Improving Prediction of Peroxide Value of Edible Oils Using Regularized Regression Models

Area of Science:

• Analytical Chemistry
• Food Science
• Spectroscopy

Background:

• Peroxide value (PV) is a critical indicator of edible oil oxidation and quality.
• Accurate and rapid PV determination is essential for food safety and quality control.
• Traditional methods for PV determination can be time-consuming and labor-intensive.

Purpose of the Study:

• To develop and evaluate advanced prediction models for edible oil peroxide values (PV).
• To explore the effectiveness of different data pre-processing methods and regression techniques.
• To establish a robust global model for PV determination using near-infrared (NIR) spectroscopy.

Main Methods:

• Near-infrared (NIR) spectra were collected from edible oil samples with varying peroxide values and natural aging.
• Spectra were acquired using two optical pathlengths (OPLs) and fused to create a third dataset.
• Four regression techniques (PLS, ridge, LASSO, elastic net) were applied with various pre-processing methods, including boxcar averaging.

Main Results:

• Global models demonstrated good agreement across different regression techniques and pre-processing methods.
• LASSO regression with raw spectral data (24 mm OPL) achieved the lowest RMSEP of 1.80.
• Boxcar averaging with small spectral window sizes improved prediction accuracy for edible oil PVs.

Conclusions:

• The study presents promising advancements in developing a comprehensive global model for edible oil PV determination.
• The integration of NIR spectroscopy, advanced regression, and optimized pre-processing offers a powerful approach for rapid oil quality assessment.
• The findings support the potential for a non-destructive, efficient method for monitoring edible oil oxidation.