 mm path length was used in these studies as a sample cell, infusion of the samples was difficult due to the high viscosity of the frying oils. Furthermore, the sample cell, which was used repeatedly, had to be washed.
The objectives of the present study were to test the use of NIR spectroscopy to determine the AV and TPC values of refined and unrefined oils used at different frying temperatures and to investigate the applicability of degradation estimation by NIR spectroscopy using separate disposable test tube cells (5 mL, 13 mm diameter), which was convenient from a practical point of view.
Sample Preparation Two types of rapeseed oil (commercially refined canola oil from the Nisshin Oillio Group Ltd, Tokyo, Japan, and unrefined rapeseed Kizakinonatane oil from Akita New Bio Farm Co., Ltd, Akita, Japan) were used as samples of frying oil. Frozen par-fried French fries in institutional packs were purchased from a local supermarket and used for deep-frying. Frying was conducted in a restaurant-style stainless steel electric fryer; model TF-40A (Taiji & Co., Ltd., Kanagawa, Japan). Using a frying temperature of 180, 200 or 220°C, batches of 100 g of frozen French fries were fried for 3 min at 22 min intervals, once the oil had reached the desired temperature, for a period of 7 h each day for 4 consecutive days. This is equivalent to frying 17 batches per day and therefore 68 batches for the whole experiment. During the frying process, 200 mL of heated oil was drawn off every 3.5 h and stored at −18°C until analysis for AV and TPC values and acquisition of NIR spectral data. The frying experiments were carried out once using canola oil and twice using rapeseed Kizakinonatane oil. A total of 156 frying oil samples, degraded to different degrees, were obtained from the food frying process.
Reference analysis The AVs of the frying oil samples were determined in triplicate using an automatic potentiometric titrator (AT-500N, Kyoto Electronics Manufacturing, Kyoto, Japan) according to AOCS Official Method Cd 3d-63 (AOCS, 1997, revised 2003). All AV analysis results were expressed as mgKOH/g oil. The TPC values of the oil samples were measured directly with a food oil monitor (FOM 310, Ebro Electronics, Ingolstadt, Germany) based on changes in the dielectric constant.
Spectral acquisition Spectroscopic data from the frying oil samples were collected using a Foss NIRSystems model 6500 scanning spectrometer (NIRSystems Division of Foss Electric, Silver Spring, MD, USA). Spectral analysis software (NSAS, version 3.53; Foss NIRSystems, Inc.) was used in this study to collect NIR spectroscopic data. As shown in Photo 1, disposable glass test tubes (5 mL, 13 mm dia.) containing the frying oil samples were used as sample cells. Transmission spectra over the 700 – 2500 nm range at a resolution of 2 nm were collected from each frying oil sample at room temperature (25°C). All samples were heated to 25°C in a water bath prior to the collection of spectra and then placed into the circular hollow of an aluminum block maintained at 25°C with an electric thermostat. An empty glass tube was used as a reference before the spectral measurements were obtained. The oil samples were scanned 32 times in the sample compartment mode for each spectrum.
NIR spectral measuring unit with glass tube as sample cell.
Because there are minute differences in the diameter and circular shape of individual glass test tubes, baseline offset and slopes in spectra may affect the development of a robust calibration model. One of the earliest methods for removing baseline offset and slope is the use of derivative spectra (Duckworth, 2004, Chen et al., 2006). The first derivative of a spectrum is simply a measure of the slope of the spectral curve at every point. The slope of the curve is not affected by baseline offsets in the spectrum, and thus the first derivative is a very effective method for removing baseline offsets. The second derivative is a measure of the change in the slope of the curve. In addition to ignoring the offset, it is not affected by any linear “tilt” that may exist in the data and is therefore a very effective method for removing both the baseline offset and slope from a spectrum. The Savitsky-Golay method (Madden, 1978), with a segment of 20 nm and gap of 0 nm, was used for both first derivative and second derivative spectrum processing in the research.
Spectral variance One useful method for observing variables in NIR spectra is the use of the spectral variance method (Galtier et al., 2011). The spectral variance represents the variance of all spectra of the sample set (n), and is calculated at each wavelength as follows (Eq. 1):
 |
where Aix is the absorbance of the spectrum i at wavelength x and Ax is the mean absorbance of all spectra at wavelength x.
Calibration development and validation The 156 oil samples obtained from the frying process were divided into calibration and validation sets as follows. Initially, samples within the parent set were sorted by the presence of each chemically determined component. Starting with the lowest-component sample, the first and third samples were assigned to the calibration set, and the second sample to the validation set. The next group of three samples was similarly assigned, and so on, until the last group. Thus, 104 samples were included in the calibration set and 52 in the validation set. Statistics for the AV and TPC values of the frying oil samples selected for the calibration and validation sets are shown in Table 1. The calibration models were created by PLS regression from the log (1/T) spectra and its first and second derivatives, where T is the transmittance of the sample at a specific wavelength. The coefficient of determination (R2), standard error of calibration (SEC) and mean square error of cross-validation (SECV) were used to determine the number of factors included in the calibration model. The models generated were validated using the validation sample set. The correlation coefficient of prediction (r) and standard error of the prediction (SEP) were used to choose the best model; SEP measured how well the model predicted sample values in the validation set.
| Calibration set (n=104) | Validation set (n=52) | |||||
|---|---|---|---|---|---|---|
| Mean | Range | SD | Mean | Range | SD | |
| AV (mgKOH/g ) | 2.82 | 0.09–8.37 | 2.06 | 2.80 | 0.09–7.47 | 2.02 |
| TPC (%) | 11.00 | 0.50–40.00 | 8.21 | 10.92 | 0.50–35.00 | 7.97 |
SD: standard deviation
NIR spectra of frying oils Figure 1 shows the raw NIR spectra (700 – 2500 nm) of samples of two types of frying oil (fine line and dotted line in Figure 1). Intense absorbance peaks from the frying oils may be observed in the NIR region depicted. The peaks evolving around 926 nm are caused by the C-H stretching 3rd overtone of CH2 groups. The peaks evolving around 1210 nm are caused by the C-H stretching 2nd overtone of CH3 and CH2 groups, and those around 1390 and 1414 nm are caused by 2 × C-H stretching and C-H deformation vibration of CH3 and CH2 groups, respectively. The peaks at 1728 and 1760 nm are due to the C-H 1st overtone of CH2 groups. A shoulder around 2144 nm can be attributed to the C-H and C = C stretching vibrations of CH = CH groups. Furthermore, the peak observed at 2178 nm is due to the asymmetric C-H stretching and C = C stretching vibrations from CH = CH groups (Osborne et al., 1993; Christy et al., 2004). Generally, the absorption peaks of oils are observed at wavelengths near 2308 nm and 2348 nm, but such peaks were not observed in this study. This may have been because the optical path of the glass test tubes was too long. This wavelength region was not used in the study.
NIR spectra of two samples of frying oil, and variance based on the spectra of all frying oil samples.
To evaluate spectral changes observed during the frying process and to estimate useful spectral regions in the NIR range for determining AV and TPC values, a variance spectrum was obtained from the spectra of all frying oil samples (see bold line plotted in Figure 1). In this plot, peaks at around 1212, 1428, 1728, 1760, 2040, 2078, 2144 and 2178 nm were observed, corresponding with those shown for the raw NIR spectrum in Figure 1. In particular, peaks and troughs were observed in the range 2000 – 2200 nm. Osborne et al. (1993) reported that absorption patterns between 2000 and 2200 nm were related to the degree of oxidation of oils. Therefore, it can be seen that there is much information about the degradation of frying oils to be obtained from the NIR spectra. We also observed that larger variance was represented in the shorter-wavelength region. This phenomenon may have been caused by the strong absorption in shorter-wavelength region, because there were significant color changes in the frying oils with degradation.
NIR models for AV PLS regression results for predicting AV in frying oils using NIR spectra are shown in Table 2. A total of 15 PLS calibration models were developed for analyzing frying oils using the calibration and validation sample sets. There was a strong correlation between the NIR predicted data and validation data, with R2 values ranging from 0.95 to 0.99 and SEP values from 0.17 to 0.48 mgKOH/g. Relatively good results were found when the results from the short-wavelength (700 – 1100 nm), middle-wavelength (1100 – 1800 nm) and long-wavelength regions (1800 – 2200 nm) were compared. The models based on the longer-wavelength region showed a relatively high correlation between the predicted values and the actual AVs for frying oils.
| F | R2 | SEC | SECV | SEP | Bias | RPD | ||
|---|---|---|---|---|---|---|---|---|
| Raw spectra | 700–1100 nm | 10 | 0.96 | 0.42 | 0.53 | 0.47 | 0.09 | 4.3 |
| 1100–1800 nm | 6 | 0.97 | 0.37 | 0.43 | 0.39 | 0.05 | 5.3 | |
| 1800–2200 nm | 6 | 0.99 | 0.17 | 0.21 | 0.17 | 0.05 | 12.0 | |
| 1100–2200 nm | 6 | 0.99 | 0.19 | 0.23 | 0.19 | 0.05 | 10.7 | |
| 700–2200 nm | 7 | 0.99 | 0.22 | 0.28 | 0.22 | 0.05 | 9.4 | |
| First derivative spectra (segment 20nm, gap Onm) | 700–1100 nm | 9 | 0.96 | 0.42 | 0.54 | 0.48 | 0.12 | 4.3 |
| 1100–1800 nm | 6 | 0.96 | 0.39 | 0.44 | 0.38 | 0.05 | 5.4 | |
| 1800–2200 nm | 5 | 0.99 | 0.15 | 0.17 | 0.17 | 0.04 | 12.8 | |
| 1100–2200 nm | 5 | 0.99 | 0.21 | 0.23 | 0.19 | 0.04 | 10.9 | |
| 700–2200 nm | 5 | 0.99 | 0.21 | 0.23 | 0.20 | 0.06 | 10.1 | |
| Second derivative spectra (segment 20nm, gap Onm) | 700–1100 nm | 10 | 0.98 | 0.27 | 0.34 | 0.46 | 0.08 | 4.6 |
| 1100–1800 nm | 7 | 0.98 | 0.32 | 0.40 | 0.35 | 0.03 | 5.9 | |
| 1800–2200 nm | 5 | 0.99 | 0.17 | 0.20 | 0.19 | 0.02 | 11.0 | |
| 1100–2200 nm | 5 | 0.99 | 0.20 | 0.23 | 0.22 | 0.02 | 9.3 | |
| 700–2200 nm | 5 | 0.99 | 0.23 | 0.26 | 0.21 | 0.03 | 9.9 |
F, number of factors; R2, coefficient of determination; SEC, standard error of calibration; SECV, standard error of cross validation; SEP, standard error of prediction; Bias, average of differences between reference value and NIR value; RPD, ratio of standard deviation of reference data in the validation set to SEP
The most accurate model involved first-derivative spectra in the wavelength range 1800 – 2200 nm. It used five PLS factors and produced a high coefficient of determination (R2), and low values for standard error of calibration (SEC), standard error of cross-validation (SECV) and standard error of prediction (SEP) of 0.99, 0.15 mgKOH/g, 0.17 mgKOH/g and 0.17 mgKOH/g, respectively. The cross-validation and prediction (validation) results, represented graphically by plotting the reference analysis AVs versus the predicted values, showed strong linearity, as shown in Figure 2. Furthermore, all of the prediction models had ratio performance deviation (RPD) values, above 4.3, with the model involving first-derivative spectra in the wavelength range 1800 – 2200 nm exhibiting an RPD value of 12.8. Generally, an RPD value above 3 indicates a useful model which allows good quantitative predictions (Williams, 2001; Chen et al., 2004, 2005 & 2007). It can be concluded that the NIR spectra provided a good estimation of AV in frying oils, particularly for spectra that showed a low SEP value and a high RPD value.
Scatter plot diagram comparing reference AV and NIR predicted values.
NIR models for TPC values PLS regression results for predicting TPC values in frying oils using NIR spectra are shown in Table 3. Similarly to the AV results, there were strong correlations between the NIR predicted data and the validation data, with R2 values ranging from 0.96 to 0.98 and SEP values ranging from 1.04 to 1.67%. Based on observation and comparison of NIR results from the short-wavelength (700 – 1100 nm), middle-wavelength (1100 – 1800 nm) and long-wavelength regions (1800 – 2200 nm), relatively good results were provided in the middle-wavelength region, which showed a relatively high association between the predicted values and the actual TPC values of frying oils.
| F | R2 | SEC | SECV | SEP | Bias | RPD | ||
|---|---|---|---|---|---|---|---|---|
| Raw spectra | 700–1100 nm | 9 | 0.98 | 1.16 | 1.82 | 1.35 | −0.04 | 6.0 |
| 1100–1800 nm | 5 | 0.98 | 1.02 | 1.25 | 1.21 | 0.19 | 6.7 | |
| 1800–2200 nm | 5 | 0.98 | 1.08 | 1.43 | 1.23 | 0.21 | 6.6 | |
| 1100–2200 nm | 5 | 0.99 | 1.01 | 1.21 | 1.21 | 0.18 | 6.7 | |
| 700–2200 nm | 5 | 0.98 | 1.16 | 1.34 | 1.33 | 0.20 | 6.1 | |
| First derivative spectra (segment 20nm, gap Onm) | 700–1100 nm | 9 | 0.98 | 1.14 | 1.51 | 1.40 | -0.01 | 5.8 |
| 1100–1800 nm | 5 | 0.98 | 1.03 | 1.25 | 1.04 | 0.05 | 7.8 | |
| 1800–2200 nm | 5 | 0.98 | 1.07 | 1.24 | 1.19 | 0.15 | 6.8 | |
| 1100–2200 nm | 3 | 0.98 | 1.05 | 1.23 | 1.04 | 0.10 | 7.7 | |
| 700–2200 nm | 3 | 0.98 | 1.12 | 1.28 | 1.28 | 0.11 | 6.3 | |
| Second derivative spectra (segment 20nm, gap Onm) | 700–1100 nm | 7 | 0.98 | 1.29 | 1.58 | 1.67 | 0.14 | 4.9 |
| 1100–1800 nm | 6 | 0.99 | 0.93 | 1.23 | 1.11 | 0.12 | 7.3 | |
| 1800–2200 nm | 6 | 0.99 | 0.94 | 1.28 | 1.19 | 0.03 | 6.8 | |
| 1100–2200 nm | 5 | 0.98 | 1.12 | 1.31 | 1.09 | 0.13 | 7.5 | |
| 700–2200 nm | 5 | 0.98 | 1.17 | 1.40 | 1.21 | 0.06 | 6.6 |
F, number of factors; R2, coefficient of determination; SEC, standard error of calibration; SECV, standard error of cross validation; SEP, standard error of prediction; Bias, average of differences between reference value and NIR value; RPD, ratio of standard deviation of reference data in the validation set to SEP
The most accurate model involved first-derivative spectra in the wavelength range 1100 – 1800 nm. It used five PLS factors and produced a high coefficient of determination (R2) and low values for standard error of calibration (SEC), standard error of cross-validation (SECV), and standard error of prediction (SEP) of 0.98, 1.03%, 1.25% and 1.04%, respectively. The cross-validation and prediction results, represented graphically by plotting the reference AVs versus the predicted values, showed strong linearity, as shown in Figure 3. Furthermore, all prediction models had ratio performance deviation (RPD) values above 4.9, with the model involving first-derivative spectra in the wavelength range 1100 – 1800 nm exhibiting an RPD value of 7.8. It can also be concluded that the NIR spectra provided a good estimation of TPC values in frying oils.
Scatter plot diagram comparing reference TPC values and NIR predicted values.
Generally, the AV of oil increases not only with increasing frying time and temperature, but also with the presence of moisture, oxygen and other contaminants transferred from the immersed food to the oil. The amounts of secondary oxidation products such as aldehydes and ketones also increase as the frying oil degrades. The measured TPC value based on the secondary oxidation products increases with frying time. There may be a high correlation between AV and TPC values in the frying oil samples, which may help in the development of NIR models for AV or TPC. In the present study, we used refined and unrefined oil samples combined with a range of frying temperatures. This could explain the low correlation value of 0.05 between the AV and TPC values of the frying oil samples, compared with a value of 0.84 in research by Ogutcu et al. (2012). Therefore, based on the present research, it may be said that the NIR models for predicting AV and TPC values are practical but are not related to each other.
Regression coefficients can be used to discuss the contributions of individual wavelengths to a PLS calibration model, since a regression coefficient spectrum shows characteristic peaks and troughs that can indicate which wavelength range is important for the calibration model (Martens et al., 1989; Chen et al., 2004, 2008). In order to determine characteristic peaks and troughs in a regression coefficient spectrum, we observed and discussed the regression coefficients of PLS regression models based on second-derivative spectra. Figure 4 show the regression coefficients (solid line) of the PLS calibration model based on the second-derivative spectra for AV and the difference spectrum (dotted line) between two average second-derivative spectra of the high-AV sample group and the low-AV sample group. Three significant peaks at wavelengths of 1820, 2040 and 2166 nm were easily observed based on consideration of the difference spectrum shown in Figure 4. The negative peak at 1820 nm could be assigned to the 2nd overtone of C = O (Osborne, 1993), which may be related to the absorption of free fatty acids, which increases with the degradation of frying oils. The negative peak at 2040 nm could be correlated to the combined absorption band of C = O stretching second overtone or a combination overtone of stretching and deformation of O-H, which may be related to the amount of free fatty acid, which increases with hydrolysis of frying oils. The negative peak at 2166 nm could be correlated to the absorption band of the combination of the fundamental stretching overtones of C = C and C-H, related to unsaturated fatty acids in frying oils.
Regression coefficients (solid line) for the PLS calibration model for AV in frying oils, and difference spectrum (dotted line) between two average second-derivative spectra of the high-AV sample group and the low-AV sample group. (All samples were sorted by AV and then divided into high-AV, middle-AV and low-AV sample groups.)
Figure 5 shows the regression coefficients of the PLS calibration models of the TPC values and the difference spectrum (dotted line) between two average second-derivative spectra of the high-TPC sample group and the low-TPC sample group. Significant peaks at wavelengths of 1428, 1685, 1714 and 1736 nm were easily observed based on consideration of the difference spectrum shown in Figure 5. The negative peak at 1428 nm could be correlated to the absorption band of O-H stretching first overtone, corresponding to a wavelength of 1430 nm (Osborne, 1993), which may be related to the absorption of carbonyl compounds and lower fatty acids such as carboxylic acids, which increase with the degradation of frying oils. Negative peaks at 1684, 1714 and 1736 nm could be correlated to the absorption band of C-H stretching first overtone corresponding to CH3 or CH2, which may be related to the absorption of hydrocarbon and carbonyl compounds such as aldehydes and ketones, which increase with the degradation of frying oils.
Regression coefficients (solid line) of the PLS calibration model for TPC values in frying oils, and difference spectrum (dotted line) between two average second-derivative spectra of the high-TPC sample group and the low-TPC sample group. (All samples were sorted by TPC value and then divided into high-TPC, middle-TPC and low-TPC sample groups.)
Taken together, these results establish PLS calibration models for the AV and TPC values of frying oils based on the absorption of free fatty acid and deterioration products such as carboxylic acid, aldehyde and ketone groups, which increase with the deterioration of frying oils. The above results demonstrate the utility of the NIR technique as a means of determining the AV and TPC values of frying oils.
NIR spectroscopy can be successfully applied to the measurement of AV and TPC values in frying oils. PLS regression analysis for spectra recorded at 1800 – 2200 nm gave a low SEP value of 0.17 mgKOH/g and a very high RPD value of 12.8 for predicting AV. PLS regression analysis for spectra recorded at 1100 – 1800 nm gave a low SEP value of 1.04% and a relatively high RPD value of 7.8 for predicting TPC values. These prediction results showed that NIR spectroscopy using separate disposable 5 mL test tube cells is a very useful method for measuring the degradation of frying oils. Furthermore, NIR spectroscopy has significant advantages over other measurement techniques; it is a fast and simple method that requires no sample preparation, so is a very practical method for measuring AV and TPC values in edible oils during the frying process.
