On Cooperative Fault Management in Multi-Domain Optical Networks Using Hybrid Learning

Abstract

This paper presents a hybrid learning approach for cooperative fault management in multi-domain optical networks (MD-ONs). The proposed approach relies on a broker-based MD-ON architecture for coordination of inter-domain service provisioning. We first propose a self-supervised learning design for soft failure detection. The self-supervised learning design makes use of a clustering algorithm for extracting normal and abnormal patterns from optical performance monitoring data and a supervised learning-based classifier trained with the learned patterns for online detection. To facilitate high soft failure detection accuracy in the absence of sufficient abnormal data for training, the proposed design estimates model uncertainties during predictions and identifies instances associated with high uncertainties as also soft failures. Then, we extend the self-supervised learning design and present a federated learning framework for the broker plane and DMs to learn cooperatively while complying with the privacy constraints of each domain. Finally, a data-driven soft failure localization scheme that operates by analyzing the patterns of data is proposed as a complement to the existing approaches. Performance evaluations indicate that the self-supervised learning design can achieve soft failure detection accuracy of up to <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\sim 97\%$</tex-math></inline-formula> with <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$0.01\%-0.04\%$</tex-math></inline-formula> false alarm rate, while federated learning enables DMs to realize <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$&gt;90\%$</tex-math></inline-formula> soft failure detection rates in the cases of highly unbalanced data distribution (two of the three domains possess zero abnormal data for training).