Common plastics THz classification via artificial neural networks: A discussion on a class of time domain features

Abstract

Pulsed Terahertz waves are widely used as non-destructive technique for dielectric characterization of dielectric materials. In the Terahertz band, a family of materials with a great applicative perspective is represented by some of the most common plastics (including Polymethyl methacrylate, Polyethylene terephthalate, Teflon etc.). In general, lenses for THz frequencies can be fabricated with this type of plastics. Refractive index, absorption and thickness of a material are generally obtained exploiting time and frequency domain measurements in the Terahertz band. These data could be well used for classification purposes combined with chemometric methods. Although this approach is still highly regarded, is gradually being flanked - and sometime even overcome - by introducing intelligent techniques based on artificial neural networks. Since the first step in classification problems is generally represented by features extraction, and this could heavily affect the classification performances, it deserves detailed discussions. In this systematic work, several tuples of features are extracted from Terahertz time domain signals reflected from a set of THz optical materials (common plastics). These tuples are suggested to be used as input data for training and testing an artificial neural network based on Multilayer Perceptron. The performance of the network for materials and thickness classification purposes is hence discussed taking into account the number and type of features, the network architecture and the proportion used to split the dataset into the training and the testing datasets. Extending this approach beyond THz optical materials, an automated methodology for the analysis and classification of potentially each unknown dielectric material could be established.