Long-term autonomous time-keeping of navigation constellations based on sparse sampling LSTM algorithm

• Published in: Satellite navigation Vol. 5; no. 1; pp. 15 - 14
• Main Authors: Yang, Shitao, Yi, Xiao, Dong, Richang, Wu, Yifan, Shuai, Tao, Zhang, Jun, Ren, Qianyi, Gong, Wenbin
• Format: Journal Article
• Language: English
• Published: Singapore Springer Nature Singapore 01.12.2024; Springer Nature B.V; SpringerOpen
• Subjects: Accuracy; Algorithms; BDS/GNSS: New Development; Clocks & watches; Constellations; Crystal oscillators; Engineering; Global positioning systems; GPS; Ground stations; Laboratories; Neural networks; Noise; Original Article; Progress and Applications; Satellites; Time series
• ISSN: 2662-9291; 2662-1363
• DOI: 10.1186/s43020-024-00137-6

The system time of the four major navigation satellite systems is mainly maintained by multiple high-performance atomic clocks at ground stations. This operational mode relies heavily on the support of ground stations. To enhance the high-precision autonomous timing capability of next-generation navigation satellites, it is necessary to autonomously generate a comprehensive space-based time scale on orbit and make long-term, high-precision predictions for the clock error of this time scale. In order to solve these two problems, this paper proposed a two-level satellite timing system, and used multiple time-keeping node satellites to generate a more stable space-based time scale. Then this paper used the sparse sampling Long Short-Term Memory (LSTM) algorithm to improve the accuracy of clock error long-term prediction on space-based time scale. After simulation, at sampling times of 300 s, 8.64 × 10 4  s, and 1 × 10 6  s, the frequency stabilities of the spaceborne timescale reach 1.35 × 10 –15 , 3.37 × 10 –16 , and 2.81 × 10 –16 , respectively. When applying the improved clock error prediction algorithm, the ten-day prediction error is 3.16 × 10 –10  s. Compared with those of the continuous sampling LSTM, Kalman filter, polynomial and quadratic polynomial models, the corresponding prediction accuracies are 1.72, 1.56, 1.83 and 1.36 times greater, respectively.