Application of machine learning on optical fibre distributed sensing for power line applications

Abstract

We present a study on the application of machine learning to optical fibre distributed sensing, with data recovered using a state-of-the-art, commercial BOTDR distributed sensing system; temperature information was extracted from the power line distribution networks that are part of the Electricity Authority of Cyprus. A machine learning approach was implemented for the prediction task of finding points of abnormal behaviour, mimicking the power cable joints that are prone to failure, along with general monitoring for unusual behaviour and potential cable fault conditions; the task is a binary classification one. Labels “0/1” were assigned to the BOTDR measurements, with “1” corresponding to data points in space and time for which the signal showcased a problematic scenario, such as that recorded by optical fibres that are collocated with power cables where the fibre’s temperature measurement increases to dangerously high values, and conversely “0” for all other scenarios. The algorithm’s base is a variation of the state-of-the-art transformer architecture, which depends solely on attention mechanisms. The field data recovered show the potential of the algorithm to predict spatiotemporally problematic points, using the temperature measurements of the collocated fibre.