Experimental Studies of Variations Reduction in Chemometric Model Transfer for FT-NIR Miniaturized Sensors

Recent technology trends to miniaturize spectrometers have opened the doors for mass production of spectrometers and for new applications that were not possible before and where the spectrometer can possibly be used as a ubiquitous spectral sensor. However, with the miniaturization from large reliable bench-top to chip-size miniaturized spectrometers and with the associated mass production, new issues have to be addressed such as spectrometers unit-to-unit variations, variations due to changing the measurement setup and variations due to changing the measurement medium. The unit-to-unit variations of the sensors usually result from changing mode of operation, aging, and production tolerances. The aim of this work is to study the issues emerging from the use of miniaturized Fourier Transform Near-Infrared (FT-NIR) spectral sensors and evaluate the influence of these issues on the multivariate classification model used in many applications. In this work, we also introduce a technique to transfer a classification model from a reference calibration sensor to other target sensors to help reducing the effect of the variations and to alleviate the degradation that occurs in the classification results. To validate the effectiveness of the model transfer technique, we developed a Gaussian Process Classification (GPC) model and Soft Independent Modeling Class Analogy (SIMCA) model both using spectral data measured from ultra-high temperature (UHT) pasteurized milk with different levels of fat content. The models aim to classify milk samples according to the percentage of their fat content. Three different experiments were conducted on the models to mimic each type of variations and to test how far they affect the models’ accuracy once the transfer technique is applied. Initially, we achieved perfect discrimination between milk classes with 100% classification accuracy. The largest retardation in accuracy appeared while changing the measuring medium reaching 45.4% in one of the cases. However, the proposed calibration transfer technique showed a significant enhancement in most of the cases and standardized the accuracy of all retarded cases to get the accuracy back to over 90%.