Explainable ML-Based Time Source Arbitration for 5G Networks: Feature Engineering, Contextual Memory, and the Stability Paradox

Abstract

The Best Master Clock Algorithm (BMCA) of IEEE 1588 selects timing sources without memory, measured quality awareness, or explainability—three properties increasingly demanded by 5G operators managing critical synchronization infrastructure. This paper presents a feature engineering and classification methodology addressing all three limitations through: (1) a multi-scale rolling statistics framework transforming raw PTP telemetry into semantically rich features capturing anomalies, stability indicators, and drift dynamics; (2) a recursive memory mechanism (prev_decision) implementing learned adaptive hysteresis via Bayesian conditional probability, eliminating flapping without manual threshold tuning; and (3) full decision explainability through Random Forest feature importance analysis (MDI). Experimental results reveal a counterintuitive stability paradox: under high-PDV conditions, the algorithm performing 1,318 source switches achieves zero ITU-T G.8271.1 mask violations, while the algorithm performing only 2 switches records 209 violations. Feature importance analysis confirms that rolling standard deviation (σ_TE) and previous decision (x_prev) dominate in adversarial scenarios, validating the hypothesis that contextual memory is the missing component in conventional synchronization management. Sensitivity analysis across oscillator classes demonstrates that ML-based arbitration with inexpensive TCXO hardware outperforms standard BMCA with rubidium oscillators, proving that algorithmic intelligence can substitute for hardware investment.