Background The LLRF system is the core of superconducting linear accelerators for hard X-ray free-electron lasers. Its reference phase largely determines beam quality and FEL output performance. Temperature fluctuations cause obvious long-term phase drift, while conventional compensation methods demand extra RF hardware and more FPGA resources, leading to higher costs.
Purpose This work aims to quantify the temperature-phase characteristics of key LLRF components, develop a deep learning-based compensation strategy that requires no additional RF hardware, and verify the effectiveness of different algorithms in phase drift suppression.
Methods Six temperature control experiments were carried out to test component sensitivity. An 8-point PT100 temperature sensor arrangement was deployed, and over 1.7 million valid data samples were collected and preprocessed. Linear regression, ridge regression, LSTM and Transformer models were established, with hyperparameters optimized via Bayesian method and time-series cross-validation.22 hour real-time compensation tests were implemented on the laboratory platform.
Results The up/down-conversion boards are the primary source of phase drift, accounting for 97.21% of the total phase variation. Among all the models, the Transformer model delivers the optimal performance on the test set. In laboratory compensation tests, it reduces the phase standard deviation by 63.70% and the phase fluctuation range by 62.96%. The model has an average inference time of 0.119 seconds, which fully meets the requirement for real-time operation.
Conclusions LLRF reference phase has a strong linear correlation with temperature. The proposed sensor layout and Transformer-based compensation method remarkably reduce phase fluctuations at low cost. It provides a feasible technical scheme for LLRF phase stabilization. Follow-up studies will conduct on-site validation on the SHINE facility and explore multi-factor coupled compensation strategies.
DownLoad: