Prediction of vegetable oil adulteration using machine learning models associated with FT-IR spectroscopy

Adulteration of vegetable oils is a common problem in the food industry that consequently compromises the quality of products. However, traditional methods of analysis to evaluate the composition and quality of vegetable oils have high operating costs and require time. In this way, Fourier transform infrared spectroscopy (FT-IR) associated with machine learning (ML) algorithms can act as fast, non-destructive and low-cost analysis alternatives. Thus, the present work aimed to develop and evaluate ML models for predicting the adulteration of vegetable oils using model mixtures composed of palm oil (OP) and soybean oil (OS) from the FT-IR spectra of the mixtures. Binary mixtures were prepared in volumetric fractions of OS (0-100% with variations of 10%). The spectra were obtained by FT-IR, in the range of 650-4000cm-1, totaling 1798 spectral points per sample. The data obtained were treated with baseline correction by Asymmetric Least Squares, smoothing by Savitzky-Golay and standardization. To increase the generalization of the models, synthetic data were generated with the addition of Gaussian noise and spectral displacement. Additionally, the data were submitted to principal component analysis (PCA) for the selection of spectral regions with significant differences (95%) allowing a reduction in the dimensionality of the data and greater precision in the modeling. Five regression models were tested: decision tree, K-nearest neighbors (KNN), support vector regression (SVR), random forest, and gradient boosting. The models were evaluated by cross-validation (Leave-One-Out), using the mean squared error (MSE) and coefficient of determination (R²). PCA was able to identify the differences between the samples before and after the addition of synthetic data, demonstrating that FT-IR, in some regions of the spectrum, is capable of identifying changes due to the addition of OS in OP. The ML models that performed best in predicting OS concentration in OP were gradient boosting (MSE = 0.0005 and R² = 0.993) and random forest (MSE = 0.0010 and R² = 0.9875). The residuals were normally distributed, indicating that the model errors are random. This work demonstrates that FTIR spectra associated with ML models present a combination that allows not only detecting but also quantifying the adulteration of vegetable oils (in the evaluated case, adulteration of palm oil with soybean oil) with high precision, being a fast and lower cost analysis for quality control in the food industry.