Continual learning for optical-network soft-failure monitoring via elastic weight consolidation

Intelligent automation of optical transport networks relies on the ability to monitor services end-to-end, detect performance degradations even in the absence of labeled failure data, and adapt seamlessly to legitimate infrastructure changes. In practice, optical telemetry is highly non-stationary due to reconfigurations, upgrades, and traffic dynamics, which makes static thresholds and fixed machine-learning models unreliable in long-lived deployments. This paper proposes a self-adaptive monitoring framework that combines unsupervised anomaly detection with continual learning. An attention-based autoencoder is used to model normal network behavior from streaming telemetry, while a mask-aware representation accommodates changes in the available telemetry feature set within a predefined feature space. When the normal operating regime shifts, the model is incrementally adapted using elastic weight consolidation to learn new legitimate conditions while limiting catastrophic forgetting. Experiments on realistic optical-network scenarios calibrated from testbed measurements show that a static detector collapses under regime shifts, producing persistent false alarms, whereas continual adaptation restores alarm reliability with false-positive rates below 3%. Moreover, regularized adaptation preserves significantly higher accuracy on previously learned operating conditions (e.g., balanced accuracy above 0.8), while naive fine-tuning shows a significant drop in performance (around 0.6), confirming the need for controlled continual learning in evolving optical networks.