부산대학교 도서관

This paper discusses the application of long short-term memory (LSTM) neural networks to maintain the synchronization of a real-time clock in holdover operation, that is, while the timing reference input of the clock is unavailable. The approach trains the LSTM network based on timestamps acquired while the slave clock is locked to its reference input coming from a master clock. When the slave clock loses its reference and enters holdover mode, the LSTM takes over and controls the clock. We evaluate the method on a testbed consisting of IEEE 1588 Precision Time Protocol (PTP) clocks based on field-programmable gate arrays (FPGA), where we collect nanosecond-accurate timestamps for offline analysis. We evaluate two oscillator stability scenarios: when the PTP clocks rely on oven-controlled crystal oscillators (OCXOs) and when they use crystal oscillators (XOs). In both cases, we demonstrate that the algorithm can sustain the clock synchronization accuracy within reasonable limits over intervals of 1000 seconds in two different temperature scenarios.