Machine learning for quantitative fibre optic sensing

Abstract

Optical fibre sensing correlates measurable properties of the light guided within an optical fibre with an external parameter being sensed. There are a number of different methods of fibre sensing, including scattering based sensing, specklegram sensing and interferometric sensing, each with their own applications and limitations. Sensing with multimode fibre offers some significant advantages over single mode fibre, but carries the inherent limitation of extracting useful information from the highly complex and sensitive process of multimode fibre transmission. Deep learning is a form of machine learning at the forefront of data analysis and processing which has solved many problems in a wide range of applications, most notably image and speech recognition. Its application in multimode fibre imaging and sensing has been brief but successful. In this thesis, deep learning is explored as a tool for understanding and quantifying the complex multimode fibre transmission process for sensing applications. Chapter 2 looks at deep learning applied to fibre specklegram sensing, demonstrating its ability to correlate the change in the specklegram output of the fibre with a parameter of the fibre's environment for temperature and refractive index sensing. The superiority of the deep learning approach over current statistical methods is demonstrated, as the deep neural networks improve upon the issues of limited dynamic range and vulnerability to specklegram misalignment that are present with the correlation method. At the same time, the first example of deep learning for regression-based sensing of a continuous variable, as opposed to discrete/classification sensing, is presented. Chapter 3 looks at deep learning applied to sensing with the wavelength spectrum output of a multimode fibre. Current methods of interferometric sensing with a wavelength spectrum and multimode fibre involve the need to inscribe resonance-producing structures within the fibre, which can be costly and time-consuming. The use of deep learning has been previously explored as a method of extracting information pertaining to an environmental parameter, specifically temperature, from a wavelength spectrum without the presence of any resonant features. This thesis looks to build upon such work by demonstrating multi-point sensing using the concept of encoding spatially resolved temperature information in a wavelength spectrum. Sapphire crystal optical fibre is used for sensing, where its highly variable fibre radius is exploited as a means of encoding such spatially resolved information in a fashion that a deep neural network can learn. It is shown that such networks trained on spectra from sapphire fibre perform far better for multi-point temperature sensing than those trained on spectra from glass silica fibres of constant radius.