Session P1b: Environmental Sensitivity of Clocks and Timing Systems

Machine Learning for Frequency Stability Estimation of High Precision Oscillators
James McKelvy, Jet Propulsion Laboratory, California Institute of Technology; Thejesh Bandi, The University of Alabama; Bryan Owings, Microchip Technology Inc.
Location: Seaview A/B
Date/Time: Tuesday, Jan. 28, 5:08 p.m.

In this work, we present recent developments in our method of using machine learning to predict the frequency stability of an active hydrogen maser in the absence of traditional phase-based comparisons. The stability of a precision oscillator can vary as a function of time. Fluctuations in an oscillator’s instability can be attributed to external environmental disturbances or long-term internal phenomena such as aging of hardware. To measure these variations with standard phase-based techniques, a timing system must operate three or more comparably stable oscillators for performing three-cornered hat comparisons. For a timing system with three oscillators, if any one oscillator fails, it is not possible to isolate glitches or variations in the frequency stability between the two remaining oscillators using phase comparisons. For NASA’s Deep Space Network (DSN), three active hydrogen masers operate in parallel for the Frequency and Timing Subsystem (FTS). Of these three masers, one maser is selected as the primary reference for the station, supporting various users for navigation and radio science applications. The other two masers serve as redundant back-ups in the event of a failure of the primary maser, and they also allow for local, three-cornered hat comparisons. If a maser goes offline for maintenance, there is no way to identify which of the remaining masers should be selected as the primary standard. However, previous work has demonstrated that fluctuations in the auxiliary telemetry data of an oscillator can be correlated to variations in the frequency stability [McKelvy, PTTI 2024]. Simple machine learning algorithms offer an effective means of extracting these correlations between telemetry and stability.
In previous work, four legacy Smithsonian Astrophysical Observatory (SAO) VLG-11 masers were studied in JPL’s Frequency Standard Test Lab (FSTL) to establish the base relationship between telemetry and stability. Phase and telemetry data from the VLG-11 masers were used with an assortment of supervised machine learning models to train a binary classifier that would predict if a given maser was above or below a set stability threshold for a given performance period. Variations of the approach used the raw telemetry data as well as telemetry data preprocessed with dimensionality reduction techniques such as principal component analysis. The best-performing model was able to classify the masers into their corresponding stability regime with an accuracy of >80%; however, this method could not be used to predict the specific Allan deviation of each maser (only the approximate stability class).
This work continues the previous efforts by replacing the binary classification system with a regression approach capable of estimating the complete Allan deviation of a given maser over a given performance period. In this study, an assortment of Microchip active hydrogen masers is evaluated. Each maser’s telemetry is recorded via the Microchip Maser Monitor program, and each maser’s phase data is measured against a reference maser with automated linear drift correction. In place of dimensionality reduction, the updated methodology involves calculating the Allan deviation of each telemetry signal for each analysis period. Various supervised machine learning techniques are then used to train a model to predict the traditional phase Allan deviation curve of each maser based on its corresponding telemetry Allan deviation. Two broad modeling topologies are considered: a maser-specific model, in which a model is trained using data from a specific maser for making future predictions of the same maser; and a maser-agnostic model, in which a model is trained on data from a set of masers and then used to make predictions of another, arbitrary maser. In the study, the maser-specific model achieved a coefficient of determination of ~0.94 when comparing the true Allan deviation against the predicted Allan deviation for Taus from 100 to 100,000 seconds, with a mean absolute error of ~2e-16. The maser-agnostic model achieved a coefficient of determination of ~0.63 with a mean absolute error of ~5e-16. While the predictive capability of the technique is still coarse, this resolution is sufficient for making decisions for frequency reference selection in the field. Once matured, the algorithms considered will enable the frequency stability estimation of an independent precision oscillator directly by its telemetry signals, when pairwise or three-cornered hat comparisons are not obtainable. However, note that this is not a replacement for direct measurements made against high precision oscillators like hydrogen masers.