ROADM-related soft failure localization and identification with data augmentation in optical networks

As a core module in elastic optical networks, reconfigurable optical add-drop multiplexers (ROADMs) are crucial for flexible resource allocation; thus, accurate and quick ROADM failure localization and identification are of paramount importance. Although machine learning (ML) holds vast potential for broad application in automated network management, differences between the training and the deployment environments, e.g., in device configurations, may lead to decreased accuracy, which poses a challenge to the quality of training data. Digital spectrum serves as effective supplementary information for soft failure localization (SFL) and identification (SFI); however, ROADM failures in wavelength division multiplexing systems can affect multiple lightpaths, complicating the strategy of employing digital spectrum. To address class imbalance for ROADM-based optical networks, this paper proposes an SFL and SFI scheme using digital spectrum based on ML. Specifically, data augmentation techniques based on the variational autoencoder (VAE) and the synthetic minority oversampling technique (SMOTE) are employed to generate synthetic samples. Subsequently, we adopt XGBoost models and leverage combinations of real and synthetic data to perform SFL and SFI. Extensive numerical analysis conducted in a four-node optical network shows that, for classes with 100× fewer data points than other classes, using SMOTE to balance the classes is more effective compared to using a VAE and yields improvements in the F1 score for localization and identification by 16.18% and 21.18%, respectively, compared to the case without data augmentation. Moreover, to simplify the model, we extract features from the digital spectrum, achieving a 1.02% improvement in the localization F1 score compared to using the digital spectrum.