Towards a Federated Learning Approach for Branched Wired Networks Prognosis

Cable prognosis approaches are necessary to monitor the health state of branched wired networks. Physical failure models are often applied first as a means of monitoring, but their complex production is limiting. To address this limitation, data-driven ones are among the most suitable approaches. This is because sensors can provide a significant amount of condition monitoring data that can be used to estimate the remaining useful life of wired networks. This paper explores the uses of machine learning and distributed reftectometry sensors in establishing the guidelines for the implementation of a wired network prognosis strategy. After realizing a reftectometry diagnosis, the acquired signals are processed to extract features that are representative of cable degradation. Then, machine learning models are used to forecast the evolution of the features for the purpose of quantifying the future global degradation state of the wired network. Finally, the remaining useful life of the wired network is estimated with this quantification and an end-of-life threshold. The next step of this project is to test the efficiency of the strategy proposed here, proving the suitability of guidelines based on data-driven models, as well as the practicality of developing a federated learning solution that addresses data quantity and privacy issues.