A Hybrid XGBoost–CNN Framework for Robust SOP Anomaly Detection in Optical Networks

Accurate detection of optical fiber anomalies is essential for reliable and secure optical network operation. This paper proposes a selective hybrid inference framework for state- of-polization (SOP) based anomaly detection, combining the efficiency of XGBoost with the one-dimensional convolutional neural networks (1D-CNNs). XGBoost performs fast baseline classification using engineered features derived from Stokes parameters, while class-specific CNNs are conditionally activated only for problematic classes. This targeted activation improves classification accuracy without incurring the full computational cost of deep learning. Experimental evaluation on a controlled SOP telemetry dataset, including vibration and tapping events under varying noise levels, shows that the proposed approach improves accuracy from 87.23% to 99.53% while maintaining low inference latency. The framework demonstrates strong robustness to noise, enabling efficient and real-time deployment in optical network monitoring systems.