Imbalance aware ensemble learning and multi domain enable optical transceiver fault diagnosis for 5G 6G and IoT transport networks

Optical transceivers are a critical building block of high-capacity transport networks that support 5G/6G and large-scale IoT services. Their in-service telemetry is noisy, high-dimensional, and strongly imbalanced toward normal operation, which complicates early fault detection. This paper presents an imbalance-aware ensemble learning framework coupled with multi-domain feature engineering for optical transceiver health classification and fault diagnosis. Using the IEEE DataPort Optical Transceiver Extracted Data dataset, we construct 187 features spanning statistical summaries, temporal trends, frequency-domain descriptors, wavelet coefficients, and physics-inspired indicators derived from power and temperature. We evaluate Random Forest, XGBoost, support vector machines, and gradient boosting, together with a weighted-vote ensemble, under stratified 80:20 and 60:40 train-test splits. Combining SMOTE with class weighting improves the minority-class F1-score from 0.72 to 0.93, while the ensemble achieves 97.1% accuracy with AUC 0.993 on the 80:20 split and 95.6% accuracy with AUC 0.986 on the 60:40 split. Ablation experiments show that the proposed feature engineering pipeline yields a 9.4-point absolute accuracy gain. Feature-importance analysis identifies optical power deviation and temperature-power coupling as key discriminators. The median inference time is 4.2 ms per sample for individual models and 12--15 ms per sample for the ensemble, enabling near real-time monitoring. These results support proactive maintenance and improved service reliability in next-generation networks.