Machine learning boosts performance of optical fiber sensors: a case study for vector bending sensing

Abstract

The spectral response produced when a high-sensitivity optical fiber sensor (OFS) is subject to an external perturbation has recently been shown to contain rich information that can be potentially exploited for multi-dimensional sensing. In this article, we propose the use of machine learning to directly and statistically learn the relation between the complex spectral response from an OFS and a measurand of interest, without knowing if there are distinct and tractable features in the spectrum. As a proof-of-concept demonstration, it is shown that a simple heterostructure-based device with a capillary tube sandwiched between two single-mode fibers without any fiber modification and complicated fabrication steps, is able to achieve directional bending sensing in a broad dynamic range with machine learning as a tool for signal analysis. It is also demonstrated that stringent requirements of the sensor interrogator, such as the wavelength and bandwidth of the light source, can be greatly relaxed due to the direct spectral mapping between the sensor and the measurand of interest, and importantly, without sacrificing the performance of the sensor. The proposed technique is highly generalizable and can be extended to any OFSs with regular or irregular characteristic spectra for sensing any measurands.