Cause-aware failure detection using an interpretable XGBoost for optical networks.

Area of Science:

• Network Engineering
• Machine Learning
• Data Science

Background:

• Failure detection is critical in network management, with machine learning (ML), particularly neural networks (NNs), widely used in optical networks.
• The 'black-box' nature of NNs hinders interpretability, making it difficult to understand failure detection mechanisms.

Purpose of the Study:

• To develop a cause-aware failure detection scheme for optical transport network (OTN) boards.
• To enhance the interpretability of ML-based failure detection in OTN.

Main Methods:

• Utilized the interpretable extreme gradient boosting (XGBoost) algorithm for failure detection.
• Applied SHapley Additive exPlanations (SHAP) to ensure consistent feature attribution and identify key failure indicators.
• Evaluated performance on two types of OTN boards using balanced and unbalanced datasets.

Main Results:

• Achieved an F1 score exceeding 98% for failure detection on both OTN board types.
• Identified average and maximum environmental temperature as the most relevant features for the two types of OTN board failures, respectively.
• Demonstrated consistent feature attribution using SHAP values.

Conclusions:

• The proposed cause-aware scheme effectively detects OTN board failures with high accuracy.
• Interpretable ML methods like XGBoost and SHAP provide valuable insights into failure causes, specifically related to environmental temperature.
• This approach enhances network reliability and facilitates proactive maintenance.