Применение машинного обучения для детектирования фазы в когерентном оптическом рефлектометре

Abstract

The work is aimed at solving the problem of phase detection in a phase-sensitive reflectometer when working with stochastic Rayleigh reflectors in the fiber. An improvement of measurement methods in coherent reflectometric systems, based on the use of machine learning algorithms, is proposed using a chirp reflectometer as an example. A chirp reflectometer with a tunable wavelength laser source was used to collect data necessary for training neural networks. The study shows that scanning by wavelength allows simulating various external effects, such as deformation or temperature changes over extended areas, thus ensuring efficient data collection for training. The use of even simple neural network algorithms leads to a significant increase in phase measurement accuracy by 43%, demonstrating the potential of this method for detecting phase of complex interferometric signals.