A Machine Learning Approach for Hydrogen Maser Frequency Stability Evaluation

James A. McKelvy, Gregory Shin, Albert Kirk, Anatoliy Savchenkov, William Diener

Abstract: In this work, we present a novel method of using auxiliary telemetry data and machine learning to predict the frequency stability regime of an active hydrogen maser. Assessing variations in the frequency stability of a precision frequency standard traditionally requires the simultaneous operation of multiple frequency standards. In the Jet Propulsion Laboratory (JPL) Frequency Standards Test Laboratory (FSTL), precision phase monitoring equipment is used to continuously measure the frequency stability of an assortment of active hydrogen masers. Glitches and performance degradation in one maser can be isolated by measuring that maser against the others and performing three-cornered-hat comparisons. The FSTL also records a variety of maser telemetry signals associated with signal strength, vacuum system integrity, and environmental regulation. Historically, these telemetry values are recorded to qualitatively assess the health of each maser and to advise preventative maintenance. This work explores the potential for repurposing this telemetry data for predicting the approximate Allan deviation of a maser in situations where a three-cornered-hat comparison is not possible. The machine learning algorithm extracts scalar features from the maser telemetry signals using principal component analysis, and these features are used to train a Random Forest binary classifier that predicts whether a maser is performing within a preset stability threshold. The highest performing model was able to predict the stability regime of select masers in the FSTL with a classification accuracy of greater than 80% and a Matthew Correlation Coefficient (MCC) of 0.45 for Taus between 104 and 105 s. Furthermore, occlusion sensitivity analyses were performed on the trained model to determine which telemetry signals were most influential in determining the stability class of the maser. The proposed methodology potentially subverts the need for the three-cornered-hat measurement approach, allowing for an approximate assessment of the maser’s stability in real time and in an isolated environment. © 2024. California Institute of Technology. Government sponsorship acknowledged.
Published in: Proceedings of the 55th Annual Precise Time and Time Interval Systems and Applications Meeting
January 22 - 25, 2024
Hyatt Regency Long Beach
Long Beach, California
Pages: 65 - 76
Cite this article: McKelvy, James A., Shin, Gregory, Kirk, Albert, Savchenkov, Anatoliy, Diener, William, "A Machine Learning Approach for Hydrogen Maser Frequency Stability Evaluation," Proceedings of the 55th Annual Precise Time and Time Interval Systems and Applications Meeting, Long Beach, California, January 2024, pp. 65-76. https://doi.org/10.33012/2024.19587
Full Paper: ION Members/Non-Members: 1 Download Credit
Sign In