Tactile discrimination of fabrics using machine learning techniques

Abstract

Data mining and machine learning methods are proposed in order to discriminate between various fabrics. In particular textile classes are distinguished, like awning, jeans, jute, pile and satin. The real signals are acquired by a laboratory setup that includes: a Cartesian robot with ability to apply controlled constant pressure and speed, a MEMS piezo capacitive sensor and a Simulink module for signal recording. A set of static and dynamic features is extracted from the data series. A novel approach to feature selection is designed, based on an iterative p-value filter, with separate runs (and results) for different pairs of classes. A set of one-to-one class classifiers (a support vector machine) is learned in corresponding feature spaces. The evaluation procedure, in terms of a ten-fold cross validation, confirmed a 100% of classification accuracy of the proposed approach on available sensor data.