Non-Supervised Neural Categorisation of near Infrared Spectra. Application to Pure Compounds

Abstract

Near infrared (NIR) spectroscopy is recognised world-wide as a powerful tool for substance quantification and identification provided that good data analysis tools are used. Most of the identification algorithms use supervised learning and require previous knowledge of existing categories to construct the mathematical models that will be later used at runtime. The use of non-supervised neural learning algorithms is not a common tool in identification of near infrared spectra, although they are widely employed as a pattern recognition technique. Problems analogous to NIR identification have already been solved by means of non-supervised neural networks. Their main advantages are the ability to learn from examples as well as the processing speed, once they are trained. We present in this work a preliminary study of a non-supervised classifier built using Kohonen self-organising maps (SOM). The result is a model useful for identification of a group of NIR spectra belonging to 15 pure products. The accuracy of the classification is discussed. The generalisation of the method for more complex data is still an open issue.