Machine Learning enabled Fault-Detection Algorithms for Optical Spectrum-as-a-Service Users

The growing usage of high-bandwidth, low-latency applications has led to a significant increase in data traffic in recent years. To meet this demand, optical network operators have begun upgrading to Elastic Optical Networks (EONs), powered by Flexible Bandwidth Variable Transceivers (Flex-BVTs). Encouraged by the disaggregation trend, where Flex-BVTs and Open Line Systems (OLS) are owned and controlled by different parties, the operators are introducing new service models like Optical Spectrum-as-a-Service (OSaaS) in their networks. The OSaaS user in this service model perceives OLS as a transparent light tunnel with no monitoring points other than the Flex-BVTs. As multiple OSaaS users share the same OLS, these networks are more susceptible to failures caused by power degradation or channel interference. To reduce system disruptions and repair costs, it is therefore crucial to detect, identify, and counter such failures timely. In this work, we investigate the methods for OSaaS users to detect and identify failures as early as possible using only the telemetry data available from the end Flex-BVTs. Deploying and monitoring five Flex-BVTs within a 400-GHz dedicated OSaaS channel on a pan-European live network for 45 days, we evaluate the applicability of two Machine Learning (ML) based algorithms for EON failure detection, namely, an Artificial Neural Network (ANN) model with dynamic threshold calculation, and a One-Class Support Vector Machine (OCSVM) model. Our results show that the ANN-based approach can detect all artificially introduced failures, with a misclassification rate of 0.01% as compared to the OCSVM-based approach which was unable to detect up to one-third of artificially introduced failures, along with a misclassification rate of 0.6%.