Study of machine learning techniques for robotic finger monitoring using image based macro-bending fiber optic sensor

Abstract

This study presents an optical-fiber system incorporating machine learning algorithms for accurately detecting a robotic arm's angular index finger position. The robotic index finger's motion monitoring is possible using only a single-mode optical fiber designed for the visible spectrum. The proposed methodology is based on image analysis applied to the light output profile emitted by the fiber opticupon experiencing curvature, which generates wavelength attenuation in the green band. The images analyzed correspond to the range motion between 0° and 85° with 22 statistical features. We evaluate several machine learning algorithms and enhance their performance by reducing the data dimensionality using neighborhood component analysis. Results show that the methodology achieves an accuracy of 92.88% and befalls a high-performance optical sensing approach for creating new competitive trends compared to traditional fiber sensing techniques. The proposed method does notuse any special optical fiber structure and the results show that the machine learning techniques allow the robotic index finger motion through simple image processing profile of an output optical fiber.