Answers are grounded in this paper and cite their sources.
Accurate alarm severity assessment in optical networks is essential for effective fault management; however, network element (NE)-assigned alarm statuses often do not reflect actual service outages reliably, particularly under highly dynamic conditions such as alarm storms. To address this limitation, this paper proposes a hybrid framework that integrates time-correlation analysis with machine learning models to improve alarm severity labels. The proposed framework refines the severity status of line card alarms by temporally correlating them with transponder-level alarms and introduces an optical channel (OCh)-scoped time-correlation approach that provides more accurate and reliable reference labels. Results indicate that NE-assigned labels achieve only 53% balanced accuracy when evaluated against these reference labels. To handle data-splitting challenges, selective under-sampling (SUS) and walk-forward retraining are applied. The Light Gradient Boosting Machine (LightGBM), Extreme Gradient Boosting (XGBoost), and Random Forest classification models are evaluated on alarm data from a large-scale production optical network. Performance evaluation demonstrates that all three models significantly outperform the NE-assigned baseline, increasing the balanced accuracy from 53% up to 76% and the macro-averaged F1 score (macro-F1) from 34% up to 66%. LightGBM delivers the most consistent and highest overall performance across both balanced accuracy and macro-F1.
No takes yet. Share an insight, caveat, or question.
Hedayatnejad et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: