High-precision identification of the actual storage periods of edible oil by FT-NIR spectroscopy combined with chemometric methods

Machine learning identification of edible vegetable oils from fatty acid compositions and hyperspectral images

Hyperspectral imaging analysis combined with machine learning was applied to identify eight edible vegetable oils, and its classification performance was compared with the chemical method based on fatty acid compositions. Furthermore, the degree of adulteration in vegetable oils was quantitatively investigated using machine learning-enabled hyperspectral approaches. The hyperspectral absorbance spectra of palm oil with a high degree of saturation were distinctly different from those of the other liquid oils. The flaxseed and olive oils exhibited the dominant hyperspectral intensities at 1170/1671 and 1212/1415 nm, respectively. Linear discriminant analysis demonstrated that two linear discriminants could explain a significant portion of the total variability, accounting for 96.0% (fatty acid compositions) and 98.9% (hyperspectral images). When the hyperspectral results were used as datasets for three machine learning models (decision tree, random forest, and k-nearest neighbor), several instances to incorrectly classify grapeseed and sunflower oils were detected, while olive, palm, and flaxseed oils were successfully identified. The machine learning models showed a great classification performance that exceeded 98.9% from the hyperspectral images of the vegetable oils, which was comparable to the fatty acid composition-based chemical method in identifying edible vegetable oils. In addition, the random forest model was the most effective in ascertaining adulteration levels in binary oil blends (R2 > 0.992 and RMSE < 2.75).

A hybrid variable selection method combining Fisher's linear discriminant combined population analysis and an improved binary cuckoo search algorithm

In this paper, a novel hybrid variable selection method for model building by near-infrared (NIR) spectroscopy is proposed for composition measurement in industrial processes. A double-layer structure is designed for variable selection by combining Fisher's linear discriminant combined population analysis (FCPA) and an improved binary cuckoo search algorithm (IBCS). The Fisher classifier combined with model population analysis is used to select the variable interval wherein the useful variables are roughly located even when strong multicollinearity exists among spectral variables. Opposition-based learning (OBL) and jumping genes (JG) are introduced to improve the binary cuckoo search algorithm for the fine selection of key variables, thus avoiding the loss of excellent solutions due to randomness and the local optimum. Different variable selection methods were used to select variables for beer, corn, and diesel fuel datasets, and the partial least squares (PLS) algorithms were used to build calibration models to predict the original extract concentration of beer, the protein and starch content of corn, and the boiling point of diesel fuel, respectively. The results showed that the proposed PLS modeling method based on FCPA-IBCS has higher fitting accuracy and smaller prediction errors.

Quantitative detection of zearalenone in wheat grains based on near-infrared spectroscopy

Zearalenone (ZEN) can easily contaminate wheat, seriously affecting the quality and safety of wheat grains. In this study, a near-infrared (NIR) spectroscopy detection method for rapid detection of ZEN in wheat grains was proposed. First, the collected original near-infrared spectra were denoised, smoothed and scatter corrected by Savitzky-Golay smoothing (SG-smoothing) and multiple scattering correction (MSC), and then normalized. Three wavelength variable selection algorithms were used to select variables from the preprocessed NIR spectra, which were random frog (RF), successive projections algorithm (SPA), least absolute shrinkage and selection operator (LASSO). Finally, based on the feature variables extracted by the above algorithms, support vector machine (SVM) models were established respectively to realize the quantitative detection of the ZEN in wheat grains. Eventually, the prediction effect of the LASSO-SVM model was the best, the prediction correlation coefficient (R_P) was 0.99, the root mean square error of prediction (RMSEP) was 2.1 μg·kg^-1, and the residual prediction deviation (RPD) was 6.0. This research shows that the NIR spectroscopy can be used for high-precision quantitative detection of the ZEN in grains, and the research gives a new technical solution for the in-situ detection of mycotoxins in stored grains.

Rapid Detection of Fatty Acids in Edible Oils Using Vis-NIR Reflectance Spectroscopy with Multivariate Methods

The composition and content of fatty acids are critical indicators to identify the quality of edible oils. This study was undertaken to establish a rapid determination method for quality detection of edible oils based on quantitative analysis of palmitic acid, stearic acid, arachidic acid, and behenic acid. Seven kinds of oils were measured to obtain Vis-NIR spectra. Multivariate methods combined with pretreatment methods were adopted to establish quantitative analysis models for the four fatty acids. The model of support vector machine (SVM) with standard normal variate (SNV) pretreatment showed the best predictive performance for the four fatty acids. For the palmitic acid, the determination coefficient of prediction (RP2) was 0.9504 and the root mean square error of prediction (RMSEP) was 0.8181. For the stearic acid, RP2 and RMSEP were 0.9636 and 0.2965. In the prediction of arachidic acid, RP2 and RMSEP were 0.9576 and 0.0577. In the prediction of behenic acid, the RP2 and RMSEP were 0.9521 and 0.1486. Furthermore, the effective wavelengths selected by successive projections algorithm (SPA) were useful for establishing simplified prediction models. The results demonstrate that Vis-NIR spectroscopy combined with multivariate methods can provide a rapid and accurate approach for fatty acids detection of edible oils.

Determination of acid value during edible oil storage using a portable NIR spectroscopy system combined with variable selection algorithms based on an MPA-based strategy

BACKGROUND

The acid value is an important indicator for evaluating the quality of edible oil during storage. This study employs a portable near-infrared (NIR) spectroscopy system to determine the acid value during edible oil storage. Four MPA-based variable selection methods, namely competitive adaptive reweighted sampling (CARS), the variable iterative space shrinkage approach (VISSA), iteratively variable subset optimization (IVSO), and bootstrapping soft shrinkage (BOSS) were introduced to optimize the preprocessed NIR spectra. Support vector machine (SVM) models based on characteristic spectra obtained by different selection methods were then established to achieve quantitative detection of the acid value during edible oil storage.

RESULTS

The results revealed that, compared with the full-spectrum SVM model, the SVM models established by the characteristic wavelengths optimized by the variable selection methods based on the MPA strategy exhibit a significant improvement in complexity and generalization performance. Furthermore, compared with the CARS, VISSA, and IVSO methods, the BOSS method obtained the least number of characteristic wavelength variables, and the SVM model established based on the optimized features of this method exhibited the optimal prediction performance. The root mean square error of prediction (RMSEP) was 0.11 mg g-1, the coefficient of determination (Rp2) was 0.92 and the ratio performance deviation (RPD) was 2.82, respectively.

CONCLUSION

The overall results indicate that the variable selection methods based on the MPA strategy can select more targeted characteristic variables. This has good application prospects in NIR spectra feature optimization. © 2020 Society of Chemical Industry.