Data-Constrained Semi-Supervised Approaches for Optical Network Fault Detection

Adryele C. Oliveira, Giovana Nascimento Silva, Andrei Nogueira Ribeiro, Fabrício Lobato, Moisés Felipe Silva, Joao Weyl Costa

DOI: 10.14209/sbrt.2025.1571144607

Evento: XLIII Simpósio Brasileiro de Telecomunicações e Processamento de Sinais (SBrT2025)

Keywords: Optical Networks Machine Learning Failure Detection Reduced Training Data

Abstract

With the emergence of 6G and IoT systems, it becomes crucial to correctly detect faults in optical networks to guarantee access to these vital systems. Although machine learning approaches show promise, most demand large datasets, often scarce in practice, posing a significant challenge for model deployment. In this work, we evaluate three semi-supervised learning approaches specifically selected for their complementary strengths in data-scarce scenarios: Principal Component Analysis (PCA) for its noise-resistant dimensionality reduction, One-Class SVM (OCSVM) for its robust boundary learning with limited normal samples, and Hierarchical Clustering (HC) for its adaptability to network operation patterns. All models are trained exclusively on normal operation data and progressive data reductions to assess their performance in resource-constrained scenarios. Experimental results using optical testbed telemetry data show that PCA, OCSVM, and HC achieve accuracies of 93\%, 91.91\%, and 74.31\%, respectively, when trained with only 5\% (544 samples).

Download