Publication Date

Spring 2026

Degree Type

Master's Project

Degree Name

Master of Science in Computer Science (MSCS)

Department

Computer Science

First Advisor

Genya Ishigaki

Second Advisor

Navrati Saxena

Third Advisor

William Andreopoulos

Keywords

Optical Networks, Soft Failure Detection, LSTM, ASE Noise, EDFA

Abstract

Optical networks, which carry a massive amount of global internet traffic, are one of the most significant digital infrastructures. Prompt failure detection in such networks is a critical challenge to guarantee their reliability and availability. Among different types of failure events, soft failures, which cause a gradual degradation of the Optical Signal to Noise Ratio (OSNR) over a period of time, are challenging to detect due to their slow and fluctuating noise patterns. This paper proposes a simulation framework of common soft failure scenarios caused by Amplified Spontaneous Emission (ASE) noise accumulation based on physics-grounded noise models and develops a Long Short-Term Memory (LSTM)-based detection solution for the soft failures. The detection quality is evaluated based on the per-path relative cumulative sum of drift features that are calculated from OSNR measurements. To verify the robustness of the LSTM-based detection model, three degradation configurations (slow, medium, and fast degradations) are simulated and analyzed. Our results indicate that the proposed LSTM-based solution achieves Matthews Correlation Coefficients (MCC) of 0.629, 0.857, and 0.869 and balanced accuracies of 83.4%, 91.0%, and 94.1% for the three configurations, respectively. Furthermore, a simple localization heuristic based on the path intersection voting achieves 60% to 86.7% of accuracy across the configurations, given the set of soft-failing paths identified by LSTM.

Download

Available for download on Saturday, May 22, 2027

Share

COinS