Proactive soft-failure prediction in optical transport networks via physics-inspired features and...

Area of Science:

• Telecommunications Engineering
• Network Reliability
• Machine Learning Applications

Background:

• Optical transport networks currently use reactive fault management, leading to service disruptions from soft failures.
• Proactive prediction of soft failures is crucial for enhancing network resilience and minimizing downtime.

Purpose of the Study:

• To develop and validate a framework for proactive soft-failure prediction in optical transport networks.
• To combine physics-inspired features, machine learning, and Infrastructure-as-Code for robust failure prediction.

Main Methods:

• Physics-inspired feature engineering including velocity, acceleration, and rolling statistics.
• Tree-ensemble machine learning (Random Forest, XGBoost) for time-to-failure prediction.
• Infrastructure-as-Code (IaC) orchestration for automated deployment and validation on simulated and real optical telemetry data.

Main Results:

• The Random Forest model achieved a Mean Absolute Error (MAE) of 17.9s on synthetic data and 73.2s on a real benchmark, outperforming heuristic methods by 6x.
• SHapley Additive exPlanations (SHAP) provided interpretable diagnostics, identifying OSNR, rolling statistics, and velocity as key predictors.
• The IaC pipeline demonstrated an end-to-end latency of 6.7s, with machine learning inference contributing less than 0.5%.

Conclusions:

• The proposed framework enables proactive soft-failure prediction in optical networks, significantly improving upon reactive measures.
• The combination of advanced feature engineering, machine learning, and IaC offers a scalable and interpretable solution for network fault management.
• Future work includes extending the framework to other failure types like ECL failures through multi-channel feature fusion.