Dynamic network-aware soft failure localization using machine learning in optical networks

With the dynamic nature of optical service provisioning and network topology reconfigurations, failure identification and management become complex, as the machine learning (ML) model is trained for a specific topology with pre-defined performance metrics. This paper proposes a hybrid ML framework for continuous monitoring and soft failure (SF) localization in a partially disaggregated optical network. The framework combines a distributed unsupervised machine learning approach for per-device monitoring and an inductive graph neural network (GNN) for SF localization. This allows the system to generalize across dynamic network conditions, including optical service reconfigurations and node additions or deletions. To support real-time data collection and provide data plane visibility in the management plane, this work proposes gNMI/gRPC-based telemetry streaming using a unified ONF-TAPI YANG data model, enabling vendor-neutral communication across multi-domain networks. The proposed telemetry streaming outperforms the existing solution by reducing traffic load by a factor of 78.4%, and the inductive GNN-based failure localization maintains an accuracy of 97.4% despite dynamic network reconfigurations.