This paper presents a practical study on early detection of optical-network failures using machine-learning algorithms, with a focus on real-world deployment constraints. We evaluate three ML models (Logistic Regression, Random Forest, and LSTM) for classifying four failure types: External Cavity Laser (ECL), Erbium-Doped Fiber Amplifier (EDFA), Nonlinear Interference (NLI), and normal operation. Our novel contribution lies in analyzing critical deployment metrics including inference time (0.03-39.2 micro seconds), memory usage (495-1665 MB), and temporal window requirements (6-12 samples) for early failure detection. Results show Random Forest achieves the best accuracyefficiency trade-off with 99.88% accuracy, 3.9 micro seconds inference time, and moderate memory consumption, making it ideal for real-time optical network monitoring.
Beyond Accuracy: Practical Aspects of Machine Learning for Early Detection of Optical Network Failures
Terzano, P.;Centofanti, C.;Battisti, G.;Marotta, A.
2026-01-01
Abstract
This paper presents a practical study on early detection of optical-network failures using machine-learning algorithms, with a focus on real-world deployment constraints. We evaluate three ML models (Logistic Regression, Random Forest, and LSTM) for classifying four failure types: External Cavity Laser (ECL), Erbium-Doped Fiber Amplifier (EDFA), Nonlinear Interference (NLI), and normal operation. Our novel contribution lies in analyzing critical deployment metrics including inference time (0.03-39.2 micro seconds), memory usage (495-1665 MB), and temporal window requirements (6-12 samples) for early failure detection. Results show Random Forest achieves the best accuracyefficiency trade-off with 99.88% accuracy, 3.9 micro seconds inference time, and moderate memory consumption, making it ideal for real-time optical network monitoring.Pubblicazioni consigliate
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
