On the estimation of sugars concentrations using Raman spectroscopy and artificial neural networks

Abstract

In this paper, we present an analysis of the performance of Raman spectroscopy, combined with feed-forward neural networks (FFNN), for the estimation of concentration percentages of glucose, sucrose, and fructose in water solutions. Indeed, we analysed our method for the estimation of sucrose in three solid industrialized food products: donuts, cereal, and cookies. Concentrations were estimated in two ways: using a non-linear fitting system, and using a classifier. Our experiments showed that both the classifier and the fitting systems performed better than a Support Vector Machine (SVM), a Linear Discriminant Analysis (LDA), a Linear Regression (LR), and interval Partial Least Squares (iPLS). The best-case obtained by an FFNN for water solutions was 93.33% of classification and 3.51% of Root Mean Square Error in Prediction (RMSEP), compared with 82.22% obtained by a LDA. Our proposed method got an RMSEP of 1% for the best-case obtained with the food products.