A Machine Learning Approach for Telemetry-Based Frequency Stability Estimation of High Precision Clocks

Frequency stability measurements require pairwise comparisons between comparably stable oscillators. It is not possible to assess the frequency stability of a lone oscillator in an isolated environment with classical phase techniques. To supplement the needs of frequency and timing practitioners in situations when phase comparisons are not available, we propose a novel technique for using an oscillator's telemetry data and supervised learning to estimate its associated frequency stability. The effectiveness of the technique is demonstrated on a study involving six commercial active hydrogen masers (AHM). Two broad modeling variants are considered: an oscillator specific model (OSM), in which a supervised model is trained to estimate the dynamic frequency stability of a specific oscillator with historical data of that oscillator; and an oscillator agnostic model (OAM), in which a supervised model is trained to estimate the stability of an arbitrary oscillator using historical data from an assortment of comparable oscillators. The best performing OSM achieved a Mean Absolute Error (MAE) of 5.6e-16 on estimations of dynamic Allan deviation in a held-out test set, while the best performing OAM achieved an MAE of 6.5e-16. The results demonstrate the viability of the proposed method for performing frequency stability assessments in an isolated environment when pairwise comparisons are not possible.