Time-Series Representation Learning in Topology Prediction for Passive Optical Network of Telecom Operators

Area of Science:

• Optical Fiber Communications
• Machine Learning

Background:

• Passive Optical Networks (PONs) are cost-effective but require manual topology identification, which is labor-intensive and error-prone.
• Accurate PON topology is crucial for network management and troubleshooting.

Purpose of the Study:

• To develop an automated methodology for predicting PON topology using optical power data.
• To improve the accuracy and efficiency of PON topology identification compared to existing methods.

Main Methods:

• Proposed PT-Predictor, a methodology leveraging representation learning on optical power data.
• Designed GCE-Scorer for feature extraction with noise-tolerant training.
• Implemented MaxMeanVoter and TransVoter for topology prediction.

Main Results:

• PT-Predictor achieved a 23.1% accuracy improvement with sufficient data and 14.8% with insufficient data.
• The method demonstrates superior performance over previous model-free approaches.

Conclusions:

• PT-Predictor offers an effective automated solution for PON topology prediction.
• Identified limitations in scenarios with non-tree structures, indicating future research directions.