Speaker
Description
The reliable operation of superconducting linear accelerators requires effective monitoring of both RF performance and cryogenic systems. This work presents machine learning developments to Low Level Radio Frequency (LLRF) fault classification and real-time thermal load estimation.
Inspired by the Time2Feat pipeline on post-mortem LLRF data, we developed automated classification of anomalies including electronic quenches versus false alarms, achieving reliable separation with a KNN classifier. Complementary neural network observers (LSTM, CNN, ensemble methods) estimate dynamic thermal loads from cryogenic sensors without requiring calorimetric measurements. Models trained on 11 cryomodules achieve prediction accuracy with < 0.5 W standard deviation using pressure, level, valve positions, and temperature data.
The integrated system enables faster fault diagnosis and provides continuous thermal monitoring. Challenges including cross-cryomodule generalization and sensor calibration robustness are addressed through multi-cryomodule training with encoded identifiers. Ongoing work focuses on exploring FPGA implementation constraints for real-time deployment, improving accelerator availability through AI-driven state observation.
| In which format do you inted to submit your paper? | LaTeX |
|---|