Anomaly detection for atomic clocks using unsupervised machine learning algorithms

• Published in: Metrologia Vol. 61; no. 5; pp. 55005 - 55017
• Main Authors: Chen, Edwin, Charbonneau, André, Gertsvolf, Marina, Wang, Yunli
• Format: Journal Article
• Language: English
• Published: Bristol IOP Publishing 01.10.2024; Bureau International des Poids et Mesures (BIPM)
• Subjects: Algorithms; Anomalies; anomaly detection; atomic clock; Atomic clocks; change point detection; Clocks & watches; Data analysis; frequency jumps; Inflection points; Machine learning; Outliers (statistics); Performance enhancement; Performance evaluation; phase jumps; Unsupervised learning
• ISSN: 0026-1394; 1681-7575; 1681-7575
• DOI: 10.1088/1681-7575/ad6b30

Atomic clocks are the most accurate timekeeping instruments in the world and play a crucial role in numerous applications, notably in satellite navigation. Nonetheless, the accuracy of these clock signals is subject to anomalies, highlighting the importance of anomaly characterization and detection. In our study, we identify three types of anomalies in the phase data of atomic clocks—outliers, phase jumps and inflection points, and also frequency jumps in the frequency data, which corresponds to the inflection points. The novelty of our work lies in the application of machine learning methods for atomic clock anomaly detection (AD), as opposed to traditional algorithms. We introduce a variety of unsupervised AD and change point detection (CPD) algorithms from the realms of statistics and machine learning, propose a customized thresholding function for AD, and evaluate AD and CPD algorithms on both synthetic and real-world atomic clock data. Our experiments reveal that the subsequence local outlier factor algorithm achieves excellent performance in detecting outliers and phase jumps, but struggles with inflection points. Crucially, we found that the application of a CPD algorithm specifically for detecting the frequency jumps in frequency data substantially improves performance. Our study adopts unsupervised machine learning methods for AD provides robust detection strategies, thereby contributing an expanded understanding of the use of machine learning for clock data monitoring.
NRC publication: Yes