Weighted Ensemble Data Labeling With Cascaded Pre-Training for Failure Management in Practical Optical Transport Networks

In practical optical transport networks, effective failure management is crucial for maintaining stable network operations and minimizing losses caused by failures. However, two data-related issues hinder effective failure management. First, large volumes of unlabeled data continuously accumulate from monitored devices, and labeling these data is both time-consuming and labor-intensive. Second, the data collected from diverse sources, which vary in measures and sample sizes, are often heterogeneous. This heterogeneity poses challenges for data fusion and the development of an accurate failure management model. To address these challenges, this paper introduces a weighted ensemble data labeling approach with cascaded pre-training for heterogeneous data fusion and labeling. We employ an ensemble unsupervised model to execute data labeling tasks, thereby reducing reliance on manual labels. The cascaded pre-training strategy is designed to capture potential correlations among multi-source heterogeneous data and facilitate data fusion at a lower cost. Additionally, weight adjustments guided by real failure information are utilized in ensemble decision-making, enhancing the adaptability of the model to real-world heterogeneous data. This entire data labeling process is visualized and discussed in an interpretable manner. Finally, we evaluate the effectiveness of our approach in a typical failure management task: failure prediction. The results achieve a prediction F1 score exceeding 0.95 across various scenarios, demonstrating the outstanding practical applicability of our labeled data.