Reduction of Variations Using Chemometric Model Transfer: A Case Study Using FT-NIR Miniaturized Sensors

Abstract

The aim of this paper is to study the unit-to-unit variations in miniaturized Fourier Transform Near-InfraRed (FT-NIR) spectral sensors and the effects of these variations on a classification model developed on a single reference calibration sensor. The paper introduces a simple technique to transfer a classification model from the reference calibration sensor to any other target sensor taking into account variations that might occur. The unit-to-unit variations of the sensors usually result from changes in the signal to noise ratio (SNR) of the sensor due to changes in the mode of operation, variations due to aging, variations due to production tolerances, or changes that occur due to the setup and usage scenario such as scanning through a different medium. To prove the effectiveness of the model transfer technique, we use a Gaussian process classification (GPC) model developed using spectral data from ultra-high temperature (UHT) pasteurized milk with different levels of fat content. The model aims to classify the milk samples based on their fat content. After the model is developed, three experiments are held to mimic each type of the variations and to test how far this will influence the GPC model accuracy after applying the transfer technique.