Cross-Task and Cross-Lightpath Failure Detection and Localization in Optical Networks Using Transfer Learning

Practical deployments of Machine-Learning(ML)-based solutions for failure management in optical networks often suffer from limited data availability, due to, especially, scarcity of labelled data describing different failure scenarios. Transfer Learning (TL) is regarded as a promising direction in cases of data scarcity, thanks to its ability to transfer knowledge from a Source Domain (SD) (e.g. SD could be a digital twin or a laboratory testbed) to a Target Domain (TD) (e.g., the infield network). In this paper, we focus on cross-lightpath and cross-task application of TL for failure localization and failure detection in optical networks. We found that, depending on the number of retrained parameters in the ML model, cross-lightpath TL for failure localization provides satisfactory accuracy (higher than 90%, in some cases) with limited amounts of TD data, and is also convenient in terms of TD retraining duration with respect to cases where TL is not used. Moreover, we found that cross-task failure detection/localization reaches up to 12% or 25% improvement in TD accuracy when considering failure localization and detection as TD task, respectively.