Machine Learning-Based Holdover for IEEE 1588 PTP Packet Fronthauling

Synchronization requirements for mobile networks in 5G/B5G have become more stringent. This work presents a solution based on machine learning (ML) to maintain synchronization between the so-called timeTransmitter and timeReceiver clocks in a scenario in which the receiver suddenly loses its time reference, i.e., it has to operate in holdover using the IEEE 1588 protocol stored samples relying on its local oscillator. Different neural network models for time series forecasting are proposed and compared to a classical ARIMA baseline model in this task of temporarily providing clock reference for the receiver. The ML models are trained using timestamps collected offline from an accurate testbed implemented with FPGA boards. The experiments are performed with two oscillators: OCXO (oven-controlled crystal oscillator) and XO (crystal oscillator), in scenarios with constant and variable temperature conditions. The results show that the ML-based holdover can significantly improve the maximum time interval error between the timeTransmitter and timeReceiver clocks, reducing the need for more expensive synchronization solutions.