Using supervised deep-learning to model edge-FBG shape sensors: a feasibility study

Abstract

Continuum robots are snake-like elastic structures that can be bent anywhere along their length hence representing ideal tools for minimally invasive surgery. To accurately control these flexible manipulators, 3D shape sensors that are small, sterile, immune to electromagnetic noise, and easy to replace are required. Fiber Bragg Grating (FBG)-based shape sensing is a promising approach for this task. The recently proposed Edge-FBG based shape sensors are particularly promising due to their high flexibility and high spatial resolution. In Edge-FBGs, the amplitude change at the Bragg wavelengths contains the strain information at sensing nodes. However, such sensors are sensitive to changes in the spectrum profile caused by undesired bending-related phenomena. As the existing theories cannot accurately predict the spectrum profile in curved optical fibers, changes in the initial intensity that each Edge-FBG receives are not precisely known. These uncontrolled variations cause inaccuracies in shape predictions and make standard characterization techniques less suitable for Edge- FBG sensors. Therefore, developing a model that distinguishes the strain signal from the changes in the spectrum profile is needed. Machine learning techniques are great tools for studying complex problems, making it possible to explore the full spectrum of the Edge-FBG sensor for identifying patterns caused by bending. In this paper, we studied the feasibility of using a low-cost interrogation system for the Edge-FBGs, considering the minimum required signal-to-noise ratio. We trained a neural network with supervised deep learning to directly extract the shape information from the Edge-FBG spectrum. The designed model can predict the shape of a fiber sensor consisting of five Edge-FBG triplets with less than 6 mm tip error.