Machine learning-aided hierarchical framework for proactive inline EDFA gain degradation detection and localization in optical networks

As the scale and complexity of fiber-optic networks increase, the automatic detection and localization of soft failures have become crucial tasks for maintaining network services and scheduling repair actions. Fortunately, in software-defined optical networks (SDONs), the availability of extensive monitoring data can facilitate the application of machine learning (ML) techniques for effective failure management. However, the detection and localization of erbium-doped fiber amplifier (EDFA)-related failures, such as EDFA gain degradation, are hard to realize without the ubiquitous deployment of additional optical performance monitoring (OPM) devices. To address this issue, we propose an ML-based hierarchical framework for proactive detection and localization of inline EDFA gain degradation in optical networks. The framework is embedded within standard digital coherent receivers and comprises three stages, namely, time series prediction, failure detection, and failure localization. In the first stage, triple exponential smoothing (TES) is used to predict future signal power time series. Next, a normalizing flow (NF)-based neural network model is developed to detect abnormalities in the predicted signal power time series. If a failure is predicted by the two above-mentioned stages, the third stage of failure localization is automatically triggered in which a multi-task adaptive classification network (MTA-CN) is used to pinpoint an EDFA facing gain degradation by analyzing the amplitude histogram of the received symbols obtained after standard digital signal processing in the receiver. The efficacy of the proposed framework is rigorously validated through extensive experiments conducted on optical links with varying numbers of spans, i.e., 3 spans, 4 spans, and 5 spans. The results demonstrate an F1 score of 0.962 for proactive failure detection, along with F1 scores of 0.958, 0.931, and 0.919 for accurately localizing the faulty EDFAs in the three respective link configurations considered.