Proactive soft-failure prediction in optical transport networks via physics-inspired features and Infrastructure-as-Code orchestration

Abstract

Abstract Optical transport networks rely on reactive fault management, which guarantees service disruption during the onset of soft failures. We present a framework for proactive soft-failure prediction that combines physics-inspired feature engineering, tree-ensemble machine learning, and Infrastructure-as-Code (IaC) orchestration. The framework is validated on (i) a multi-physics stochastic simulation spanning five degradation modes (Ornstein–Uhlenbeck, exponential, Weibull, step, oscillatory) and (ii) a publicly available real optical telemetry benchmark (Ghosh & Adhya, 2025) comprising 756 lightpaths $$\times$$ 4 failure classes $$\times$$ 900 samples (2.72 M records). A Random Forest regressor augmented with velocity, acceleration, and rolling-statistic features predicts time-to-failure with 17.9 s mean absolute error (MAE) on synthetic test data and 73.2 $$\,\pm \,$$ 0.03 s MAE (95% CI, $$n{=}10$$ seeds) on the real benchmark, outperforming heuristic baselines by 6 $$\times$$ and matching a tuned XGBoost ( $$72.7\pm 0.05$$  s) while surpassing LSTM and 1D-CNN sequence models trained under identical conditions. A trajectory-level train/test split eliminates temporal leakage. SHapley Additive exPlanations (SHAP) applied to four operational case studies (EDFA-aging, NLI-accelerating, stable-link, and false-alarm trajectories) show that alarm decisions are driven primarily by current OSNR, rolling-window statistics, and velocity, yielding interpretable diagnostics at the moment of alert. An end-to-end latency budget of the proposed IaC pipeline, measured stage-by-stage, totals 6.7 s mean wall-clock, dominated by Kubernetes reconciliation and Terraform apply; machine-learning inference contributes < 0.5%. The framework is scoped to gradual OSNR-degrading failures (EDFA pump aging, nonlinear-interference drift); extension to laser-current-visible ECL failures through multi-channel feature fusion is identified as future work.