Digital-twin-assisted meta learning for soft-failure localization in ROADM-based optical networks

Abstract

Reconfigurable optical add/drop multiplexer (ROADM) nodes are evolving towards high-degree architectures to support growing traffic and enable flexible network connectivity. Due to the complex composition of high-degree ROADMs, soft failures may occur between both inter- and intra-node components, like wavelength selective switches and fiber spans. The intricate ROADM structure significantly contributes to the challenge of localizing inter-/intra-node soft failures in ROADM-based optical networks. Machine learning (ML) has shown to be a promising solution to the problem of soft-failure localization, enabling network operators to take accurate and swift measures to overcome such challenges. However, data scarcity is a main hindrance when using ML for soft-failure localization, especially in the complex scenario of inter- and intra-node soft failures. In this work, we propose a digital-twin-assisted meta-learning framework to localize inter-/intra-node soft failures with limited samples. In our proposed framework, we construct several mirror models using a digital twin of the physical optical network and then generate multiple training tasks. These training tasks serve as pretraining data for the meta learner. Then, we use real data for fine-tuning and testing of the meta learner. The proposed framework is compared with the rule-based reasoning method, transfer-learning-based method, and artificial-neural-network-based method with no pretraining. Experimental results indicate that the proposed framework improves localization accuracy by over 15%, 33%, and 54%, on average, compared to benchmark approaches, respectively.